add hip code

hip/utils/cublas_wrapper.h

+// !!! This is a file automatically generated by hipify!!!
+#include <ATen/dtk_macros.h>
+#include <iostream>
+#ifndef CUBLAS_WRAPPER_H
+#define CUBLAS_WRAPPER_H
+//#include </opt/dtk/include/rocblas/internal/rocblas-types.h>
+
+#include <hipblas/hipblas.h>
+#include <c10/util/Half.h>
+#include <c10/util/BFloat16.h>
+#include </opt/dtk/hip/include/hip/amd_detail/amd_hip_bf16.h>
+//#include </opt/dtk/include/rocblas/internal/rocblas-types.h>
+inline hipblasStatus_t cublasXgemmBatched(hipblasHandle_t handle,
+                                  hipblasOperation_t transa,
+                                  hipblasOperation_t transb,
+                                  int m, int n, int k,
+                                  const float           *alpha,
+                                  const float           *Aarray[], int lda,
+                                  const float           *Barray[], int ldb,
+                                  const float           *beta,
+                                  float           *Carray[], int ldc,
+                                  int batchCount) {
+    return hipblasSgemmBatched(handle, transa, transb, m, n, k, alpha, Aarray, lda, Barray, ldb, beta, Carray, ldc, batchCount);
+}
+
+inline hipblasStatus_t cublasXgemmBatched(hipblasHandle_t handle,
+                                  hipblasOperation_t transa,
+                                  hipblasOperation_t transb,
+                                  int m, int n, int k,
+                                  const double           *alpha,
+                                  const double           *Aarray[], int lda,
+                                  const double           *Barray[], int ldb,
+                                  const double           *beta,
+                                  double           *Carray[], int ldc,
+                                  int batchCount) {
+    return hipblasDgemmBatched(handle, transa, transb, m, n, k, alpha, Aarray, lda, Barray, ldb, beta, Carray, ldc, batchCount);
+}
+
+inline hipblasStatus_t cublasXgemmBatched(hipblasHandle_t handle,
+                                  hipblasOperation_t transa,
+                                  hipblasOperation_t transb,
+                                  int m, int n, int k,
+                                  const __half           *alpha,
+                                  const __half           *Aarray[], int lda,
+                                  const __half           *Barray[], int ldb,
+                                  const __half           *beta,
+                                  __half           *Carray[], int ldc,
+                                  int batchCount) {
+#if defined (FMOE_USE_HIP) && defined(__CUDA_MIX_HIP__)
+//#ifdef FMOE_USE_HIP
+    return hipblasHgemmBatched(handle, transa, transb, m, n, k, (const rocblas_half*)alpha, (const rocblas_half* const*)Aarray, lda, (const rocblas_half* const*)Barray, ldb, (const rocblas_half*)beta, (rocblas_half* const*)Carray, ldc, batchCount);
+#else
+//    return hipblasHgemmBatched(handle, transa, transb, m, n, k, alpha, Aarray, lda, Barray, ldb, beta, Carray, ldc, batchCount);
+    return hipblasHgemmBatched(handle, transa, transb, m, n, k, (const hipblasHalf*)alpha, (const hipblasHalf* const*)Aarray, lda, (const hipblasHalf* const*)Barray, ldb, (const hipblasHalf*)beta, (hipblasHalf* const*)Carray, ldc, batchCount);
+#endif
+}
+
+
+inline hipblasStatus_t cublasXgemm(hipblasHandle_t handle,
+                                hipblasOperation_t transa, hipblasOperation_t transb,
+                                int m, int n, int k,
+                                const float           *alpha,
+                                const float           *A, int lda,
+                                const float           *B, int ldb,
+                                const float           *beta,
+                                float           *C, int ldc) {
+    return hipblasSgemm(handle, transa, transb, m, n, k, alpha, A, lda, B, ldb, beta, C, ldc);
+}
+
+inline hipblasStatus_t cublasXgemm(hipblasHandle_t handle,
+                                hipblasOperation_t transa, hipblasOperation_t transb,
+                                int m, int n, int k,
+                                const double          *alpha,
+                                const double          *A, int lda,
+                                const double          *B, int ldb,
+                                const double          *beta,
