| Scope tree | Memory map | Name | sum | gtx540 | gtx660 | gtx750ti | gtx-titan | tesla-c2075 |
|---|---|---|---|---|---|---|---|---|
| P0 |cta P1 | x: global, y: global | 2+2W+membar.cta+po | 1.1K/300K | 0/50K | 192/100K | 2/50K | 956/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | 2+2W+membar.cta+po | 242/300K | 0/50K | 0/100K | 242/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | 2+2W+membar.cta+po | 75/300K | 0/50K | 0/100K | 75/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | 2+2W+membar.ctas | 925/300K | 0/50K | 199/100K | 0/50K | 726/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | 2+2W+membar.gl+po | 8/300K | 0/50K | 0/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | 2+2W+membar.gl+po | 6/300K | 0/50K | 0/100K | 6/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | 2+2W | 1.2K/300K | 0/50K | 196/100K | 13/50K | 967/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | 2+2W | 1.5K/300K | 0/50K | 0/100K | 1.5K/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.2W+membar.cta+membar.cta+po | 30/300K | 0/50K | 1/100K | 0/50K | 29/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.2W+membar.cta+membar.cta+po | 2/300K | 0/50K | 0/100K | 1/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.2W+membar.cta+membar.cta+po | 41/300K | 0/50K | 6/100K | 0/50K | 35/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.2W+membar.cta+membar.cta+po | 26/300K | 0/50K | 2/100K | 0/50K | 24/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.2W+membar.cta+membar.cta+po | 43/300K | 0/50K | 4/100K | 0/50K | 39/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.2W+membar.cta+membar.cta+po | 53/300K | 0/50K | 5/100K | 0/50K | 48/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.2W+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.2W+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.2W+membar.cta+po+membar.gl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.2W+membar.cta+po+po | 28/300K | 0/50K | 2/100K | 0/50K | 26/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.2W+membar.cta+po+po | 91/300K | 0/50K | 0/100K | 90/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.2W+membar.cta+po+po | 39/300K | 0/50K | 1/100K | 0/50K | 38/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.2W+membar.cta+po+po | 47/300K | 0/50K | 5/100K | 18/50K | 24/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | 3.2W+membar.cta+po+po | 38/300K | 0/50K | 0/100K | 38/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | 3.2W+membar.cta+po+po | 25/300K | 0/50K | 0/100K | 25/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | 3.2W+membar.cta+po+po | 17/250K | --- | 0/100K | 17/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | 3.2W+membar.cta+po+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.2W+membar.cta+po+po | 14/250K | --- | 0/100K | 14/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.2W+membar.cta+po+po | 32/300K | 0/50K | 2/100K | 0/50K | 30/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.2W+membar.cta+po+po | 49/300K | 0/50K | 4/100K | 1/50K | 44/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.2W+membar.ctas | 24/300K | 0/50K | 2/100K | 0/50K | 22/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.2W+membar.ctas | 26/300K | 0/50K | 1/100K | 0/50K | 25/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.2W+membar.ctas | 32/300K | 0/50K | 5/100K | 0/50K | 27/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.2W+membar.gl+po+po | 24/300K | 0/50K | 0/100K | 24/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.2W+membar.gl+po+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | 3.2W+membar.gl+po+po | 13/300K | 0/50K | 0/100K | 13/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.2W+membar.gl+po+po | 10/250K | --- | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.2W | 28/300K | 0/50K | 3/100K | 0/50K | 25/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.2W | 43/300K | 0/50K | 2/100K | 0/50K | 41/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | 3.2W | 208/300K | 0/50K | 0/100K | 208/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.2W | 175/250K | --- | 0/100K | 175/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.2W | 286/300K | 0/50K | 5/100K | 228/50K | 53/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+addr+addr+po | 4/300K | 0/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+addr+addr+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+addr+addr+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+addr+addr+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+addr+addr+po | 86/300K | 0/50K | 3/100K | 60/50K | 23/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+addr+ctrl+po | 8/300K | 0/50K | 0/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+addr+ctrl+po | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+addr+ctrl+po | 1/300K | 0/50K | 1/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+addr+ctrl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+addr+ctrl+po | 81/300K | 0/50K | 0/100K | 61/50K | 20/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+addr+data+po | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+addr+data+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+addr+data+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+addr+data+po | 8/300K | 0/50K | 2/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+addr+data+po | 89/300K | 0/50K | 2/100K | 68/50K | 19/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+addr+po+ctrl | 65/300K | 0/50K | 0/100K | 63/50K | 0/50K | 2/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+addr+po+data | 105/300K | 0/50K | 0/100K | 105/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | 3.LB+addr+po+data | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+addr+po+po | 14/300K | 0/50K | 0/100K | 0/50K | 14/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+addr+po+po | 189/300K | 0/50K | 0/100K | 188/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+addr+po+po | 13/300K | 0/50K | 0/100K | 0/50K | 13/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+addr+po+po | 2/300K | 0/50K | 0/100K | 1/50K | 1/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | 3.LB+addr+po+po | 8/300K | 0/50K | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | 3.LB+addr+po+po | 9/300K | 0/50K | 0/100K | 9/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | 3.LB+addr+po+po | 62/300K | 0/50K | 0/100K | 62/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | 3.LB+addr+po+po | 22/250K | --- | 0/100K | 22/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.LB+addr+po+po | 7/250K | --- | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+addr+po+po | 25/300K | 0/50K | 4/100K | 0/50K | 21/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+addr+po+po | 224/300K | 0/50K | 7/100K | 159/50K | 57/50K | 1/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+ctrl+ctrl+po | 8/300K | 0/50K | 0/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+ctrl+ctrl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+ctrl+ctrl+po | 3/300K | 0/50K | 1/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+ctrl+ctrl+po | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+ctrl+ctrl+po | 76/300K | 0/50K | 0/100K | 55/50K | 21/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+ctrl+po+po | 13/300K | 0/50K | 0/100K | 0/50K | 13/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | 3.LB+ctrl+po+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+ctrl+po+po | 181/300K | 0/50K | 0/100K | 178/50K | 0/50K | 3/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+ctrl+po+po | 21/300K | 0/50K | 2/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+ctrl+po+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | 3.LB+ctrl+po+po | 12/300K | 0/50K | 0/100K | 12/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | 3.LB+ctrl+po+po | 5/300K | 0/50K | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | 3.LB+ctrl+po+po | 56/300K | 0/50K | 0/100K | 56/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | 3.LB+ctrl+po+po | 4/250K | --- | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.LB+ctrl+po+po | 12/250K | --- | 0/100K | 12/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+ctrl+po+po | 21/300K | 0/50K | 2/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+ctrl+po+po | 214/300K | 0/50K | 5/100K | 139/50K | 70/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+data+ctrl+po | 4/300K | 0/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+data+ctrl+po | 8/300K | 0/50K | 1/100K | 0/50K | 7/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+data+ctrl+po | 2/300K | 0/50K | 1/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+data+ctrl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+data+ctrl+po | 103/300K | 0/50K | 4/100K | 80/50K | 19/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+data+data+po | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+data+data+po | 6/300K | 0/50K | 1/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+data+data+po | 4/300K | 0/50K | 3/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.LB+data+data+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+data+data+po | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+data+data+po | 112/300K | 0/50K | 1/100K | 89/50K | 22/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+data+po+ctrl | 95/300K | 0/50K | 0/100K | 93/50K | 0/50K | 2/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | 3.LB+data+po+ctrl | 5/250K | --- | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+data+po+po | 17/300K | 0/50K | 2/100K | 0/50K | 15/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | 3.LB+data+po+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+data+po+po | 184/300K | 0/50K | 0/100K | 184/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+data+po+po | 17/300K | 0/50K | 0/100K | 0/50K | 17/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+data+po+po | 3/300K | 0/50K | 1/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | 3.LB+data+po+po | 8/300K | 0/50K | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | 3.LB+data+po+po | 7/300K | 0/50K | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | 3.LB+data+po+po | 83/300K | 0/50K | 0/100K | 83/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | 3.LB+data+po+po | 43/250K | --- | 0/100K | 43/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.LB+data+po+po | 8/250K | --- | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+data+po+po | 21/300K | 0/50K | 1/100K | 0/50K | 20/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+data+po+po | 256/300K | 0/50K | 4/100K | 186/50K | 66/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.cta+addr+po | 24/300K | 0/50K | 0/100K | 0/50K | 24/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.cta+addr+po | 22/300K | 0/50K | 2/100K | 0/50K | 20/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+membar.cta+addr+po | 66/300K | 0/50K | 2/100K | 0/50K | 64/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+membar.cta+addr+po | 12/300K | 0/50K | 1/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+membar.cta+addr+po | 34/300K | 0/50K | 2/100K | 0/50K | 32/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.cta+ctrl+po | 20/300K | 0/50K | 1/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.cta+ctrl+po | 20/300K | 0/50K | 3/100K | 0/50K | 17/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+membar.cta+ctrl+po | 58/300K | 0/50K | 1/100K | 0/50K | 57/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+membar.cta+ctrl+po | 14/300K | 0/50K | 0/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+membar.cta+ctrl+po | 26/300K | 0/50K | 0/100K | 0/50K | 26/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.cta+data+po | 21/300K | 0/50K | 0/100K | 0/50K | 21/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.cta+data+po | 25/300K | 0/50K | 3/100K | 0/50K | 22/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+membar.cta+data+po | 72/300K | 0/50K | 6/100K | 0/50K | 66/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+membar.cta+data+po | 14/300K | 0/50K | 1/100K | 0/50K | 13/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+membar.cta+data+po | 32/300K | 0/50K | 4/100K | 0/50K | 28/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | 3.LB+membar.cta+membar.cta+addr | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.cta+membar.cta+po | 37/300K | 0/50K | 0/100K | 0/50K | 37/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | 3.LB+membar.cta+membar.cta+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.cta+membar.cta+po | 65/300K | 0/50K | 5/100K | 0/50K | 60/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+membar.cta+membar.cta+po | 61/300K | 0/50K | 4/100K | 0/50K | 57/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+membar.cta+membar.cta+po | 49/300K | 0/50K | 4/100K | 0/50K | 45/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+membar.cta+membar.cta+po | 77/300K | 0/50K | 9/100K | 0/50K | 68/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+membar.cta+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+membar.cta+po+addr | 9/300K | 0/50K | 0/100K | 8/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+membar.cta+po+ctrl | 14/300K | 0/50K | 0/100K | 13/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+membar.cta+po+data | 11/300K | 0/50K | 0/100K | 10/50K | 0/50K | 1/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.cta+po+po | 49/300K | 0/50K | 4/100K | 0/50K | 45/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | 3.LB+membar.cta+po+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+membar.cta+po+po | 24/300K | 0/50K | 0/100K | 22/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.cta+po+po | 55/300K | 0/50K | 3/100K | 0/50K | 52/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.LB+membar.cta+po+po | 64/300K | 0/50K | 3/100K | 0/50K | 61/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | 3.LB+membar.cta+po+po | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | 3.LB+membar.cta+po+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.LB+membar.cta+po+po | 2/250K | --- | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.LB+membar.cta+po+po | 51/300K | 0/50K | 6/100K | 0/50K | 45/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+membar.cta+po+po | 88/300K | 0/50K | 6/100K | 1/50K | 81/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.ctas | 39/300K | 0/50K | 3/100K | 0/50K | 36/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB+membar.ctas | 37/300K | 0/50K | 1/100K | 0/50K | 36/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB+membar.ctas | 54/300K | 0/50K | 5/100K | 0/50K | 49/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+membar.gl+po+addr | 4/300K | 0/50K | 0/100K | 3/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+membar.gl+po+ctrl | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+membar.gl+po+data | 10/300K | 0/50K | 0/100K | 8/50K | 0/50K | 2/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.LB+membar.gl+po+po | 8/300K | 0/50K | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | 3.LB+membar.gl+po+po | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.LB+membar.gl+po+po | 2/250K | --- | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.LB | 49/300K | 0/50K | 1/100K | 0/50K | 46/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.LB | 77/300K | 0/50K | 9/100K | 0/50K | 68/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | 3.LB | 448/300K | 0/50K | 0/100K | 448/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.LB | 60/250K | --- | 0/100K | 60/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.LB | 359/300K | 0/50K | 11/100K | 271/50K | 77/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.SB+membar.cta+membar.cta+po | 61/300K | 0/50K | 5/100K | 0/50K | 56/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.SB+membar.cta+membar.cta+po | 8/300K | 0/50K | 0/100K | 7/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.SB+membar.cta+membar.cta+po | 76/300K | 0/50K | 3/100K | 0/50K | 73/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.SB+membar.cta+membar.cta+po | 56/300K | 0/50K | 10/100K | 0/50K | 46/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | 3.SB+membar.cta+membar.cta+po | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.SB+membar.cta+membar.cta+po | 76/300K | 0/50K | 5/100K | 0/50K | 69/50K | 2/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.SB+membar.cta+membar.cta+po | 86/300K | 0/50K | 10/100K | 0/50K | 76/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.SB+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.SB+membar.cta+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.SB+membar.cta+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.SB+membar.cta+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.SB+membar.cta+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.SB+membar.cta+po+membar.gl | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.SB+membar.cta+po+po | 69/300K | 0/50K | 5/100K | 0/50K | 63/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.SB+membar.cta+po+po | 129/300K | 0/50K | 0/100K | 127/50K | 1/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.SB+membar.cta+po+po | 98/300K | 0/50K | 11/100K | 0/50K | 87/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.SB+membar.cta+po+po | 89/300K | 0/50K | 5/100K | 30/50K | 54/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | 3.SB+membar.cta+po+po | 55/300K | 0/50K | 0/100K | 55/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | 3.SB+membar.cta+po+po | 42/300K | 0/50K | 0/100K | 42/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | 3.SB+membar.cta+po+po | 29/250K | --- | 0/100K | 29/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.SB+membar.cta+po+po | 27/250K | --- | 0/100K | 27/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | 3.SB+membar.cta+po+po | 66/300K | 0/50K | 9/100K | 0/50K | 55/50K | 2/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.SB+membar.cta+po+po | 82/300K | 0/50K | 19/100K | 1/50K | 62/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.SB+membar.ctas | 53/300K | 0/50K | 3/100K | 0/50K | 50/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.SB+membar.ctas | 69/300K | 0/50K | 6/100K | 0/50K | 63/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.SB+membar.ctas | 56/300K | 0/50K | 8/100K | 0/50K | 48/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | 3.SB+membar.gl+po+po | 54/300K | 0/50K | 0/100K | 54/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | 3.SB+membar.gl+po+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | 3.SB+membar.gl+po+po | 13/300K | 0/50K | 0/100K | 13/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.SB+membar.gl+po+po | 22/250K | --- | 0/100K | 22/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | 3.SB | 72/300K | 0/50K | 7/100K | 0/50K | 64/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | 3.SB | 113/300K | 0/50K | 14/100K | 0/50K | 99/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | 3.SB | 539/300K | 0/50K | 0/100K | 539/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | 3.SB | 469/250K | --- | 0/100K | 469/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | 3.SB | 412/300K | 0/50K | 12/100K | 304/50K | 96/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRIW+addr+po | 153/300K | 35/50K | 54/100K | 0/50K | 45/50K | 19/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRIW+addr+po | 180/300K | 25/50K | 80/100K | 0/50K | 52/50K | 23/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRIW+addr+po | 93/300K | 22/50K | 47/100K | 0/50K | 12/50K | 12/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: global | IRIW+addr+po | 87/300K | 35/50K | 32/100K | 0/50K | 2/50K | 18/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: shared | IRIW+addr+po | 81/300K | 58/50K | 9/100K | 0/50K | 0/50K | 14/50K |
| P0 |cta P1 |warp P3 |cta P2 | x: global, y: global | IRIW+addr+po | 89/300K | 21/50K | 49/100K | 0/50K | 2/50K | 17/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRIW+addr+po | 199/300K | 41/50K | 72/100K | 0/50K | 63/50K | 23/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRIW+addr+po | 243/300K | 42/50K | 79/100K | 0/50K | 90/50K | 32/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRIW+addr+po | 127/300K | 43/50K | 41/100K | 0/50K | 22/50K | 21/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: global | IRIW+addr+po | 86/300K | 29/50K | 35/100K | 0/50K | 1/50K | 21/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: shared | IRIW+addr+po | 170/300K | 130/50K | 10/100K | 0/50K | 0/50K | 30/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRIW+addr+po | 173/300K | 7/50K | 107/100K | 0/50K | 39/50K | 20/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: shared | IRIW+addr+po | 190/250K | --- | 124/100K | 0/50K | 17/50K | 49/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: global, y: global | IRIW+addr+po | 114/300K | 48/50K | 49/100K | 0/50K | 2/50K | 15/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRIW+addr+po | 294/300K | 49/50K | 166/100K | 2/50K | 50/50K | 27/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRIW+addr+po | 159/300K | 22/50K | 67/100K | 0/50K | 51/50K | 19/50K |
| P0 |warp P2 |cta P1 |warp P3 | x: global, y: global | IRIW+addr+po | 117/300K | 37/50K | 57/100K | 0/50K | 1/50K | 22/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRIW+addr+po | 263/300K | 20/50K | 106/100K | 0/50K | 108/50K | 29/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRIW+addr+po | 152/300K | 26/50K | 72/100K | 0/50K | 32/50K | 22/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRIW+addr+po | 148/300K | 50/50K | 66/100K | 0/50K | 15/50K | 17/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRIW+ctrl+po | 134/300K | 18/50K | 60/100K | 0/50K | 36/50K | 20/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRIW+ctrl+po | 167/300K | 18/50K | 61/100K | 0/50K | 58/50K | 30/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRIW+ctrl+po | 113/300K | 29/50K | 54/100K | 0/50K | 12/50K | 18/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: global | IRIW+ctrl+po | 76/300K | 32/50K | 27/100K | 0/50K | 2/50K | 15/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: shared | IRIW+ctrl+po | 62/300K | 41/50K | 5/100K | 0/50K | 0/50K | 16/50K |
| P0 |cta P1 |warp P3 |cta P2 | x: global, y: global | IRIW+ctrl+po | 78/300K | 20/50K | 50/100K | 0/50K | 1/50K | 7/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRIW+ctrl+po | 175/300K | 31/50K | 50/100K | 0/50K | 74/50K | 20/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRIW+ctrl+po | 213/300K | 45/50K | 60/100K | 0/50K | 79/50K | 29/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRIW+ctrl+po | 129/300K | 47/50K | 50/100K | 0/50K | 19/50K | 13/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: global | IRIW+ctrl+po | 79/300K | 24/50K | 24/100K | 0/50K | 5/50K | 26/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: shared | IRIW+ctrl+po | 197/300K | 156/50K | 7/100K | 0/50K | 0/50K | 34/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRIW+ctrl+po | 189/300K | 8/50K | 105/100K | 0/50K | 59/50K | 17/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: shared | IRIW+ctrl+po | 233/250K | --- | 175/100K | 0/50K | 13/50K | 45/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: global, y: global | IRIW+ctrl+po | 101/300K | 37/50K | 43/100K | 0/50K | 2/50K | 19/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRIW+ctrl+po | 246/300K | 34/50K | 122/100K | 0/50K | 55/50K | 35/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRIW+ctrl+po | 163/300K | 18/50K | 60/100K | 0/50K | 57/50K | 28/50K |
| P0 |warp P2 |cta P1 |warp P3 | x: global, y: global | IRIW+ctrl+po | 116/300K | 40/50K | 56/100K | 0/50K | 2/50K | 18/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRIW+ctrl+po | 206/300K | 22/50K | 74/100K | 0/50K | 81/50K | 29/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRIW+ctrl+po | 159/300K | 15/50K | 78/100K | 0/50K | 47/50K | 19/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRIW+ctrl+po | 125/300K | 43/50K | 53/100K | 0/50K | 10/50K | 19/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRIW+membar.cta+addr | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRIW+membar.cta+ctrl | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRIW+membar.cta+po | 441/300K | 124/50K | 131/100K | 0/50K | 157/50K | 29/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRIW+membar.cta+po | 471/300K | 91/50K | 170/100K | 0/50K | 175/50K | 35/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRIW+membar.cta+po | 422/300K | 133/50K | 130/100K | 0/50K | 127/50K | 32/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: global | IRIW+membar.cta+po | 333/300K | 141/50K | 128/100K | 0/50K | 43/50K | 21/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: shared | IRIW+membar.cta+po | 337/300K | 203/50K | 105/100K | 0/50K | 11/50K | 18/50K |
| P0 |cta P1 |warp P3 |cta P2 | x: global, y: global | IRIW+membar.cta+po | 379/300K | 124/50K | 185/100K | 0/50K | 37/50K | 33/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRIW+membar.cta+po | 579/300K | 152/50K | 201/100K | 0/50K | 200/50K | 26/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRIW+membar.cta+po | 647/300K | 199/50K | 178/100K | 0/50K | 240/50K | 30/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRIW+membar.cta+po | 437/300K | 153/50K | 135/100K | 0/50K | 119/50K | 30/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: global | IRIW+membar.cta+po | 337/300K | 118/50K | 143/100K | 0/50K | 50/50K | 26/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: shared | IRIW+membar.cta+po | 544/300K | 363/50K | 108/100K | 0/50K | 10/50K | 63/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRIW+membar.cta+po | 308/300K | 13/50K | 166/100K | 0/50K | 106/50K | 23/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: shared | IRIW+membar.cta+po | 304/250K | --- | 157/100K | 0/50K | 56/50K | 91/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: global, y: global | IRIW+membar.cta+po | 400/300K | 151/50K | 173/100K | 0/50K | 40/50K | 36/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRIW+membar.cta+po | 501/300K | 66/50K | 288/100K | 0/50K | 104/50K | 43/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRIW+membar.cta+po | 472/300K | 110/50K | 173/100K | 0/50K | 157/50K | 32/50K |
| P0 |warp P2 |cta P1 |warp P3 | x: global, y: global | IRIW+membar.cta+po | 401/300K | 170/50K | 164/100K | 0/50K | 33/50K | 34/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRIW+membar.cta+po | 600/300K | 93/50K | 235/100K | 0/50K | 237/50K | 35/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRIW+membar.cta+po | 550/300K | 108/50K | 240/100K | 0/50K | 164/50K | 38/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRIW+membar.cta+po | 441/300K | 98/50K | 163/100K | 0/50K | 156/50K | 24/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRIW+membar.ctas | 65/300K | 0/50K | 3/100K | 0/50K | 62/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRIW+membar.ctas | 43/300K | 0/50K | 2/100K | 0/50K | 41/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRIW+membar.ctas | 68/300K | 0/50K | 3/100K | 0/50K | 65/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRIW+membar.ctas | 53/300K | 0/50K | 3/100K | 0/50K | 50/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRIW+membar.ctas | 34/300K | 0/50K | 3/100K | 0/50K | 31/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRIW+membar.ctas | 56/300K | 0/50K | 2/100K | 0/50K | 54/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRIW+membar.ctas | 37/300K | 0/50K | 2/100K | 0/50K | 35/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRIW+membar.gl+po | 21/300K | 21/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRIW+membar.gl+po | 21/300K | 17/50K | 0/100K | 0/50K | 3/50K | 1/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRIW+membar.gl+po | 29/300K | 29/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: global | IRIW+membar.gl+po | 16/300K | 15/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: shared | IRIW+membar.gl+po | 66/300K | 52/50K | 0/100K | 0/50K | 0/50K | 14/50K |
| P0 |cta P1 |warp P3 |cta P2 | x: global, y: global | IRIW+membar.gl+po | 28/300K | 28/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRIW+membar.gl+po | 45/300K | 42/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRIW+membar.gl+po | 47/300K | 36/50K | 0/100K | 0/50K | 9/50K | 2/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRIW+membar.gl+po | 26/300K | 26/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: global | IRIW+membar.gl+po | 17/300K | 16/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: shared | IRIW+membar.gl+po | 151/300K | 113/50K | 0/100K | 0/50K | 0/50K | 38/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRIW+membar.gl+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: shared | IRIW+membar.gl+po | 16/250K | --- | 0/100K | 0/50K | 0/50K | 16/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: global, y: global | IRIW+membar.gl+po | 37/300K | 37/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRIW+membar.gl+po | 13/300K | 13/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRIW+membar.gl+po | 29/300K | 29/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 |warp P3 | x: global, y: global | IRIW+membar.gl+po | 22/300K | 21/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRIW+membar.gl+po | 25/300K | 18/50K | 0/100K | 0/50K | 5/50K | 2/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRIW+membar.gl+po | 23/300K | 21/50K | 0/100K | 0/50K | 1/50K | 1/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRIW+membar.gl+po | 27/300K | 27/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRIW | 3.6K/300K | 923/50K | 1.6K/100K | 0/50K | 645/50K | 430/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRIW | 3.9K/300K | 966/50K | 1.6K/100K | 0/50K | 839/50K | 480/50K |
