Benchmarks - Support T4 and A10 in GEMM benchmark (#294)

Support T4 and A10 in GEMM benchmark.

superbench/benchmarks/micro_benchmarks/cuda_gemm_flops_performance.py

@@ -24,7 +24,7 @@ def __init__(self, name, parameters=''):
 
         self._bin_name = 'cutlass_profiler'
 
-        # TODO - To support more architecutres, currently only support compute capability = 7.0 and 8.0
+        # TODO - To support more architecutres, currently only support compute capability = 7.0, 7.5, 8.0, 8.6
         self.__kernel_map = {
             7.0: {
                 'fp64': 'cutlass_simt_dgemm_128x128_8x2_*',
@@ -44,6 +44,9 @@ def __init__(self, name, parameters=''):
                 'int4_tc': 'cutlass_tensorop_s4_i16864gemm_s4_256x128_128x3_*',
             }
         }
+        # Skip FP64 for RTX Turing/Ampere and Tesla T4/GA10x due to very limited FP64 TFLOP rate
+        self.__kernel_map[7.5] = {k: self.__kernel_map[7.0][k] for k in self.__kernel_map[7.0] if 'fp64' not in k}
+        self.__kernel_map[8.6] = {k: self.__kernel_map[8.0][k] for k in self.__kernel_map[8.0] if 'fp64' not in k}
         self.__parse_logline = [
             'gemm,cutlass_simt_dgemm_128x128_8x2', 'gemm,cutlass_simt_sgemm_128x128_8x2',
             'gemm,cutlass_simt_hgemm_256x128_8x2', 'gemm,cutlass_tensorop_d884gemm_128x128_16x3',
@@ -70,8 +73,8 @@ def _preprocess(self):
             super()._preprocess()
             self._result.set_return_code(ReturnCode.MICROBENCHMARK_UNSUPPORTED_ARCHITECTURE)
             logger.error(
-                'Unsupported architecture - benchmark: {}, compute capability: {}, expected: 7.0 or 8.0'.format(
-                    self._name, capability
+                'Unsupported architecture - benchmark: {}, compute capability: {}, supports {}'.format(
+                    self._name, capability, ' '.join(sorted([str(k) for k in self.__kernel_map]))
                 )
             )
             return False

tests/benchmarks/micro_benchmarks/test_cuda_gemm_flops_performance.py

@@ -35,7 +35,7 @@ def test_flops_performance_cuda(self):
         )
 
         ret = benchmark._preprocess()
-        if dm.device_manager.get_device_compute_capability() not in [7.0, 8.0]:
+        if dm.device_manager.get_device_compute_capability() not in benchmark._CudaGemmFlopsBenchmark__kernel_map:
             assert (ret is False)
             assert (benchmark.return_code == ReturnCode.MICROBENCHMARK_UNSUPPORTED_ARCHITECTURE)
         else:

third_party/Makefile

@@ -27,7 +27,7 @@ sb_micro_path:
 cuda_cutlass:
 ifneq (,$(wildcard cutlass/CMakeLists.txt))
 	cmake -DCMAKE_INSTALL_BINDIR=$(SB_MICRO_PATH)/bin -DCMAKE_INSTALL_LIBDIR=$(SB_MICRO_PATH)/lib -DCMAKE_BUILD_TYPE=Release \
-		-DCUTLASS_NVCC_ARCHS='70;80' -DCUTLASS_ENABLE_EXAMPLES=OFF -DCUTLASS_ENABLE_TESTS=OFF -S ./cutlass -B ./cutlass/build
+		-DCUTLASS_NVCC_ARCHS='70;75;80;86' -DCUTLASS_ENABLE_EXAMPLES=OFF -DCUTLASS_ENABLE_TESTS=OFF -S ./cutlass -B ./cutlass/build
 	cmake --build ./cutlass/build -j 8 --target install
 endif