+                                double          *C, int ldc) {
+    return hipblasDgemm(handle, transa, transb, m, n, k, alpha, A, lda, B, ldb, beta, C, ldc);
+}
+
+inline hipblasStatus_t cublasXgemm(hipblasHandle_t handle,
+                                hipblasOperation_t transa, hipblasOperation_t transb,
+                                int m, int n, int k,
+                                const __half *alpha,
+                                const __half *A, int lda,
+                                const __half *B, int ldb,
+                                const __half *beta,
+                                __half *C, int ldc) {
+//#ifdef FMOE_USE_HIP
+#if defined (FMOE_USE_HIP) && defined(__CUDA_MIX_HIP__)
+    return hipblasHgemm(handle, transa, transb, m, n, k, (const rocblas_half*)alpha, (const rocblas_half* )A, lda, (const rocblas_half* )B, ldb, (const rocblas_half*)beta, (rocblas_half* )C, ldc);
+#else
+//    return hipblasHgemm(handle, transa, transb, m, n, k, alpha, A, lda, B, ldb, beta, C, ldc);
+    return hipblasHgemm(handle, transa, transb, m, n, k, (const hipblasHalf*)alpha, (const hipblasHalf*)A, lda, (const hipblasHalf*)B, ldb, (const hipblasHalf*)beta, (hipblasHalf*)C, ldc);
+#endif
+}
+
+inline hipblasStatus_t cublasXgemm(hipblasHandle_t handle,
+                                hipblasOperation_t transa, hipblasOperation_t transb,
+                                int m, int n, int k,
+                                const c10::Half *alpha,
+                                const c10::Half *A, int lda,
+                                const c10::Half *B, int ldb,
+                                const c10::Half *beta,
+                                c10::Half *C, int ldc) {
+//#ifdef FMOE_USE_HIP
+#if defined (FMOE_USE_HIP) && defined(__CUDA_MIX_HIP__)
+    return hipblasHgemm(handle, transa, transb, m, n, k,
+            (const rocblas_half*)alpha,
+            (const rocblas_half*)A, lda,
+            (const rocblas_half*)B, ldb,
+            (const rocblas_half*)beta,
+            (rocblas_half*)C, ldc);
+#else
+    return hipblasHgemm(handle, transa, transb, m, n, k,
+            //(const __half*)alpha,
+            (const hipblasHalf*)alpha,
+            //(const __half*)A, lda,
+            (const hipblasHalf*)A, lda,
+            //(const __half*)B, ldb,
+            (const hipblasHalf*)B, ldb,
+            //(const __half*)beta,
+            (const hipblasHalf*)beta,
+            //(__half*)C, ldc);
+            (hipblasHalf*)C, ldc);
+#endif
+}
+
+inline hipblasStatus_t cublasXgemm(hipblasHandle_t handle,
+                                hipblasOperation_t transa, hipblasOperation_t transb,
+                                int m, int n, int k,
+                                const c10::BFloat16 *alpha,
+				//const void *alpha,
+                                const c10::BFloat16 *A, int lda,
+                                const c10::BFloat16 *B, int ldb,
+                                const c10::BFloat16 *beta,
+				//const void *beta,
+                                c10::BFloat16 *C, int ldc) {
+//#ifdef FMOE_USE_HIP
+#if defined (FMOE_USE_HIP) && defined(__CUDA_MIX_HIP__)
+    // TODO: Support bf16 for HIP
+    assert(false);
+#else
+    //const float alpha_fp32(*alpha), beta_fp32(*beta);
+
+    hipblasDatatype_t datatype_C =  HIPBLAS_R_16B;
+    float alpha_ = static_cast<float>(*alpha);
+    float beta_ = static_cast<float>(*beta);
+    return hipblasGemmEx(handle, transa, transb, m, n, k,
+            //(const float*)&alpha_fp32,
+            //(const void*)A, datatype_C, lda,
+            //(const void*)B, datatype_C, ldb,
+            //(const float*)&beta_fp32,
+            //(void*)C, datatype_C, ldc,
+	    reinterpret_cast<const float*>(&alpha_),
+	    //alpha,
+            reinterpret_cast<const void*>(A), datatype_C, lda,
+            reinterpret_cast<const void*>(B), datatype_C, ldb,
+	    reinterpret_cast<const float*>(&beta_),
+	    //beta,
+            reinterpret_cast<void*>(C), datatype_C, ldc,
+             HIPBLAS_R_32F,
+             HIPBLAS_GEMM_DEFAULT);
+#endif
+}
+#endif  // CUBLAS_WRAPPER_H
+

hip/utils/fmoe_utils.h

+// !!! This is a file automatically generated by hipify!!!
+#include <ATen/dtk_macros.h>
+#ifndef FMOE_UTILS_H
+#define FMOE_UTILS_H
+
+#define CHECK_CUDA(x) AT_ASSERTM(x.device().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)
+
+#define CEIL(_x_,_y_) (((_x_)-1)/(_y_)+1)
+
+#endif  // FMOE_UTILS_H

hip/utils/helper_cuda.h

+// !!! This is a file automatically generated by hipify!!!
