Merge pull request #55 from ROCmSoftwarePlatform/IFU-master-2021-10-15

IFU-2021-10-15 (+ remove redundant defines + C10_CUDA_CHECK)

Merge pull request #55 from ROCmSoftwarePlatform/IFU-master-2021-10-15
IFU-2021-10-15 (+ remove redundant defines + C10_CUDA_CHECK)
cc92a4b4 · Jithun Nair · GitHub · 1e0f9bc6 · fec3141c · cc92a4b4
Unverified Commit cc92a4b4 authored Dec 08, 2021 by Jithun Nair Committed by GitHub Dec 08, 2021
20 changed files
--- a/apex/contrib/csrc/fmha/src/fmha_dgrad_fp16_256_64_kernel.sm80.cu
+++ b/apex/contrib/csrc/fmha/src/fmha_dgrad_fp16_256_64_kernel.sm80.cu
--- a/apex/contrib/csrc/fmha/src/fmha_dgrad_fp16_384_64_kernel.sm80.cu
+++ b/apex/contrib/csrc/fmha/src/fmha_dgrad_fp16_384_64_kernel.sm80.cu
--- a/apex/contrib/csrc/fmha/src/fmha_dgrad_fp16_512_64_kernel.sm80.cu
+++ b/apex/contrib/csrc/fmha/src/fmha_dgrad_fp16_512_64_kernel.sm80.cu
--- a/apex/contrib/csrc/fmha/src/fmha_dgrad_kernel_1xN_reload.h
+++ b/apex/contrib/csrc/fmha/src/fmha_dgrad_kernel_1xN_reload.h
--- a/apex/contrib/csrc/fmha/src/fmha_dgrad_kernel_1xN_reload_nl.h
+++ b/apex/contrib/csrc/fmha/src/fmha_dgrad_kernel_1xN_reload_nl.h
--- a/apex/contrib/csrc/fmha/src/fmha_fprop_fp16_128_64_kernel.sm80.cu
+++ b/apex/contrib/csrc/fmha/src/fmha_fprop_fp16_128_64_kernel.sm80.cu
--- a/apex/contrib/csrc/fmha/src/fmha_fprop_fp16_256_64_kernel.sm80.cu
+++ b/apex/contrib/csrc/fmha/src/fmha_fprop_fp16_256_64_kernel.sm80.cu
--- a/apex/contrib/csrc/fmha/src/fmha_fprop_fp16_384_64_kernel.sm80.cu
+++ b/apex/contrib/csrc/fmha/src/fmha_fprop_fp16_384_64_kernel.sm80.cu
--- a/apex/contrib/csrc/fmha/src/fmha_fprop_fp16_512_64_kernel.sm80.cu
+++ b/apex/contrib/csrc/fmha/src/fmha_fprop_fp16_512_64_kernel.sm80.cu
--- a/apex/contrib/csrc/fmha/src/fmha_fprop_kernel_1xN.h
+++ b/apex/contrib/csrc/fmha/src/fmha_fprop_kernel_1xN.h
--- a/apex/contrib/csrc/fmha/src/fmha_fprop_kernel_1xN_nl.h
+++ b/apex/contrib/csrc/fmha/src/fmha_fprop_kernel_1xN_nl.h
--- a/apex/contrib/csrc/fmha/src/fmha_fprop_kernel_1xN_reload_v.h
+++ b/apex/contrib/csrc/fmha/src/fmha_fprop_kernel_1xN_reload_v.h
--- a/apex/contrib/csrc/fmha/src/fmha_kernel.h
+++ b/apex/contrib/csrc/fmha/src/fmha_kernel.h
--- a/apex/contrib/csrc/fmha/src/fmha_noloop_reduce.cu
+++ b/apex/contrib/csrc/fmha/src/fmha_noloop_reduce.cu
--- a/apex/contrib/csrc/fmha/src/fmha_utils.h
+++ b/apex/contrib/csrc/fmha/src/fmha_utils.h
--- a/apex/contrib/csrc/groupbn/batch_norm.cu
+++ b/apex/contrib/csrc/groupbn/batch_norm.cu
--- a/apex/contrib/csrc/groupbn/batch_norm_add_relu.cu
+++ b/apex/contrib/csrc/groupbn/batch_norm_add_relu.cu
--- a/apex/contrib/csrc/groupbn/ipc.cu
+++ b/apex/contrib/csrc/groupbn/ipc.cu
 #include <ATen/ATen.h>
 #include <ATen/cuda/CUDAContext.h>
-#include <THC/THCNumerics.cuh>
-
-#include "THC/THC.h"

 #include <cuda.h>


--- a/apex/contrib/csrc/multihead_attn/additive_masked_softmax_dropout_cuda.cu
+++ b/apex/contrib/csrc/multihead_attn/additive_masked_softmax_dropout_cuda.cu
--- a/apex/contrib/csrc/multihead_attn/dropout.h
+++ b/apex/contrib/csrc/multihead_attn/dropout.h