ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-15 18:42:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

clean up

Browse Source 
[ROCm/composable_kernel commit: bd7a230006]
This commit is contained in:
Chao Liu
2019-09-12 14:55:46 -05:00
parent bea9be7a3c
commit 79e13d9ded
10 changed files with 578 additions and 102 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
driver/include/device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded.hpp
View File

@@ -6,15 +6,15 @@
using namespace ck;
template <class T, class InDesc, class WeiDesc, class OutDesc, class LowerPads, class UpperPads>
template <typename T, class InDesc, class WeiDesc, class OutDesc, class LeftPads, class RightPads>
void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded(InDesc,
const Tensor<T>& in_nchw,
WeiDesc,
const Tensor<T>& wei_kcyx,
OutDesc,
Tensor<T>& out_nkhw,
LowerPads,
UpperPads,
LeftPads,
RightPads,
index_t nrepeat)
{
constexpr auto I0 = Number<0>{};
@@ -131,8 +131,8 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded(InDesc,
decltype(in_chwn_desc),
decltype(wei_cyxk_desc),
decltype(out_khwn_desc),
LowerPads,
UpperPads,
LeftPads,
RightPads,
NPerBlock,
KPerBlock,
CPerBlock,

13
driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp
View File

@@ -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,
@@ -33,18 +33,11 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
constexpr auto wei_kcyx_desc = WeiDesc{};
constexpr auto out_nkhw_desc = OutDesc{};
constexpr index_t Hi = in_nchw_desc.GetLength(I2);
constexpr index_t Wi = in_nchw_desc.GetLength(I3);
constexpr index_t N = out_nkhw_desc.GetLength(I0);
constexpr index_t K = out_nkhw_desc.GetLength(I1);
constexpr index_t Ho = out_nkhw_desc.GetLength(I2);
constexpr index_t Wo = out_nkhw_desc.GetLength(I3);
constexpr index_t K = wei_kcyx_desc.GetLength(I0);
constexpr index_t C = wei_kcyx_desc.GetLength(I1);
constexpr index_t Y = wei_kcyx_desc.GetLength(I2);
constexpr index_t X = wei_kcyx_desc.GetLength(I3);
std::size_t data_sz = sizeof(T);
DeviceMem in_nchw_device_buf(data_sz * in_nchw.mDesc.GetElementSpace());
DeviceMem wei_kcyx_device_buf(data_sz * wei_kcyx.mDesc.GetElementSpace());
@@ -143,7 +136,7 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(InDesc,
for(index_t i = 0; i < nrepeat; ++i)
{
constexpr auto gridwise_conv =
#if 0
#if 1
GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw
#else
GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_lds_double_buffer

159
driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded.hpp Normal file
View File