+#include <ATen/dtk_macros.h>
+/* Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+////////////////////////////////////////////////////////////////////////////////
+// These are CUDA Helper functions for initialization and error checking
+// This file is clipped from the original header file by laekov
+
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#ifndef HELPER_CUDA_H
+#define HELPER_CUDA_H
+
+static const char *_cudaGetErrorEnum(hipError_t error) {
+  return hipGetErrorName(error);
+}
+
+#ifdef CUDA_DRIVER_API
+// CUDA Driver API errors
+static const char *_cudaGetErrorEnum(hipError_t error) {
+  static char unknown[] = "<unknown>";
+  const char *ret = NULL;
+  hipGetErrorName(error, &ret);
+  return ret ? ret : unknown;
+}
+#endif
+
+#if defined(FMOE_USE_HIP)  && defined(__CUDA_MIX_HIP__)
+static const char *_cudaGetErrorEnum(hipblasStatus_t error) {
+  switch (error) {
+
+    case HIPBLAS_STATUS_SUCCESS:
+      return "HIPBLAS_STATUS_SUCCESS";
+
+    case HIPBLAS_STATUS_NOT_INITIALIZED:
+      return "HIPBLAS_STATUS_NOT_INITIALIZED";
+
+    case HIPBLAS_STATUS_ARCH_MISMATCH:
+      return "HIPBLAS_STATUS_ARCH_MISMATCH";
+
+    case HIPBLAS_STATUS_INVALID_VALUE:
+      return "HIPBLAS_STATUS_INVALID_VALUE:";
+
+    case rocblas_status_invalid_size:
+      return "rocblas_status_invalid_size";
+
+    case HIPBLAS_STATUS_ALLOC_FAILED:
+      return "HIPBLAS_STATUS_ALLOC_FAILED";
+
+    case HIPBLAS_STATUS_INTERNAL_ERROR:
+      return "HIPBLAS_STATUS_INTERNAL_ERROR";
+
+    case rocblas_status_perf_degraded:
+      return "rocblas_status_perf_degraded";
+
+    case rocblas_status_size_query_mismatch:
+      return "rocblas_status_size_query_mismatch";
+
+    case rocblas_status_size_increased:
+      return "rocblas_status_size_increased";
+
+    case rocblas_status_size_unchanged:
+      return "rocblas_status_size_unchanged";
+
+    case rocblas_status_invalid_value:
+      return "rocblas_status_invalid_value";
+
+    case rocblas_status_continue:
+      return "rocblas_status_continue";
+  }
+
+  return "<unknown>";
+}
+#else
+// cuBLAS API errors
+static const char *_cudaGetErrorEnum(hipblasStatus_t error) {
+  switch (error) {
+    case HIPBLAS_STATUS_SUCCESS:
+      return "HIPBLAS_STATUS_SUCCESS";
+
+    case HIPBLAS_STATUS_NOT_INITIALIZED:
+      return "HIPBLAS_STATUS_NOT_INITIALIZED";
+
+    case HIPBLAS_STATUS_ALLOC_FAILED:
+      return "HIPBLAS_STATUS_ALLOC_FAILED";
+
+    case HIPBLAS_STATUS_INVALID_VALUE:
+      return "HIPBLAS_STATUS_INVALID_VALUE";
+
+    case HIPBLAS_STATUS_ARCH_MISMATCH:
+      return "HIPBLAS_STATUS_ARCH_MISMATCH";
+
+    case HIPBLAS_STATUS_MAPPING_ERROR:
+      return "HIPBLAS_STATUS_MAPPING_ERROR";
+
+    case HIPBLAS_STATUS_EXECUTION_FAILED:
+      return "HIPBLAS_STATUS_EXECUTION_FAILED";
+
+    case HIPBLAS_STATUS_INTERNAL_ERROR:
+      return "HIPBLAS_STATUS_INTERNAL_ERROR";
+
+    case HIPBLAS_STATUS_NOT_SUPPORTED:
+      return "HIPBLAS_STATUS_NOT_SUPPORTED";
+
+    //case CUBLAS_STATUS_LICENSE_ERROR:
+     // return "HIPBLAS_STATUS_INTERNAL_ERROR";
