From ec5ea792fc10d4c03aa80e0f88522c81569c5464 Mon Sep 17 00:00:00 2001
From: Alejandro Acosta <alejandro.acosta@codeplay.com>
Date: Fri, 31 May 2024 16:03:36 +0200
Subject: [PATCH] Run sycl ampere benchmarks on cuda (#68)

* Add generic example runner

* Init d and ref_d with different values

* Move runner to benchmark folder

* Add generic example runner

* Add Ampere half-float example

* Add generic example runner

* Add Ampere half-float example

* Add Ampere half-float example

* Add Ampere bfloat-float example

* Add generic example runner

* Add Ampere half-float example

* Add Ampere bfloat-float example

* Run sycl ampere example on cuda
---
 CMakeLists.txt                                |  5 +----
 benchmarks/CMakeLists.txt                     | 22 ++++++++++---------
 benchmarks/ampere/CMakeLists.txt              |  4 ++--
 ...ere_gemm_bf16_bf16_fp32_tensor_op_fp32.cu} |  0
 ...ere_gemm_fp16_fp16_fp32_tensor_op_fp32.cu} |  0
 5 files changed, 15 insertions(+), 16 deletions(-)
 rename benchmarks/ampere/{bench_ampere_gemm_bf16_bf16_fp32_tensor_op_fp32.cpp => bench_ampere_gemm_bf16_bf16_fp32_tensor_op_fp32.cu} (100%)
 rename benchmarks/ampere/{bench_ampere_gemm_fp16_fp16_fp32_tensor_op_fp32.cpp => bench_ampere_gemm_fp16_fp16_fp32_tensor_op_fp32.cu} (100%)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 91408ddaa9..46b95fbaa6 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -159,10 +159,7 @@ set(CUTLASS_ENABLE_LIBRARY ${CUTLASS_ENABLE_LIBRARY_INIT} CACHE BOOL "Enable CUT
 set(CUTLASS_ENABLE_PROFILER ${CUTLASS_ENABLE_LIBRARY} CACHE BOOL "Enable CUTLASS Profiler")
 set(CUTLASS_ENABLE_PERFORMANCE ${CUTLASS_ENABLE_PROFILER} CACHE BOOL "Enable CUTLASS Performance")
 option(CUTLASS_ENABLE_DEBUG_PRINTS "Whether or not to enable debug prints in CUTLASS kernels" OFF)
-
-if (CUTLASS_ENABLE_SYCL)
-  set(CUTLASS_ENABLE_BENCHMARKS ON CACHE BOOL "Enable CUTLASS Benchmarks")
-endif()
+set(CUTLASS_ENABLE_BENCHMARKS ON CACHE BOOL "Enable CUTLASS Benchmarks")
 
 set(CUTLASS_ENABLE_TESTS ${CUTLASS_ENABLE_TESTS_INIT} CACHE BOOL "Enable CUTLASS Tests")
 set(CUTLASS_ENABLE_GTEST_UNIT_TESTS ${CUTLASS_ENABLE_TESTS} CACHE BOOL "Enable CUTLASS GTest-based Unit Tests")
diff --git a/benchmarks/CMakeLists.txt b/benchmarks/CMakeLists.txt
index d499f839a9..e4ec52b63f 100644
--- a/benchmarks/CMakeLists.txt
+++ b/benchmarks/CMakeLists.txt
@@ -43,30 +43,32 @@ function(cutlass_benchmark_add_executable NAME)
 
   add_dependencies(cutlass_benchmarks ${NAME})
 
+  if (NOT CUTLASS_ENABLE_SYCL)
+    SET(ADD_CUDA ON)
+  endif()
+
   target_link_libraries(
     ${NAME}
     PRIVATE
     CUTLASS
     cutlass_tools_util_includes
-    )
-
-  target_include_directories(
-    ${NAME}
-    PRIVATE
-    ${CUTLASS_BENCHMARKS_COMMON_SOURCE_DIR}
-    )
+    $<$<BOOL:${CUTLASS_ENABLE_CUBLAS}>:nvidia::cublas>
+    $<$<BOOL:${ADD_CUDA}>:cuda>
+  )
 
-  add_sycl_to_target(TARGET ${NAME})
+  if (CUTLASS_ENABLE_SYCL)
+    add_sycl_to_target(TARGET ${NAME})
+  endif()
 
   install(
     TARGETS ${NAME}
     RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
-    )
+  )
 endfunction()
 
 if(SYCL_INTEL_TARGET)
   add_subdirectory(pvc)
 endif()
-if (SYCL_NVIDIA_TARGET)
+if(SYCL_NVIDIA_TARGET OR NOT CUTLASS_ENABLE_SYCL)
   add_subdirectory(ampere)
 endif()
diff --git a/benchmarks/ampere/CMakeLists.txt b/benchmarks/ampere/CMakeLists.txt
index 666d9cac60..a77901594b 100644
--- a/benchmarks/ampere/CMakeLists.txt
+++ b/benchmarks/ampere/CMakeLists.txt
@@ -29,10 +29,10 @@
 
 cutlass_benchmark_add_executable(
   bench_ampere_gemm_fp16_fp16_fp32_tensor_op_fp32
-  bench_ampere_gemm_fp16_fp16_fp32_tensor_op_fp32.cpp
+  bench_ampere_gemm_fp16_fp16_fp32_tensor_op_fp32.cu
 )
 
 cutlass_benchmark_add_executable(
   bench_ampere_gemm_bf16_bf16_fp32_tensor_op_fp32
-  bench_ampere_gemm_bf16_bf16_fp32_tensor_op_fp32.cpp
+  bench_ampere_gemm_bf16_bf16_fp32_tensor_op_fp32.cu
 )
diff --git a/benchmarks/ampere/bench_ampere_gemm_bf16_bf16_fp32_tensor_op_fp32.cpp b/benchmarks/ampere/bench_ampere_gemm_bf16_bf16_fp32_tensor_op_fp32.cu
similarity index 100%
rename from benchmarks/ampere/bench_ampere_gemm_bf16_bf16_fp32_tensor_op_fp32.cpp
rename to benchmarks/ampere/bench_ampere_gemm_bf16_bf16_fp32_tensor_op_fp32.cu
diff --git a/benchmarks/ampere/bench_ampere_gemm_fp16_fp16_fp32_tensor_op_fp32.cpp b/benchmarks/ampere/bench_ampere_gemm_fp16_fp16_fp32_tensor_op_fp32.cu
similarity index 100%
rename from benchmarks/ampere/bench_ampere_gemm_fp16_fp16_fp32_tensor_op_fp32.cpp
rename to benchmarks/ampere/bench_ampere_gemm_fp16_fp16_fp32_tensor_op_fp32.cu