ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 17:00:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

another version of direct conv

Browse Source
This commit is contained in:
Chao Liu
2018-12-18 03:22:12 -06:00
parent 20423a3583
commit 39775d484c
12 changed files with 596 additions and 157 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
driver/conv.cu
View File

@@ -7,6 +7,7 @@
#include "constant_tensor_descriptor.cuh"
#include "device_direct_convolution_1.cuh"
#include "device_direct_convolution_2.cuh"
#include "device_direct_convolution_3.cuh"
//#include "device_winograd_convolution.cuh"
struct GeneratorTensor_1
@@ -338,7 +339,7 @@ int main()
constexpr unsigned K = 1;
constexpr unsigned S = 3;
constexpr unsigned R = 3;
#elif 0
#elif 1
constexpr unsigned N = 64;
constexpr unsigned C = 256;
constexpr unsigned HI = 34;
@@ -346,7 +347,7 @@ int main()
constexpr unsigned K = 64;
constexpr unsigned S = 3;
constexpr unsigned R = 3;
#elif 1
#elif 0
constexpr unsigned N = 64;
constexpr unsigned C = 64;
constexpr unsigned HI = 56;
@@ -387,11 +388,15 @@ int main()
wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
#endif
for(int i = 0; i < 20; ++i)
for(int i = 0; i < 40; ++i)
{
#if 1
device_direct_convolution_1(in_desc, in, wei_desc, wei, out_desc, out_device);
#else
#elif 0
device_direct_convolution_2(in_desc, in, wei_desc, wei, out_desc, out_device);
#elif 0
device_direct_convolution_3(in_desc, in, wei_desc, wei, out_desc, out_device);
#elif 0
device_winograd_convolution(in_desc, in, wei_desc, wei, out_desc, out_device);
#endif
}

2
driver/device_direct_convolution_1.cuh
View File

@@ -25,7 +25,7 @@ void device_direct_convolution_1(
constexpr auto wei_desc = WeiDesc{};
constexpr auto out_desc = OutDesc{};
#if 0
#if 1
constexpr unsigned OutTileSizeH = 2;
constexpr unsigned OutTileSizeW = 2;
constexpr unsigned NPerBlock = 2;

6
driver/device_direct_convolution_2.cuh
View File

@@ -25,7 +25,7 @@ void device_direct_convolution_2(
constexpr auto wei_desc = WeiDesc{};
constexpr auto out_desc = OutDesc{};
#if 0
#if 1
constexpr unsigned OutTileSizeH = 2;
constexpr unsigned OutTileSizeW = 2;
constexpr unsigned NPerBlock = 2;
@@ -39,7 +39,7 @@ void device_direct_convolution_2(
constexpr unsigned CPerThread = 2;
constexpr unsigned BlockSize = 128;
#elif 1
#elif 0
constexpr unsigned OutTileSizeH = 2;
constexpr unsigned OutTileSizeW = 2;
constexpr unsigned NPerBlock = 2;
@@ -53,7 +53,7 @@ void device_direct_convolution_2(
constexpr unsigned CPerThread = 2;
constexpr unsigned BlockSize = 216;
#elif 1
#elif 0
constexpr unsigned OutTileSizeH = 2;
constexpr unsigned OutTileSizeW = 2;
constexpr unsigned NPerBlock = 2;

120
driver/device_direct_convolution_3.cuh Normal file
View File

@@ -0,0 +1,120 @@
#pragma once
#include "gridwise_direct_convolution_3.cuh"
template <class T, class InDesc, class WeiDesc, class OutDesc>
void device_direct_convolution_3(
InDesc, const Tensor<T>& in, WeiDesc, const Tensor<T>& wei, OutDesc, Tensor<T>& out)
{
std::size_t data_sz = sizeof(T);
DeviceMem in_device_buf(data_sz * in.mDesc.GetElementSpace());
DeviceMem wei_device_buf(data_sz * wei.mDesc.GetElementSpace());
DeviceMem out_device_buf(data_sz * out.mDesc.GetElementSpace());
int num_thread = std::thread::hardware_concurrency();
in_device_buf.ToDevice(in.mData.data());
wei_device_buf.ToDevice(wei.mData.data());
out_device_buf.ToDevice(out.mData.data());
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_desc = InDesc{};
constexpr auto wei_desc = WeiDesc{};
constexpr auto out_desc = OutDesc{};
#if 1
constexpr unsigned OutTileSizeH = 2;
constexpr unsigned OutTileSizeW = 2;
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 32;
constexpr unsigned CPerBlock = 4;
constexpr unsigned YPerBlock = 1;
constexpr unsigned XPerBlock = 16;
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 4;
constexpr unsigned CPerThread = 2;
constexpr unsigned BlockSize = 128;
#elif 0
constexpr unsigned OutTileSizeH = 2;
constexpr unsigned OutTileSizeW = 2;
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 32;
constexpr unsigned CPerBlock = 4;
constexpr unsigned YPerBlock = 1;
constexpr unsigned XPerBlock = 27;
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 4;
constexpr unsigned CPerThread = 2;
constexpr unsigned BlockSize = 216;
#elif 0
constexpr unsigned OutTileSizeH = 2;
constexpr unsigned OutTileSizeW = 2;
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 32;
constexpr unsigned CPerBlock = 4;
constexpr unsigned YPerBlock = 1;
constexpr unsigned XPerBlock = 32;
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 4;
constexpr unsigned CPerThread = 2;
constexpr unsigned BlockSize = 256;
#endif
constexpr unsigned GridSize = (out_desc.GetLength(I0) / NPerBlock) *
(out_desc.GetLength(I1) / KPerBlock) *
(out_desc.GetLength(I2) / (OutTileSizeH * YPerBlock)) *
(out_desc.GetLength(I3) / (OutTileSizeW * XPerBlock));
dim3 block_dim(BlockSize);
dim3 grid_dim(GridSize);
printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
cudaEvent_t start, stop;
float elapsedTime;
cudaEventCreate(&start);
cudaEventRecord(start, 0);
gridwise_direct_convolution_3<T,
InDesc,
WeiDesc,
OutDesc,
OutTileSizeH,
OutTileSizeW,
NPerBlock,
KPerBlock,
CPerBlock,
YPerBlock,
XPerBlock,
NPerThread,
KPerThread,
CPerThread,
BlockSize,
GridSize>
<<<grid_dim, block_dim>>>(InDesc{},
static_cast<T*>(in_device_buf.GetDeviceBuffer()),
WeiDesc{},
static_cast<T*>(wei_device_buf.GetDeviceBuffer()),
OutDesc{},
static_cast<T*>(out_device_buf.GetDeviceBuffer()));
cudaEventCreate(&stop);
cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
cudaEventElapsedTime(&elapsedTime, start, stop);
printf("Elapsed time : %f ms\n", elapsedTime);
checkCudaErrors(cudaGetLastError());
out_device_buf.FromDevice(out.mData.data());
}