+  }
+
+  return "<unknown>";
+}
+#endif
+
+#ifdef _CUFFT_H_
+// cuFFT API errors
+static const char *_cudaGetErrorEnum(hipfftResult error) {
+  switch (error) {
+    case HIPFFT_SUCCESS:
+      return "HIPFFT_SUCCESS";
+
+    case HIPFFT_INVALID_PLAN:
+      return "HIPFFT_INVALID_PLAN";
+
+    case HIPFFT_ALLOC_FAILED:
+      return "HIPFFT_ALLOC_FAILED";
+
+    case HIPFFT_INVALID_TYPE:
+      return "HIPFFT_INVALID_TYPE";
+
+    case HIPFFT_INVALID_VALUE:
+      return "HIPFFT_INVALID_VALUE";
+
+    case HIPFFT_INTERNAL_ERROR:
+      return "HIPFFT_INTERNAL_ERROR";
+
+    case HIPFFT_EXEC_FAILED:
+      return "HIPFFT_EXEC_FAILED";
+
+    case HIPFFT_SETUP_FAILED:
+      return "HIPFFT_SETUP_FAILED";
+
+    case HIPFFT_INVALID_SIZE:
+      return "HIPFFT_INVALID_SIZE";
+
+    case HIPFFT_UNALIGNED_DATA:
+      return "HIPFFT_UNALIGNED_DATA";
+
+    case HIPFFT_INCOMPLETE_PARAMETER_LIST:
+      return "HIPFFT_INCOMPLETE_PARAMETER_LIST";
+
+    case HIPFFT_INVALID_DEVICE:
+      return "HIPFFT_INVALID_DEVICE";
+
+    case HIPFFT_PARSE_ERROR:
+      return "HIPFFT_PARSE_ERROR";
+
+    case HIPFFT_NO_WORKSPACE:
+      return "HIPFFT_NO_WORKSPACE";
+
+    case HIPFFT_NOT_IMPLEMENTED:
+      return "HIPFFT_NOT_IMPLEMENTED";
+
+    case HIPFFT_LICENSE_ERROR:
+      return "HIPFFT_LICENSE_ERROR";
+
+    case HIPFFT_NOT_SUPPORTED:
+      return "HIPFFT_NOT_SUPPORTED";
+  }
+
+  return "<unknown>";
+}
+#endif
+
+#ifdef CUSPARSEAPI
+// cuSPARSE API errors
+static const char *_cudaGetErrorEnum(hipsparseStatus_t error) {
+  switch (error) {
+    case HIPSPARSE_STATUS_SUCCESS:
+      return "HIPSPARSE_STATUS_SUCCESS";
+
+    case HIPSPARSE_STATUS_NOT_INITIALIZED:
+      return "HIPSPARSE_STATUS_NOT_INITIALIZED";
+
+    case HIPSPARSE_STATUS_ALLOC_FAILED:
+      return "HIPSPARSE_STATUS_ALLOC_FAILED";
+
+    case HIPSPARSE_STATUS_INVALID_VALUE:
+      return "HIPSPARSE_STATUS_INVALID_VALUE";
+
+    case HIPSPARSE_STATUS_ARCH_MISMATCH:
+      return "HIPSPARSE_STATUS_ARCH_MISMATCH";
+
+    case HIPSPARSE_STATUS_MAPPING_ERROR:
+      return "HIPSPARSE_STATUS_MAPPING_ERROR";
+
+    case HIPSPARSE_STATUS_EXECUTION_FAILED:
+      return "HIPSPARSE_STATUS_EXECUTION_FAILED";
+
+    case HIPSPARSE_STATUS_INTERNAL_ERROR:
+      return "HIPSPARSE_STATUS_INTERNAL_ERROR";
+
+    case HIPSPARSE_STATUS_MATRIX_TYPE_NOT_SUPPORTED:
+      return "HIPSPARSE_STATUS_MATRIX_TYPE_NOT_SUPPORTED";
+  }
+
+  return "<unknown>";
+}
+#endif
+
+#ifdef CUSOLVER_COMMON_H_
+// cuSOLVER API errors
+static const char *_cudaGetErrorEnum(cusolverStatus_t error) {
+  switch (error) {
+    case CUSOLVER_STATUS_SUCCESS:
+      return "CUSOLVER_STATUS_SUCCESS";
+    case CUSOLVER_STATUS_NOT_INITIALIZED:
+      return "CUSOLVER_STATUS_NOT_INITIALIZED";
+    case CUSOLVER_STATUS_ALLOC_FAILED:
+      return "CUSOLVER_STATUS_ALLOC_FAILED";
+    case CUSOLVER_STATUS_INVALID_VALUE:
+      return "CUSOLVER_STATUS_INVALID_VALUE";
+    case CUSOLVER_STATUS_ARCH_MISMATCH:
+      return "CUSOLVER_STATUS_ARCH_MISMATCH";
+    case CUSOLVER_STATUS_MAPPING_ERROR:
+      return "CUSOLVER_STATUS_MAPPING_ERROR";
+    case CUSOLVER_STATUS_EXECUTION_FAILED:
+      return "CUSOLVER_STATUS_EXECUTION_FAILED";
+    case CUSOLVER_STATUS_INTERNAL_ERROR:
+      return "CUSOLVER_STATUS_INTERNAL_ERROR";
+    case CUSOLVER_STATUS_MATRIX_TYPE_NOT_SUPPORTED:
+      return "CUSOLVER_STATUS_MATRIX_TYPE_NOT_SUPPORTED";
+    case CUSOLVER_STATUS_NOT_SUPPORTED:
+      return "CUSOLVER_STATUS_NOT_SUPPORTED ";
+    case CUSOLVER_STATUS_ZERO_PIVOT:
+      return "CUSOLVER_STATUS_ZERO_PIVOT";
+    case CUSOLVER_STATUS_INVALID_LICENSE:
+      return "CUSOLVER_STATUS_INVALID_LICENSE";
+  }
+
+  return "<unknown>";
+}
+#endif
+
+#ifdef CURAND_H_
+// cuRAND API errors
+static const char *_cudaGetErrorEnum(hiprandStatus_t error) {
+  switch (error) {
+    case HIPRAND_STATUS_SUCCESS:
+      return "HIPRAND_STATUS_SUCCESS";
+
+    case HIPRAND_STATUS_VERSION_MISMATCH:
+      return "HIPRAND_STATUS_VERSION_MISMATCH";
+
+    case HIPRAND_STATUS_NOT_INITIALIZED:
+      return "HIPRAND_STATUS_NOT_INITIALIZED";
+
+    case HIPRAND_STATUS_ALLOCATION_FAILED:
+      return "HIPRAND_STATUS_ALLOCATION_FAILED";
+
+    case HIPRAND_STATUS_TYPE_ERROR:
+      return "HIPRAND_STATUS_TYPE_ERROR";
+
+    case HIPRAND_STATUS_OUT_OF_RANGE:
+      return "HIPRAND_STATUS_OUT_OF_RANGE";
+
+    case HIPRAND_STATUS_LENGTH_NOT_MULTIPLE:
+      return "HIPRAND_STATUS_LENGTH_NOT_MULTIPLE";
+
+    case HIPRAND_STATUS_DOUBLE_PRECISION_REQUIRED:
+      return "HIPRAND_STATUS_DOUBLE_PRECISION_REQUIRED";
+
+    case HIPRAND_STATUS_LAUNCH_FAILURE:
+      return "HIPRAND_STATUS_LAUNCH_FAILURE";
+
+    case HIPRAND_STATUS_PREEXISTING_FAILURE:
+      return "HIPRAND_STATUS_PREEXISTING_FAILURE";
+
+    case HIPRAND_STATUS_INITIALIZATION_FAILED:
+      return "HIPRAND_STATUS_INITIALIZATION_FAILED";
+
+    case HIPRAND_STATUS_ARCH_MISMATCH:
+      return "HIPRAND_STATUS_ARCH_MISMATCH";
+
+    case HIPRAND_STATUS_INTERNAL_ERROR:
+      return "HIPRAND_STATUS_INTERNAL_ERROR";
+  }
+
+  return "<unknown>";
+}
+#endif
+
+#ifdef NVJPEGAPI
+// nvJPEG API errors
+static const char *_cudaGetErrorEnum(nvjpegStatus_t error) {
+  switch (error) {
+    case NVJPEG_STATUS_SUCCESS:
+      return "NVJPEG_STATUS_SUCCESS";
+
+    case NVJPEG_STATUS_NOT_INITIALIZED:
+      return "NVJPEG_STATUS_NOT_INITIALIZED";
+
+    case NVJPEG_STATUS_INVALID_PARAMETER:
+      return "NVJPEG_STATUS_INVALID_PARAMETER";
+
+    case NVJPEG_STATUS_BAD_JPEG:
+      return "NVJPEG_STATUS_BAD_JPEG";
+
+    case NVJPEG_STATUS_JPEG_NOT_SUPPORTED:
+      return "NVJPEG_STATUS_JPEG_NOT_SUPPORTED";
+
+    case NVJPEG_STATUS_ALLOCATOR_FAILURE:
+      return "NVJPEG_STATUS_ALLOCATOR_FAILURE";
+
+    case NVJPEG_STATUS_EXECUTION_FAILED:
+      return "NVJPEG_STATUS_EXECUTION_FAILED";
+
+    case NVJPEG_STATUS_ARCH_MISMATCH:
+      return "NVJPEG_STATUS_ARCH_MISMATCH";
+
+    case NVJPEG_STATUS_INTERNAL_ERROR:
+      return "NVJPEG_STATUS_INTERNAL_ERROR";
+  }
+
+  return "<unknown>";
+}
+#endif
+
+#ifdef NV_NPPIDEFS_H
+// NPP API errors