| P0 |cta P1 |warp P3 |cta P2 | x: global, y: global | IRIW | 3.3K/300K | 871/50K | 1.5K/100K | 0/50K | 442/50K | 429/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRIW | 3.6K/300K | 852/50K | 1.6K/100K | 0/50K | 739/50K | 422/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRIW | 3.4K/300K | 876/50K | 1.4K/100K | 0/50K | 668/50K | 464/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRIW | 3.7K/300K | 905/50K | 1.4K/100K | 0/50K | 904/50K | 549/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: global | IRIW | 3.4K/300K | 720/50K | 1.4K/100K | 0/50K | 669/50K | 640/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRIW | 2.9K/300K | 724/50K | 1.2K/100K | 87/50K | 324/50K | 539/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: shared | IRIW | 2.2K/250K | --- | 1.2K/100K | 0/50K | 230/50K | 771/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: global, y: global | IRIW | 3.1K/300K | 727/50K | 1.5K/100K | 0/50K | 412/50K | 453/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRIW | 2.7K/300K | 705/50K | 1.3K/100K | 88/50K | 274/50K | 380/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRIW | 3.5K/300K | 859/50K | 1.5K/100K | 0/50K | 587/50K | 498/50K |
| P0 |warp P2 |cta P1 |warp P3 | x: global, y: global | IRIW | 3.2K/300K | 895/50K | 1.4K/100K | 0/50K | 397/50K | 500/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRIW | 3.6K/300K | 728/50K | 1.5K/100K | 0/50K | 826/50K | 479/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+addr+membar.cta | 10/300K | 0/50K | 0/100K | 0/50K | 10/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+addr+membar.cta | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+addr+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+addr+membar.cta | 18/300K | 0/50K | 0/100K | 0/50K | 18/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+addr+membar.cta | 15/300K | 0/50K | 0/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+addr+membar.cta | 7/300K | 0/50K | 2/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW+addr+membar.cta | 6/300K | 0/50K | 1/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW+addr+membar.cta | 8/300K | 0/50K | 1/100K | 0/50K | 7/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW+addr+membar.cta | 12/300K | 0/50K | 0/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW+addr+membar.cta | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+addr+po | 13/300K | 0/50K | 0/100K | 0/50K | 13/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+addr+po | 5/300K | 0/50K | 0/100K | 0/50K | 4/50K | 1/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+addr+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+addr+po | 20/300K | 0/50K | 1/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+addr+po | 15/300K | 0/50K | 0/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+addr+po | 5/300K | 0/50K | 2/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW+addr+po | 13/300K | 0/50K | 2/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW+addr+po | 16/300K | 0/50K | 0/100K | 0/50K | 16/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW+addr+po | 16/300K | 0/50K | 1/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW+addr+po | 6/300K | 0/50K | 1/100K | 0/50K | 5/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+ctrl+membar.cta | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+ctrl+membar.cta | 7/300K | 0/50K | 2/100K | 0/50K | 5/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+ctrl+membar.cta | 3/300K | 0/50K | 1/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+ctrl+membar.cta | 20/300K | 0/50K | 1/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+ctrl+membar.cta | 14/300K | 0/50K | 0/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+ctrl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW+ctrl+membar.cta | 11/300K | 0/50K | 0/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW+ctrl+membar.cta | 4/300K | 0/50K | 1/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW+ctrl+membar.cta | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW+ctrl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+ctrl+po | 9/300K | 0/50K | 0/100K | 0/50K | 9/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+ctrl+po | 9/300K | 0/50K | 2/100K | 0/50K | 7/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+ctrl+po | 4/300K | 0/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+ctrl+po | 17/300K | 0/50K | 0/100K | 0/50K | 17/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+ctrl+po | 13/300K | 0/50K | 1/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+ctrl+po | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW+ctrl+po | 19/300K | 0/50K | 0/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW+ctrl+po | 15/300K | 0/50K | 1/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW+ctrl+po | 13/300K | 0/50K | 0/100K | 0/50K | 13/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW+ctrl+po | 8/300K | 0/50K | 2/100K | 0/50K | 6/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+membar.cta+po | 43/300K | 0/50K | 0/100K | 0/50K | 43/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+membar.cta+po | 41/300K | 0/50K | 3/100K | 0/50K | 38/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+membar.cta+po | 32/300K | 0/50K | 1/100K | 0/50K | 31/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+membar.cta+po | 61/300K | 0/50K | 4/100K | 0/50K | 56/50K | 1/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+membar.cta+po | 51/300K | 0/50K | 1/100K | 0/50K | 50/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+membar.cta+po | 26/300K | 0/50K | 0/100K | 0/50K | 26/50K | 0/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRRWIW+membar.cta+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRRWIW+membar.cta+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW+membar.cta+po | 49/300K | 0/50K | 2/100K | 0/50K | 47/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW+membar.cta+po | 27/300K | 0/50K | 3/100K | 0/50K | 24/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW+membar.cta+po | 32/300K | 0/50K | 1/100K | 0/50K | 31/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW+membar.cta+po | 45/300K | 0/50K | 0/100K | 0/50K | 45/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+membar.ctas | 28/300K | 0/50K | 1/100K | 0/50K | 27/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+membar.ctas | 25/300K | 0/50K | 2/100K | 0/50K | 23/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+membar.ctas | 23/300K | 0/50K | 0/100K | 0/50K | 23/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+membar.ctas | 45/300K | 0/50K | 4/100K | 0/50K | 41/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+membar.ctas | 30/300K | 0/50K | 2/100K | 0/50K | 28/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+membar.ctas | 27/300K | 0/50K | 1/100K | 0/50K | 26/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW+membar.ctas | 49/300K | 0/50K | 1/100K | 0/50K | 48/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW+membar.ctas | 26/300K | 0/50K | 2/100K | 0/50K | 24/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW+membar.ctas | 19/300K | 0/50K | 1/100K | 0/50K | 18/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW+membar.ctas | 23/300K | 0/50K | 2/100K | 0/50K | 21/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+po+addr | 86/300K | 23/50K | 63/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+po+addr | 90/300K | 38/50K | 52/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+po+addr | 92/300K | 18/50K | 72/100K | 0/50K | 0/50K | 2/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW+po+addr | 99/300K | 25/50K | 72/100K | 0/50K | 1/50K | 1/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW+po+addr | 215/300K | 19/50K | 125/100K | 1/50K | 51/50K | 19/50K |
| P0 |cta P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW+po+addr | 76/300K | 30/50K | 43/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+po+addr | 91/300K | 19/50K | 68/100K | 0/50K | 1/50K | 3/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+po+addr | 75/300K | 24/50K | 49/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+po+addr | 97/300K | 13/50K | 81/100K | 0/50K | 2/50K | 1/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW+po+addr | 91/300K | 25/50K | 66/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW+po+addr | 171/300K | 3/50K | 121/100K | 0/50K | 41/50K | 6/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRRWIW+po+addr | 129/250K | 118/50K | 7/100K | --- | 0/50K | 4/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: shared | IRRWIW+po+addr | 98/200K | --- | 38/100K | --- | 11/50K | 49/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW+po+addr | 69/300K | 16/50K | 52/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRRWIW+po+addr | 58/250K | 54/50K | 2/100K | --- | 0/50K | 2/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW+po+addr | 74/300K | 21/50K | 52/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P2 |cta P1 |warp P3 | x: global, y: global | IRRWIW+po+addr | 86/300K | 36/50K | 46/100K | 0/50K | 3/50K | 1/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW+po+addr | 74/300K | 27/50K | 45/100K | 0/50K | 1/50K | 1/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW+po+addr | 80/300K | 31/50K | 49/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW+po+addr | 79/300K | 20/50K | 59/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+po+ctrl | 60/300K | 24/50K | 34/100K | 0/50K | 1/50K | 1/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+po+ctrl | 81/300K | 44/50K | 37/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+po+ctrl | 91/300K | 24/50K | 67/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW+po+ctrl | 70/300K | 25/50K | 45/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW+po+ctrl | 152/300K | 16/50K | 84/100K | 0/50K | 30/50K | 22/50K |
| P0 |cta P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW+po+ctrl | 70/300K | 26/50K | 42/100K | 0/50K | 1/50K | 1/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+po+ctrl | 79/300K | 20/50K | 58/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+po+ctrl | 59/300K | 32/50K | 27/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+po+ctrl | 87/300K | 15/50K | 68/100K | 0/50K | 2/50K | 2/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW+po+ctrl | 79/300K | 21/50K | 57/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW+po+ctrl | 158/300K | 1/50K | 109/100K | 0/50K | 36/50K | 12/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRRWIW+po+ctrl | 136/300K | 124/50K | 5/100K | 2/50K | 0/50K | 5/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: shared | IRRWIW+po+ctrl | 83/250K | --- | 37/100K | 0/50K | 8/50K | 38/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW+po+ctrl | 93/300K | 40/50K | 51/100K | 0/50K | 0/50K | 2/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRRWIW+po+ctrl | 57/300K | 51/50K | 4/100K | 1/50K | 0/50K | 1/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW+po+ctrl | 63/300K | 24/50K | 37/100K | 0/50K | 1/50K | 1/50K |
| P0 |warp P2 |cta P1 |warp P3 | x: global, y: global | IRRWIW+po+ctrl | 82/300K | 30/50K | 52/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW+po+ctrl | 70/300K | 23/50K | 46/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW+po+ctrl | 68/300K | 20/50K | 47/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW+po+ctrl | 86/300K | 27/50K | 54/100K | 0/50K | 4/50K | 1/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+po+data | 79/300K | 19/50K | 57/100K | 0/50K | 2/50K | 1/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+po+data | 109/300K | 60/50K | 42/100K | 0/50K | 5/50K | 2/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+po+data | 112/300K | 31/50K | 81/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW+po+data | 86/300K | 31/50K | 55/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW+po+data | 230/300K | 23/50K | 134/100K | 9/50K | 42/50K | 22/50K |
| P0 |cta P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW+po+data | 87/300K | 27/50K | 59/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+po+data | 72/300K | 21/50K | 50/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+po+data | 97/300K | 34/50K | 58/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+po+data | 100/300K | 20/50K | 74/100K | 0/50K | 4/50K | 2/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW+po+data | 108/300K | 30/50K | 77/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW+po+data | 162/300K | 8/50K | 107/100K | 0/50K | 45/50K | 2/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRRWIW+po+data | 149/300K | 124/50K | 11/100K | 1/50K | 0/50K | 13/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: shared | IRRWIW+po+data | 114/250K | --- | 51/100K | 0/50K | 16/50K | 47/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW+po+data | 81/300K | 29/50K | 49/100K | 0/50K | 1/50K | 2/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRRWIW+po+data | 75/300K | 64/50K | 8/100K | 0/50K | 0/50K | 3/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW+po+data | 71/300K | 20/50K | 46/100K | 0/50K | 3/50K | 2/50K |
| P0 |warp P2 |cta P1 |warp P3 | x: global, y: global | IRRWIW+po+data | 83/300K | 26/50K | 56/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW+po+data | 74/300K | 30/50K | 40/100K | 0/50K | 2/50K | 2/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW+po+data | 86/300K | 31/50K | 50/100K | 0/50K | 1/50K | 4/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW+po+data | 78/300K | 20/50K | 57/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+po+membar.cta | 300/300K | 80/50K | 123/100K | 0/50K | 96/50K | 1/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+po+membar.cta | 295/300K | 97/50K | 144/100K | 0/50K | 54/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+po+membar.cta | 405/300K | 60/50K | 181/100K | 0/50K | 162/50K | 2/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW+po+membar.cta | 323/300K | 96/50K | 181/100K | 0/50K | 42/50K | 4/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW+po+membar.cta | 213/300K | 14/50K | 136/100K | 0/50K | 50/50K | 13/50K |
| P0 |cta P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW+po+membar.cta | 227/300K | 81/50K | 117/100K | 0/50K | 28/50K | 1/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+po+membar.cta | 313/300K | 66/50K | 124/100K | 0/50K | 119/50K | 4/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+po+membar.cta | 282/300K | 77/50K | 133/100K | 0/50K | 69/50K | 3/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+po+membar.cta | 444/300K | 49/50K | 207/100K | 0/50K | 187/50K | 1/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW+po+membar.cta | 294/300K | 72/50K | 187/100K | 0/50K | 32/50K | 3/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW+po+membar.cta | 164/300K | 3/50K | 114/100K | 0/50K | 44/50K | 3/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRRWIW+po+membar.cta | 370/300K | 248/50K | 110/100K | 1/50K | 3/50K | 8/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: shared | IRRWIW+po+membar.cta | 117/250K | --- | 47/100K | 0/50K | 12/50K | 58/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW+po+membar.cta | 197/300K | 52/50K | 112/100K | 0/50K | 30/50K | 3/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRRWIW+po+membar.cta | 213/300K | 130/50K | 77/100K | 0/50K | 3/50K | 3/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW+po+membar.cta | 273/300K | 59/50K | 116/100K | 0/50K | 96/50K | 2/50K |
| P0 |warp P2 |cta P1 |warp P3 | x: global, y: global | IRRWIW+po+membar.cta | 226/300K | 85/50K | 121/100K | 0/50K | 19/50K | 1/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW+po+membar.cta | 254/300K | 77/50K | 126/100K | 0/50K | 51/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW+po+membar.cta | 256/300K | 69/50K | 103/100K | 0/50K | 83/50K | 1/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW+po+membar.cta | 419/300K | 68/50K | 188/100K | 0/50K | 160/50K | 3/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+po+membar.gl | 11/300K | 11/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+po+membar.gl | 35/300K | 35/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+po+membar.gl | 13/300K | 13/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW+po+membar.gl | 23/300K | 23/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW+po+membar.gl | 18/300K | 13/50K | 0/100K | 0/50K | 0/50K | 5/50K |
| P0 |cta P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW+po+membar.gl | 14/300K | 14/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW+po+membar.gl | 14/300K | 14/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW+po+membar.gl | 25/300K | 25/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW+po+membar.gl | 10/300K | 10/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW+po+membar.gl | 23/300K | 22/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW+po+membar.gl | 3/300K | 2/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRRWIW+po+membar.gl | 62/300K | 62/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: shared | IRRWIW+po+membar.gl | 13/250K | --- | 0/100K | 0/50K | 0/50K | 13/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW+po+membar.gl | 9/300K | 9/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRRWIW+po+membar.gl | 29/300K | 29/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW+po+membar.gl | 25/300K | 25/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 |warp P3 | x: global, y: global | IRRWIW+po+membar.gl | 20/300K | 20/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW+po+membar.gl | 27/300K | 27/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW+po+membar.gl | 11/300K | 11/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW+po+membar.gl | 22/300K | 22/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW | 384/300K | 76/50K | 168/100K | 0/50K | 112/50K | 28/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW | 345/300K | 90/50K | 159/100K | 0/50K | 66/50K | 30/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW | 573/300K | 78/50K | 222/100K | 0/50K | 235/50K | 38/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW | 451/300K | 94/50K | 260/100K | 0/50K | 38/50K | 59/50K |
| P0 |cta P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW | 350/300K | 16/50K | 167/100K | 48/50K | 89/50K | 30/50K |
| P0 |cta P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW | 327/300K | 95/50K | 170/100K | 0/50K | 40/50K | 22/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRRWIW | 426/300K | 57/50K | 164/100K | 0/50K | 187/50K | 18/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRRWIW | 374/300K | 89/50K | 148/100K | 0/50K | 102/50K | 35/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRRWIW | 620/300K | 74/50K | 214/100K | 0/50K | 274/50K | 58/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: global | IRRWIW | 390/300K | 65/50K | 226/100K | 0/50K | 57/50K | 42/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: global, y: shared | IRRWIW | 318/300K | 5/50K | 166/100K | 55/50K | 78/50K | 14/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRRWIW | 594/300K | 253/50K | 176/100K | 62/50K | 7/50K | 96/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: shared | IRRWIW | 183/250K | --- | 94/100K | 0/50K | 30/50K | 59/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: global, y: global | IRRWIW | 226/300K | 63/50K | 120/100K | 0/50K | 31/50K | 12/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRRWIW | 330/300K | 129/50K | 93/100K | 74/50K | 5/50K | 29/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRRWIW | 358/300K | 65/50K | 135/100K | 0/50K | 124/50K | 34/50K |
| P0 |warp P2 |cta P1 |warp P3 | x: global, y: global | IRRWIW | 318/300K | 94/50K | 168/100K | 0/50K | 25/50K | 31/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRRWIW | 337/300K | 72/50K | 143/100K | 0/50K | 85/50K | 37/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRRWIW | 354/300K | 70/50K | 155/100K | 0/50K | 108/50K | 21/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRRWIW | 559/300K | 77/50K | 217/100K | 0/50K | 231/50K | 34/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRWIW+addr+po | 11/300K | 0/50K | 0/100K | 0/50K | 11/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRWIW+addr+po | 7/300K | 0/50K | 0/100K | 0/50K | 7/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRWIW+addr+po | 4/300K | 0/50K | 1/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRWIW+addr+po | 12/300K | 0/50K | 1/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRWIW+addr+po | 12/300K | 0/50K | 0/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRWIW+addr+po | 4/300K | 0/50K | 2/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRWIW+addr+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRWIW+addr+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRWIW+addr+po | 23/300K | 0/50K | 0/100K | 0/50K | 23/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRWIW+addr+po | 12/300K | 0/50K | 1/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRWIW+addr+po | 8/300K | 0/50K | 0/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRWIW+addr+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRWIW+ctrl+po | 11/300K | 0/50K | 0/100K | 0/50K | 11/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRWIW+ctrl+po | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRWIW+ctrl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRWIW+ctrl+po | 15/300K | 0/50K | 0/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRWIW+ctrl+po | 15/300K | 0/50K | 0/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRWIW+ctrl+po | 7/300K | 0/50K | 0/100K | 0/50K | 7/50K | 0/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRWIW+ctrl+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRWIW+ctrl+po | 20/300K | 0/50K | 0/100K | 0/50K | 20/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRWIW+ctrl+po | 12/300K | 0/50K | 0/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRWIW+ctrl+po | 11/300K | 0/50K | 0/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRWIW+ctrl+po | 3/300K | 0/50K | 1/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRWIW+data+po | 9/300K | 0/50K | 0/100K | 0/50K | 9/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRWIW+data+po | 10/300K | 0/50K | 1/100K | 0/50K | 9/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRWIW+data+po | 9/300K | 0/50K | 1/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRWIW+data+po | 20/300K | 0/50K | 2/100K | 0/50K | 18/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRWIW+data+po | 16/300K | 0/50K | 1/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRWIW+data+po | 4/300K | 0/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRWIW+data+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRWIW+data+po | 29/300K | 0/50K | 0/100K | 29/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRWIW+data+po | 19/300K | 0/50K | 1/100K | 0/50K | 18/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRWIW+data+po | 10/300K | 0/50K | 2/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRWIW+data+po | 11/300K | 0/50K | 2/100K | 0/50K | 9/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRWIW+data+po | 7/300K | 0/50K | 1/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRWIW+membar.cta+data | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRWIW+membar.cta+po | 39/300K | 0/50K | 0/100K | 0/50K | 39/50K | 0/50K |
| P0 |cta P1 |cta P2 |warp P3 | x: global, y: global | IRWIW+membar.cta+po | 21/300K | 0/50K | 2/100K | 0/50K | 19/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRWIW+membar.cta+po | 22/300K | 0/50K | 0/100K | 0/50K | 22/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRWIW+membar.cta+po | 44/300K | 0/50K | 1/100K | 0/50K | 43/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRWIW+membar.cta+po | 30/300K | 0/50K | 3/100K | 0/50K | 27/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRWIW+membar.cta+po | 21/300K | 0/50K | 1/100K | 0/50K | 20/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRWIW+membar.cta+po | 39/300K | 0/50K | 0/100K | 0/50K | 39/50K | 0/50K |
| P0 |warp P2 |warp P3 |cta P1 | x: global, y: global | IRWIW+membar.cta+po | 37/300K | 0/50K | 1/100K | 0/50K | 36/50K | 0/50K |
| P0 |warp P3 |cta P1 |cta P2 | x: global, y: global | IRWIW+membar.cta+po | 34/300K | 0/50K | 2/100K | 0/50K | 32/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRWIW+membar.cta+po | 25/300K | 0/50K | 2/100K | 0/50K | 23/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRWIW+membar.ctas | 28/300K | 0/50K | 1/100K | 0/50K | 27/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRWIW+membar.ctas | 17/300K | 0/50K | 0/100K | 0/50K | 17/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRWIW+membar.ctas | 38/300K | 0/50K | 2/100K | 0/50K | 36/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRWIW+membar.ctas | 28/300K | 0/50K | 0/100K | 0/50K | 28/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRWIW+membar.ctas | 14/300K | 0/50K | 0/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRWIW+membar.ctas | 36/300K | 0/50K | 0/100K | 0/50K | 36/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRWIW+membar.ctas | 10/300K | 0/50K | 0/100K | 0/50K | 10/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRWIW+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 |cta P3 | x: global, y: global | IRWIW | 50/300K | 0/50K | 1/100K | 0/50K | 49/50K | 0/50K |
| P0 |cta P1 |warp P2 |cta P3 | x: global, y: global | IRWIW | 28/300K | 0/50K | 1/100K | 0/50K | 27/50K | 0/50K |
| P0 |warp P1 |cta P2 |cta P3 | x: global, y: global | IRWIW | 54/300K | 0/50K | 1/100K | 0/50K | 53/50K | 0/50K |
| P0 |warp P1 |cta P2 |warp P3 | x: global, y: global | IRWIW | 42/300K | 0/50K | 1/100K | 0/50K | 41/50K | 0/50K |
| P0 |warp P1 |warp P2 |cta P3 | x: global, y: global | IRWIW | 35/300K | 0/50K | 0/100K | 0/50K | 35/50K | 0/50K |
| P0 |warp P1 |warp P2 |warp P3 | x: shared, y: global | IRWIW | 48/300K | 0/50K | 0/100K | 48/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P3 |cta P2 | x: shared, y: global | IRWIW | 71/300K | 0/50K | 0/100K | 71/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 |cta P3 | x: global, y: global | IRWIW | 52/300K | 0/50K | 1/100K | 0/50K | 51/50K | 0/50K |
