mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-12 09:16:52 +00:00
clean up
This commit is contained in:
@@ -3,7 +3,7 @@
|
||||
#include "device.hpp"
|
||||
#include "tensor.hpp"
|
||||
#include "gridwise_convolution_kernel_wrapper.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp"
|
||||
//#include "gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer.hpp"
|
||||
|
||||
template <class T,
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
#include "device.hpp"
|
||||
#include "tensor.hpp"
|
||||
#include "gridwise_convolution_kernel_wrapper.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp"
|
||||
//#include "gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp"
|
||||
#include "gridwise_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_lds_double_buffer.hpp"
|
||||
|
||||
using namespace ck;
|
||||
@@ -164,7 +164,7 @@ void device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(InDesc,
|
||||
printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
|
||||
|
||||
constexpr auto gridwise_conv =
|
||||
#if 1
|
||||
#if 0
|
||||
GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw
|
||||
#else
|
||||
GridwiseConvolutionImplicitGemm_v4r4_nchw_kcyx_nkhw_lds_double_buffer
|
||||
|
||||
@@ -51,7 +51,7 @@ void device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_padded(InDesc,
|
||||
wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
|
||||
out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
|
||||
|
||||
#if 0
|
||||
#if 1
|
||||
constexpr index_t BlockSize = 256;
|
||||
|
||||
constexpr index_t BPerBlock = 128;
|
||||
|
||||
Reference in New Issue
Block a user