+static const char *_cudaGetErrorEnum(NppStatus error) {
+  switch (error) {
+    case NPP_NOT_SUPPORTED_MODE_ERROR:
+      return "NPP_NOT_SUPPORTED_MODE_ERROR";
+
+    case NPP_ROUND_MODE_NOT_SUPPORTED_ERROR:
+      return "NPP_ROUND_MODE_NOT_SUPPORTED_ERROR";
+
+    case NPP_RESIZE_NO_OPERATION_ERROR:
+      return "NPP_RESIZE_NO_OPERATION_ERROR";
+
+    case NPP_NOT_SUFFICIENT_COMPUTE_CAPABILITY:
+      return "NPP_NOT_SUFFICIENT_COMPUTE_CAPABILITY";
+
+#if ((NPP_VERSION_MAJOR << 12) + (NPP_VERSION_MINOR << 4)) <= 0x5000
+
+    case NPP_BAD_ARG_ERROR:
+      return "NPP_BAD_ARGUMENT_ERROR";
+
+    case NPP_COEFF_ERROR:
+      return "NPP_COEFFICIENT_ERROR";
+
+    case NPP_RECT_ERROR:
+      return "NPP_RECTANGLE_ERROR";
+
+    case NPP_QUAD_ERROR:
+      return "NPP_QUADRANGLE_ERROR";
+
+    case NPP_MEM_ALLOC_ERR:
+      return "NPP_MEMORY_ALLOCATION_ERROR";
+
+    case NPP_HISTO_NUMBER_OF_LEVELS_ERROR:
+      return "NPP_HISTOGRAM_NUMBER_OF_LEVELS_ERROR";
+
+    case NPP_INVALID_INPUT:
+      return "NPP_INVALID_INPUT";
+
+    case NPP_POINTER_ERROR:
+      return "NPP_POINTER_ERROR";
+
+    case NPP_WARNING:
+      return "NPP_WARNING";
+
+    case NPP_ODD_ROI_WARNING:
+      return "NPP_ODD_ROI_WARNING";
+#else
+
+    // These are for CUDA 5.5 or higher
+    case NPP_BAD_ARGUMENT_ERROR:
+      return "NPP_BAD_ARGUMENT_ERROR";
+
+    case NPP_COEFFICIENT_ERROR:
+      return "NPP_COEFFICIENT_ERROR";
+
+    case NPP_RECTANGLE_ERROR:
+      return "NPP_RECTANGLE_ERROR";
+
+    case NPP_QUADRANGLE_ERROR:
+      return "NPP_QUADRANGLE_ERROR";
+
+    case NPP_MEMORY_ALLOCATION_ERR:
+      return "NPP_MEMORY_ALLOCATION_ERROR";
+
+    case NPP_HISTOGRAM_NUMBER_OF_LEVELS_ERROR:
+      return "NPP_HISTOGRAM_NUMBER_OF_LEVELS_ERROR";
+
+    case NPP_INVALID_HOST_POINTER_ERROR:
+      return "NPP_INVALID_HOST_POINTER_ERROR";
+
+    case NPP_INVALID_DEVICE_POINTER_ERROR:
+      return "NPP_INVALID_DEVICE_POINTER_ERROR";
+#endif
+
+    case NPP_LUT_NUMBER_OF_LEVELS_ERROR:
+      return "NPP_LUT_NUMBER_OF_LEVELS_ERROR";
+
+    case NPP_TEXTURE_BIND_ERROR:
+      return "NPP_TEXTURE_BIND_ERROR";
+
+    case NPP_WRONG_INTERSECTION_ROI_ERROR:
+      return "NPP_WRONG_INTERSECTION_ROI_ERROR";
+
+    case NPP_NOT_EVEN_STEP_ERROR:
+      return "NPP_NOT_EVEN_STEP_ERROR";
+
+    case NPP_INTERPOLATION_ERROR:
+      return "NPP_INTERPOLATION_ERROR";
+
+    case NPP_RESIZE_FACTOR_ERROR:
+      return "NPP_RESIZE_FACTOR_ERROR";
+
+    case NPP_HAAR_CLASSIFIER_PIXEL_MATCH_ERROR:
+      return "NPP_HAAR_CLASSIFIER_PIXEL_MATCH_ERROR";
+
+#if ((NPP_VERSION_MAJOR << 12) + (NPP_VERSION_MINOR << 4)) <= 0x5000
+
+    case NPP_MEMFREE_ERR:
+      return "NPP_MEMFREE_ERR";
+
+    case NPP_MEMSET_ERR:
+      return "NPP_MEMSET_ERR";
+
+    case NPP_MEMCPY_ERR:
+      return "NPP_MEMCPY_ERROR";
+
+    case NPP_MIRROR_FLIP_ERR:
+      return "NPP_MIRROR_FLIP_ERR";
+#else
+
+    case NPP_MEMFREE_ERROR:
+      return "NPP_MEMFREE_ERROR";
+
+    case NPP_MEMSET_ERROR:
+      return "NPP_MEMSET_ERROR";
+
+    case NPP_MEMCPY_ERROR:
+      return "NPP_MEMCPY_ERROR";