| P0 |warp P3 |cta P1 |warp P2 | x: global, y: global | IRWIW | 28/300K | 0/50K | 1/100K | 0/50K | 27/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+addr+membar.cta | 25/300K | 0/50K | 1/100K | 0/50K | 24/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+addr+membar.cta | 16/300K | 0/50K | 1/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+addr+membar.cta | 37/300K | 0/50K | 1/100K | 0/50K | 36/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.cta+addr+membar.cta | 15/300K | 0/50K | 1/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.cta+addr+membar.cta | 16/300K | 0/50K | 2/100K | 0/50K | 14/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+addr+po | 137/300K | 16/50K | 64/100K | 0/50K | 51/50K | 6/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+addr+po | 59/300K | 21/50K | 35/100K | 0/50K | 1/50K | 2/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+addr+po | 147/300K | 133/50K | 4/100K | 0/50K | 0/50K | 10/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+addr+po | 108/300K | 44/50K | 33/100K | 0/50K | 26/50K | 5/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+addr+po | 317/300K | 59/50K | 156/100K | 0/50K | 79/50K | 23/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+addr+po | 94/300K | 44/50K | 40/100K | 0/50K | 2/50K | 8/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+addr+po | 158/300K | 124/50K | 9/100K | 0/50K | 0/50K | 25/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+membar.cta+addr+po | 232/300K | 8/50K | 179/100K | 0/50K | 20/50K | 25/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+membar.cta+addr+po | 54/250K | --- | 21/100K | 0/50K | 1/50K | 32/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+membar.cta+addr+po | 378/300K | 245/50K | 64/100K | 0/50K | 9/50K | 60/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+membar.cta+addr+po | 173/250K | --- | 76/100K | 0/50K | 0/50K | 97/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+membar.cta+addr+po | 203/250K | --- | 120/100K | 0/50K | 79/50K | 4/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.cta+addr+po | 108/150K | --- | --- | 0/50K | 7/50K | 101/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.cta+addr+po | 146/300K | 26/50K | 79/100K | 0/50K | 33/50K | 8/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.cta+addr+po | 403/300K | 182/50K | 139/100K | 0/50K | 76/50K | 6/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+ctrl+membar.cta | 14/300K | 0/50K | 0/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+ctrl+membar.cta | 14/300K | 0/50K | 0/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+ctrl+membar.cta | 39/300K | 0/50K | 3/100K | 0/50K | 36/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.cta+ctrl+membar.cta | 12/300K | 0/50K | 2/100K | 0/50K | 10/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.cta+ctrl+membar.cta | 18/300K | 0/50K | 1/100K | 0/50K | 17/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+ctrl+po | 104/300K | 29/50K | 41/100K | 0/50K | 32/50K | 2/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+ctrl+po | 45/300K | 28/50K | 16/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+ctrl+po | 131/300K | 121/50K | 3/100K | 0/50K | 0/50K | 7/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+ctrl+po | 97/300K | 29/50K | 33/100K | 0/50K | 34/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+ctrl+po | 299/300K | 63/50K | 143/100K | 0/50K | 78/50K | 15/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+ctrl+po | 67/300K | 20/50K | 38/100K | 0/50K | 1/50K | 8/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+ctrl+po | 377/300K | 349/50K | 2/100K | 0/50K | 0/50K | 26/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+membar.cta+ctrl+po | 234/300K | 18/50K | 178/100K | 0/50K | 15/50K | 23/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+membar.cta+ctrl+po | 51/250K | --- | 8/100K | 1/50K | 1/50K | 41/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+membar.cta+ctrl+po | 381/300K | 243/50K | 63/100K | 0/50K | 13/50K | 62/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+membar.cta+ctrl+po | 166/250K | --- | 58/100K | 0/50K | 0/50K | 108/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+membar.cta+ctrl+po | 229/250K | --- | 144/100K | 0/50K | 84/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.cta+ctrl+po | 196/150K | --- | --- | 0/50K | 4/50K | 192/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.cta+ctrl+po | 118/300K | 21/50K | 54/100K | 0/50K | 37/50K | 6/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.cta+ctrl+po | 415/300K | 177/50K | 140/100K | 0/50K | 91/50K | 7/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+data+membar.cta | 22/300K | 0/50K | 1/100K | 0/50K | 21/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+data+membar.cta | 19/300K | 0/50K | 1/100K | 0/50K | 18/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+data+membar.cta | 42/300K | 0/50K | 5/100K | 0/50K | 37/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.cta+data+membar.cta | 15/300K | 0/50K | 1/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.cta+data+membar.cta | 20/300K | 0/50K | 2/100K | 0/50K | 18/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+data+po | 122/300K | 36/50K | 50/100K | 0/50K | 31/50K | 5/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+data+po | 42/300K | 22/50K | 19/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+data+po | 113/300K | 99/50K | 7/100K | 0/50K | 0/50K | 7/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+data+po | 134/300K | 45/50K | 41/100K | 0/50K | 42/50K | 6/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+data+po | 325/300K | 68/50K | 163/100K | 0/50K | 78/50K | 16/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+data+po | 95/300K | 39/50K | 45/100K | 0/50K | 2/50K | 9/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+data+po | 363/300K | 330/50K | 11/100K | 0/50K | 0/50K | 22/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+membar.cta+data+po | 248/300K | 8/50K | 200/100K | 0/50K | 13/50K | 27/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+membar.cta+data+po | 55/250K | --- | 24/100K | 0/50K | 2/50K | 29/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+membar.cta+data+po | 413/300K | 273/50K | 65/100K | 0/50K | 12/50K | 63/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+membar.cta+data+po | 176/250K | --- | 75/100K | 0/50K | 0/50K | 101/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+membar.cta+data+po | 214/250K | --- | 139/100K | 0/50K | 68/50K | 7/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.cta+data+po | 140/150K | --- | --- | 0/50K | 9/50K | 131/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.cta+data+po | 126/300K | 21/50K | 71/100K | 0/50K | 29/50K | 5/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.cta+data+po | 386/300K | 142/50K | 139/100K | 0/50K | 98/50K | 7/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+membar.cta+addr | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+membar.cta+po | 239/300K | 42/50K | 75/100K | 0/50K | 120/50K | 2/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+membar.cta+po | 122/300K | 57/50K | 41/100K | 0/50K | 17/50K | 7/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+membar.cta+po | 234/300K | 169/50K | 53/100K | 0/50K | 3/50K | 9/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+membar.cta+po | 274/300K | 71/50K | 65/100K | 0/50K | 135/50K | 3/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+membar.cta+po | 321/300K | 55/50K | 172/100K | 0/50K | 81/50K | 13/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+membar.cta+po | 187/300K | 100/50K | 68/100K | 0/50K | 9/50K | 10/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+membar.cta+po | 546/300K | 466/50K | 60/100K | 0/50K | 1/50K | 19/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+membar.cta+membar.cta+po | 245/300K | 10/50K | 198/100K | 0/50K | 14/50K | 23/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+membar.cta+membar.cta+po | 52/250K | --- | 21/100K | 0/50K | 2/50K | 29/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+membar.cta+membar.cta+po | 681/300K | 402/50K | 129/100K | 0/50K | 85/50K | 65/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+membar.cta+membar.cta+po | 418/250K | --- | 306/100K | 0/50K | 13/50K | 99/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+membar.cta+membar.cta+po | 215/250K | --- | 123/100K | 0/50K | 84/50K | 8/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.cta+membar.cta+po | 117/150K | --- | --- | 0/50K | 12/50K | 105/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.cta+membar.cta+po | 270/300K | 35/50K | 117/100K | 0/50K | 114/50K | 4/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.cta+membar.cta+po | 702/300K | 231/50K | 240/100K | 0/50K | 227/50K | 4/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+membar.gl+po | 14/300K | 13/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+membar.gl+po | 18/300K | 18/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+membar.gl+po | 110/300K | 105/50K | 0/100K | 0/50K | 0/50K | 5/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+membar.gl+po | 23/300K | 22/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+membar.gl+po | 27/300K | 26/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+membar.gl+po | 36/300K | 35/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+membar.gl+po | 142/300K | 123/50K | 0/100K | 0/50K | 0/50K | 19/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+membar.cta+membar.gl+po | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+membar.cta+membar.gl+po | 15/250K | --- | 0/100K | 0/50K | 0/50K | 15/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+membar.cta+membar.gl+po | 190/300K | 189/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+membar.cta+membar.gl+po | 88/250K | --- | 0/100K | 0/50K | 0/50K | 88/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.cta+membar.gl+po | 71/150K | --- | --- | 0/50K | 0/50K | 71/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.cta+membar.gl+po | 12/300K | 12/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.cta+membar.gl+po | 141/300K | 138/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+po+addr | 76/300K | 0/50K | 0/100K | 74/50K | 0/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+po+addr | 14/300K | 0/50K | 0/100K | 14/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+po+addr | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+po+ctrl | 68/300K | 0/50K | 0/100K | 65/50K | 0/50K | 3/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+po+ctrl | 8/300K | 0/50K | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+po+ctrl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+po+membar.cta | 56/300K | 0/50K | 2/100K | 0/50K | 54/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+po+membar.cta | 2/300K | 0/50K | 0/100K | 1/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+po+membar.cta | 47/300K | 0/50K | 3/100K | 0/50K | 44/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+po+membar.cta | 38/300K | 0/50K | 4/100K | 1/50K | 33/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.cta+po+membar.cta | 52/300K | 0/50K | 3/100K | 0/50K | 49/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.cta+po+membar.cta | 57/300K | 0/50K | 10/100K | 0/50K | 47/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+po+membar.gl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+po+po | 262/300K | 44/50K | 89/100K | 0/50K | 117/50K | 12/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+po+po | 132/300K | 49/50K | 55/100K | 0/50K | 14/50K | 14/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+po+po | 427/300K | 216/50K | 57/100K | 108/50K | 3/50K | 43/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.cta+po+po | 405/300K | 94/50K | 105/100K | 0/50K | 189/50K | 17/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.cta+po+po | 367/300K | 65/50K | 176/100K | 15/50K | 84/50K | 27/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.cta+po+po | 202/300K | 89/50K | 76/100K | 0/50K | 8/50K | 29/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.cta+po+po | 907/300K | 725/50K | 60/100K | 26/50K | 0/50K | 96/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+membar.cta+po+po | 293/300K | 13/50K | 199/100K | 27/50K | 24/50K | 30/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+membar.cta+po+po | 65/250K | --- | 15/100K | 2/50K | 5/50K | 43/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+membar.cta+po+po | 775/300K | 403/50K | 127/100K | 1/50K | 66/50K | 178/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+membar.cta+po+po | 676/250K | --- | 336/100K | 0/50K | 29/50K | 311/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+membar.cta+po+po | 264/250K | --- | 145/100K | 17/50K | 101/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.cta+po+po | 177/150K | --- | --- | 0/50K | 17/50K | 160/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.cta+po+po | 308/300K | 45/50K | 130/100K | 0/50K | 120/50K | 13/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.cta+po+po | 800/300K | 283/50K | 258/100K | 1/50K | 245/50K | 13/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.ctas | 58/300K | 0/50K | 2/100K | 0/50K | 56/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.ctas | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.ctas | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.ctas | 45/300K | 0/50K | 1/100K | 0/50K | 44/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.ctas | 44/300K | 0/50K | 4/100K | 0/50K | 40/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.ctas | 55/300K | 0/50K | 4/100K | 0/50K | 51/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.ctas | 73/300K | 0/50K | 9/100K | 0/50K | 64/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+addr+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.gl+addr+po | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+membar.gl+addr+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+membar.gl+addr+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.gl+addr+po | 4/150K | --- | --- | 0/50K | 0/50K | 4/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+ctrl+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.gl+ctrl+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+ctrl+po | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.gl+ctrl+po | 1/150K | --- | --- | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+data+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.gl+data+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.gl+data+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+data+po | 4/300K | 4/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+membar.gl+data+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+membar.gl+data+po | 1/250K | --- | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+membar.gl+data+po | 3/300K | 3/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.gl+data+po | 1/150K | --- | --- | 0/50K | 0/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.gl+data+po | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.gl+membar.cta+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+membar.cta+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.gl+membar.cta+po | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.gl+membar.cta+po | 6/300K | 6/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.gl+membar.cta+po | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+membar.cta+po | 4/300K | 4/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+membar.gl+membar.cta+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+membar.gl+membar.cta+po | 2/250K | --- | 0/100K | 0/50K | 0/50K | 2/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+membar.gl+membar.cta+po | 4/300K | 4/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+membar.gl+membar.cta+po | 1/250K | --- | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.gl+membar.cta+po | 2/150K | --- | --- | 0/50K | 0/50K | 2/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.gl+membar.cta+po | 3/300K | 3/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+membar.gl+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+membar.gl+membar.gl+po | 1/250K | --- | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.gl+membar.gl+po | 4/150K | --- | --- | 0/50K | 0/50K | 4/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+po+addr | 16/300K | 0/50K | 0/100K | 16/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+po+ctrl | 15/300K | 0/50K | 0/100K | 15/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+membar.gl+po+po | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+membar.gl+po+po | 3/300K | 3/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+po+po | 44/300K | 2/50K | 0/100K | 36/50K | 0/50K | 6/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+membar.gl+po+po | 4/300K | 4/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+membar.gl+po+po | 2/300K | 1/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+membar.gl+po+po | 16/300K | 4/50K | 0/100K | 12/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+membar.gl+po+po | 1/250K | --- | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+membar.gl+po+po | 6/300K | 6/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+membar.gl+po+po | 3/250K | --- | 0/100K | 0/50K | 0/50K | 3/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+membar.gl+po+po | 16/250K | --- | 0/100K | 16/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+membar.gl+po+po | 4/150K | --- | --- | 0/50K | 0/50K | 4/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+membar.gl+po+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+membar.gl+po+po | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+addr+addr | 9/300K | 0/50K | 0/100K | 9/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+addr+ctrl | 6/300K | 0/50K | 0/100K | 6/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+po+addr+membar.cta | 21/300K | 0/50K | 2/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+po+addr+membar.cta | 25/300K | 0/50K | 3/100K | 0/50K | 22/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+addr+membar.cta | 64/300K | 0/50K | 4/100K | 2/50K | 58/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+po+addr+membar.cta | 16/300K | 0/50K | 0/100K | 0/50K | 16/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+po+addr+membar.cta | 25/300K | 0/50K | 1/100K | 0/50K | 24/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+po+addr+po | 174/300K | 28/50K | 77/100K | 0/50K | 53/50K | 16/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+po+addr+po | 86/300K | 29/50K | 39/100K | 0/50K | 9/50K | 9/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+addr+po | 192/300K | 150/50K | 17/100K | 0/50K | 0/50K | 25/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+po+addr+po | 188/300K | 41/50K | 76/100K | 0/50K | 60/50K | 11/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+addr+po | 449/300K | 86/50K | 165/100K | 30/50K | 144/50K | 24/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+po+addr+po | 144/300K | 60/50K | 50/100K | 0/50K | 4/50K | 30/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+addr+po | 294/300K | 184/50K | 11/100K | 0/50K | 0/50K | 99/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+po+addr+po | 353/300K | 12/50K | 222/100K | 50/50K | 35/50K | 34/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+po+addr+po | 128/250K | --- | 24/100K | 16/50K | 6/50K | 82/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+po+addr+po | 503/300K | 255/50K | 92/100K | 10/50K | 26/50K | 120/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+po+addr+po | 367/250K | --- | 132/100K | 0/50K | 0/50K | 235/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+po+addr+po | 273/250K | --- | 155/100K | 0/50K | 107/50K | 11/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+po+addr+po | 189/150K | --- | --- | 0/50K | 29/50K | 160/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+po+addr+po | 214/300K | 29/50K | 112/100K | 0/50K | 51/50K | 22/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+po+addr+po | 738/300K | 180/50K | 215/100K | 209/50K | 114/50K | 20/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+ctrl+addr | 10/300K | 0/50K | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+ctrl+ctrl | 10/300K | 0/50K | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+po+ctrl+membar.cta | 31/300K | 0/50K | 2/100K | 0/50K | 29/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+po+ctrl+membar.cta | 22/300K | 0/50K | 1/100K | 0/50K | 21/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+ctrl+membar.cta | 50/300K | 0/50K | 7/100K | 1/50K | 42/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+po+ctrl+membar.cta | 17/300K | 0/50K | 1/100K | 0/50K | 16/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+po+ctrl+membar.cta | 31/300K | 0/50K | 2/100K | 0/50K | 29/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+po+ctrl+po | 159/300K | 29/50K | 71/100K | 0/50K | 52/50K | 7/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+po+ctrl+po | 83/300K | 41/50K | 32/100K | 0/50K | 2/50K | 8/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+ctrl+po | 179/300K | 141/50K | 13/100K | 0/50K | 0/50K | 25/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+po+ctrl+po | 170/300K | 36/50K | 57/100K | 0/50K | 68/50K | 9/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+ctrl+po | 456/300K | 74/50K | 174/100K | 23/50K | 151/50K | 34/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+po+ctrl+po | 111/300K | 24/50K | 49/100K | 0/50K | 7/50K | 31/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+ctrl+po | 541/300K | 417/50K | 12/100K | 0/50K | 0/50K | 112/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+po+ctrl+po | 316/300K | 13/50K | 187/100K | 55/50K | 25/50K | 36/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+po+ctrl+po | 152/250K | --- | 21/100K | 12/50K | 1/50K | 118/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+po+ctrl+po | 512/300K | 278/50K | 100/100K | 10/50K | 21/50K | 103/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+po+ctrl+po | 303/250K | --- | 93/100K | 1/50K | 0/50K | 209/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+po+ctrl+po | 271/250K | --- | 168/100K | 0/50K | 97/50K | 6/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+po+ctrl+po | 215/150K | --- | --- | 0/50K | 26/50K | 189/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+po+ctrl+po | 219/300K | 22/50K | 122/100K | 0/50K | 45/50K | 30/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+po+ctrl+po | 697/300K | 189/50K | 199/100K | 179/50K | 121/50K | 9/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+data+addr | 25/300K | 0/50K | 0/100K | 25/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+data+ctrl | 14/300K | 0/50K | 0/100K | 14/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+po+data+membar.cta | 20/300K | 0/50K | 0/100K | 0/50K | 20/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+po+data+membar.cta | 26/300K | 0/50K | 2/100K | 0/50K | 24/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+data+membar.cta | 67/300K | 0/50K | 5/100K | 4/50K | 58/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+po+data+membar.cta | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+po+data+membar.cta | 20/300K | 0/50K | 2/100K | 0/50K | 18/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+po+data+membar.cta | 19/300K | 0/50K | 2/100K | 0/50K | 17/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+data+membar.gl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+po+data+po | 179/300K | 35/50K | 67/100K | 0/50K | 66/50K | 11/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+po+data+po | 99/300K | 38/50K | 46/100K | 0/50K | 2/50K | 13/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+data+po | 170/300K | 119/50K | 27/100K | 0/50K | 0/50K | 24/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+po+data+po | 194/300K | 50/50K | 76/100K | 0/50K | 58/50K | 10/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+data+po | 487/300K | 75/50K | 194/100K | 48/50K | 130/50K | 40/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+po+data+po | 165/300K | 57/50K | 69/100K | 0/50K | 6/50K | 33/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+data+po | 513/300K | 386/50K | 21/100K | 0/50K | 0/50K | 106/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+po+data+po | 381/300K | 20/50K | 229/100K | 74/50K | 33/50K | 25/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+po+data+po | 151/250K | --- | 25/100K | 21/50K | 2/50K | 103/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+po+data+po | 556/300K | 280/50K | 128/100K | 11/50K | 23/50K | 114/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+po+data+po | 319/250K | --- | 120/100K | 0/50K | 0/50K | 199/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+po+data+po | 317/250K | --- | 175/100K | 15/50K | 120/50K | 7/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+po+data+po | 196/150K | --- | --- | 0/50K | 30/50K | 166/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+po+data+po | 235/300K | 31/50K | 113/100K | 0/50K | 50/50K | 41/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+po+data+po | 771/300K | 222/50K | 209/100K | 182/50K | 143/50K | 15/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+po+membar.cta+addr | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+po+membar.cta+ctrl | 2/250K | --- | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+po+membar.cta+membar.cta | 61/300K | 0/50K | 2/100K | 0/50K | 59/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+membar.cta+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+po+membar.cta+membar.cta | 61/300K | 0/50K | 2/100K | 0/50K | 59/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+membar.cta+membar.cta | 67/300K | 0/50K | 5/100K | 0/50K | 62/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+po+membar.cta+membar.cta | 71/300K | 0/50K | 4/100K | 0/50K | 67/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+po+membar.cta+membar.cta | 80/300K | 0/50K | 4/100K | 0/50K | 76/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+po+membar.cta+po | 358/300K | 59/50K | 146/100K | 0/50K | 147/50K | 6/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+po+membar.cta+po | 155/300K | 52/50K | 63/100K | 0/50K | 26/50K | 14/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+membar.cta+po | 286/300K | 155/50K | 103/100K | 0/50K | 8/50K | 20/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+po+membar.cta+po | 398/300K | 67/50K | 124/100K | 0/50K | 192/50K | 15/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+membar.cta+po | 465/300K | 77/50K | 195/100K | 1/50K | 158/50K | 34/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+po+membar.cta+po | 281/300K | 102/50K | 115/100K | 0/50K | 25/50K | 39/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+membar.cta+po | 776/300K | 593/50K | 75/100K | 0/50K | 5/50K | 103/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+po+membar.cta+po | 335/300K | 14/50K | 240/100K | 1/50K | 41/50K | 39/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+po+membar.cta+po | 124/250K | --- | 25/100K | 6/50K | 9/50K | 84/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+po+membar.cta+po | 837/300K | 433/50K | 189/100K | 0/50K | 97/50K | 118/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+po+membar.cta+po | 722/250K | --- | 468/100K | 0/50K | 22/50K | 232/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+po+membar.cta+po | 311/250K | --- | 162/100K | 0/50K | 139/50K | 10/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+po+membar.cta+po | 190/150K | --- | --- | 0/50K | 26/50K | 164/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+po+membar.cta+po | 451/300K | 59/50K | 192/100K | 0/50K | 165/50K | 35/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+po+membar.cta+po | 946/300K | 274/50K | 300/100K | 26/50K | 335/50K | 11/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+po+membar.gl+po | 19/300K | 19/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+po+membar.gl+po | 28/300K | 27/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+membar.gl+po | 115/300K | 95/50K | 0/100K | 0/50K | 0/50K | 20/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+po+membar.gl+po | 47/300K | 42/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+membar.gl+po | 41/300K | 39/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2+po+membar.gl+po | 37/300K | 35/50K | 0/100K | 0/50K | 0/50K | 2/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+membar.gl+po | 246/300K | 161/50K | 0/100K | 0/50K | 0/50K | 85/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2+po+membar.gl+po | 10/300K | 9/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+po+membar.gl+po | 50/250K | --- | 0/100K | 1/50K | 0/50K | 49/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+po+membar.gl+po | 194/300K | 193/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+po+membar.gl+po | 211/250K | --- | 0/100K | 0/50K | 0/50K | 211/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2+po+membar.gl+po | 86/150K | --- | --- | 0/50K | 0/50K | 86/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+po+membar.gl+po | 24/300K | 24/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+po+membar.gl+po | 136/300K | 118/50K | 0/100K | 12/50K | 6/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+po+addr | 250/300K | 0/50K | 0/100K | 248/50K | 0/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+po+addr | 56/300K | 0/50K | 0/100K | 56/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+po+addr | 115/300K | 0/50K | 0/100K | 115/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+po+po+addr | 11/250K | --- | 0/100K | 11/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+po+po+addr | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+po+po+addr | 13/250K | --- | 0/100K | 13/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+po+po+addr | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+po+ctrl | 214/300K | 0/50K | 0/100K | 211/50K | 1/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+po+ctrl | 89/300K | 0/50K | 0/100K | 89/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+po+ctrl | 69/300K | 0/50K | 0/100K | 69/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+po+po+ctrl | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2+po+po+ctrl | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+po+po+ctrl | 7/250K | --- | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+po+po+ctrl | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2+po+po+membar.cta | 61/300K | 0/50K | 4/100K | 0/50K | 57/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2+po+po+membar.cta | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+po+membar.cta | 7/300K | 0/50K | 0/100K | 1/50K | 1/50K | 5/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2+po+po+membar.cta | 74/300K | 0/50K | 6/100K | 0/50K | 68/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+po+membar.cta | 82/300K | 0/50K | 5/100K | 7/50K | 70/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2+po+po+membar.cta | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2+po+po+membar.cta | 12/250K | --- | 0/100K | 12/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2+po+po+membar.cta | 5/250K | --- | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+po+po+membar.cta | 10/250K | --- | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2+po+po+membar.cta | 79/300K | 0/50K | 4/100K | 0/50K | 75/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2+po+po+membar.cta | 84/300K | 0/50K | 5/100K | 0/50K | 79/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2+po+po+membar.gl | 8/300K | 0/50K | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2+po+po+membar.gl | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | ISA2 | 371/300K | 56/50K | 133/100K | 0/50K | 161/50K | 21/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | ISA2 | 273/300K | 99/50K | 100/100K | 0/50K | 35/50K | 39/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | ISA2 | 870/300K | 231/50K | 115/100K | 438/50K | 14/50K | 72/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | ISA2 | 480/300K | 114/50K | 141/100K | 0/50K | 189/50K | 36/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | ISA2 | 611/300K | 72/50K | 190/100K | 159/50K | 150/50K | 40/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | ISA2 | 337/300K | 97/50K | 128/100K | 0/50K | 28/50K | 84/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | ISA2 | 2.0K/300K | 813/50K | 106/100K | 798/50K | 8/50K | 250/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | ISA2 | 604/300K | 28/50K | 245/100K | 226/50K | 57/50K | 48/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | ISA2 | 536/250K | --- | 37/100K | 374/50K | 10/50K | 115/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | ISA2 | 1.4K/300K | 412/50K | 214/100K | 360/50K | 108/50K | 295/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | ISA2 | 1.2K/250K | --- | 456/100K | 186/50K | 51/50K | 551/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | ISA2 | 432/250K | --- | 186/100K | 106/50K | 127/50K | 13/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | ISA2 | 198/150K | --- | --- | 0/50K | 51/50K | 147/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | ISA2 | 529/300K | 59/50K | 220/100K | 0/50K | 192/50K | 58/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | ISA2 | 1.4K/300K | 281/50K | 370/100K | 308/50K | 361/50K | 50/50K |
