tests: - name: ComputeLoads inputs: - { M: 128, N: 128, K: 64, gpu_arch: 942, expected: 8192 } - name: EstimateL2Hit inputs: - { 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} - name: ComputeNumMatrixInstructions inputs: - { 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 - { 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 - name: ComputeMTComputeLatency inputs: - { 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 - { 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 - { 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 - { 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 - { 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 - { 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 - { 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 - { 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 - name: ComputeMemoryLatency inputs: - { 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} - name: ComputeTileLatency inputs: - { 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} - name: ComputeWaveLatency inputs: - { 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} - name: ComputeTotalLatency inputs: - { 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} - name: ComputePerfGflops inputs: - { 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} - name: EstimateMallHit inputs: - { 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} - name: HardwareArchEnum inputs: - { gpu_arch: 942} - name: BestGridSize inputs: - { 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} - name: BestMacroTileSize inputs: - { 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} - name: BestWGM inputs: - { 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} - name: UtilsTFlopsFromLatency inputs: - { M: 10, N: 10, K: 10, batch: 1, latency_cycles: 256, clock_GHz: 256, gpu_arch: 942}