bug fix for wrong usage of checkGPU && port to windows with msvc

Author: BTNC

lib/THC/CMakeLists.txt

@@ -116,7 +116,11 @@ ENDIF()
 INCLUDE_DIRECTORIES("${CMAKE_CURRENT_BINARY_DIR}")
 CONFIGURE_FILE(THCGeneral.h.in "${CMAKE_CURRENT_BINARY_DIR}/THCGeneral.h")
 
-SET(CMAKE_C_FLAGS "-std=c99 ${CMAKE_C_FLAGS}")
+IF(MSVC)
+  LIST(APPEND CUDA_NVCC_FLAGS "-Xcompiler /wd4819")
+ELSE()
+  SET(CMAKE_C_FLAGS "-std=c99 ${CMAKE_C_FLAGS}")
+ENDIF()
 
 SET(src
     THCAllocator.c

lib/THC/THCHalf.h

@@ -20,8 +20,8 @@
 
 THC_EXTERNC void THCFloat2Half(THCState *state, half *out, float *in, long len);
 THC_EXTERNC void THCHalf2Float(THCState *state, float *out, half *in, long len);
-THC_EXTERNC half THC_float2half(float a);
-THC_EXTERNC float THC_half2float(half a);
+THC_API half THC_float2half(float a);
+THC_API float THC_half2float(half a);
 
 /* Check for native fp16 support on the current device (CC 5.3+) */
 THC_EXTERNC int THC_nativeHalfInstructions(THCState *state);

lib/THC/THCReduceApplyUtils.cu

@@ -4,7 +4,7 @@
 #include <stdlib.h>
 
 // Maximum size per grid dimension that we assume (compute capability >= 2.0)
-#define MAX_GRID_SIZE 65535L
+#define MAX_GRID_SIZE 65535LL
 
 void THCCheckTensorDims(THCState* state, THCudaTensor* tensor, int arg) {
   long dims = THCudaTensor_nDimension(state, tensor);
@@ -21,11 +21,11 @@ bool THC_getGridFromTiles(long gridTiles, dim3& grid) {
   long gridZ = 1;
 
   if (gridTiles > MAX_GRID_SIZE) {
-    gridTiles = THCCeilDiv(gridTiles, MAX_GRID_SIZE);
+    gridTiles = THCCeilDiv(gridTiles, (long) MAX_GRID_SIZE);
     gridY = gridTiles > MAX_GRID_SIZE ? MAX_GRID_SIZE : gridTiles;
 
     if (gridTiles > MAX_GRID_SIZE) {
-      gridTiles = THCCeilDiv(gridTiles, MAX_GRID_SIZE);
+      gridTiles = THCCeilDiv(gridTiles, (long) MAX_GRID_SIZE);
       gridZ = gridTiles > MAX_GRID_SIZE ? MAX_GRID_SIZE : gridTiles;
     }
   }

lib/THC/THCTensorMath2.cu

@@ -270,7 +270,7 @@ __host__ void THCudaTensor_varOuterDim(THCState *state, THCudaTensor *tgt, THCud
   unsigned ndim = THCudaTensor_nDimension(state, src);
   // Treat all outer dimensions (i.e. dim < dimension) as one.
   unsigned num_orows = 1;
-  for (unsigned dim = 0; dim < dimension; dim++) {
+  for (long dim = 0; dim < dimension; dim++) {
     num_orows *= THCudaTensor_size(state, src, dim);
   }
   unsigned row_size = THCudaTensor_size(state, src, dimension);

lib/THC/THCTensorMathReduce.cuh

@@ -148,7 +148,7 @@ THC_transformReduceOuterDimIndex(THCState *state,
                                  BinaryFunction binary_op) {
   unsigned ndim = TensorUtils<TensorTypeK>::getDims(state, src);
   unsigned num_orows = 1;
-  for (unsigned dim = 0; dim < rdim; dim++) {
+  for (long dim = 0; dim < rdim; dim++) {
     num_orows *= TensorUtils<TensorTypeK>::getSize(state, src, dim);
   }
   unsigned row_size = TensorUtils<TensorTypeK>::getSize(state, src, rdim);

lib/THC/THCTensorMathScan.cu

@@ -47,7 +47,7 @@ __host__ void THCudaTensor_scanOuterDim(THCState *state, THCudaTensor *tgt, THCu
   unsigned ndim = THCudaTensor_nDimension(state, src);
   // Treat all outer dimensions (i.e. dim < dimension) as one.
   unsigned num_orows = 1;
-  for (unsigned dim = 0; dim < dimension; dim++) {
+  for (long dim = 0; dim < dimension; dim++) {
     num_orows *= THCudaTensor_size(state, src, dim);
   }
   unsigned row_size = THCudaTensor_size(state, src, dimension);

lib/THC/THCTensorSort.cu

@@ -8,7 +8,9 @@ unsigned long nextHighestPowerOf2(unsigned long n) {
   n |= n >> 4;
   n |= n >> 8;
   n |= n >> 16;
+#ifndef _MSC_VER
   n |= n >> 32;
+#endif
   n++;
 
   return n;

lib/THC/THCThreadLocal.h

@@ -2,6 +2,7 @@
 #define THC_THREAD_LOCAL_INC
 
 #ifdef _WIN32
+#include <windows.h>
 typedef DWORD THCThreadLocal;
 #else
 #include <pthread.h>

lib/THC/cmake/select_compute_arch.cmake

@@ -69,6 +69,10 @@ function(CUDA_DETECT_INSTALLED_GPUS OUT_VARIABLE)
                     ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
 
     if(nvcc_res EQUAL 0)
+      # only keep the last line of nvcc_out
+      STRING(REGEX REPLACE ";" "\\\\;" nvcc_out "${nvcc_out}")
+      STRING(REGEX REPLACE "\n" ";" nvcc_out "${nvcc_out}")
+      list(GET nvcc_out -1 nvcc_out)
       string(REPLACE "2.1" "2.1(2.0)" nvcc_out "${nvcc_out}")
       set(CUDA_GPU_DETECT_OUTPUT ${nvcc_out} CACHE INTERNAL "Returned GPU architetures from detect_gpus tool" FORCE)
     endif()

lib/THC/generic/THCTensorIndex.cu

@@ -16,7 +16,7 @@ void THCTensor_(indexCopy_long)(THCState *state, THCTensor *dst, int dim, THLong
 
 void THCTensor_(indexCopy)(THCState *state, THCTensor *dst, int dim, THCudaLongTensor *indices, THCTensor *src)
 {
-  THAssert(THCTensor_(checkGPU)(state, 3, dst, src));
+  THAssert(THCTensor_(checkGPU)(state, 2, dst, src));
   THAssert(THCudaLongTensor_checkGPU(state, 1, indices));
 
   long dims = THCTensor_(nDimension)(state, dst);
@@ -144,7 +144,7 @@ void THCTensor_(indexAdd_long)(THCState *state, THCTensor *dst, int dim, THLongT
 
 void THCTensor_(indexAdd)(THCState *state, THCTensor *dst, int dim, THCudaLongTensor *indices, THCTensor *src)
 {
-  THAssert(THCTensor_(checkGPU)(state, 3, dst, src));
+  THAssert(THCTensor_(checkGPU)(state, 2, dst, src));
   THAssert(THCudaLongTensor_checkGPU(state, 1, indices));
 
   long dims = THCTensor_(nDimension)(state, dst);