| P0 |cta P1 | x: global, y: global | LB+addr+po | 513/300K | 0/50K | 166/100K | 0/50K | 347/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | LB+addr+po | 5/300K | 0/50K | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | LB+addr+po | 799/300K | 0/50K | 0/100K | 799/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | LB+ctrl+po | 491/300K | 0/50K | 152/100K | 0/50K | 339/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | LB+ctrl+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | LB+ctrl+po | 859/300K | 0/50K | 0/100K | 859/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | LB+data+po | 517/300K | 0/50K | 168/100K | 0/50K | 349/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | LB+data+po | 8/300K | 0/50K | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | LB+data+po | 897/300K | 0/50K | 0/100K | 897/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | LB+membar.cta+po | 972/300K | 0/50K | 282/100K | 0/50K | 690/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | LB+membar.cta+po | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | LB+membar.cta+po | 23/300K | 0/50K | 0/100K | 23/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | LB+membar.ctas | 780/300K | 0/50K | 194/100K | 0/50K | 586/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | LB+membar.gl+po | 10/300K | 0/50K | 0/100K | 0/50K | 10/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | LB | 979/300K | 0/50K | 255/100K | 18/50K | 706/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | LB | 1.3K/300K | 0/50K | 0/100K | 1.3K/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | MP+membar.cta+po | 8.9K/300K | 1.8K/50K | 3.8K/100K | 4/50K | 2.4K/50K | 877/50K |
| P0 |warp P1 | x: global, y: global | MP+membar.cta+po | 7.4K/300K | 1.6K/50K | 3.0K/100K | 0/50K | 1.5K/50K | 1.4K/50K |
| P0 |warp P1 | x: global, y: shared | MP+membar.cta+po | 9.0K/300K | 2.7K/50K | 3.5K/100K | 291/50K | 331/50K | 2.2K/50K |
| P0 |warp P1 | x: shared, y: global | MP+membar.cta+po | 9.1K/300K | 378/50K | 4.3K/100K | 42/50K | 3.6K/50K | 743/50K |
| P0 |warp P1 | x: shared, y: shared | MP+membar.cta+po | 7.4K/250K | --- | 4.3K/100K | 0/50K | 844/50K | 2.2K/50K |
| P0 |cta P1 | x: global, y: global | MP+membar.ctas | 1.2K/300K | 0/50K | 321/100K | 0/50K | 908/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | MP+membar.gl+membar.cta | 9/300K | 0/50K | 0/100K | 0/50K | 9/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | MP+membar.gl+po | 899/300K | 862/50K | 0/100K | 0/50K | 26/50K | 11/50K |
| P0 |warp P1 | x: global, y: global | MP+membar.gl+po | 807/300K | 786/50K | 0/100K | 0/50K | 0/50K | 21/50K |
| P0 |warp P1 | x: global, y: shared | MP+membar.gl+po | 2.0K/300K | 1.3K/50K | 0/100K | 9/50K | 0/50K | 692/50K |
| P0 |warp P1 | x: shared, y: global | MP+membar.gl+po | 212/300K | 181/50K | 0/100K | 0/50K | 0/50K | 31/50K |
| P0 |warp P1 | x: shared, y: shared | MP+membar.gl+po | 1.1K/250K | --- | 0/100K | 0/50K | 0/50K | 1.1K/50K |
| P0 |warp P1 | x: global, y: shared | MP+po+addr | 273/300K | 0/50K | 0/100K | 273/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | MP+po+ctrl | 226/300K | 0/50K | 0/100K | 226/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | MP+po+membar.cta | 1.3K/300K | 0/50K | 302/100K | 0/50K | 984/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | MP+po+membar.cta | 152/300K | 0/50K | 0/100K | 152/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | MP | 9.8K/300K | 1.9K/50K | 4.1K/100K | 16/50K | 2.7K/50K | 1.1K/50K |
| P0 |warp P1 | x: global, y: global | MP | 7.8K/300K | 1.6K/50K | 3.1K/100K | 0/50K | 1.7K/50K | 1.4K/50K |
| P0 |warp P1 | x: global, y: shared | MP | 13K/300K | 2.9K/50K | 4.1K/100K | 2.7K/50K | 503/50K | 2.4K/50K |
| P0 |warp P1 | x: shared, y: global | MP | 12K/300K | 432/50K | 4.7K/100K | 2.0K/50K | 4.0K/50K | 855/50K |
| P0 |warp P1 | x: shared, y: shared | MP | 8.4K/250K | --- | 4.8K/100K | 0/50K | 1.1K/50K | 2.5K/50K |
| P0 |cta P1 | x: global, y: global | R+membar.cta+po | 1.2K/300K | 0/50K | 230/100K | 6/50K | 1.0K/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | R+membar.cta+po | 266/300K | 0/50K | 0/100K | 266/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | R+membar.cta+po | 51/300K | 0/50K | 0/100K | 51/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | R+membar.ctas | 1.1K/300K | 0/50K | 230/100K | 1/50K | 885/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | R+membar.gl+membar.cta | 9/300K | 0/50K | 0/100K | 0/50K | 9/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | R+membar.gl+po | 4/300K | 0/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | R+membar.gl+po | 9/300K | 0/50K | 0/100K | 9/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | R+po+membar.cta | 1.1K/300K | 0/50K | 226/100K | 4/50K | 906/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | R+po+membar.cta | 215/300K | 0/50K | 0/100K | 215/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | R+po+membar.cta | 157/300K | 0/50K | 0/100K | 157/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | R | 1.2K/300K | 0/50K | 245/100K | 12/50K | 920/50K | 1/50K |
| P0 |warp P1 | x: global, y: shared | R | 3.6K/300K | 0/50K | 0/100K | 3.6K/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | R | 2.1K/300K | 0/50K | 0/100K | 2.1K/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | RWC+addr+membar.cta | 69/300K | 0/50K | 2/100K | 0/50K | 67/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | RWC+addr+membar.cta | 115/300K | 0/50K | 4/100K | 0/50K | 111/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | RWC+addr+membar.cta | 61/300K | 0/50K | 5/100K | 0/50K | 56/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | RWC+addr+po | 89/300K | 0/50K | 3/100K | 0/50K | 86/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | RWC+addr+po | 126/300K | 0/50K | 10/100K | 0/50K | 116/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | RWC+addr+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | RWC+addr+po | 92/300K | 0/50K | 8/100K | 0/50K | 84/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | RWC+ctrl+membar.cta | 65/300K | 0/50K | 3/100K | 0/50K | 62/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | RWC+ctrl+membar.cta | 98/300K | 0/50K | 6/100K | 0/50K | 92/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | RWC+ctrl+membar.cta | 63/300K | 0/50K | 5/100K | 0/50K | 58/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | RWC+ctrl+po | 92/300K | 0/50K | 1/100K | 0/50K | 91/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | RWC+ctrl+po | 119/300K | 0/50K | 7/100K | 0/50K | 112/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | RWC+ctrl+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | RWC+ctrl+po | 76/300K | 0/50K | 5/100K | 0/50K | 71/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | RWC+membar.cta+po | 234/300K | 0/50K | 7/100K | 0/50K | 227/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | RWC+membar.cta+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | RWC+membar.cta+po | 298/300K | 0/50K | 18/100K | 0/50K | 280/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | RWC+membar.cta+po | 220/300K | 0/50K | 21/100K | 0/50K | 199/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | RWC+membar.ctas | 187/300K | 0/50K | 7/100K | 0/50K | 180/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | RWC+membar.ctas | 261/300K | 0/50K | 18/100K | 0/50K | 243/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | RWC+membar.ctas | 182/300K | 0/50K | 9/100K | 0/50K | 173/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | RWC+membar.gl+membar.cta | 11/300K | 0/50K | 0/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | RWC+membar.gl+membar.cta | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | RWC+membar.gl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | RWC+membar.gl+po | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | RWC+membar.gl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | RWC+po+membar.cta | 2.3K/300K | 563/50K | 975/100K | 0/50K | 464/50K | 267/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | RWC+po+membar.cta | 2.5K/300K | 555/50K | 1.2K/100K | 0/50K | 368/50K | 371/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | RWC+po+membar.cta | 2.4K/300K | 255/50K | 1.2K/100K | 0/50K | 491/50K | 490/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | RWC+po+membar.cta | 2.3K/300K | 519/50K | 966/100K | 0/50K | 571/50K | 254/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | RWC+po+membar.cta | 2.2K/300K | 530/50K | 1.0K/100K | 0/50K | 387/50K | 306/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | RWC+po+membar.cta | 1.8K/300K | 50/50K | 862/100K | 4/50K | 468/50K | 397/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | RWC+po+membar.cta | 1.5K/300K | 526/50K | 583/100K | 15/50K | 36/50K | 312/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | RWC+po+membar.cta | 1.8K/250K | --- | 947/100K | 0/50K | 205/50K | 644/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | RWC+po+membar.cta | 2.1K/300K | 541/50K | 900/100K | 0/50K | 403/50K | 264/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | RWC+po+membar.gl | 138/300K | 136/50K | 0/100K | 0/50K | 0/50K | 2/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | RWC+po+membar.gl | 148/300K | 148/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | RWC+po+membar.gl | 220/300K | 104/50K | 0/100K | 0/50K | 0/50K | 116/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | RWC+po+membar.gl | 175/300K | 175/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | RWC+po+membar.gl | 169/300K | 167/50K | 0/100K | 0/50K | 0/50K | 2/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | RWC+po+membar.gl | 76/300K | 11/50K | 0/100K | 0/50K | 0/50K | 65/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | RWC+po+membar.gl | 177/300K | 173/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | RWC+po+membar.gl | 202/250K | --- | 0/100K | 0/50K | 0/50K | 202/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | RWC+po+membar.gl | 154/300K | 153/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | RWC | 2.5K/300K | 594/50K | 1.1K/100K | 0/50K | 536/50K | 297/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | RWC | 2.7K/300K | 548/50K | 1.4K/100K | 0/50K | 431/50K | 397/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | RWC | 3.4K/300K | 304/50K | 1.4K/100K | 570/50K | 614/50K | 570/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | RWC | 2.5K/300K | 508/50K | 1.0K/100K | 1/50K | 688/50K | 286/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | RWC | 2.5K/300K | 529/50K | 1.1K/100K | 0/50K | 485/50K | 377/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | RWC | 3.3K/300K | 242/50K | 982/100K | 971/50K | 578/50K | 522/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | RWC | 2.2K/300K | 541/50K | 690/100K | 531/50K | 58/50K | 348/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | RWC | 2.1K/250K | --- | 1.2K/100K | 0/50K | 229/50K | 684/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | RWC | 2.2K/300K | 519/50K | 921/100K | 0/50K | 481/50K | 276/50K |
| P0 |cta P1 | x: global, y: global | S+membar.cta+po | 1.2K/300K | 0/50K | 234/100K | 1/50K | 1.0K/50K | 2/50K |
| P0 |warp P1 | x: global, y: shared | S+membar.cta+po | 216/300K | 0/50K | 0/100K | 216/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | S+membar.cta+po | 51/300K | 0/50K | 0/100K | 51/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | S+membar.ctas | 1.1K/300K | 0/50K | 222/100K | 0/50K | 924/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | S+membar.gl+membar.cta | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | S+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | S+membar.gl+po | 10/300K | 0/50K | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | S+po+addr | 244/300K | 0/50K | 0/100K | 244/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | S+po+ctrl | 213/300K | 0/50K | 0/100K | 213/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | S+po+data | 470/300K | 0/50K | 0/100K | 470/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | S+po+membar.cta | 1.2K/300K | 0/50K | 224/100K | 0/50K | 936/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | S+po+membar.cta | 130/300K | 0/50K | 0/100K | 130/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | S | 1.3K/300K | 0/50K | 276/100K | 9/50K | 1.0K/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | S | 1.6K/300K | 0/50K | 0/100K | 1.6K/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | S | 1.6K/300K | 0/50K | 0/100K | 1.6K/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | SB+membar.cta+po | 1.3K/300K | 0/50K | 295/100K | 3/50K | 960/50K | 2/50K |
| P0 |warp P1 | x: global, y: shared | SB+membar.cta+po | 405/300K | 0/50K | 0/100K | 405/50K | 0/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | SB+membar.cta+po | 62/300K | 0/50K | 0/100K | 62/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | SB+membar.ctas | 1.2K/300K | 0/50K | 257/100K | 0/50K | 904/50K | 1/50K |
| P0 |cta P1 | x: global, y: global | SB+membar.gl+po | 12/300K | 0/50K | 0/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P1 | x: global, y: shared | SB+membar.gl+po | 16/300K | 0/50K | 0/100K | 16/50K | 0/50K | 0/50K |
| P0 |cta P1 | x: global, y: global | SB | 1.3K/300K | 0/50K | 286/100K | 27/50K | 1.0K/50K | 0/50K |
| P0 |warp P1 | x: shared, y: global | SB | 3.2K/300K | 0/50K | 0/100K | 3.2K/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+addr+membar.cta | 7/300K | 0/50K | 0/100K | 0/50K | 7/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+addr+membar.cta | 14/300K | 0/50K | 0/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+membar.cta+addr+membar.cta | 32/300K | 0/50K | 4/100K | 0/50K | 28/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+membar.cta+addr+membar.cta | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.cta+addr+membar.cta | 13/300K | 0/50K | 3/100K | 0/50K | 10/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+addr+po | 9/300K | 0/50K | 0/100K | 0/50K | 9/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+addr+po | 25/300K | 0/50K | 4/100K | 0/50K | 21/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+membar.cta+addr+po | 34/300K | 0/50K | 1/100K | 0/50K | 33/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+membar.cta+addr+po | 14/300K | 0/50K | 0/100K | 0/50K | 13/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.cta+addr+po | 17/300K | 0/50K | 1/100K | 0/50K | 16/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+ctrl+membar.cta | 15/300K | 0/50K | 1/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+ctrl+membar.cta | 4/300K | 0/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+membar.cta+ctrl+membar.cta | 35/300K | 0/50K | 1/100K | 0/50K | 34/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+membar.cta+ctrl+membar.cta | 7/300K | 0/50K | 0/100K | 0/50K | 7/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.cta+ctrl+membar.cta | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+ctrl+po | 11/300K | 0/50K | 1/100K | 0/50K | 10/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+ctrl+po | 6/300K | 0/50K | 1/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+membar.cta+ctrl+po | 37/300K | 0/50K | 1/100K | 0/50K | 36/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+membar.cta+ctrl+po | 7/300K | 0/50K | 0/100K | 0/50K | 7/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.cta+ctrl+po | 10/300K | 0/50K | 3/100K | 0/50K | 7/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+membar.cta+po | 31/300K | 0/50K | 2/100K | 0/50K | 29/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | W+RWC+membar.cta+membar.cta+po | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+membar.cta+po | 32/300K | 0/50K | 2/100K | 0/50K | 30/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+membar.cta+membar.cta+po | 26/300K | 0/50K | 1/100K | 0/50K | 25/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+membar.cta+membar.cta+po | 24/300K | 0/50K | 2/100K | 0/50K | 22/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.cta+membar.cta+po | 25/300K | 0/50K | 4/100K | 0/50K | 21/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+po+membar.cta | 341/300K | 87/50K | 140/100K | 0/50K | 69/50K | 45/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | W+RWC+membar.cta+po+membar.cta | 473/300K | 146/50K | 196/100K | 0/50K | 44/50K | 87/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.cta+po+membar.cta | 597/300K | 20/50K | 347/100K | 1/50K | 130/50K | 99/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+po+membar.cta | 294/300K | 86/50K | 95/100K | 0/50K | 88/50K | 25/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+membar.cta+po+membar.cta | 532/300K | 302/50K | 116/100K | 14/50K | 51/50K | 49/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | W+RWC+membar.cta+po+membar.cta | 217/300K | 58/50K | 86/100K | 0/50K | 30/50K | 43/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.cta+po+membar.cta | 497/300K | 5/50K | 325/100K | 1/50K | 84/50K | 82/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+membar.cta+po+membar.cta | 853/300K | 503/50K | 167/100K | 6/50K | 1/50K | 176/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+membar.cta+po+membar.cta | 1.1K/250K | --- | 512/100K | 0/50K | 18/50K | 598/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | W+RWC+membar.cta+po+membar.cta | 438/300K | 9/50K | 372/100K | 0/50K | 38/50K | 19/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | W+RWC+membar.cta+po+membar.cta | 242/250K | --- | 173/100K | 0/50K | 64/50K | 5/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | W+RWC+membar.cta+po+membar.cta | 98/250K | --- | 44/100K | 4/50K | 3/50K | 47/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | W+RWC+membar.cta+po+membar.cta | 196/150K | --- | --- | 0/50K | 32/50K | 164/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+membar.cta+po+membar.cta | 320/300K | 140/50K | 84/100K | 0/50K | 67/50K | 29/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.cta+po+membar.cta | 451/300K | 41/50K | 312/100K | 0/50K | 66/50K | 32/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+po+membar.gl | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.cta+po+membar.gl | 8/300K | 0/50K | 0/100K | 0/50K | 0/50K | 8/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+membar.cta+po+membar.gl | 8/300K | 3/50K | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | W+RWC+membar.cta+po+membar.gl | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.cta+po+membar.gl | 2/300K | 0/50K | 0/100K | 0/50K | 0/50K | 2/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+membar.cta+po+membar.gl | 3/300K | 3/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+membar.cta+po+membar.gl | 1/250K | --- | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | W+RWC+membar.cta+po+membar.gl | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | W+RWC+membar.cta+po+membar.gl | 2/150K | --- | --- | 0/50K | 0/50K | 2/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+membar.cta+po+membar.gl | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.cta+po+membar.gl | 5/300K | 5/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+po+po | 452/300K | 100/50K | 194/100K | 0/50K | 92/50K | 66/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | W+RWC+membar.cta+po+po | 609/300K | 126/50K | 270/100K | 0/50K | 100/50K | 113/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.cta+po+po | 851/300K | 17/50K | 417/100K | 151/50K | 158/50K | 108/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.cta+po+po | 340/300K | 86/50K | 119/100K | 0/50K | 98/50K | 37/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+membar.cta+po+po | 586/300K | 307/50K | 163/100K | 25/50K | 33/50K | 58/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | W+RWC+membar.cta+po+po | 351/300K | 57/50K | 169/100K | 0/50K | 54/50K | 71/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.cta+po+po | 780/300K | 1/50K | 447/100K | 43/50K | 156/50K | 133/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+membar.cta+po+po | 973/300K | 508/50K | 219/100K | 30/50K | 7/50K | 209/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+membar.cta+po+po | 1.3K/250K | --- | 610/100K | 7/50K | 29/50K | 634/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | W+RWC+membar.cta+po+po | 491/300K | 21/50K | 368/100K | 0/50K | 68/50K | 34/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | W+RWC+membar.cta+po+po | 335/250K | --- | 201/100K | 2/50K | 118/50K | 14/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | W+RWC+membar.cta+po+po | 193/250K | --- | 76/100K | 41/50K | 3/50K | 73/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | W+RWC+membar.cta+po+po | 256/150K | --- | --- | 0/50K | 63/50K | 193/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+membar.cta+po+po | 351/300K | 92/50K | 130/100K | 0/50K | 98/50K | 31/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.cta+po+po | 541/300K | 53/50K | 345/100K | 5/50K | 90/50K | 48/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.ctas | 25/300K | 0/50K | 1/100K | 0/50K | 24/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.ctas | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.ctas | 28/300K | 0/50K | 1/100K | 0/50K | 27/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+membar.ctas | 45/300K | 0/50K | 3/100K | 0/50K | 42/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+membar.ctas | 28/300K | 0/50K | 0/100K | 0/50K | 28/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.ctas | 26/300K | 0/50K | 5/100K | 0/50K | 21/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.gl+po+membar.cta | 43/300K | 43/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | W+RWC+membar.gl+po+membar.cta | 60/300K | 59/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.gl+po+membar.cta | 39/300K | 11/50K | 0/100K | 0/50K | 0/50K | 28/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.gl+po+membar.cta | 49/300K | 49/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+membar.gl+po+membar.cta | 166/300K | 166/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | W+RWC+membar.gl+po+membar.cta | 20/300K | 20/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.gl+po+membar.cta | 26/300K | 3/50K | 0/100K | 0/50K | 0/50K | 23/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+membar.gl+po+membar.cta | 256/300K | 251/50K | 0/100K | 1/50K | 0/50K | 4/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+membar.gl+po+membar.cta | 256/250K | --- | 0/100K | 0/50K | 0/50K | 256/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | W+RWC+membar.gl+po+membar.cta | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | W+RWC+membar.gl+po+membar.cta | 1/250K | --- | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | W+RWC+membar.gl+po+membar.cta | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | W+RWC+membar.gl+po+membar.cta | 81/150K | --- | --- | 0/50K | 0/50K | 81/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+membar.gl+po+membar.cta | 63/300K | 63/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.gl+po+membar.cta | 27/300K | 27/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.gl+po+membar.gl | 1/300K | 0/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+membar.gl+po+membar.gl | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+membar.gl+po+membar.gl | 1/250K | --- | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | W+RWC+membar.gl+po+membar.gl | 1/150K | --- | --- | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.gl+po+po | 43/300K | 42/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | W+RWC+membar.gl+po+po | 59/300K | 59/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.gl+po+po | 81/300K | 13/50K | 0/100K | 43/50K | 0/50K | 25/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+membar.gl+po+po | 37/300K | 36/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+membar.gl+po+po | 158/300K | 155/50K | 0/100K | 1/50K | 0/50K | 2/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | W+RWC+membar.gl+po+po | 20/300K | 20/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | W+RWC+membar.gl+po+po | 40/300K | 2/50K | 0/100K | 0/50K | 0/50K | 38/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+membar.gl+po+po | 300/300K | 280/50K | 0/100K | 20/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+membar.gl+po+po | 256/250K | --- | 0/100K | 0/50K | 0/50K | 256/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | W+RWC+membar.gl+po+po | 6/300K | 6/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | W+RWC+membar.gl+po+po | 6/250K | --- | 0/100K | 0/50K | 0/50K | 6/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | W+RWC+membar.gl+po+po | 34/250K | --- | 0/100K | 34/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | W+RWC+membar.gl+po+po | 115/150K | --- | --- | 0/50K | 0/50K | 115/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+membar.gl+po+po | 33/300K | 33/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+membar.gl+po+po | 43/300K | 42/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+addr+membar.cta | 14/300K | 0/50K | 3/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+addr+membar.cta | 15/300K | 0/50K | 1/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+po+addr+membar.cta | 60/300K | 0/50K | 2/100K | 20/50K | 38/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+po+addr+membar.cta | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+po+addr+membar.cta | 2/250K | --- | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+po+addr+membar.cta | 8/300K | 0/50K | 0/100K | 0/50K | 7/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+po+addr+membar.cta | 18/300K | 0/50K | 1/100K | 7/50K | 10/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+po+addr+membar.gl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+addr+po | 11/300K | 0/50K | 1/100K | 0/50K | 10/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+addr+po | 19/300K | 0/50K | 2/100K | 0/50K | 17/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+po+addr+po | 66/300K | 0/50K | 2/100K | 31/50K | 33/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+po+addr+po | 55/300K | 0/50K | 0/100K | 55/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+po+addr+po | 20/250K | --- | 0/100K | 20/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | W+RWC+po+addr+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+po+addr+po | 8/300K | 0/50K | 1/100K | 0/50K | 6/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+po+addr+po | 184/300K | 0/50K | 4/100K | 164/50K | 16/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+ctrl+membar.cta | 12/300K | 0/50K | 0/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+ctrl+membar.cta | 9/300K | 0/50K | 1/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+po+ctrl+membar.cta | 52/300K | 0/50K | 2/100K | 10/50K | 40/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+po+ctrl+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+po+ctrl+membar.cta | 8/300K | 0/50K | 0/100K | 0/50K | 7/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+po+ctrl+membar.cta | 16/300K | 0/50K | 0/100K | 11/50K | 5/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+po+ctrl+membar.gl | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+ctrl+po | 19/300K | 0/50K | 0/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+ctrl+po | 24/300K | 0/50K | 1/100K | 0/50K | 23/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+po+ctrl+po | 78/300K | 0/50K | 4/100K | 33/50K | 41/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+po+ctrl+po | 64/300K | 0/50K | 0/100K | 64/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+po+ctrl+po | 7/250K | --- | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | W+RWC+po+ctrl+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | W+RWC+po+ctrl+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+po+ctrl+po | 15/300K | 0/50K | 0/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+po+ctrl+po | 152/300K | 0/50K | 0/100K | 138/50K | 14/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+membar.cta+membar.cta | 26/300K | 0/50K | 1/100K | 0/50K | 24/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC+po+membar.cta+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+membar.cta+membar.cta | 47/300K | 0/50K | 3/100K | 0/50K | 44/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+po+membar.cta+membar.cta | 44/300K | 0/50K | 4/100K | 0/50K | 40/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+po+membar.cta+membar.cta | 28/300K | 0/50K | 4/100K | 0/50K | 24/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+po+membar.cta+membar.cta | 29/300K | 0/50K | 1/100K | 1/50K | 27/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+membar.cta+po | 45/300K | 0/50K | 3/100K | 0/50K | 42/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | W+RWC+po+membar.cta+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC+po+membar.cta+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+membar.cta+po | 67/300K | 0/50K | 6/100K | 0/50K | 61/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+po+membar.cta+po | 53/300K | 0/50K | 9/100K | 1/50K | 43/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+po+membar.cta+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+po+membar.cta+po | 10/250K | --- | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | W+RWC+po+membar.cta+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+po+membar.cta+po | 57/300K | 0/50K | 3/100K | 0/50K | 53/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+po+membar.cta+po | 73/300K | 0/50K | 5/100K | 31/50K | 37/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+po+membar.gl+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+po+membar.gl+po | 12/300K | 0/50K | 0/100K | 12/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+po+membar.cta | 353/300K | 98/50K | 139/100K | 0/50K | 70/50K | 