+
+    case NPP_MIRROR_FLIP_ERROR:
+      return "NPP_MIRROR_FLIP_ERROR";
+#endif
+
+    case NPP_ALIGNMENT_ERROR:
+      return "NPP_ALIGNMENT_ERROR";
+
+    case NPP_STEP_ERROR:
+      return "NPP_STEP_ERROR";
+
+    case NPP_SIZE_ERROR:
+      return "NPP_SIZE_ERROR";
+
+    case NPP_NULL_POINTER_ERROR:
+      return "NPP_NULL_POINTER_ERROR";
+
+    case NPP_CUDA_KERNEL_EXECUTION_ERROR:
+      return "NPP_CUDA_KERNEL_EXECUTION_ERROR";
+
+    case NPP_NOT_IMPLEMENTED_ERROR:
+      return "NPP_NOT_IMPLEMENTED_ERROR";
+
+    case NPP_ERROR:
+      return "NPP_ERROR";
+
+    case NPP_SUCCESS:
+      return "NPP_SUCCESS";
+
+    case NPP_WRONG_INTERSECTION_QUAD_WARNING:
+      return "NPP_WRONG_INTERSECTION_QUAD_WARNING";
+
+    case NPP_MISALIGNED_DST_ROI_WARNING:
+      return "NPP_MISALIGNED_DST_ROI_WARNING";
+
+    case NPP_AFFINE_QUAD_INCORRECT_WARNING:
+      return "NPP_AFFINE_QUAD_INCORRECT_WARNING";
+
+    case NPP_DOUBLE_SIZE_WARNING:
+      return "NPP_DOUBLE_SIZE_WARNING";
+
+    case NPP_WRONG_INTERSECTION_ROI_WARNING:
+      return "NPP_WRONG_INTERSECTION_ROI_WARNING";
+
+#if ((NPP_VERSION_MAJOR << 12) + (NPP_VERSION_MINOR << 4)) >= 0x6000
+    /* These are 6.0 or higher */
+    case NPP_LUT_PALETTE_BITSIZE_ERROR:
+      return "NPP_LUT_PALETTE_BITSIZE_ERROR";
+
+    case NPP_ZC_MODE_NOT_SUPPORTED_ERROR:
+      return "NPP_ZC_MODE_NOT_SUPPORTED_ERROR";
+
+    case NPP_QUALITY_INDEX_ERROR:
+      return "NPP_QUALITY_INDEX_ERROR";
+
+    case NPP_CHANNEL_ORDER_ERROR:
+      return "NPP_CHANNEL_ORDER_ERROR";
+
+    case NPP_ZERO_MASK_VALUE_ERROR:
+      return "NPP_ZERO_MASK_VALUE_ERROR";
+
+    case NPP_NUMBER_OF_CHANNELS_ERROR:
+      return "NPP_NUMBER_OF_CHANNELS_ERROR";
+
+    case NPP_COI_ERROR:
+      return "NPP_COI_ERROR";
+
+    case NPP_DIVISOR_ERROR:
+      return "NPP_DIVISOR_ERROR";
+
+    case NPP_CHANNEL_ERROR:
+      return "NPP_CHANNEL_ERROR";
+
+    case NPP_STRIDE_ERROR:
+      return "NPP_STRIDE_ERROR";
+
+    case NPP_ANCHOR_ERROR:
+      return "NPP_ANCHOR_ERROR";
+
+    case NPP_MASK_SIZE_ERROR:
+      return "NPP_MASK_SIZE_ERROR";
+
+    case NPP_MOMENT_00_ZERO_ERROR:
+      return "NPP_MOMENT_00_ZERO_ERROR";
+
+    case NPP_THRESHOLD_NEGATIVE_LEVEL_ERROR:
+      return "NPP_THRESHOLD_NEGATIVE_LEVEL_ERROR";
+
+    case NPP_THRESHOLD_ERROR:
+      return "NPP_THRESHOLD_ERROR";
+
+    case NPP_CONTEXT_MATCH_ERROR:
+      return "NPP_CONTEXT_MATCH_ERROR";
+
+    case NPP_FFT_FLAG_ERROR:
+      return "NPP_FFT_FLAG_ERROR";
+
+    case NPP_FFT_ORDER_ERROR:
+      return "NPP_FFT_ORDER_ERROR";
+
+    case NPP_SCALE_RANGE_ERROR:
+      return "NPP_SCALE_RANGE_ERROR";
+
+    case NPP_DATA_TYPE_ERROR:
+      return "NPP_DATA_TYPE_ERROR";
+
+    case NPP_OUT_OFF_RANGE_ERROR:
+      return "NPP_OUT_OFF_RANGE_ERROR";
+
+    case NPP_DIVIDE_BY_ZERO_ERROR:
+      return "NPP_DIVIDE_BY_ZERO_ERROR";
+
+    case NPP_RANGE_ERROR:
+      return "NPP_RANGE_ERROR";
+
+    case NPP_NO_MEMORY_ERROR:
+      return "NPP_NO_MEMORY_ERROR";
+
+    case NPP_ERROR_RESERVED:
