fix merge conflicts

xadupre · Jun 12, 2024 · 3846169 · 3846169
1 parent 6da49bf
commit 3846169
Show file tree

Hide file tree

Showing 4 changed files with 16 additions and 14 deletions.
diff --git a/operators/cuda/cuda_ops.cc b/operators/cuda/cuda_ops.cc
@@ -8,16 +8,12 @@
 #include "cuda/fast_gelu.h"
 #include "cuda/mul_sigmoid.h"
 #include "cuda/negxplus1.h"
-<<<<<<< HEAD
 #include "cuda/replace_zero.h"
-=======
 #include "cuda/scatter_nd_of_shape.h"
->>>>>>> f5055466d5376059c2ea74e3cea46e16a537bc0d
 #include "cuda/transpose_cast.h"
 #endif
 
 FxLoadCustomOpFactory LoadCustomOpClasses_Contrib = []() -> CustomOpArray& {
-
   using AddSharedInputFloat32Type = typename contrib::AddOrMulSharedInput<float, true>;
   using MulSharedInputFloat32Type = typename contrib::AddOrMulSharedInput<float, false>;
 
@@ -28,7 +24,6 @@ FxLoadCustomOpFactory LoadCustomOpClasses_Contrib = []() -> CustomOpArray& {
   using Transpose2DCastFloat16ToFloat32Type = typename contrib::Transpose2DCast<ortc::MFloat16, float>;
 #endif
 
-
   static OrtOpLoader op_loader(
       []() { return nullptr; }
 #ifdef USE_CUDA

diff --git a/operators/cuda/replace_zero.h b/operators/cuda/replace_zero.h
@@ -13,6 +13,9 @@ namespace contrib {
 *
 * Y = X.copy()
 * X[X == 0] = c
+*
+* This operation usually appears when a tensor is updated with an operator Equal and Where.
+* This kernel avoids the creation of one null tensor.
 */
 template <typename T>
 struct ReplaceZero {

diff --git a/operators/cuda/replace_zero_impl.cu b/operators/cuda/replace_zero_impl.cu
@@ -12,11 +12,13 @@
 
 using namespace Ort::Custom;
 
-template <typename T> __device__ __inline__ T _replace_zero(const T x, const T by) {
+template <typename T>
+__device__ __inline__ T _replace_zero(const T x, const T by) {
   return x == (T)0 ? by : x;
 }
 
-template <> __device__ __inline__ half _replace_zero(const half x, const half by) {
+template <>
+__device__ __inline__ half _replace_zero(const half x, const half by) {
 #if __CUDA_ARCH__ < 700
   return __half2float(x) == 0 ? by : x;
 #else
@@ -25,31 +27,33 @@ template <> __device__ __inline__ half _replace_zero(const half x, const half by
 }
 
 template <typename T>
-__global__ void ReplaceZeroKernel(T *output_data, const T *input_data, CUDA_LONG N, const T by) {
+__global__ void ReplaceZeroKernel(T* output_data, const T* input_data, CUDA_LONG N, const T by) {
   CUDA_LONG id = blockDim.x * blockIdx.x + threadIdx.x;
   if (id >= N)
     return;
   output_data[id] = _replace_zero(input_data[id], by);
 }
 
-template <typename T> T _cvt(float value) { return (T)value; }
+template <typename T>
+T _cast(float value) { return (T)value; }
 
-template <> half _cvt(float value) { return __float2half(value); }
+template <>
+half _cast(float value) { return __float2half(value); }
 
 template <typename T>
 cudaError_t _LaunchReplaceZeroKernel(cudaStream_t stream, int input_length, const T* input_data, T* output_data, float by) {
   if (input_length == 0)
-      return cudaGetLastError();
+    return cudaGetLastError();
   using TT = typename contrib::CudaT<T>::MappedType;
 
   CUDA_LONG N = static_cast<CUDA_LONG>(input_length);
 
   const int num_threads_per_block = 256;
   const int num_elements_per_thread = (N + num_threads_per_block - 1) / num_threads_per_block;
 
-  TT cby = _cvt<TT>(by);
+  TT cby = _cast<TT>(by);
   ReplaceZeroKernel<TT><<<num_elements_per_thread, num_threads_per_block, 0, stream>>>(
-    reinterpret_cast<TT*>(output_data), reinterpret_cast<const TT*>(input_data), N, cby);
+      reinterpret_cast<TT*>(output_data), reinterpret_cast<const TT*>(input_data), N, cby);
   return cudaGetLastError();
 }
 

diff --git a/operators/cuda/replace_zero_impl.cuh b/operators/cuda/replace_zero_impl.cuh
@@ -6,4 +6,4 @@
 #include <cuda_runtime.h>
 
 template <typename T>
-cudaError_t LaunchReplaceZeroKernel(cudaStream_t stream, int input_length,  const T* input_data, T* output_data, float by);
+cudaError_t LaunchReplaceZeroKernel(cudaStream_t stream, int input_length, const T* input_data, T* output_data, float by);