46/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | W+RWC+po+po+membar.cta | 529/300K | 126/50K | 242/100K | 0/50K | 77/50K | 84/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC+po+po+membar.cta | 617/300K | 16/50K | 361/100K | 2/50K | 132/50K | 106/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+po+membar.cta | 344/300K | 86/50K | 122/100K | 0/50K | 103/50K | 33/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+po+po+membar.cta | 638/300K | 317/50K | 121/100K | 84/50K | 51/50K | 65/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | W+RWC+po+po+membar.cta | 272/300K | 65/50K | 111/100K | 0/50K | 43/50K | 53/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | W+RWC+po+po+membar.cta | 605/300K | 2/50K | 413/100K | 3/50K | 93/50K | 94/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+po+po+membar.cta | 959/300K | 542/50K | 186/100K | 34/50K | 8/50K | 189/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+po+po+membar.cta | 1.3K/250K | --- | 643/100K | 13/50K | 46/50K | 615/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | W+RWC+po+po+membar.cta | 478/300K | 19/50K | 358/100K | 9/50K | 62/50K | 30/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | W+RWC+po+po+membar.cta | 320/250K | --- | 209/100K | 7/50K | 98/50K | 6/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | W+RWC+po+po+membar.cta | 130/250K | --- | 73/100K | 16/50K | 4/50K | 37/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | W+RWC+po+po+membar.cta | 210/150K | --- | --- | 0/50K | 38/50K | 172/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+po+po+membar.cta | 371/300K | 138/50K | 104/100K | 0/50K | 80/50K | 49/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+po+po+membar.cta | 543/300K | 70/50K | 317/100K | 20/50K | 92/50K | 44/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC+po+po+membar.gl | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC+po+po+membar.gl | 13/300K | 0/50K | 0/100K | 0/50K | 0/50K | 13/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC+po+po+membar.gl | 38/300K | 5/50K | 0/100K | 33/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | W+RWC+po+po+membar.gl | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC+po+po+membar.gl | 9/300K | 9/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC+po+po+membar.gl | 5/250K | --- | 0/100K | 0/50K | 0/50K | 5/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | W+RWC+po+po+membar.gl | 7/300K | 4/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | W+RWC+po+po+membar.gl | 5/250K | --- | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | W+RWC+po+po+membar.gl | 1/150K | --- | --- | 0/50K | 0/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC+po+po+membar.gl | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC+po+po+membar.gl | 4/300K | 4/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | W+RWC | 526/300K | 83/50K | 252/100K | 0/50K | 116/50K | 75/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | W+RWC | 666/300K | 111/50K | 330/100K | 0/50K | 111/50K | 114/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | W+RWC | 1.3K/300K | 14/50K | 435/100K | 537/50K | 174/50K | 117/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | W+RWC | 394/300K | 73/50K | 147/100K | 0/50K | 139/50K | 35/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | W+RWC | 842/300K | 330/50K | 191/100K | 196/50K | 63/50K | 62/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | W+RWC | 403/300K | 54/50K | 197/100K | 0/50K | 77/50K | 75/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | W+RWC | 1.9K/300K | 5/50K | 489/100K | 1.1K/50K | 156/50K | 166/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | W+RWC | 1.4K/300K | 559/50K | 254/100K | 350/50K | 13/50K | 269/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | W+RWC | 1.9K/250K | --- | 638/100K | 471/50K | 51/50K | 697/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | W+RWC | 869/300K | 28/50K | 432/100K | 302/50K | 69/50K | 38/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | W+RWC | 653/250K | --- | 247/100K | 252/50K | 130/50K | 24/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | W+RWC | 428/250K | --- | 91/100K | 271/50K | 8/50K | 58/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | W+RWC | 325/150K | --- | --- | 0/50K | 95/50K | 230/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | W+RWC | 414/300K | 124/50K | 146/100K | 0/50K | 95/50K | 49/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | W+RWC | 889/300K | 76/50K | 351/100K | 286/50K | 133/50K | 43/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC+addr+membar.cta | 161/300K | 0/50K | 8/100K | 0/50K | 153/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+addr+membar.cta | 192/300K | 0/50K | 15/100K | 0/50K | 177/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+addr+membar.cta | 144/300K | 0/50K | 8/100K | 0/50K | 136/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC+addr+po | 989/300K | 203/50K | 368/100K | 0/50K | 375/50K | 43/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | WRC+addr+po | 552/300K | 218/50K | 256/100K | 0/50K | 30/50K | 48/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRC+addr+po | 953/300K | 565/50K | 245/100K | 0/50K | 0/50K | 143/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+addr+po | 1.2K/300K | 293/50K | 397/100K | 0/50K | 465/50K | 63/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | WRC+addr+po | 710/300K | 248/50K | 307/100K | 0/50K | 41/50K | 114/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRC+addr+po | 2.0K/300K | 1.3K/50K | 317/100K | 0/50K | 0/50K | 419/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRC+addr+po | 1.0K/300K | 37/50K | 597/100K | 2/50K | 362/50K | 31/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | WRC+addr+po | 1.1K/250K | --- | 437/100K | 0/50K | 170/50K | 444/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+addr+po | 1.1K/300K | 183/50K | 529/100K | 0/50K | 346/50K | 55/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC+ctrl+membar.cta | 161/300K | 0/50K | 2/100K | 0/50K | 159/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+ctrl+membar.cta | 208/300K | 0/50K | 10/100K | 0/50K | 198/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+ctrl+membar.cta | 146/300K | 0/50K | 11/100K | 0/50K | 135/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC+ctrl+po | 985/300K | 202/50K | 393/100K | 0/50K | 348/50K | 42/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | WRC+ctrl+po | 526/300K | 217/50K | 224/100K | 0/50K | 33/50K | 52/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRC+ctrl+po | 1.0K/300K | 680/50K | 201/100K | 0/50K | 0/50K | 155/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+ctrl+po | 1.1K/300K | 283/50K | 334/100K | 0/50K | 453/50K | 52/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | WRC+ctrl+po | 671/300K | 207/50K | 322/100K | 0/50K | 50/50K | 92/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRC+ctrl+po | 1.8K/300K | 1.2K/50K | 243/100K | 0/50K | 0/50K | 333/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRC+ctrl+po | 1.1K/300K | 50/50K | 603/100K | 1/50K | 383/50K | 38/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | WRC+ctrl+po | 1.0K/250K | --- | 360/100K | 0/50K | 171/50K | 469/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+ctrl+po | 1.2K/300K | 197/50K | 500/100K | 0/50K | 390/50K | 65/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC+data+membar.cta | 160/300K | 0/50K | 11/100K | 0/50K | 149/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+data+membar.cta | 193/300K | 0/50K | 16/100K | 0/50K | 177/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+data+membar.cta | 163/300K | 0/50K | 10/100K | 0/50K | 153/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC+data+po | 1.0K/300K | 190/50K | 409/100K | 0/50K | 364/50K | 59/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | WRC+data+po | 643/300K | 257/50K | 277/100K | 0/50K | 45/50K | 64/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRC+data+po | 1.1K/300K | 633/50K | 253/100K | 0/50K | 0/50K | 175/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+data+po | 1.3K/300K | 274/50K | 427/100K | 0/50K | 501/50K | 81/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | WRC+data+po | 731/300K | 219/50K | 336/100K | 0/50K | 55/50K | 121/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRC+data+po | 1.2K/300K | 424/50K | 331/100K | 0/50K | 0/50K | 415/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRC+data+po | 1.2K/300K | 37/50K | 602/100K | 56/50K | 427/50K | 41/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | WRC+data+po | 1.1K/250K | --- | 408/100K | 0/50K | 228/50K | 445/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+data+po | 1.1K/300K | 164/50K | 539/100K | 0/50K | 329/50K | 70/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+membar.cta+ctrl | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+membar.cta+ctrl | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC+membar.cta+po | 1.9K/300K | 267/50K | 716/100K | 0/50K | 819/50K | 56/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | WRC+membar.cta+po | 1.0K/300K | 415/50K | 351/100K | 0/50K | 193/50K | 66/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRC+membar.cta+po | 2.0K/300K | 1.0K/50K | 702/100K | 0/50K | 71/50K | 161/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+membar.cta+po | 2.1K/300K | 439/50K | 584/100K | 0/50K | 968/50K | 64/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | WRC+membar.cta+po | 1.3K/300K | 412/50K | 559/100K | 0/50K | 214/50K | 125/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRC+membar.cta+po | 2.2K/300K | 694/50K | 943/100K | 1/50K | 61/50K | 456/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRC+membar.cta+po | 1.1K/300K | 34/50K | 640/100K | 0/50K | 377/50K | 46/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | WRC+membar.cta+po | 1.1K/250K | --- | 392/100K | 0/50K | 222/50K | 467/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+membar.cta+po | 2.0K/300K | 256/50K | 840/100K | 0/50K | 816/50K | 71/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC+membar.ctas | 307/300K | 0/50K | 18/100K | 0/50K | 289/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+membar.ctas | 456/300K | 0/50K | 24/100K | 0/50K | 432/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+membar.ctas | 279/300K | 0/50K | 21/100K | 0/50K | 258/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC+membar.gl+membar.cta | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+membar.gl+membar.cta | 13/300K | 0/50K | 0/100K | 0/50K | 13/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+membar.gl+membar.cta | 9/300K | 0/50K | 0/100K | 0/50K | 9/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC+membar.gl+po | 137/300K | 128/50K | 0/100K | 0/50K | 9/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | WRC+membar.gl+po | 168/300K | 165/50K | 0/100K | 0/50K | 0/50K | 3/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRC+membar.gl+po | 622/300K | 480/50K | 0/100K | 0/50K | 0/50K | 142/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+membar.gl+po | 225/300K | 209/50K | 0/100K | 0/50K | 16/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | WRC+membar.gl+po | 170/300K | 164/50K | 0/100K | 0/50K | 0/50K | 6/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRC+membar.gl+po | 1.2K/300K | 873/50K | 0/100K | 0/50K | 0/50K | 291/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRC+membar.gl+po | 8/300K | 8/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | WRC+membar.gl+po | 220/250K | --- | 0/100K | 0/50K | 0/50K | 220/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+membar.gl+po | 143/300K | 122/50K | 0/100K | 0/50K | 20/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRC+po+addr | 295/300K | 0/50K | 0/100K | 295/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRC+po+addr | 89/300K | 0/50K | 0/100K | 89/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRC+po+addr | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRC+po+ctrl | 262/300K | 0/50K | 0/100K | 262/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRC+po+ctrl | 80/300K | 0/50K | 0/100K | 80/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRC+po+ctrl | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+po+ctrl | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC+po+membar.cta | 327/300K | 0/50K | 18/100K | 0/50K | 309/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRC+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC+po+membar.cta | 389/300K | 0/50K | 23/100K | 0/50K | 366/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRC+po+membar.cta | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC+po+membar.cta | 301/300K | 0/50K | 23/100K | 0/50K | 278/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRC+po+membar.gl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRC | 2.0K/300K | 277/50K | 696/100K | 0/50K | 859/50K | 132/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | WRC | 1.1K/300K | 384/50K | 412/100K | 0/50K | 190/50K | 124/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRC | 2.8K/300K | 1.1K/50K | 806/100K | 434/50K | 79/50K | 400/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRC | 2.3K/300K | 430/50K | 612/100K | 0/50K | 1.0K/50K | 180/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | WRC | 1.5K/300K | 374/50K | 621/100K | 0/50K | 261/50K | 276/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRC | 3.6K/300K | 830/50K | 1.1K/100K | 704/50K | 86/50K | 885/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRC | 1.5K/300K | 53/50K | 738/100K | 200/50K | 447/50K | 41/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | WRC | 1.4K/250K | --- | 472/100K | 0/50K | 314/50K | 583/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRC | 2.2K/300K | 291/50K | 856/100K | 1/50K | 883/50K | 166/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRR+2W+addr+membar.cta | 52/300K | 0/50K | 3/100K | 0/50K | 49/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRR+2W+addr+membar.cta | 56/300K | 0/50K | 3/100K | 0/50K | 53/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRR+2W+addr+membar.cta | 44/300K | 0/50K | 5/100K | 0/50K | 39/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRR+2W+addr+po | 66/300K | 0/50K | 4/100K | 0/50K | 62/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRR+2W+addr+po | 73/300K | 0/50K | 5/100K | 0/50K | 68/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRR+2W+addr+po | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRR+2W+addr+po | 54/300K | 0/50K | 4/100K | 0/50K | 50/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRR+2W+ctrl+membar.cta | 47/300K | 0/50K | 2/100K | 0/50K | 45/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRR+2W+ctrl+membar.cta | 58/300K | 0/50K | 3/100K | 0/50K | 55/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRR+2W+ctrl+membar.cta | 43/300K | 0/50K | 5/100K | 0/50K | 38/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRR+2W+ctrl+po | 72/300K | 0/50K | 3/100K | 0/50K | 69/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRR+2W+ctrl+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRR+2W+ctrl+po | 85/300K | 0/50K | 6/100K | 0/50K | 79/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRR+2W+ctrl+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRR+2W+ctrl+po | 74/300K | 0/50K | 9/100K | 0/50K | 65/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRR+2W+membar.cta+po | 178/300K | 0/50K | 17/100K | 0/50K | 161/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRR+2W+membar.cta+po | 213/300K | 0/50K | 8/100K | 0/50K | 205/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRR+2W+membar.cta+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRR+2W+membar.cta+po | 171/300K | 0/50K | 15/100K | 0/50K | 156/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRR+2W+membar.ctas | 140/300K | 0/50K | 8/100K | 0/50K | 132/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRR+2W+membar.ctas | 169/300K | 0/50K | 13/100K | 0/50K | 156/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRR+2W+membar.ctas | 113/300K | 0/50K | 13/100K | 0/50K | 100/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRR+2W+membar.gl+membar.cta | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRR+2W+membar.gl+membar.cta | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRR+2W+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRR+2W+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRR+2W+membar.gl+po | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRR+2W+membar.gl+po | 4/300K | 0/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRR+2W+po+membar.cta | 1.5K/300K | 386/50K | 707/100K | 0/50K | 279/50K | 173/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | WRR+2W+po+membar.cta | 1.7K/300K | 422/50K | 882/100K | 0/50K | 217/50K | 214/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRR+2W+po+membar.cta | 1.5K/300K | 186/50K | 746/100K | 4/50K | 326/50K | 270/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRR+2W+po+membar.cta | 1.6K/300K | 380/50K | 643/100K | 0/50K | 410/50K | 133/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | WRR+2W+po+membar.cta | 1.7K/300K | 389/50K | 790/100K | 0/50K | 245/50K | 248/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRR+2W+po+membar.cta | 1.4K/300K | 255/50K | 602/100K | 2/50K | 299/50K | 206/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRR+2W+po+membar.cta | 1.1K/250K | 448/50K | 453/100K | --- | 25/50K | 215/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | WRR+2W+po+membar.cta | 1.2K/200K | --- | 677/100K | --- | 98/50K | 458/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRR+2W+po+membar.cta | 1.6K/300K | 454/50K | 630/100K | 0/50K | 295/50K | 195/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRR+2W+po+membar.gl | 124/300K | 124/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | WRR+2W+po+membar.gl | 125/300K | 124/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRR+2W+po+membar.gl | 152/300K | 89/50K | 0/100K | 0/50K | 0/50K | 63/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRR+2W+po+membar.gl | 113/300K | 113/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | WRR+2W+po+membar.gl | 102/300K | 101/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRR+2W+po+membar.gl | 87/300K | 63/50K | 0/100K | 0/50K | 0/50K | 24/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRR+2W+po+membar.gl | 144/250K | 143/50K | 0/100K | --- | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | WRR+2W+po+membar.gl | 149/200K | --- | 0/100K | --- | 0/50K | 149/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRR+2W+po+membar.gl | 133/300K | 133/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRR+2W | 1.9K/300K | 424/50K | 858/100K | 0/50K | 407/50K | 212/50K |
| P0 |cta P1 |warp P2 | x: global, y: global | WRR+2W | 2.0K/300K | 425/50K | 1.0K/100K | 0/50K | 294/50K | 288/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRR+2W | 2.5K/300K | 183/50K | 999/100K | 524/50K | 474/50K | 320/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRR+2W | 1.9K/300K | 386/50K | 816/100K | 0/50K | 505/50K | 230/50K |
| P0 |warp P1 |warp P2 | x: global, y: global | WRR+2W | 2.0K/300K | 418/50K | 859/100K | 0/50K | 366/50K | 356/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRR+2W | 2.5K/300K | 188/50K | 733/100K | 705/50K | 449/50K | 415/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRR+2W | 1.4K/250K | 435/50K | 600/100K | --- | 63/50K | 286/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared | WRR+2W | 1.5K/200K | --- | 802/100K | --- | 180/50K | 559/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRR+2W | 1.8K/300K | 441/50K | 732/100K | 0/50K | 402/50K | 217/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W+addr+membar.cta | 53/300K | 0/50K | 2/100K | 0/50K | 51/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W+addr+membar.cta | 52/300K | 0/50K | 3/100K | 0/50K | 49/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+2W+addr+membar.cta | 39/300K | 0/50K | 4/100K | 0/50K | 35/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W+addr+po | 63/300K | 0/50K | 5/100K | 0/50K | 58/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W+addr+po | 85/300K | 0/50K | 8/100K | 0/50K | 77/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+2W+addr+po | 62/300K | 0/50K | 5/100K | 0/50K | 57/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W+ctrl+membar.cta | 66/300K | 0/50K | 1/100K | 0/50K | 65/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W+ctrl+membar.cta | 73/300K | 0/50K | 5/100K | 0/50K | 68/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+2W+ctrl+membar.cta | 35/300K | 0/50K | 2/100K | 0/50K | 33/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W+ctrl+po | 69/300K | 0/50K | 1/100K | 0/50K | 68/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W+ctrl+po | 83/300K | 0/50K | 4/100K | 0/50K | 79/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+2W+ctrl+po | 48/300K | 0/50K | 3/100K | 0/50K | 45/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W+data+membar.cta | 49/300K | 0/50K | 1/100K | 0/50K | 48/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W+data+membar.cta | 68/300K | 0/50K | 1/100K | 0/50K | 67/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+2W+data+membar.cta | 38/300K | 0/50K | 0/100K | 0/50K | 38/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W+data+po | 55/300K | 0/50K | 3/100K | 0/50K | 52/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRW+2W+data+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W+data+po | 75/300K | 0/50K | 6/100K | 0/50K | 69/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRW+2W+data+po | 14/300K | 0/50K | 0/100K | 14/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+2W+data+po | 68/300K | 0/50K | 1/100K | 0/50K | 67/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W+membar.cta+po | 146/300K | 0/50K | 10/100K | 0/50K | 136/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W+membar.cta+po | 208/300K | 0/50K | 9/100K | 0/50K | 199/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+2W+membar.cta+po | 126/300K | 0/50K | 12/100K | 0/50K | 114/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W+membar.ctas | 106/300K | 0/50K | 6/100K | 0/50K | 100/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W+membar.ctas | 155/300K | 0/50K | 7/100K | 0/50K | 148/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+2W+membar.ctas | 122/300K | 0/50K | 12/100K | 0/50K | 110/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W+membar.gl+membar.cta | 7/300K | 0/50K | 0/100K | 0/50K | 7/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W+membar.gl+po | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W+po+membar.cta | 93/300K | 0/50K | 4/100K | 0/50K | 89/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W+po+membar.cta | 145/300K | 0/50K | 12/100K | 0/50K | 133/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRW+2W+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRW+2W+po+membar.cta | 18/300K | 0/50K | 0/100K | 18/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+2W+po+membar.cta | 89/300K | 0/50K | 13/100K | 0/50K | 76/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRW+2W+po+membar.gl | 10/300K | 0/50K | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+2W | 173/300K | 0/50K | 4/100K | 0/50K | 169/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRW+2W | 406/300K | 0/50K | 0/100K | 406/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+2W | 221/300K | 0/50K | 10/100K | 0/50K | 211/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRW+2W | 429/300K | 0/50K | 0/100K | 429/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRW+2W | 271/300K | 0/50K | 0/100K | 271/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+2W | 140/300K | 0/50K | 17/100K | 0/50K | 123/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR+addr+membar.cta | 75/300K | 0/50K | 7/100K | 0/50K | 68/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR+addr+membar.cta | 81/300K | 0/50K | 5/100K | 0/50K | 76/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR+addr+membar.cta | 79/300K | 0/50K | 8/100K | 0/50K | 71/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR+addr+po | 72/300K | 0/50K | 4/100K | 0/50K | 68/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR+addr+po | 119/300K | 0/50K | 10/100K | 0/50K | 109/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRW+WR+addr+po | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR+addr+po | 80/300K | 0/50K | 3/100K | 0/50K | 77/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR+ctrl+membar.cta | 60/300K | 0/50K | 2/100K | 0/50K | 58/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR+ctrl+membar.cta | 93/300K | 0/50K | 6/100K | 0/50K | 87/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR+ctrl+membar.cta | 70/300K | 0/50K | 3/100K | 0/50K | 67/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR+ctrl+po | 82/300K | 0/50K | 2/100K | 0/50K | 80/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR+ctrl+po | 124/300K | 0/50K | 6/100K | 0/50K | 118/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRW+WR+ctrl+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR+ctrl+po | 103/300K | 0/50K | 7/100K | 0/50K | 96/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR+data+membar.cta | 77/300K | 0/50K | 4/100K | 0/50K | 73/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR+data+membar.cta | 93/300K | 0/50K | 5/100K | 0/50K | 88/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR+data+membar.cta | 59/300K | 0/50K | 3/100K | 0/50K | 56/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR+data+po | 85/300K | 0/50K | 4/100K | 0/50K | 81/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR+data+po | 121/300K | 0/50K | 8/100K | 0/50K | 113/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRW+WR+data+po | 22/300K | 0/50K | 0/100K | 22/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR+data+po | 93/300K | 0/50K | 5/100K | 0/50K | 88/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR+membar.cta+po | 199/300K | 0/50K | 13/100K | 0/50K | 186/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR+membar.cta+po | 273/300K | 0/50K | 20/100K | 0/50K | 253/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRW+WR+membar.cta+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR+membar.cta+po | 215/300K | 0/50K | 16/100K | 0/50K | 199/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR+membar.ctas | 192/300K | 0/50K | 8/100K | 0/50K | 184/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR+membar.ctas | 236/300K | 0/50K | 10/100K | 0/50K | 226/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR+membar.ctas | 167/300K | 0/50K | 13/100K | 0/50K | 154/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR+membar.gl+po | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR+membar.gl+po | 10/300K | 0/50K | 0/100K | 0/50K | 10/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR+po+membar.cta | 180/300K | 0/50K | 11/100K | 0/50K | 169/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRW+WR+po+membar.cta | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR+po+membar.cta | 240/300K | 0/50K | 16/100K | 0/50K | 224/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRW+WR+po+membar.cta | 28/300K | 0/50K | 0/100K | 28/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR+po+membar.cta | 179/300K | 0/50K | 9/100K | 0/50K | 170/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRW+WR+po+membar.gl | 7/300K | 0/50K | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WRW+WR | 197/300K | 0/50K | 9/100K | 0/50K | 188/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WRW+WR | 475/300K | 0/50K | 0/100K | 475/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WRW+WR | 260/300K | 0/50K | 16/100K | 0/50K | 244/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WRW+WR | 414/300K | 0/50K | 0/100K | 414/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WRW+WR | 339/300K | 0/50K | 0/100K | 339/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WRW+WR | 223/300K | 0/50K | 21/100K | 0/50K | 202/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC+addr+membar.cta | 83/300K | 0/50K | 2/100K | 0/50K | 81/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+addr+membar.cta | 99/300K | 0/50K | 6/100K | 0/50K | 93/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC+addr+membar.cta | 83/300K | 0/50K | 5/100K | 0/50K | 78/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC+addr+po | 126/300K | 0/50K | 6/100K | 0/50K | 120/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+addr+po | 161/300K | 0/50K | 13/100K | 0/50K | 148/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WWC+addr+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WWC+addr+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC+addr+po | 119/300K | 0/50K | 5/100K | 0/50K | 114/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC+ctrl+membar.cta | 75/300K | 0/50K | 2/100K | 0/50K | 73/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+ctrl+membar.cta | 119/300K | 0/50K | 5/100K | 0/50K | 114/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC+ctrl+membar.cta | 76/300K | 0/50K | 3/100K | 0/50K | 73/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC+ctrl+po | 130/300K | 0/50K | 11/100K | 0/50K | 119/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+ctrl+po | 146/300K | 0/50K | 6/100K | 0/50K | 140/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WWC+ctrl+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WWC+ctrl+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC+ctrl+po | 121/300K | 0/50K | 9/100K | 0/50K | 112/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC+data+membar.cta | 72/300K | 0/50K | 5/100K | 0/50K | 67/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+data+membar.cta | 115/300K | 0/50K | 5/100K | 0/50K | 110/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC+data+membar.cta | 80/300K | 0/50K | 5/100K | 0/50K | 75/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC+data+po | 129/300K | 0/50K | 8/100K | 0/50K | 121/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+data+po | 169/300K | 0/50K | 11/100K | 0/50K | 158/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WWC+data+po | 41/300K | 0/50K | 0/100K | 41/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC+data+po | 126/300K | 0/50K | 8/100K | 0/50K | 118/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC+membar.cta+po | 297/300K | 0/50K | 17/100K | 0/50K | 280/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+membar.cta+po | 348/300K | 0/50K | 13/100K | 0/50K | 335/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WWC+membar.cta+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC+membar.cta+po | 283/300K | 0/50K | 19/100K | 0/50K | 264/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC+membar.ctas | 229/300K | 0/50K | 14/100K | 0/50K | 215/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+membar.ctas | 243/300K | 0/50K | 9/100K | 0/50K | 234/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC+membar.ctas | 184/300K | 0/50K | 9/100K | 0/50K | 175/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC+membar.gl+membar.cta | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+membar.gl+membar.cta | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC+membar.gl+membar.cta | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC+membar.gl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+membar.gl+po | 14/300K | 0/50K | 0/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC+membar.gl+po | 4/300K | 0/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WWC+po+addr | 233/300K | 0/50K | 0/100K | 233/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WWC+po+addr | 90/300K | 0/50K | 0/100K | 90/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WWC+po+addr | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WWC+po+ctrl | 248/300K | 0/50K | 0/100K | 248/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WWC+po+ctrl | 49/300K | 0/50K | 0/100K | 49/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WWC+po+ctrl | 5/300K | 0/50K | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WWC+po+data | 284/300K | 0/50K | 0/100K | 284/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+po+data | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WWC+po+data | 84/300K | 0/50K | 0/100K | 84/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WWC+po+data | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC+po+membar.cta | 226/300K | 0/50K | 7/100K | 0/50K | 219/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WWC+po+membar.cta | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC+po+membar.cta | 269/300K | 0/50K | 9/100K | 0/50K | 260/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WWC+po+membar.cta | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC+po+membar.cta | 230/300K | 0/50K | 11/100K | 0/50K | 219/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global | WWC | 293/300K | 0/50K | 15/100K | 0/50K | 278/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: shared | WWC | 394/300K | 0/50K | 0/100K | 394/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global | WWC | 360/300K | 0/50K | 16/100K | 0/50K | 344/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared | WWC | 502/300K | 0/50K | 0/100K | 502/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global | WWC | 204/300K | 0/50K | 0/100K | 204/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global | WWC | 268/300K | 0/50K | 27/100K | 0/50K | 238/50K | 3/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+addr+membar.cta | 7/300K | 0/50K | 0/100K | 0/50K | 7/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+addr+membar.cta | 13/300K | 0/50K | 0/100K | 0/50K | 13/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.cta+addr+membar.cta | 33/300K | 0/50K | 0/100K | 0/50K | 33/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+membar.cta+addr+membar.cta | 9/300K | 0/50K | 1/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.cta+addr+membar.cta | 11/300K | 0/50K | 0/100K | 0/50K | 11/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+addr+po | 10/300K | 0/50K | 0/100K | 0/50K | 10/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+addr+po | 9/300K | 0/50K | 0/100K | 0/50K | 9/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.cta+addr+po | 41/300K | 0/50K | 6/100K | 0/50K | 35/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+membar.cta+addr+po | 4/300K | 0/50K | 1/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.cta+addr+po | 10/300K | 0/50K | 2/100K | 0/50K | 8/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+ctrl+membar.cta | 10/300K | 0/50K | 1/100K | 0/50K | 9/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+ctrl+membar.cta | 7/300K | 0/50K | 1/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.cta+ctrl+membar.cta | 34/300K | 0/50K | 2/100K | 0/50K | 32/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+membar.cta+ctrl+membar.cta | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.cta+ctrl+membar.cta | 11/300K | 0/50K | 2/100K | 0/50K | 9/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+ctrl+po | 13/300K | 0/50K | 1/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+ctrl+po | 10/300K | 0/50K | 1/100K | 0/50K | 9/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.cta+ctrl+po | 40/300K | 0/50K | 3/100K | 0/50K | 37/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.0+membar.cta+ctrl+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+membar.cta+ctrl+po | 4/300K | 0/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.cta+ctrl+po | 9/300K | 0/50K | 1/100K | 0/50K | 8/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+data+membar.cta | 12/300K | 0/50K | 0/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+data+membar.cta | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.cta+data+membar.cta | 34/300K | 0/50K | 4/100K | 0/50K | 30/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+membar.cta+data+membar.cta | 8/300K | 0/50K | 0/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.cta+data+membar.cta | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+data+po | 12/300K | 0/50K | 0/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+data+po | 7/300K | 0/50K | 0/100K | 0/50K | 7/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.cta+data+po | 41/300K | 0/50K | 3/100K | 0/50K | 38/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+membar.cta+data+po | 8/300K | 0/50K | 0/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.cta+data+po | 10/300K | 0/50K | 2/100K | 0/50K | 8/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+membar.cta+po | 28/300K | 0/50K | 0/100K | 0/50K | 28/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+membar.cta+po | 43/300K | 0/50K | 2/100K | 0/50K | 41/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.cta+membar.cta+po | 35/300K | 0/50K | 0/100K | 0/50K | 35/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+membar.cta+membar.cta+po | 38/300K | 0/50K | 1/100K | 0/50K | 37/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.cta+membar.cta+po | 32/300K | 0/50K | 3/100K | 0/50K | 29/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+po+membar.cta | 28/300K | 0/50K | 1/100K | 0/50K | 27/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.0+membar.cta+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.0+membar.cta+po+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 0/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+po+membar.cta | 32/300K | 0/50K | 1/100K | 0/50K | 31/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.cta+po+membar.cta | 48/300K | 0/50K | 3/100K | 11/50K | 34/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0+membar.cta+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.0+membar.cta+po+membar.cta | 5/250K | --- | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+membar.cta+po+membar.cta | 22/300K | 0/50K | 2/100K | 0/50K | 20/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.cta+po+membar.cta | 23/300K | 0/50K | 2/100K | 0/50K | 21/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.cta+po+membar.gl | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+po+po | 35/300K | 0/50K | 1/100K | 0/50K | 34/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.0+membar.cta+po+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.0+membar.cta+po+po | 133/300K | 0/50K | 0/100K | 132/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.cta+po+po | 30/300K | 0/50K | 4/100K | 0/50K | 26/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.cta+po+po | 53/300K | 0/50K | 1/100K | 15/50K | 37/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.0+membar.cta+po+po | 22/300K | 0/50K | 0/100K | 22/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0+membar.cta+po+po | 40/300K | 0/50K | 0/100K | 40/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.0+membar.cta+po+po | 10/250K | --- | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.0+membar.cta+po+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.0+membar.cta+po+po | 38/250K | --- | 0/100K | 38/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+membar.cta+po+po | 39/300K | 0/50K | 1/100K | 0/50K | 38/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.cta+po+po | 31/300K | 0/50K | 3/100K | 3/50K | 25/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.ctas | 28/300K | 0/50K | 0/100K | 0/50K | 28/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.0+membar.ctas | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+membar.ctas | 23/300K | 0/50K | 3/100K | 0/50K | 20/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.ctas | 31/300K | 0/50K | 1/100K | 0/50K | 30/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+membar.ctas | 18/300K | 0/50K | 1/100K | 0/50K | 17/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+membar.ctas | 17/300K | 0/50K | 3/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0+membar.gl+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.0+membar.gl+po+po | 40/300K | 0/50K | 0/100K | 40/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+membar.gl+po+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0+membar.gl+po+po | 34/300K | 0/50K | 0/100K | 34/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.0+membar.gl+po+po | 28/250K | --- | 0/100K | 28/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+addr+membar.cta | 11/300K | 0/50K | 0/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+addr+membar.cta | 18/300K | 0/50K | 1/100K | 0/50K | 17/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+addr+membar.cta | 59/300K | 0/50K | 3/100K | 16/50K | 40/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0+po+addr+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.0+po+addr+membar.cta | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.0+po+addr+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+po+addr+membar.cta | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+po+addr+membar.cta | 11/300K | 0/50K | 0/100K | 5/50K | 6/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+addr+membar.gl | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+addr+po | 12/300K | 0/50K | 1/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+addr+po | 23/300K | 0/50K | 2/100K | 0/50K | 21/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+addr+po | 65/300K | 0/50K | 1/100K | 29/50K | 35/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0+po+addr+po | 61/300K | 0/50K | 0/100K | 61/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.0+po+addr+po | 31/250K | --- | 0/100K | 31/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.0+po+addr+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.0+po+addr+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+po+addr+po | 11/300K | 0/50K | 2/100K | 0/50K | 9/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+po+addr+po | 137/300K | 0/50K | 1/100K | 123/50K | 13/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+ctrl+membar.cta | 14/300K | 0/50K | 0/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+ctrl+membar.cta | 15/300K | 0/50K | 1/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+ctrl+membar.cta | 52/300K | 0/50K | 1/100K | 11/50K | 40/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0+po+ctrl+membar.cta | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.0+po+ctrl+membar.cta | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+po+ctrl+membar.cta | 11/300K | 0/50K | 0/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+po+ctrl+membar.cta | 19/300K | 0/50K | 3/100K | 10/50K | 6/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+ctrl+membar.gl | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+ctrl+po | 12/300K | 0/50K | 0/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+ctrl+po | 19/300K | 0/50K | 0/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+ctrl+po | 65/300K | 0/50K | 3/100K | 21/50K | 41/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.0+po+ctrl+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0+po+ctrl+po | 47/300K | 0/50K | 0/100K | 47/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.0+po+ctrl+po | 6/250K | --- | 0/100K | 6/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.0+po+ctrl+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+po+ctrl+po | 16/300K | 0/50K | 1/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+po+ctrl+po | 117/300K | 0/50K | 1/100K | 104/50K | 12/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+data+membar.cta | 13/300K | 0/50K | 1/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+data+membar.cta | 10/300K | 0/50K | 0/100K | 0/50K | 10/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+data+membar.cta | 76/300K | 0/50K | 2/100K | 30/50K | 44/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0+po+data+membar.cta | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+po+data+membar.cta | 8/300K | 0/50K | 0/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+po+data+membar.cta | 12/300K | 0/50K | 0/100K | 6/50K | 6/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+data+membar.gl | 7/300K | 0/50K | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+data+po | 10/300K | 0/50K | 2/100K | 0/50K | 8/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.0+po+data+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+data+po | 11/300K | 0/50K | 0/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+data+po | 93/300K | 0/50K | 2/100K | 50/50K | 41/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0+po+data+po | 66/300K | 0/50K | 0/100K | 66/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.0+po+data+po | 52/250K | --- | 0/100K | 52/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.0+po+data+po | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.0+po+data+po | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+po+data+po | 9/300K | 0/50K | 0/100K | 0/50K | 9/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+po+data+po | 151/300K | 0/50K | 2/100K | 143/50K | 6/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+membar.cta+membar.cta | 37/300K | 0/50K | 1/100K | 0/50K | 36/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+membar.cta+membar.cta | 34/300K | 0/50K | 3/100K | 0/50K | 31/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+membar.cta+membar.cta | 45/300K | 0/50K | 1/100K | 3/50K | 41/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+po+membar.cta+membar.cta | 38/300K | 0/50K | 3/100K | 0/50K | 35/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+po+membar.cta+membar.cta | 18/300K | 0/50K | 0/100K | 1/50K | 17/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+membar.cta+po | 40/300K | 0/50K | 4/100K | 0/50K | 36/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.0+po+membar.cta+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.0+po+membar.cta+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+membar.cta+po | 34/300K | 0/50K | 2/100K | 0/50K | 32/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+membar.cta+po | 44/300K | 0/50K | 0/100K | 0/50K | 44/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.0+po+membar.cta+po | 10/250K | --- | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+po+membar.cta+po | 43/300K | 0/50K | 5/100K | 0/50K | 38/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+po+membar.cta+po | 59/300K | 0/50K | 1/100K | 17/50K | 41/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+po+membar.gl+po | 7/300K | 0/50K | 0/100K | 6/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+po+membar.cta | 33/300K | 0/50K | 1/100K | 0/50K | 32/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.0+po+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.0+po+po+membar.cta | 8/300K | 0/50K | 1/100K | 0/50K | 4/50K | 3/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0+po+po+membar.cta | 32/300K | 0/50K | 3/100K | 0/50K | 29/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+po+membar.cta | 85/300K | 0/50K | 1/100K | 57/50K | 27/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0+po+po+membar.cta | 34/300K | 0/50K | 0/100K | 34/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.0+po+po+membar.cta | 8/250K | --- | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.0+po+po+membar.cta | 18/300K | 0/50K | 0/100K | 18/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.0+po+po+membar.cta | 5/250K | --- | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.0+po+po+membar.cta | 12/250K | --- | 0/100K | 12/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0+po+po+membar.cta | 30/300K | 0/50K | 1/100K | 0/50K | 29/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0+po+po+membar.cta | 46/300K | 0/50K | 2/100K | 18/50K | 26/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0+po+po+membar.gl | 15/300K | 0/50K | 0/100K | 15/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.0+po+po+membar.gl | 8/250K | --- | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.0 | 50/300K | 0/50K | 1/100K | 0/50K | 49/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.0 | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.0 | 463/300K | 0/50K | 0/100K | 459/50K | 2/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.0 | 45/300K | 0/50K | 4/100K | 0/50K | 41/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.0 | 198/300K | 0/50K | 5/100K | 160/50K | 33/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.0 | 866/300K | 0/50K | 0/100K | 866/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.0 | 277/300K | 0/50K | 0/100K | 277/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.0 | 524/250K | --- | 0/100K | 524/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.0 | 277/300K | 0/50K | 0/100K | 277/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.0 | 186/250K | --- | 0/100K | 186/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.0 | 202/250K | --- | 0/100K | 202/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.0 | 46/300K | 0/50K | 3/100K | 0/50K | 43/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.0 | 262/300K | 0/50K | 3/100K | 229/50K | 30/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.1+membar.cta+membar.cta+ctrl | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+membar.cta+data | 1/300K | 0/50K | 1/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.1+membar.cta+membar.cta+po | 72/300K | 0/50K | 1/100K | 0/50K | 69/50K | 2/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+membar.cta+po | 3/300K | 0/50K | 0/100K | 2/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.1+membar.cta+membar.cta+po | 81/300K | 0/50K | 6/100K | 0/50K | 75/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+membar.cta+membar.cta+po | 51/300K | 0/50K | 4/100K | 0/50K | 47/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.1+membar.cta+membar.cta+po | 80/300K | 0/50K | 5/100K | 0/50K | 74/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+membar.cta+membar.cta+po | 76/300K | 0/50K | 5/100K | 1/50K | 70/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.1+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+membar.cta+membar.gl+membar.cta | 5/300K | 0/50K | 0/100K | 0/50K | 5/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.1+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.1+membar.cta+membar.gl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+membar.cta+membar.gl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+po+addr | 15/300K | 0/50K | 0/100K | 15/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+membar.cta+po+addr | 5/300K | 0/50K | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+po+addr | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+po+ctrl | 12/300K | 0/50K | 0/100K | 11/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+membar.cta+po+ctrl | 10/300K | 0/50K | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+po+ctrl | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+po+data | 35/300K | 0/50K | 0/100K | 35/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+membar.cta+po+data | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+po+data | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.1+membar.cta+po+membar.cta | 47/300K | 0/50K | 1/100K | 0/50K | 46/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.1+membar.cta+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.1+membar.cta+po+membar.cta | 56/300K | 0/50K | 1/100K | 0/50K | 55/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+membar.cta+po+membar.cta | 55/300K | 0/50K | 2/100K | 1/50K | 52/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.1+membar.cta+po+membar.cta | 67/300K | 0/50K | 1/100K | 0/50K | 66/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+membar.cta+po+membar.cta | 46/300K | 0/50K | 3/100K | 0/50K | 43/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+membar.cta+po+membar.gl | 5/300K | 0/50K | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.1+membar.cta+po+po | 70/300K | 0/50K | 2/100K | 0/50K | 68/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+po+po | 77/300K | 0/50K | 1/100K | 74/50K | 1/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.1+membar.cta+po+po | 79/300K | 0/50K | 10/100K | 0/50K | 69/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+membar.cta+po+po | 55/300K | 0/50K | 4/100K | 19/50K | 32/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.cta+po+po | 23/300K | 0/50K | 0/100K | 23/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.1+membar.cta+po+po | 28/300K | 0/50K | 0/100K | 28/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.1+membar.cta+po+po | 11/250K | --- | 0/100K | 11/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.1+membar.cta+po+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.1+membar.cta+po+po | 8/250K | --- | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.1+membar.cta+po+po | 95/300K | 0/50K | 6/100K | 0/50K | 88/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+membar.cta+po+po | 62/300K | 0/50K | 10/100K | 0/50K | 52/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.1+membar.ctas | 45/300K | 0/50K | 1/100K | 0/50K | 44/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.ctas | 1/300K | 0/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.1+membar.ctas | 51/300K | 0/50K | 1/100K | 0/50K | 50/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+membar.ctas | 35/300K | 0/50K | 2/100K | 0/50K | 33/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.1+membar.ctas | 57/300K | 0/50K | 1/100K | 0/50K | 56/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+membar.ctas | 46/300K | 0/50K | 8/100K | 0/50K | 38/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.gl+po+addr | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.gl+po+ctrl | 4/300K | 0/50K | 0/100K | 3/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.gl+po+data | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+membar.gl+po+po | 17/300K | 0/50K | 0/100K | 17/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+membar.gl+po+po | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.1+membar.gl+po+po | 19/300K | 0/50K | 0/100K | 19/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.1+membar.gl+po+po | 2/250K | --- | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.1+po+membar.cta+addr | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.1+po+membar.cta+ctrl | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.1+po+membar.cta+ctrl | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.1+po+membar.cta+membar.cta | 50/300K | 0/50K | 2/100K | 0/50K | 48/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.1+po+membar.cta+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.1+po+membar.cta+membar.cta | 66/300K | 0/50K | 3/100K | 0/50K | 63/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+po+membar.cta+membar.cta | 68/300K | 0/50K | 7/100K | 0/50K | 61/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.1+po+membar.cta+membar.cta | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.1+po+membar.cta+membar.cta | 76/300K | 0/50K | 4/100K | 0/50K | 72/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+po+membar.cta+membar.cta | 62/300K | 0/50K | 9/100K | 0/50K | 53/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.1+po+membar.cta+po | 80/300K | 0/50K | 2/100K | 0/50K | 77/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.1+po+membar.cta+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+po+membar.cta+po | 6/300K | 0/50K | 0/100K | 4/50K | 0/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.1+po+membar.cta+po | 98/300K | 0/50K | 5/100K | 0/50K | 93/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+po+membar.cta+po | 79/300K | 0/50K | 11/100K | 1/50K | 66/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.1+po+membar.cta+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.1+po+membar.cta+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.1+po+membar.cta+po | 5/250K | --- | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.1+po+membar.cta+po | 10/300K | 0/50K | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.1+po+membar.cta+po | 8/250K | --- | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.1+po+membar.cta+po | 106/300K | 0/50K | 4/100K | 0/50K | 100/50K | 2/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+po+membar.cta+po | 199/300K | 0/50K | 13/100K | 110/50K | 76/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.1+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.1+po+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.1+po+membar.gl+po | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+po+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.1+po+membar.gl+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.1+po+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+po+membar.gl+po | 37/300K | 0/50K | 0/100K | 32/50K | 5/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+po+po+addr | 47/300K | 0/50K | 0/100K | 46/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+po+po+addr | 86/300K | 0/50K | 0/100K | 86/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.1+po+po+addr | 25/300K | 0/50K | 0/100K | 25/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.1+po+po+addr | 4/250K | --- | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.1+po+po+addr | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.1+po+po+addr | 6/250K | --- | 0/100K | 6/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.1+po+po+ctrl | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+po+po+ctrl | 26/300K | 0/50K | 0/100K | 26/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+po+po+ctrl | 69/300K | 0/50K | 0/100K | 69/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.1+po+po+ctrl | 12/300K | 0/50K | 0/100K | 12/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.1+po+po+ctrl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.1+po+po+ctrl | 7/250K | --- | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.1+po+po+ctrl | 5/300K | 0/50K | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.1+po+po+ctrl | 10/250K | --- | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.1+po+po+ctrl | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+po+po+ctrl | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+po+po+data | 88/300K | 0/50K | 0/100K | 85/50K | 1/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+po+po+data | 87/300K | 0/50K | 0/100K | 87/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.1+po+po+data | 24/300K | 0/50K | 0/100K | 24/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.1+po+po+data | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.1+po+po+data | 34/250K | --- | 0/100K | 34/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.1+po+po+data | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.1+po+po+data | 15/250K | --- | 0/100K | 15/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.1+po+po+data | 7/250K | --- | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.1+po+po+membar.cta | 49/300K | 0/50K | 6/100K | 0/50K | 43/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1+po+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.1+po+po+membar.cta | 59/300K | 0/50K | 3/100K | 0/50K | 56/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+po+po+membar.cta | 79/300K | 0/50K | 2/100K | 16/50K | 61/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.1+po+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.1+po+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.1+po+po+membar.cta | 14/250K | --- | 0/100K | 14/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.1+po+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.1+po+po+membar.cta | 5/250K | --- | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.1+po+po+membar.cta | 9/250K | --- | 0/100K | 9/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.1+po+po+membar.cta | 78/300K | 