Fix types and warp sizes for ROCm (ROCm 256)

* Correct the warp size for current AMD GPUs.

* Fix copy paste error located by Jithun Nair.

* Correct the wrong typing explicitly.

Author: iotamudelta

aten/src/THC/generic/THCTensorIndex.cu

@@ -560,16 +560,17 @@ void THCTensor_(indexSelect)(THCState *state, THCTensor *dst, THCTensor *src, in
   indexSelectSmallIndex<TENSOR_TYPE, TYPE, DST_DIM, SRC_DIM, IDX_DIM>     \
     <<<smallIndexGrid, smallIndexBlock, 0, stream>>>(           \
       dstInfo, srcInfo, indicesInfo,                            \
-      dstSelectDim, srcSelectDim, sliceSize, srcSelectDimSize);
+      dstSelectDim, srcSelectDim, static_cast<TYPE>(sliceSize), \
+      srcSelectDimSize);
 
 #define LARGE_INDEX(TENSOR_TYPE, TYPE,                           \
                     DST_DIM, SRC_DIM, IDX_DIM, IDX_IS_MAJOR)     \
   indexSelectLargeIndex<TENSOR_TYPE, TYPE,                       \
                         DST_DIM, SRC_DIM, IDX_DIM, IDX_IS_MAJOR> \
     <<<largeIndexGrid, largeIndexBlock, 0, stream>>>(            \
       dstInfo, srcInfo, indicesInfo,                             \
-      dstSelectDim, srcSelectDim, dstTotalSize,                  \
-      (IDX_IS_MAJOR) ? sliceSize : numIndices,                   \
+      dstSelectDim, srcSelectDim, static_cast<TYPE>(dstTotalSize), \
+      static_cast<TYPE>((IDX_IS_MAJOR) ? sliceSize : numIndices),  \
       srcSelectDimSize);
 
   dim3 smallIndexGrid(std::min(THCCeilDiv(sliceSize, (ptrdiff_t)128), (ptrdiff_t)(mpc * 8)));

aten/src/THCUNN/LookupTableBag.cu

@@ -15,7 +15,11 @@
 #include "THCHalfAutoNumerics.cuh"
 #include "THCTensorSort.cuh"
 
+#if defined(__HIP_PLATFORM_HCC__)
+const int WARP_SIZE = 64;
+#else
 const int WARP_SIZE = 32;
+#endif
 const int MODE_SUM = 0;
 const int MODE_MEAN = 1;
 

cmake/public/LoadHIP.cmake

@@ -47,7 +47,7 @@ ENDIF()
 
 # ROCFFT_PATH
 IF(NOT DEFINED ENV{ROCFFT_PATH})
-  SET(ROCBLAS_PATH ${ROCM_PATH}/rocfft)
+  SET(ROCFFT_PATH ${ROCM_PATH}/rocfft)
 ELSE()
   SET(ROCFFT_PATH $ENV{ROCFFT_PATH})
 ENDIF()