+      return "NPP_ERROR_RESERVED";
+
+    case NPP_NO_OPERATION_WARNING:
+      return "NPP_NO_OPERATION_WARNING";
+
+    case NPP_DIVIDE_BY_ZERO_WARNING:
+      return "NPP_DIVIDE_BY_ZERO_WARNING";
+#endif
+
+#if ((NPP_VERSION_MAJOR << 12) + (NPP_VERSION_MINOR << 4)) >= 0x7000
+    /* These are 7.0 or higher */
+    case NPP_OVERFLOW_ERROR:
+      return "NPP_OVERFLOW_ERROR";
+
+    case NPP_CORRUPTED_DATA_ERROR:
+      return "NPP_CORRUPTED_DATA_ERROR";
+#endif
+  }
+
+  return "<unknown>";
+}
+#endif
+
+
+template <typename T>
+void check(T result, char const *const func, const char *const file,
+           int const line) {
+  if (result) {
+    fprintf(stderr, "CUDA error at %s:%d code=%d(%s) \"%s\" \n", file, line,
+            static_cast<unsigned int>(result), _cudaGetErrorEnum(result), func);
+    exit(EXIT_FAILURE);
+  }
+}
+
+// This will output the proper CUDA error strings in the event
+// that a CUDA host call returns an error
+#define checkCudaErrors(val) check((val), #val, __FILE__, __LINE__)
+
+#endif  // HELPER_CUDA_H
+

setup.py

@@ -2,7 +2,8 @@ import setuptools
 from torch.utils.cpp_extension import BuildExtension, CUDAExtension
 import os
 import torch
-
+from fastpt import CUDAExtension
+import get_dcu_version
 cxx_flags = []
 ext_libs = []
 
@@ -37,11 +38,28 @@ if is_rocm_pytorch:
 else:
     define_macros=[]
 
+def get_package_version():
+    __version__ = '1.1.0'
+    cwd = os.path.dirname(os.path.abspath(__file__))
+    version_path = os.path.join(cwd, "fmoe", "version.py")
+    with open(version_path, "w") as f:
+        f.write(f"__version__ = '{__version__}'\n")
+
+    return __version__
+
+def write_version_file():
+    cwd = os.path.dirname(os.path.abspath(__file__))
+    version_path = os.path.join(cwd, "fmoe", "version.py")
+    version = get_dcu_version.dcu_version()
+    with open(version_path, "w") as f:
+        f.write(f"__dcu_version__ = '{version}'\n")
+write_version_file()
 
 if __name__ == '__main__':
     setuptools.setup(
         name='fastmoe',
-        version='1.1.0',
+        #version='1.1.0',
+        version = get_package_version() + get_dcu_version.moe_whl_name(),
         description='An efficient Mixture-of-Experts system for PyTorch',
         author=', '.join(authors),
         author_email='hja20@mails.tsinghua.edu.cn',

tests/test_numerical.py

@@ -61,8 +61,10 @@ def _assert_numerical(names, moe_out_list, raw_out_list, rank, precision=1e-3):
         if err > precision:
             sys.stderr.write(f"=========== {name} moe out ==============\n")
             sys.stderr.write("{}\n".format(mo))
+            sys.stderr.write(f'------------->>>>>>>>>>>> mo dtype: {mo.dtype}\n')
             sys.stderr.write(f"=========== {name} raw out ==============\n")
             sys.stderr.write("{}\n".format(ro))
+            sys.stderr.write(f'------------->>>>>>>>>>>> ro dtype: {ro.dtype}\n')
             sys.stderr.write(f"=========== {name} diff ==============\n")
             sys.stderr.write("{}\n{}\n".format((mo - ro).abs(), err))
             assert False