0/50K | 5/100K | 0/50K | 73/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1+po+po+membar.cta | 49/300K | 0/50K | 9/100K | 0/50K | 40/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1+po+po+membar.gl | 11/300K | 0/50K | 0/100K | 11/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.1+po+po+membar.gl | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.1 | 77/300K | 0/50K | 7/100K | 0/50K | 70/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.1 | 232/300K | 0/50K | 0/100K | 230/50K | 1/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.1 | 109/300K | 0/50K | 5/100K | 0/50K | 104/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.1 | 265/300K | 0/50K | 7/100K | 165/50K | 93/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.1 | 273/300K | 0/50K | 0/100K | 273/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.1 | 323/300K | 0/50K | 0/100K | 323/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.1 | 335/250K | --- | 0/100K | 335/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.1 | 303/300K | 0/50K | 0/100K | 303/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.1 | 73/250K | --- | 0/100K | 73/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.1 | 91/250K | --- | 0/100K | 91/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.1 | 124/300K | 0/50K | 7/100K | 0/50K | 113/50K | 4/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.1 | 398/300K | 0/50K | 10/100K | 312/50K | 76/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+addr+membar.cta | 11/300K | 0/50K | 1/100K | 0/50K | 10/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+addr+membar.cta | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+addr+membar.cta | 29/300K | 0/50K | 1/100K | 0/50K | 28/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+membar.cta+addr+membar.cta | 8/300K | 0/50K | 0/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+membar.cta+addr+membar.cta | 15/300K | 0/50K | 0/100K | 0/50K | 15/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+addr+po | 19/300K | 0/50K | 0/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+addr+po | 15/300K | 0/50K | 0/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+addr+po | 44/300K | 0/50K | 4/100K | 0/50K | 40/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+membar.cta+addr+po | 13/300K | 0/50K | 0/100K | 0/50K | 13/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+membar.cta+addr+po | 23/300K | 0/50K | 1/100K | 0/50K | 22/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+ctrl+membar.cta | 13/300K | 0/50K | 0/100K | 0/50K | 13/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+ctrl+membar.cta | 7/300K | 0/50K | 1/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+ctrl+membar.cta | 30/300K | 0/50K | 1/100K | 0/50K | 29/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+membar.cta+ctrl+membar.cta | 8/300K | 0/50K | 1/100K | 0/50K | 7/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+membar.cta+ctrl+membar.cta | 9/300K | 0/50K | 0/100K | 0/50K | 9/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+ctrl+po | 18/300K | 0/50K | 0/100K | 0/50K | 18/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+ctrl+po | 14/300K | 0/50K | 0/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+ctrl+po | 28/300K | 0/50K | 2/100K | 0/50K | 26/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+membar.cta+ctrl+po | 11/300K | 0/50K | 1/100K | 0/50K | 10/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+membar.cta+ctrl+po | 23/300K | 0/50K | 3/100K | 0/50K | 20/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+data+membar.cta | 9/300K | 0/50K | 0/100K | 0/50K | 9/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+data+membar.cta | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+data+membar.cta | 44/300K | 0/50K | 1/100K | 0/50K | 43/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+membar.cta+data+membar.cta | 7/300K | 0/50K | 1/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+membar.cta+data+membar.cta | 11/300K | 0/50K | 2/100K | 0/50K | 9/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+data+po | 15/300K | 0/50K | 0/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+data+po | 20/300K | 0/50K | 2/100K | 0/50K | 18/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+data+po | 48/300K | 0/50K | 3/100K | 0/50K | 45/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+membar.cta+data+po | 15/300K | 0/50K | 1/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+membar.cta+data+po | 30/300K | 0/50K | 1/100K | 0/50K | 29/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.cta+membar.cta+data | 1/300K | 0/50K | 1/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+membar.cta+po | 38/300K | 0/50K | 1/100K | 0/50K | 37/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.2+membar.cta+membar.cta+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.cta+membar.cta+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+membar.cta+po | 58/300K | 0/50K | 5/100K | 0/50K | 53/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+membar.cta+po | 33/300K | 0/50K | 2/100K | 0/50K | 31/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+membar.cta+membar.cta+po | 31/300K | 0/50K | 3/100K | 0/50K | 27/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+membar.cta+membar.cta+po | 101/300K | 0/50K | 7/100K | 0/50K | 94/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+membar.cta+membar.gl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.cta+po+addr | 61/300K | 0/50K | 0/100K | 58/50K | 0/50K | 3/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+po+addr | 11/300K | 0/50K | 0/100K | 11/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.cta+po+addr | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+membar.cta+po+addr | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.cta+po+ctrl | 53/300K | 0/50K | 0/100K | 53/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+po+ctrl | 9/300K | 0/50K | 0/100K | 9/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.cta+po+data | 54/300K | 0/50K | 0/100K | 52/50K | 1/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+po+data | 8/300K | 0/50K | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.cta+po+data | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+po+membar.cta | 32/300K | 0/50K | 0/100K | 0/50K | 32/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.2+membar.cta+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.cta+po+membar.cta | 3/300K | 0/50K | 0/100K | 2/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+po+membar.cta | 30/300K | 0/50K | 2/100K | 0/50K | 28/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+po+membar.cta | 37/300K | 0/50K | 3/100K | 1/50K | 33/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.cta+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+membar.cta+po+membar.cta | 38/300K | 0/50K | 2/100K | 0/50K | 36/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+membar.cta+po+membar.cta | 34/300K | 0/50K | 7/100K | 0/50K | 27/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+po+membar.gl | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+po+po | 39/300K | 0/50K | 0/100K | 0/50K | 39/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.cta+po+po | 106/300K | 0/50K | 0/100K | 104/50K | 0/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.cta+po+po | 54/300K | 0/50K | 5/100K | 0/50K | 49/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.cta+po+po | 57/300K | 0/50K | 1/100K | 24/50K | 32/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.cta+po+po | 20/300K | 0/50K | 0/100K | 20/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.2+membar.cta+po+po | 33/300K | 0/50K | 0/100K | 33/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+membar.cta+po+po | 4/250K | --- | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.2+membar.cta+po+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.2+membar.cta+po+po | 15/250K | --- | 0/100K | 15/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+membar.cta+po+po | 49/300K | 0/50K | 6/100K | 0/50K | 42/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+membar.cta+po+po | 83/300K | 0/50K | 8/100K | 1/50K | 74/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.ctas | 31/300K | 0/50K | 2/100K | 0/50K | 29/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+membar.ctas | 36/300K | 0/50K | 6/100K | 0/50K | 30/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+membar.ctas | 39/300K | 0/50K | 2/100K | 0/50K | 37/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+membar.ctas | 34/300K | 0/50K | 3/100K | 0/50K | 31/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+membar.ctas | 50/300K | 0/50K | 4/100K | 0/50K | 46/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.gl+po+addr | 15/300K | 0/50K | 0/100K | 13/50K | 0/50K | 2/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.gl+po+ctrl | 13/300K | 0/50K | 0/100K | 12/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.gl+po+data | 15/300K | 0/50K | 0/100K | 14/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+membar.gl+po+po | 26/300K | 0/50K | 0/100K | 26/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.2+membar.gl+po+po | 12/300K | 0/50K | 0/100K | 12/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.2+membar.gl+po+po | 12/250K | --- | 0/100K | 12/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+addr+addr | 11/300K | 0/50K | 0/100K | 11/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+addr+ctrl | 18/300K | 0/50K | 0/100K | 18/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+addr+ctrl | 4/250K | --- | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+addr+data | 11/300K | 0/50K | 0/100K | 11/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+addr+membar.cta | 13/300K | 0/50K | 0/100K | 0/50K | 13/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+addr+membar.cta | 15/300K | 0/50K | 4/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+addr+membar.cta | 49/300K | 0/50K | 4/100K | 0/50K | 45/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+addr+membar.cta | 2/250K | --- | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+po+addr+membar.cta | 18/300K | 0/50K | 1/100K | 0/50K | 17/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+po+addr+membar.cta | 12/300K | 0/50K | 0/100K | 0/50K | 12/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+addr+po | 23/300K | 0/50K | 3/100K | 0/50K | 18/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+addr+po | 21/300K | 0/50K | 2/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+addr+po | 70/300K | 0/50K | 6/100K | 20/50K | 44/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.2+po+addr+po | 33/300K | 0/50K | 0/100K | 33/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+addr+po | 5/250K | --- | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.2+po+addr+po | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.2+po+addr+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+po+addr+po | 19/300K | 0/50K | 2/100K | 0/50K | 17/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+po+addr+po | 173/300K | 0/50K | 1/100K | 138/50K | 34/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+ctrl+addr | 12/300K | 0/50K | 0/100K | 12/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+ctrl+addr | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+ctrl+ctrl | 8/300K | 0/50K | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+ctrl+data | 7/300K | 0/50K | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+ctrl+data | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+ctrl+membar.cta | 23/300K | 0/50K | 3/100K | 0/50K | 20/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+ctrl+membar.cta | 15/300K | 0/50K | 1/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+ctrl+membar.cta | 49/300K | 0/50K | 2/100K | 2/50K | 45/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+ctrl+membar.cta | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+po+ctrl+membar.cta | 8/300K | 0/50K | 2/100K | 0/50K | 6/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+po+ctrl+membar.cta | 14/300K | 0/50K | 2/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+ctrl+membar.gl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+ctrl+po | 27/300K | 0/50K | 1/100K | 0/50K | 26/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+ctrl+po | 15/300K | 0/50K | 0/100K | 0/50K | 15/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+ctrl+po | 86/300K | 0/50K | 2/100K | 28/50K | 56/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.2+po+ctrl+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.2+po+ctrl+po | 51/300K | 0/50K | 0/100K | 51/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+ctrl+po | 9/250K | --- | 0/100K | 9/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.2+po+ctrl+po | 14/300K | 0/50K | 0/100K | 14/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+po+ctrl+po | 13/300K | 0/50K | 2/100K | 0/50K | 11/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+po+ctrl+po | 207/300K | 0/50K | 5/100K | 165/50K | 37/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+data+addr | 14/300K | 0/50K | 0/100K | 14/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+data+addr | 2/250K | --- | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+data+ctrl | 26/300K | 0/50K | 0/100K | 26/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+data+ctrl | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+data+data | 24/300K | 0/50K | 0/100K | 24/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+data+data | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+data+membar.cta | 12/300K | 0/50K | 0/100K | 0/50K | 12/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+data+membar.cta | 18/300K | 0/50K | 2/100K | 0/50K | 16/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+data+membar.cta | 51/300K | 0/50K | 3/100K | 6/50K | 42/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+data+membar.cta | 2/250K | --- | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+po+data+membar.cta | 9/300K | 0/50K | 2/100K | 0/50K | 7/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+po+data+membar.cta | 17/300K | 0/50K | 1/100K | 0/50K | 16/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+data+membar.gl | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+data+po | 20/300K | 0/50K | 1/100K | 0/50K | 19/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+data+po | 27/300K | 0/50K | 2/100K | 0/50K | 25/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+data+po | 108/300K | 0/50K | 4/100K | 51/50K | 53/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.2+po+data+po | 70/300K | 0/50K | 0/100K | 70/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+data+po | 19/250K | --- | 0/100K | 19/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.2+po+data+po | 10/300K | 0/50K | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.2+po+data+po | 2/250K | --- | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.2+po+data+po | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+po+data+po | 17/300K | 0/50K | 3/100K | 0/50K | 14/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+po+data+po | 179/300K | 0/50K | 1/100K | 144/50K | 34/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.2+po+membar.cta+addr | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+membar.cta+membar.cta | 40/300K | 0/50K | 1/100K | 0/50K | 39/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.2+po+membar.cta+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+membar.cta+membar.cta | 41/300K | 0/50K | 2/100K | 0/50K | 39/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+membar.cta+membar.cta | 41/300K | 0/50K | 2/100K | 0/50K | 39/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+membar.cta+membar.cta | 2/250K | --- | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+po+membar.cta+membar.cta | 34/300K | 0/50K | 2/100K | 0/50K | 32/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+po+membar.cta+membar.cta | 50/300K | 0/50K | 4/100K | 0/50K | 46/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+membar.cta+po | 47/300K | 0/50K | 2/100K | 0/50K | 45/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.2+po+membar.cta+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+membar.cta+po | 65/300K | 0/50K | 3/100K | 0/50K | 62/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+membar.cta+po | 52/300K | 0/50K | 6/100K | 0/50K | 46/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.2+po+membar.cta+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+membar.cta+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+po+membar.cta+po | 56/300K | 0/50K | 3/100K | 0/50K | 53/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+po+membar.cta+po | 146/300K | 0/50K | 10/100K | 33/50K | 103/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+po+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+membar.gl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+po+membar.gl+po | 9/300K | 0/50K | 0/100K | 8/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+po+po+addr | 164/300K | 0/50K | 0/100K | 164/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+po+addr | 50/300K | 0/50K | 0/100K | 50/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.2+po+po+addr | 82/300K | 0/50K | 0/100K | 82/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.2+po+po+addr | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+po+addr | 10/250K | --- | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.2+po+po+addr | 11/250K | --- | 0/100K | 11/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.2+po+po+addr | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+po+po+ctrl | 174/300K | 0/50K | 0/100K | 171/50K | 2/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+po+ctrl | 72/300K | 0/50K | 0/100K | 72/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.2+po+po+ctrl | 52/300K | 0/50K | 0/100K | 52/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+po+ctrl | 11/250K | --- | 0/100K | 11/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.2+po+po+ctrl | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.2+po+po+ctrl | 12/250K | --- | 0/100K | 12/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.2+po+po+ctrl | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+po+po+data | 255/300K | 0/50K | 1/100K | 250/50K | 0/50K | 4/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+po+data | 85/300K | 0/50K | 0/100K | 85/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.2+po+po+data | 90/300K | 0/50K | 0/100K | 90/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+po+data | 30/250K | --- | 0/100K | 30/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.2+po+po+data | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.2+po+po+data | 11/250K | --- | 0/100K | 11/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.2+po+po+data | 4/250K | --- | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+po+membar.cta | 35/300K | 0/50K | 1/100K | 0/50K | 34/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.2+po+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2+po+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2+po+po+membar.cta | 32/300K | 0/50K | 0/100K | 0/50K | 32/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+po+membar.cta | 67/300K | 0/50K | 4/100K | 14/50K | 49/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2+po+po+membar.cta | 10/250K | --- | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.2+po+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.2+po+po+membar.cta | 4/250K | --- | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.2+po+po+membar.cta | 6/250K | --- | 0/100K | 6/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2+po+po+membar.cta | 38/300K | 0/50K | 3/100K | 0/50K | 35/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2+po+po+membar.cta | 47/300K | 0/50K | 7/100K | 0/50K | 40/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2+po+po+membar.gl | 5/300K | 0/50K | 0/100K | 5/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.2+po+po+membar.gl | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.2 | 49/300K | 0/50K | 1/100K | 0/50K | 47/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.2 | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.2 | 367/300K | 0/50K | 0/100K | 363/50K | 0/50K | 4/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.2 | 55/300K | 0/50K | 2/100K | 0/50K | 53/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.2 | 183/300K | 0/50K | 6/100K | 142/50K | 35/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.2 | 591/300K | 0/50K | 0/100K | 591/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.2 | 204/300K | 0/50K | 0/100K | 204/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.2 | 216/250K | --- | 0/100K | 216/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.2 | 299/300K | 0/50K | 0/100K | 299/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.2 | 136/250K | --- | 0/100K | 136/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.2 | 77/250K | --- | 0/100K | 77/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.2 | 74/300K | 0/50K | 6/100K | 0/50K | 64/50K | 4/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.2 | 388/300K | 0/50K | 6/100K | 303/50K | 79/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.cta+membar.cta+po | 350/300K | 49/50K | 104/100K | 0/50K | 173/50K | 24/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.3+membar.cta+membar.cta+po | 160/300K | 75/50K | 51/100K | 0/50K | 12/50K | 22/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.cta+membar.cta+po | 275/300K | 150/50K | 77/100K | 3/50K | 2/50K | 43/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.cta+membar.cta+po | 425/300K | 109/50K | 73/100K | 0/50K | 229/50K | 14/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+membar.cta+membar.cta+po | 427/250K | 50/50K | 230/100K | --- | 131/50K | 16/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | Z6.3+membar.cta+membar.cta+po | 205/300K | 80/50K | 72/100K | 0/50K | 10/50K | 43/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.cta+membar.cta+po | 680/300K | 507/50K | 48/100K | 0/50K | 0/50K | 125/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.3+membar.cta+membar.cta+po | 346/250K | 15/50K | 269/100K | --- | 37/50K | 25/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.3+membar.cta+membar.cta+po | 64/200K | --- | 13/100K | --- | 4/50K | 47/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.3+membar.cta+membar.cta+po | 734/300K | 364/50K | 145/100K | 0/50K | 69/50K | 156/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.3+membar.cta+membar.cta+po | 622/250K | --- | 257/100K | 0/50K | 11/50K | 354/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.3+membar.cta+membar.cta+po | 310/200K | --- | 212/100K | --- | 95/50K | 3/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | Z6.3+membar.cta+membar.cta+po | 185/100K | --- | --- | --- | 15/50K | 170/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3+membar.cta+membar.cta+po | 426/300K | 53/50K | 132/100K | 0/50K | 195/50K | 46/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+membar.cta+membar.cta+po | 899/300K | 271/50K | 266/100K | 0/50K | 334/50K | 28/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.cta+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.cta+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+membar.cta+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.cta+membar.gl+po | 19/300K | 18/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.3+membar.cta+membar.gl+po | 35/300K | 35/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.cta+membar.gl+po | 99/300K | 85/50K | 0/100K | 0/50K | 0/50K | 14/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.cta+membar.gl+po | 53/300K | 50/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+membar.cta+membar.gl+po | 37/250K | 31/50K | 0/100K | --- | 6/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | Z6.3+membar.cta+membar.gl+po | 40/300K | 40/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.cta+membar.gl+po | 318/300K | 283/50K | 0/100K | 0/50K | 0/50K | 35/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.3+membar.cta+membar.gl+po | 3/250K | 2/50K | 0/100K | --- | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.3+membar.cta+membar.gl+po | 20/200K | --- | 0/100K | --- | 0/50K | 20/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.3+membar.cta+membar.gl+po | 161/300K | 161/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.3+membar.cta+membar.gl+po | 126/250K | --- | 0/100K | 0/50K | 0/50K | 126/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | Z6.3+membar.cta+membar.gl+po | 104/100K | --- | --- | --- | 0/50K | 104/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3+membar.cta+membar.gl+po | 26/300K | 22/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+membar.cta+membar.gl+po | 126/300K | 114/50K | 0/100K | 0/50K | 12/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.cta+po+addr | 15/300K | 0/50K | 1/100K | 12/50K | 0/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+membar.cta+po+addr | 7/300K | 0/50K | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.cta+po+addr | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.cta+po+ctrl | 11/300K | 0/50K | 0/100K | 10/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+membar.cta+po+ctrl | 6/300K | 0/50K | 0/100K | 6/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.cta+po+ctrl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.cta+po+membar.cta | 78/300K | 0/50K | 4/100K | 0/50K | 74/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.cta+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.cta+po+membar.cta | 82/300K | 0/50K | 4/100K | 0/50K | 78/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+membar.cta+po+membar.cta | 69/300K | 0/50K | 6/100K | 4/50K | 59/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3+membar.cta+po+membar.cta | 91/300K | 0/50K | 5/100K | 0/50K | 86/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+membar.cta+po+membar.cta | 96/300K | 0/50K | 9/100K | 0/50K | 87/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+membar.cta+po+membar.gl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.cta+po+po | 351/300K | 37/50K | 89/100K | 0/50K | 192/50K | 33/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.3+membar.cta+po+po | 172/300K | 80/50K | 55/100K | 0/50K | 16/50K | 21/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.cta+po+po | 410/300K | 174/50K | 76/100K | 98/50K | 3/50K | 59/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.cta+po+po | 410/300K | 88/50K | 72/100K | 0/50K | 236/50K | 14/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+membar.cta+po+po | 499/250K | 73/50K | 267/100K | --- | 138/50K | 21/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | Z6.3+membar.cta+po+po | 207/300K | 67/50K | 86/100K | 0/50K | 22/50K | 32/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.cta+po+po | 687/300K | 454/50K | 76/100K | 47/50K | 2/50K | 108/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.3+membar.cta+po+po | 370/250K | 8/50K | 308/100K | --- | 38/50K | 16/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.3+membar.cta+po+po | 74/200K | --- | 26/100K | --- | 2/50K | 46/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.3+membar.cta+po+po | 760/300K | 378/50K | 138/100K | 0/50K | 79/50K | 165/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.3+membar.cta+po+po | 715/250K | --- | 310/100K | 0/50K | 18/50K | 387/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.3+membar.cta+po+po | 344/200K | --- | 220/100K | --- | 117/50K | 7/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | Z6.3+membar.cta+po+po | 221/100K | --- | --- | --- | 24/50K | 197/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3+membar.cta+po+po | 455/300K | 53/50K | 145/100K | 0/50K | 211/50K | 46/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+membar.cta+po+po | 889/300K | 247/50K | 276/100K | 1/50K | 339/50K | 26/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.ctas | 75/300K | 0/50K | 2/100K | 0/50K | 73/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.3+membar.ctas | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.ctas | 2/300K | 0/50K | 1/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.ctas | 75/300K | 0/50K | 4/100K | 0/50K | 71/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+membar.ctas | 53/300K | 0/50K | 4/100K | 0/50K | 49/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3+membar.ctas | 74/300K | 0/50K | 6/100K | 0/50K | 68/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+membar.ctas | 74/300K | 0/50K | 7/100K | 0/50K | 67/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.gl+membar.cta+po | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.3+membar.gl+membar.cta+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.gl+membar.cta+po | 6/300K | 1/50K | 0/100K | 1/50K | 0/50K | 4/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+membar.gl+membar.cta+po | 2/250K | 2/50K | 0/100K | --- | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | Z6.3+membar.gl+membar.cta+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.gl+membar.cta+po | 19/300K | 17/50K | 0/100K | 0/50K | 0/50K | 2/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.3+membar.gl+membar.cta+po | 2/250K | 2/50K | 0/100K | --- | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.3+membar.gl+membar.cta+po | 2/300K | 2/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.3+membar.gl+membar.cta+po | 7/250K | --- | 0/100K | 0/50K | 0/50K | 7/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | Z6.3+membar.gl+membar.cta+po | 4/100K | --- | --- | --- | 0/50K | 4/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+membar.gl+membar.cta+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+membar.gl+membar.gl+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+membar.gl+membar.gl+po | 1/250K | 