@@ -0,0 +1,159 @@
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "tensor.hpp"
#include "gridwise_convolution_kernel_wrapper.hpp"
#include "gridwise_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded.hpp"
template <typename T,
typename InDesc,
typename WeiDesc,
typename OutDesc,
typename ConvStrides,
typename ConvDilations,
typename LeftPads,
typename RightPads>
void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded(InDesc,
const Tensor<T>& in_nchw,
WeiDesc,
const Tensor<T>& wei_kcyx,
OutDesc,
Tensor<T>& out_nkhw,
ConvStrides,
ConvDilations,
LeftPads,
RightPads,
index_t nrepeat)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_nchw_desc = InDesc{};
constexpr auto wei_kcyx_desc = WeiDesc{};
constexpr auto out_nkhw_desc = OutDesc{};
constexpr index_t N = out_nkhw_desc.GetLength(I0);
constexpr index_t K = out_nkhw_desc.GetLength(I1);
constexpr index_t Ho = out_nkhw_desc.GetLength(I2);
constexpr index_t Wo = out_nkhw_desc.GetLength(I3);
std::size_t data_sz = sizeof(T);
DeviceMem in_nchw_device_buf(data_sz * in_nchw.mDesc.GetElementSpace());
DeviceMem wei_kcyx_device_buf(data_sz * wei_kcyx.mDesc.GetElementSpace());
DeviceMem out_nkhw_device_buf(data_sz * out_nkhw.mDesc.GetElementSpace());
in_nchw_device_buf.ToDevice(in_nchw.mData.data());
wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
#if 1
// each thread hold 64 data
constexpr index_t BlockSize = 256;
constexpr index_t BPerBlock = 16;
constexpr index_t KPerBlock = 128;
constexpr index_t EPerBlock = 8;
constexpr index_t GemmNRepeat = 2;
constexpr index_t GemmMPerThreadSubC = 4;
constexpr index_t GemmNPerThreadSubC = 4;
constexpr index_t GemmMLevel0Cluster = 4;
constexpr index_t GemmNLevel0Cluster = 4;
constexpr index_t GemmMLevel1Cluster = 4;
constexpr index_t GemmNLevel1Cluster = 4;
constexpr index_t GemmKPerThreadLoop = 1;
constexpr index_t GemmDataPerReadA = 4;
constexpr index_t GemmDataPerReadB = 4;
using InBlockCopySubLengths_E_N1_B_N2 = Sequence<1, 1, 1, 4>;
using InBlockCopyClusterLengths_E_N1_B_N2 = Sequence<8, 2, 16, 1>;
using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopySrcAccessOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
using InBlockCopyDstAccessOrder = Sequence<0, 1, 2, 3>; // [E, N1, B, N2]
constexpr index_t InBlockCopySrcDataPerRead_B = 1;
constexpr index_t InBlockCopyDstDataPerWrite_N2 = 4;
using WeiBlockCopySubLengths_E_K = Sequence<4, 1>;
using WeiBlockCopyClusterLengths_E_K = Sequence<2, 128>;
using WeiBlockCopyThreadClusterArrangeOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopySrcAccessOrder = Sequence<1, 0>; // [K, E]
using WeiBlockCopyDstAccessOrder = Sequence<0, 1>; // [E, K]
constexpr index_t WeiBlockCopySrcDataPerRead_E = 4;
constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
#endif
constexpr index_t N1 = GemmNRepeat;
constexpr index_t N2 = GemmNPerThreadSubC;
constexpr index_t B = (N * Ho * Wo) / (N1 * N2);
constexpr index_t GridSize =
((B + BPerBlock - 1) / BPerBlock) * ((K + KPerBlock - 1) / KPerBlock);
printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
for(index_t i = 0; i < nrepeat; ++i)
{
constexpr auto gridwise_conv = GridwiseConvolutionImplicitGemm_v4r1_nchw_kcyx_nkhw_padded<
GridSize,
BlockSize,
T,
decltype(in_nchw_desc),
decltype(wei_kcyx_desc),
decltype(out_nkhw_desc),
ConvStrides,
ConvDilations,
LeftPads,
RightPads,
BPerBlock,
KPerBlock,
EPerBlock,
GemmNRepeat,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
GemmDataPerReadA,
GemmDataPerReadB,
InBlockCopySubLengths_E_N1_B_N2,
InBlockCopyClusterLengths_E_N1_B_N2,
InBlockCopyThreadClusterArrangeOrder,
InBlockCopySrcAccessOrder,
InBlockCopyDstAccessOrder,
InBlockCopySrcDataPerRead_B,
InBlockCopyDstDataPerWrite_N2,
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>{};
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
dim3(GridSize),
dim3(BlockSize),
0,
static_cast<T*>(in_nchw_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_kcyx_device_buf.GetDeviceBuffer()),
static_cast<T*>(out_nkhw_device_buf.GetDeviceBuffer()));
printf("Elapsed time : %f ms, %f TFlop/s\n",
time,
(float)calculate_convolution_flops(InDesc{}, WeiDesc{}, OutDesc{}) /
(std::size_t(1000) * 1000 * 1000) / time);
usleep(std::min(time * 1000, float(10000)));
}
out_nkhw_device_buf.FromDevice(out_nkhw.mData.data());
}