Remove hipblaslt on Windows (#1)

* Remove hipblaslt on Windows

* Additional changes with hipblaslt skipped

* changed Blas.cpp

* fix format

Author: tvukovic-amd

aten/src/ATen/cuda/CUDABlas.cpp

@@ -346,7 +346,7 @@ void bgemm<at::BFloat16>(CUDABLAS_BGEMM_ARGTYPES(at::BFloat16)) {
   const float fbeta = beta;
   _cublasAdjustLdLevel3(transa, transb, m, n, k, &lda, &ldb, &ldc);
 
-#if defined(USE_ROCM) && ROCM_VERSION >= 60000
+#if defined(USE_ROCM) && ROCM_VERSION >= 60000 && !defined(_WIN32)
   auto compute_type = CUBLAS_COMPUTE_32F;
 #else
   auto compute_type = CUDA_R_32F;
@@ -529,7 +529,7 @@ void gemm<at::BFloat16>(CUDABLAS_GEMM_ARGTYPES(at::BFloat16)) {
     cublas_flags = static_cast<cublasMath_t>(cublas_flags | CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION);
   }
 #endif
-#if defined(USE_ROCM) && ROCM_VERSION >= 60000
+#if defined(USE_ROCM) && !defined(_WIN32) && ROCM_VERSION >= 60000
   auto compute_type = CUBLAS_COMPUTE_32F;
 #else
   auto compute_type = CUDA_R_32F;
@@ -558,7 +558,7 @@ void gemm<at::BFloat16>(CUDABLAS_GEMM_ARGTYPES(at::BFloat16)) {
   TORCH_CUDABLAS_CHECK(cublasSetMathMode(handle, CUBLAS_DEFAULT_MATH));
 }
 
-#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && ROCM_VERSION >= 50700)
+#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && !defined(_WIN32) && ROCM_VERSION >= 50700)
 
 #if defined(USE_ROCM) && ROCM_VERSION >= 50700 && ROCM_VERSION < 60000
 // only for rocm 5.7 where we first supported hipblaslt, it was difficult
@@ -1102,7 +1102,7 @@ void int8_gemm(
   TORCH_CHECK(false, "int8_gemm is only supported for ROCm 6.0 and above");
 #endif // !defined(USE_ROCM) || (defined(USE_ROCM) && ROCM_VERSION >= 60000)
 }
-#endif // (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && ROCM_VERSION >= 50700)
+#endif // (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && !defined(_WIN32) && ROCM_VERSION >= 50700)
 
 // ROCm 5.6 hipblas matches the const Dtype *A API, but prior hipblas does not.
 #if defined(USE_ROCM) && ROCM_VERSION < 50600

aten/src/ATen/cuda/CUDABlas.h

@@ -62,7 +62,7 @@ void gemm<at::Half>(CUDABLAS_GEMM_ARGTYPES(at::Half));
 template <>
 void gemm<at::BFloat16>(CUDABLAS_GEMM_ARGTYPES(at::BFloat16));
 
-#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && ROCM_VERSION >= 50700)
+#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && !defined(_WIN32) && ROCM_VERSION >= 50700)
 enum GEMMAndBiasActivationEpilogue {
   None,
   RELU,

aten/src/ATen/cuda/CUDAContextLight.h

@@ -9,7 +9,7 @@
 
 // cublasLT was introduced in CUDA 10.1 but we enable only for 11.1 that also
 // added bf16 support
-#if (!defined(_MSC_VER) && (!defined(USE_ROCM) || ROCM_VERSION >= 50700))
+#if (!defined(_MSC_VER) && (!defined(USE_ROCM) && !defined(_WIN32) || ROCM_VERSION >= 50700))
 #include <cublasLt.h>
 #endif
 

aten/src/ATen/cuda/CublasHandlePool.cpp

@@ -28,7 +28,7 @@ namespace at::cuda {
 
 namespace {
 
-#if defined(USE_ROCM) && ROCM_VERSION >= 50700
+#if defined(USE_ROCM) && !defined(_WIN32) && ROCM_VERSION >= 50700
 void createCublasLtHandle(cublasLtHandle_t *handle) {
   TORCH_CUDABLAS_CHECK(cublasLtCreate(handle));
 }
@@ -177,7 +177,7 @@ cublasHandle_t getCurrentCUDABlasHandle() {
   return handle;
 }
 
-#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && ROCM_VERSION >= 50700)
+#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && !defined(_MSC_VER) && ROCM_VERSION >= 50700)
 cublasLtHandle_t getCurrentCUDABlasLtHandle() {
 #ifdef USE_ROCM
   int device;

aten/src/ATen/native/cuda/Blas.cpp

@@ -153,7 +153,7 @@ enum class Activation {
   GELU,
 };
 
-#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && ROCM_VERSION >= 50700)
+#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && !defined(_WIN32) && ROCM_VERSION >= 50700)
 cuda::blas::GEMMAndBiasActivationEpilogue activation_to_gemm_and_blas_arg(Activation a) {
   switch (a) {
     case Activation::None:
@@ -320,7 +320,7 @@ Tensor& addmm_out_cuda_impl(Tensor& result, const Tensor& self, const Tensor& ma
 
   TORCH_INTERNAL_ASSERT_DEBUG_ONLY(!args.result->is_conj());
 
-#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && ROCM_VERSION >= 50700)
+#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && !defined(_WIN32) && ROCM_VERSION >= 50700)
   if (useLtInterface) {
     AT_DISPATCH_FLOATING_TYPES_AND2(
         at::ScalarType::Half,

cmake/Dependencies.cmake

@@ -1312,7 +1312,7 @@ if(USE_ROCM)
     if(UNIX)
       list(APPEND Caffe2 ${ROCM_ROCTX_LIB})
     endif(UNIX)
-    if(ROCM_VERSION_DEV VERSION_GREATER_EQUAL "5.7.0")
+    if(ROCM_VERSION_DEV VERSION_GREATER_EQUAL "5.7.0" AND NOT WIN32)
       list(APPEND Caffe2_PUBLIC_HIP_DEPENDENCY_LIBS ${hipblaslt_LIBRARIES})
     endif()