1/50K | 0/100K | --- | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.gl+membar.gl+po | 3/300K | 3/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.gl+po+addr | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.gl+po+ctrl | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.gl+po+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 0/50K | 2/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.3+membar.gl+po+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.gl+po+po | 24/300K | 0/50K | 0/100K | 20/50K | 0/50K | 4/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+membar.gl+po+po | 14/300K | 12/50K | 0/100K | 0/50K | 0/50K | 2/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.3+membar.gl+po+po | 3/250K | 3/50K | 0/100K | --- | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.3+membar.gl+po+po | 1/200K | --- | 0/100K | --- | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.3+membar.gl+po+po | 3/300K | 3/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.3+membar.gl+po+po | 5/250K | --- | 0/100K | 0/50K | 0/50K | 5/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | Z6.3+membar.gl+po+po | 10/100K | --- | --- | --- | 0/50K | 10/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3+membar.gl+po+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+membar.gl+po+po | 1/300K | 1/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.3+po+membar.cta+addr | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.3+po+membar.cta+addr | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+po+membar.cta+membar.cta | 72/300K | 0/50K | 3/100K | 0/50K | 69/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+po+membar.cta+membar.cta | 3/300K | 0/50K | 0/100K | 0/50K | 1/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+po+membar.cta+membar.cta | 106/300K | 0/50K | 9/100K | 0/50K | 97/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+po+membar.cta+membar.cta | 79/300K | 0/50K | 4/100K | 0/50K | 75/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3+po+membar.cta+membar.cta | 79/300K | 0/50K | 9/100K | 0/50K | 70/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+po+membar.cta+membar.cta | 111/300K | 0/50K | 7/100K | 0/50K | 104/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+po+membar.cta+po | 457/300K | 54/50K | 145/100K | 0/50K | 222/50K | 36/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.3+po+membar.cta+po | 254/300K | 87/50K | 87/100K | 0/50K | 30/50K | 50/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+po+membar.cta+po | 415/300K | 195/50K | 113/100K | 5/50K | 4/50K | 98/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+po+membar.cta+po | 544/300K | 100/50K | 104/100K | 0/50K | 310/50K | 30/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+po+membar.cta+po | 615/250K | 73/50K | 292/100K | --- | 225/50K | 25/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | Z6.3+po+membar.cta+po | 267/300K | 81/50K | 88/100K | 0/50K | 35/50K | 63/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+po+membar.cta+po | 876/300K | 565/50K | 90/100K | 4/50K | 5/50K | 212/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.3+po+membar.cta+po | 433/250K | 10/50K | 344/100K | --- | 55/50K | 24/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.3+po+membar.cta+po | 141/200K | --- | 39/100K | --- | 13/50K | 89/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.3+po+membar.cta+po | 914/300K | 397/50K | 203/100K | 17/50K | 119/50K | 178/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.3+po+membar.cta+po | 785/250K | --- | 367/100K | 3/50K | 21/50K | 394/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.3+po+membar.cta+po | 407/200K | --- | 248/100K | --- | 153/50K | 6/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | Z6.3+po+membar.cta+po | 267/100K | --- | --- | --- | 42/50K | 225/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3+po+membar.cta+po | 512/300K | 46/50K | 187/100K | 0/50K | 213/50K | 66/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+po+membar.cta+po | 1.2K/300K | 272/50K | 391/100K | 135/50K | 411/50K | 40/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+po+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+po+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+po+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+po+membar.gl+membar.cta | 6/300K | 0/50K | 0/100K | 0/50K | 6/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+po+membar.gl+po | 23/300K | 20/50K | 0/100K | 0/50K | 2/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.3+po+membar.gl+po | 40/300K | 40/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+po+membar.gl+po | 138/300K | 98/50K | 0/100K | 0/50K | 0/50K | 40/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+po+membar.gl+po | 61/300K | 53/50K | 0/100K | 0/50K | 8/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+po+membar.gl+po | 25/250K | 23/50K | 0/100K | --- | 2/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | Z6.3+po+membar.gl+po | 33/300K | 33/50K | 0/100K | 0/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+po+membar.gl+po | 322/300K | 257/50K | 0/100K | 0/50K | 0/50K | 65/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.3+po+membar.gl+po | 1/250K | 1/50K | 0/100K | --- | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.3+po+membar.gl+po | 32/200K | --- | 0/100K | --- | 0/50K | 32/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.3+po+membar.gl+po | 161/300K | 159/50K | 0/100K | 0/50K | 0/50K | 2/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.3+po+membar.gl+po | 128/250K | --- | 0/100K | 0/50K | 0/50K | 128/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | Z6.3+po+membar.gl+po | 138/100K | --- | --- | --- | 0/50K | 138/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3+po+membar.gl+po | 23/300K | 21/50K | 0/100K | 0/50K | 1/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+po+membar.gl+po | 152/300K | 102/50K | 0/100K | 47/50K | 3/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+po+po+addr | 58/300K | 0/50K | 0/100K | 56/50K | 0/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+po+po+addr | 92/300K | 0/50K | 0/100K | 92/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+po+po+addr | 27/300K | 0/50K | 0/100K | 27/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.3+po+po+addr | 4/250K | --- | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.3+po+po+addr | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.3+po+po+addr | 4/250K | --- | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.3+po+po+addr | 4/250K | --- | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+po+po+ctrl | 49/300K | 0/50K | 0/100K | 49/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+po+po+ctrl | 77/300K | 0/50K | 0/100K | 77/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+po+po+ctrl | 13/300K | 0/50K | 0/100K | 13/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.3+po+po+ctrl | 15/250K | --- | 0/100K | 15/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.3+po+po+ctrl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.3+po+po+ctrl | 4/250K | --- | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.3+po+po+ctrl | 7/250K | --- | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3+po+po+membar.cta | 90/300K | 0/50K | 4/100K | 0/50K | 86/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3+po+po+membar.cta | 3/300K | 0/50K | 0/100K | 2/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3+po+po+membar.cta | 106/300K | 0/50K | 3/100K | 0/50K | 103/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+po+po+membar.cta | 97/300K | 0/50K | 10/100K | 15/50K | 72/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3+po+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.3+po+po+membar.cta | 14/250K | --- | 0/100K | 14/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.3+po+po+membar.cta | 8/250K | --- | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3+po+po+membar.cta | 76/300K | 0/50K | 6/100K | 0/50K | 70/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3+po+po+membar.cta | 110/300K | 0/50K | 8/100K | 1/50K | 101/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3+po+po+membar.gl | 9/300K | 0/50K | 0/100K | 9/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.3+po+po+membar.gl | 2/250K | --- | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.3 | 497/300K | 75/50K | 154/100K | 0/50K | 221/50K | 47/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.3 | 265/300K | 94/50K | 78/100K | 0/50K | 34/50K | 59/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.3 | 762/300K | 204/50K | 124/100K | 302/50K | 13/50K | 119/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.3 | 566/300K | 126/50K | 97/100K | 0/50K | 307/50K | 36/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.3 | 640/250K | 88/50K | 311/100K | --- | 208/50K | 33/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: global | Z6.3 | 330/300K | 98/50K | 114/100K | 0/50K | 49/50K | 69/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.3 | 1.2K/300K | 515/50K | 121/100K | 386/50K | 5/50K | 217/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.3 | 510/250K | 12/50K | 397/100K | --- | 77/50K | 24/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.3 | 148/200K | --- | 37/100K | --- | 14/50K | 97/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.3 | 1.4K/300K | 424/50K | 225/100K | 433/50K | 123/50K | 205/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.3 | 1.0K/250K | --- | 414/100K | 126/50K | 30/50K | 437/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.3 | 457/200K | --- | 266/100K | --- | 186/50K | 5/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: shared | Z6.3 | 309/100K | --- | --- | --- | 66/50K | 243/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.3 | 613/300K | 64/50K | 214/100K | 0/50K | 256/50K | 79/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.3 | 1.5K/300K | 264/50K | 343/100K | 379/50K | 459/50K | 46/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.4+membar.cta+membar.cta+po | 46/300K | 0/50K | 3/100K | 0/50K | 43/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.4+membar.cta+membar.cta+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.4+membar.cta+membar.cta+po | 8/300K | 0/50K | 0/100K | 6/50K | 0/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.4+membar.cta+membar.cta+po | 65/300K | 0/50K | 5/100K | 0/50K | 60/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+membar.cta+membar.cta+po | 46/300K | 0/50K | 4/100K | 0/50K | 42/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.4+membar.cta+membar.cta+po | 44/300K | 0/50K | 1/100K | 0/50K | 42/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4+membar.cta+membar.cta+po | 43/300K | 0/50K | 6/100K | 0/50K | 36/50K | 1/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.4+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.4+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+membar.cta+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.4+membar.cta+po+membar.cta | 35/300K | 0/50K | 2/100K | 0/50K | 33/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.4+membar.cta+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.4+membar.cta+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.4+membar.cta+po+membar.cta | 48/300K | 0/50K | 3/100K | 0/50K | 45/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+membar.cta+po+membar.cta | 61/300K | 0/50K | 4/100K | 17/50K | 40/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.4+membar.cta+po+membar.cta | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.4+membar.cta+po+membar.cta | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.4+membar.cta+po+membar.cta | 33/300K | 0/50K | 1/100K | 0/50K | 30/50K | 2/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4+membar.cta+po+membar.cta | 42/300K | 0/50K | 7/100K | 0/50K | 35/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+membar.cta+po+membar.gl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.4+membar.cta+po+po | 63/300K | 0/50K | 2/100K | 0/50K | 60/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.4+membar.cta+po+po | 110/300K | 0/50K | 0/100K | 109/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.4+membar.cta+po+po | 66/300K | 0/50K | 4/100K | 0/50K | 62/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+membar.cta+po+po | 83/300K | 0/50K | 5/100K | 26/50K | 52/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.4+membar.cta+po+po | 50/300K | 0/50K | 0/100K | 50/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.4+membar.cta+po+po | 45/300K | 0/50K | 0/100K | 45/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.4+membar.cta+po+po | 10/250K | --- | 0/100K | 10/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.4+membar.cta+po+po | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.4+membar.cta+po+po | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.4+membar.cta+po+po | 41/250K | --- | 0/100K | 41/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.4+membar.cta+po+po | 59/300K | 0/50K | 3/100K | 0/50K | 55/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4+membar.cta+po+po | 42/300K | 0/50K | 8/100K | 0/50K | 34/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.4+membar.ctas | 35/300K | 0/50K | 3/100K | 0/50K | 31/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.4+membar.ctas | 3/300K | 0/50K | 0/100K | 0/50K | 0/50K | 3/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.4+membar.ctas | 42/300K | 0/50K | 1/100K | 0/50K | 41/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+membar.ctas | 32/300K | 0/50K | 0/100K | 0/50K | 32/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.4+membar.ctas | 45/300K | 0/50K | 0/100K | 0/50K | 45/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4+membar.ctas | 42/300K | 0/50K | 12/100K | 0/50K | 30/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.4+membar.gl+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+membar.gl+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.4+membar.gl+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.4+membar.gl+po+membar.cta | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.4+membar.gl+po+po | 32/300K | 0/50K | 0/100K | 32/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+membar.gl+po+po | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.4+membar.gl+po+po | 28/300K | 0/50K | 0/100K | 28/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.4+membar.gl+po+po | 26/250K | --- | 0/100K | 26/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.4+po+membar.cta+membar.cta | 50/300K | 0/50K | 3/100K | 0/50K | 46/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.4+po+membar.cta+membar.cta | 62/300K | 0/50K | 5/100K | 0/50K | 57/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+po+membar.cta+membar.cta | 57/300K | 0/50K | 4/100K | 0/50K | 53/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.4+po+membar.cta+membar.cta | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.4+po+membar.cta+membar.cta | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.4+po+membar.cta+membar.cta | 58/300K | 0/50K | 8/100K | 0/50K | 49/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4+po+membar.cta+membar.cta | 56/300K | 0/50K | 9/100K | 10/50K | 37/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.4+po+membar.cta+po | 81/300K | 0/50K | 6/100K | 0/50K | 75/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.4+po+membar.cta+po | 24/300K | 0/50K | 0/100K | 23/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.4+po+membar.cta+po | 79/300K | 0/50K | 5/100K | 0/50K | 74/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+po+membar.cta+po | 76/300K | 0/50K | 13/100K | 2/50K | 61/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.4+po+membar.cta+po | 7/300K | 0/50K | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.4+po+membar.cta+po | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.4+po+membar.cta+po | 15/250K | --- | 0/100K | 15/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.4+po+membar.cta+po | 24/300K | 0/50K | 0/100K | 24/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.4+po+membar.cta+po | 24/250K | --- | 0/100K | 24/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.4+po+membar.cta+po | 62/300K | 0/50K | 5/100K | 0/50K | 56/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4+po+membar.cta+po | 159/300K | 0/50K | 11/100K | 104/50K | 44/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.4+po+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.4+po+membar.gl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.4+po+membar.gl+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.4+po+membar.gl+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4+po+membar.gl+po | 35/300K | 0/50K | 0/100K | 33/50K | 2/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.4+po+po+membar.cta | 40/300K | 0/50K | 2/100K | 0/50K | 38/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.4+po+po+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.4+po+po+membar.cta | 3/300K | 0/50K | 0/100K | 0/50K | 0/50K | 3/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.4+po+po+membar.cta | 60/300K | 0/50K | 4/100K | 0/50K | 56/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+po+po+membar.cta | 150/300K | 0/50K | 9/100K | 96/50K | 45/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.4+po+po+membar.cta | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.4+po+po+membar.cta | 26/300K | 0/50K | 0/100K | 26/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.4+po+po+membar.cta | 18/250K | --- | 0/100K | 18/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.4+po+po+membar.cta | 11/300K | 0/50K | 0/100K | 11/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.4+po+po+membar.cta | 4/250K | --- | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.4+po+po+membar.cta | 26/250K | --- | 0/100K | 26/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.4+po+po+membar.cta | 56/300K | 0/50K | 3/100K | 0/50K | 52/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4+po+po+membar.cta | 67/300K | 0/50K | 15/100K | 29/50K | 23/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4+po+po+membar.gl | 13/300K | 0/50K | 0/100K | 13/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.4+po+po+membar.gl | 13/250K | --- | 0/100K | 13/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.4 | 70/300K | 0/50K | 3/100K | 0/50K | 67/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.4 | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.4 | 382/300K | 0/50K | 0/100K | 380/50K | 0/50K | 2/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.4 | 86/300K | 0/50K | 10/100K | 0/50K | 76/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.4 | 314/300K | 0/50K | 7/100K | 261/50K | 46/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.4 | 639/300K | 0/50K | 0/100K | 639/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.4 | 510/300K | 0/50K | 0/100K | 510/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.4 | 696/250K | --- | 0/100K | 696/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.4 | 379/300K | 0/50K | 0/100K | 379/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.4 | 355/250K | --- | 0/100K | 355/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.4 | 462/250K | --- | 0/100K | 462/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.4 | 69/300K | 0/50K | 3/100K | 0/50K | 64/50K | 2/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.4 | 326/300K | 0/50K | 8/100K | 278/50K | 40/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.5+membar.cta+membar.cta+po | 41/300K | 0/50K | 3/100K | 0/50K | 38/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.5+membar.cta+membar.cta+po | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.5+membar.cta+membar.cta+po | 59/300K | 0/50K | 5/100K | 0/50K | 54/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+membar.cta+membar.cta+po | 41/300K | 0/50K | 3/100K | 0/50K | 38/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.5+membar.cta+membar.cta+po | 46/300K | 0/50K | 5/100K | 0/50K | 41/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.5+membar.cta+membar.cta+po | 67/300K | 0/50K | 5/100K | 0/50K | 62/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.5+membar.cta+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.5+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+membar.cta+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.5+membar.cta+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+membar.cta+membar.gl+po | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.5+membar.cta+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.5+membar.cta+membar.gl+po | 3/300K | 0/50K | 0/100K | 0/50K | 3/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.5+membar.cta+po+membar.cta | 28/300K | 0/50K | 2/100K | 0/50K | 26/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.5+membar.cta+po+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.5+membar.cta+po+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 1/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.5+membar.cta+po+membar.cta | 32/300K | 0/50K | 1/100K | 0/50K | 31/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+membar.cta+po+membar.cta | 46/300K | 0/50K | 1/100K | 9/50K | 36/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.5+membar.cta+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.5+membar.cta+po+membar.cta | 2/300K | 0/50K | 0/100K | 2/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.5+membar.cta+po+membar.cta | 3/250K | --- | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.5+membar.cta+po+membar.cta | 56/300K | 0/50K | 3/100K | 0/50K | 53/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.5+membar.cta+po+membar.cta | 43/300K | 0/50K | 7/100K | 0/50K | 36/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+membar.cta+po+membar.gl | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.5+membar.cta+po+po | 57/300K | 0/50K | 4/100K | 0/50K | 53/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.5+membar.cta+po+po | 88/300K | 0/50K | 0/100K | 85/50K | 2/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.5+membar.cta+po+po | 33/300K | 0/50K | 2/100K | 0/50K | 31/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+membar.cta+po+po | 58/300K | 0/50K | 3/100K | 27/50K | 28/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.5+membar.cta+po+po | 30/300K | 0/50K | 0/100K | 30/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.5+membar.cta+po+po | 33/300K | 0/50K | 0/100K | 33/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.5+membar.cta+po+po | 24/250K | --- | 0/100K | 24/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.5+membar.cta+po+po | 16/250K | --- | 0/100K | 16/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.5+membar.cta+po+po | 54/300K | 0/50K | 3/100K | 0/50K | 50/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.5+membar.cta+po+po | 70/300K | 0/50K | 6/100K | 3/50K | 61/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.5+membar.ctas | 30/300K | 0/50K | 1/100K | 0/50K | 28/50K | 1/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.5+membar.ctas | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.5+membar.ctas | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.5+membar.ctas | 34/300K | 0/50K | 0/100K | 0/50K | 34/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+membar.ctas | 36/300K | 0/50K | 4/100K | 0/50K | 32/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.5+membar.ctas | 48/300K | 0/50K | 3/100K | 0/50K | 44/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.5+membar.ctas | 38/300K | 0/50K | 5/100K | 0/50K | 33/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.5+membar.gl+membar.cta+po | 1/300K | 0/50K | 0/100K | 0/50K | 0/50K | 1/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.5+membar.gl+po+membar.cta | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.5+membar.gl+po+po | 35/300K | 0/50K | 0/100K | 35/50K | 0/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+membar.gl+po+po | 3/300K | 0/50K | 0/100K | 3/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.5+membar.gl+po+po | 27/300K | 0/50K | 0/100K | 27/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.5+membar.gl+po+po | 11/250K | --- | 0/100K | 11/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.5+po+membar.cta+membar.cta | 35/300K | 0/50K | 2/100K | 0/50K | 33/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.5+po+membar.cta+membar.cta | 51/300K | 0/50K | 2/100K | 0/50K | 49/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+po+membar.cta+membar.cta | 58/300K | 0/50K | 6/100K | 3/50K | 49/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.5+po+membar.cta+membar.cta | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.5+po+membar.cta+membar.cta | 56/300K | 0/50K | 6/100K | 0/50K | 49/50K | 1/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.5+po+membar.cta+membar.cta | 52/300K | 0/50K | 2/100K | 10/50K | 40/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.5+po+membar.cta+po | 52/300K | 0/50K | 3/100K | 0/50K | 49/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.5+po+membar.cta+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.5+po+membar.cta+po | 8/300K | 0/50K | 0/100K | 6/50K | 1/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.5+po+membar.cta+po | 68/300K | 0/50K | 2/100K | 0/50K | 66/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+po+membar.cta+po | 61/300K | 0/50K | 6/100K | 2/50K | 53/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.5+po+membar.cta+po | 4/300K | 0/50K | 0/100K | 4/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.5+po+membar.cta+po | 15/250K | --- | 0/100K | 15/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.5+po+membar.cta+po | 22/300K | 0/50K | 0/100K | 22/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.5+po+membar.cta+po | 6/250K | --- | 0/100K | 6/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.5+po+membar.cta+po | 77/300K | 0/50K | 4/100K | 0/50K | 73/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.5+po+membar.cta+po | 171/300K | 0/50K | 8/100K | 90/50K | 73/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.5+po+membar.gl+membar.cta | 4/300K | 0/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+po+membar.gl+membar.cta | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.5+po+membar.gl+membar.cta | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.5+po+membar.gl+membar.cta | 5/300K | 0/50K | 0/100K | 2/50K | 3/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.5+po+membar.gl+po | 4/300K | 0/50K | 0/100K | 0/50K | 4/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+po+membar.gl+po | 1/300K | 0/50K | 0/100K | 0/50K | 1/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.5+po+membar.gl+po | 1/300K | 0/50K | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.5+po+membar.gl+po | 21/300K | 0/50K | 0/100K | 20/50K | 1/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.5+po+po+membar.cta | 42/300K | 0/50K | 2/100K | 0/50K | 40/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.5+po+po+membar.cta | 3/300K | 0/50K | 0/100K | 2/50K | 0/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.5+po+po+membar.cta | 48/300K | 0/50K | 4/100K | 0/50K | 44/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+po+po+membar.cta | 140/300K | 0/50K | 3/100K | 95/50K | 42/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.5+po+po+membar.cta | 65/300K | 0/50K | 0/100K | 65/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.5+po+po+membar.cta | 13/250K | --- | 0/100K | 13/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.5+po+po+membar.cta | 7/300K | 0/50K | 0/100K | 7/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.5+po+po+membar.cta | 6/250K | --- | 0/100K | 6/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.5+po+po+membar.cta | 25/250K | --- | 0/100K | 25/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.5+po+po+membar.cta | 61/300K | 0/50K | 4/100K | 0/50K | 57/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.5+po+po+membar.cta | 88/300K | 0/50K | 8/100K | 31/50K | 49/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5+po+po+membar.gl | 27/300K | 0/50K | 0/100K | 27/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.5+po+po+membar.gl | 1/250K | --- | 0/100K | 1/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.5+po+po+membar.gl | 8/250K | --- | 0/100K | 8/50K | 0/50K | 0/50K |
| P0 |cta P1 |cta P2 | x: global, y: global, z: global | Z6.5 | 42/300K | 0/50K | 2/100K | 0/50K | 40/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: global | Z6.5 | 2/300K | 0/50K | 0/100K | 0/50K | 2/50K | 0/50K |
| P0 |cta P1 |warp P2 | x: global, y: global, z: shared | Z6.5 | 325/300K | 0/50K | 0/100K | 322/50K | 2/50K | 1/50K |
| P0 |warp P1 |cta P2 | x: global, y: global, z: global | Z6.5 | 58/300K | 0/50K | 3/100K | 0/50K | 55/50K | 0/50K |
| P0 |warp P1 |cta P2 | x: global, y: shared, z: global | Z6.5 | 263/300K | 0/50K | 6/100K | 200/50K | 57/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: global, z: shared | Z6.5 | 345/300K | 0/50K | 0/100K | 345/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: global | Z6.5 | 390/300K | 0/50K | 0/100K | 390/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: global, y: shared, z: shared | Z6.5 | 703/250K | --- | 0/100K | 703/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: global | Z6.5 | 336/300K | 0/50K | 0/100K | 336/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: global, z: shared | Z6.5 | 108/250K | --- | 0/100K | 108/50K | 0/50K | 0/50K |
| P0 |warp P1 |warp P2 | x: shared, y: shared, z: global | Z6.5 | 273/250K | --- | 0/100K | 273/50K | 0/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: global, y: global, z: global | Z6.5 | 81/300K | 0/50K | 4/100K | 0/50K | 77/50K | 0/50K |
| P0 |warp P2 |cta P1 | x: shared, y: global, z: global | Z6.5 | 305/300K | 0/50K | 6/100K | 243/50K | 56/50K | 0/50K |