| 1 |
tests: |
| 2 |
- name: ComputeLoads |
| 3 |
inputs: |
| 4 |
- { M: 128, N: 128, K: 64, gpu_arch: 942, expected: 8192 } |
| 5 |
|
| 6 |
- name: EstimateL2Hit |
| 7 |
inputs: |
| 8 |
- { M: 4096, N: 4096, K: 1024, batch: 1, MT_M: 256, MT_N: 256, MT_K: 64, element_size: 16, splittingFactor: 3, gpu_arch: 942, expected_gt: 0.0, expected_lt: 1.0} |
| 9 |
|
| 10 |
- name: ComputeNumMatrixInstructions |
| 11 |
inputs: |
| 12 |
- { MT_M: 128, MT_N: 128, MT_K: 64, MI_M: 16, MI_N: 16, MI_K: 16, gpu_arch: 942, expected: 256} # 8 * 8 * 4 |
| 13 |
- { MT_M: 16, MT_N: 16, MT_K: 64, MI_M: 16, MI_N: 16, MI_K: 32, gpu_arch: 942, expected: 2} # 1 * 1 * 2 |
| 14 |
|
| 15 |
- name: ComputeMTComputeLatency |
| 16 |
inputs: |
| 17 |
- { M: 4096, N: 4096, K: 1024, transA: 1, transB: 0, MT_M: 128, MT_N: 128, MT_K: 64, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, gpu_arch: 942, expected: 4096} # 8 * 8 * 4 |
| 18 |
- { M: 4096, N: 4096, K: 1024, transA: 0, transB: 1, MT_M: 128, MT_N: 128, MT_K: 64, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, gpu_arch: 942, expected: 4096} # 8 * 8 * 4 |
| 19 |
- { M: 4096, N: 4096, K: 1024, transA: 0, transB: 0, MT_M: 128, MT_N: 128, MT_K: 64, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, gpu_arch: 942, expected: 4096} # 8 * 8 * 4 |
| 20 |
- { M: 4096, N: 4096, K: 1024, transA: 1, transB: 1, MT_M: 128, MT_N: 128, MT_K: 64, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, gpu_arch: 942, expected: 4096} # 8 * 8 * 4 |
| 21 |
- { M: 4096, N: 4096, K: 1024, transA: 1, transB: 1, MT_M: 128, MT_N: 128, MT_K: 32, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, gpu_arch: 942, expected_gt: 2048} # 8 * 8 * 4 |
| 22 |
- { M: 4096, N: 4096, K: 1024, transA: 1, transB: 1, MT_M: 224, MT_N: 224, MT_K: 64, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, gpu_arch: 942, expected_gt: 12543} # 8 * 8 * 4 |
| 23 |
- { M: 4096, N: 4096, K: 1024, transA: 1, transB: 1, MT_M: 128, MT_N: 32, MT_K: 32, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, gpu_arch: 942, expected_gt: 512} # 8 * 8 * 4 |
| 24 |
- { M: 4096, N: 4096, K: 1024, transA: 1, transB: 1, MT_M: 32, MT_N: 128, MT_K: 32, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, gpu_arch: 942, expected_gt: 512} # 8 * 8 * 4 |
| 25 |
|
| 26 |
- name: ComputeMemoryLatency |
| 27 |
inputs: |
| 28 |
- { M: 4096, N: 4096, K: 1024, transA: 1, transB: 0, batch: 1, MT_M: 128, MT_N: 128, MT_K: 64, element_size_A: 16, element_size_B: 16, mx_block_size: 0, wgm: 8, numActiveCUs: 304, splittingFactor: 2, gpu_arch: 942} |
| 29 |
|
| 30 |
- name: ComputeTileLatency |
| 31 |
inputs: |
| 32 |
- { M: 4096, N: 4096, K: 1024, transA: 1, transB: 0, batch: 2, MT_M: 128, MT_N: 128, MT_K: 64, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, element_size_out: 32, mx_block_size: 0, wgm: 6, numActiveCUs: 304, splittingFactor: 3, gpu_arch: 942} |
| 33 |
|
| 34 |
- name: ComputeWaveLatency |
| 35 |
inputs: |
| 36 |
- { M: 4096, N: 4096, K: 1024, transA: 1, transB: 0, batch: 2, MT_M: 128, MT_N: 128, MT_K: 64, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, element_size_out: 32, mx_block_size: 0, wgm: 8, numActiveCUs: 304, splittingFactor: 4, gpu_arch: 942} |
| 37 |
|
| 38 |
- name: ComputeTotalLatency |
| 39 |
inputs: |
| 40 |
- { M: 4096, N: 4096, K: 1024, transA: 1, transB: 0, batch: 2, MT_M: 128, MT_N: 128, MT_K: 64, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, element_size_out: 32, mx_block_size: 0, wgm: 1, splittingFactor: 6, gpu_arch: 942} |
| 41 |
|
| 42 |
- name: ComputePerfGflops |
| 43 |
inputs: |
| 44 |
- { M: 4096, N: 4096, K: 1024, transA: 1, transB: 0, batch: 2, MT_M: 128, MT_N: 128, MT_K: 64, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, element_size_out: 32, WGM: 1, gpu_arch: 942} |
| 45 |
|
| 46 |
- name: EstimateMallHit |
| 47 |
inputs: |
| 48 |
- { M: 4096, N: 4096, K: 1024, batch: 1, MT_M: 256, MT_N: 256, MT_K: 64, numActiveCUs: 304, splittingFactor: 8, gpu_arch: 942, expected_gt: 0} |
| 49 |
|
| 50 |
- name: HardwareArchEnum |
| 51 |
inputs: |
| 52 |
- { gpu_arch: 942} |
| 53 |
|
| 54 |
- name: BestGridSize |
| 55 |
inputs: |
| 56 |
- { M: 1024, N: 1024, K: 4096, batch: 1, transA: 1, transB: 0, MT_M: 256, MT_N: 256, MT_K: 64, MI_M: 32, MI_N: 32, MI_K: 8, element_size_A: 16, element_size_B: 16, element_size_out: 32, mx_block_size: 0, H_L2: 0.0, WGM: 1, biggest_allowable_split: 20, gpu_arch: 942, expected_gt: 16} |
| 57 |
|
| 58 |
- name: BestMacroTileSize |
| 59 |
inputs: |
| 60 |
- { M: 1024, N: 1024, K: 4096, batch: 1, transA: 1, transB: 0, element_size_A: 16, element_size_B: 16, element_size_out: 32, mx_block_size: 0, H_L2: 0.0, WGM: 1, gpu_arch: 942} |
| 61 |
|
| 62 |
- name: BestWGM |
| 63 |
inputs: |
| 64 |
- { M: 4096, N: 4096, K: 8192, batch: 1, MT_M: 256, MT_N: 256, MT_K: 32, MI_M: 32, MI_N: 32, MI_K: 8, element_size: 16, H_L2: 0.0, gpu_arch: 942} |
| 65 |
|
| 66 |
- name: UtilsTFlopsFromLatency |
| 67 |
inputs: |
| 68 |
- { M: 10, N: 10, K: 10, batch: 1, latency_cycles: 256, clock_GHz: 256, gpu_arch: 942} |