ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 01:10:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into implicit_gemm_fp16

Browse Source
This commit is contained in:
Chao Liu
2019-03-09 13:46:47 -06:00
parent 8edbc659b8 f54cad7d4f
commit ce0182ce05
9 changed files with 182 additions and 182 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
driver/device_implicit_gemm_convolution_1_chwn_csrk_khwn.hpp → driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp
View File

@@ -1,13 +1,13 @@
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn.hip.hpp"
#include "gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn.hip.hpp"
template <class T, class InDesc, class WeiDesc, class OutDesc>
void device_implicit_gemm_convolution_1_chwn_csrk_khwn(InDesc,
void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
const Tensor<T>& in_nchw,
WeiDesc,
const Tensor<T>& wei_kcsr,
const Tensor<T>& wei_kcyx,
OutDesc,
Tensor<T>& out_nkhw,
unsigned nrepeat)
@@ -18,7 +18,7 @@ void device_implicit_gemm_convolution_1_chwn_csrk_khwn(InDesc,
constexpr auto I3 = Number<3>{};
constexpr auto in_nchw_desc = InDesc{};
constexpr auto wei_kcsr_desc = WeiDesc{};
constexpr auto wei_kcyx_desc = WeiDesc{};
constexpr auto out_nkhw_desc = OutDesc{};
constexpr unsigned Hi = in_nchw_desc.GetLength(I2);
@@ -28,22 +28,22 @@ void device_implicit_gemm_convolution_1_chwn_csrk_khwn(InDesc,
constexpr unsigned Ho = out_nkhw_desc.GetLength(I2);
constexpr unsigned Wo = out_nkhw_desc.GetLength(I3);
constexpr unsigned K = wei_kcsr_desc.GetLength(I0);
constexpr unsigned C = wei_kcsr_desc.GetLength(I1);
constexpr unsigned Y = wei_kcsr_desc.GetLength(I2);
constexpr unsigned X = wei_kcsr_desc.GetLength(I3);
constexpr unsigned K = wei_kcyx_desc.GetLength(I0);
constexpr unsigned C = wei_kcyx_desc.GetLength(I1);
constexpr unsigned Y = wei_kcyx_desc.GetLength(I2);
constexpr unsigned X = wei_kcyx_desc.GetLength(I3);
// reorder weight
auto wei_csrk_desc = make_ConstantTensorDescriptor(Sequence<C, Y, X, K>{});
ostream_ConstantTensorDescriptor(wei_csrk_desc, std::cout << "wei_csrk_desc: ");
auto wei_cyxk_desc = make_ConstantTensorDescriptor(Sequence<C, Y, X, K>{});
ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
Tensor<T> wei_csrk(make_TensorDescriptor(wei_csrk_desc));
Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
auto f_reorder_kcsr2csrk = [&](auto k, auto c, auto s, auto r) {
wei_csrk(c, s, r, k) = wei_kcsr(k, c, s, r);
auto f_reorder_kcyx2cyxk = [&](auto k, auto c, auto y, auto x) {
wei_cyxk(c, y, x, k) = wei_kcyx(k, c, y, x);
};
make_ParallelTensorFunctor(f_reorder_kcsr2csrk, K, C, Y, X)(
make_ParallelTensorFunctor(f_reorder_kcyx2cyxk, K, C, Y, X)(
std::thread::hardware_concurrency());
// reorder input
@@ -67,11 +67,11 @@ void device_implicit_gemm_convolution_1_chwn_csrk_khwn(InDesc,
std::size_t data_sz = sizeof(T);
DeviceMem in_chwn_device_buf(data_sz * in_chwn.mDesc.GetElementSpace());
DeviceMem wei_csrk_device_buf(data_sz * wei_csrk.mDesc.GetElementSpace());
DeviceMem wei_cyxk_device_buf(data_sz * wei_cyxk.mDesc.GetElementSpace());
DeviceMem out_khwn_device_buf(data_sz * out_khwn.mDesc.GetElementSpace());
in_chwn_device_buf.ToDevice(in_chwn.mData.data());
wei_csrk_device_buf.ToDevice(wei_csrk.mData.data());
wei_cyxk_device_buf.ToDevice(wei_cyxk.mData.data());
out_khwn_device_buf.ToDevice(out_khwn.mData.data());
#if 1
@@ -257,11 +257,11 @@ void device_implicit_gemm_convolution_1_chwn_csrk_khwn(InDesc,
for(unsigned i = 0; i < nrepeat; ++i)
{
float time = launch_kernel(
gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn<GridSize,
gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn<GridSize,
BlockSize,
T,
decltype(in_chwn_desc),
decltype(wei_csrk_desc),
decltype(wei_cyxk_desc),
decltype(out_khwn_desc),
NPerBlock,
KPerBlock,
@@ -289,7 +289,7 @@ void device_implicit_gemm_convolution_1_chwn_csrk_khwn(InDesc,
dim3(GridSize),
dim3(BlockSize),
static_cast<T*>(in_chwn_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_csrk_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_cyxk_device_buf.GetDeviceBuffer()),
static_cast<T*>(out_khwn_device_buf.GetDeviceBuffer()));
printf("Elapsed time : %f ms\n", time);

38
driver/device_implicit_gemm_convolution_1_chwn_csrk_khwn_padded.hpp → driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded.hpp
View File

@@ -1,13 +1,13 @@
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded.hip.hpp"
#include "gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded.hip.hpp"
template <class T, class InDesc, class WeiDesc, class OutDesc, class LowerPads, class UpperPads>
void device_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(InDesc,
void device_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded(InDesc,
const Tensor<T>& in_nchw,
WeiDesc,
const Tensor<T>& wei_kcsr,
const Tensor<T>& wei_kcyx,
OutDesc,
Tensor<T>& out_nkhw,
LowerPads,
@@ -20,7 +20,7 @@ void device_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(InDesc,
constexpr auto I3 = Number<3>{};
constexpr auto in_nchw_desc = InDesc{};
constexpr auto wei_kcsr_desc = WeiDesc{};
constexpr auto wei_kcyx_desc = WeiDesc{};
constexpr auto out_nkhw_desc = OutDesc{};
constexpr unsigned Hi = in_nchw_desc.GetLength(I2);
@@ -30,22 +30,22 @@ void device_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(InDesc,
constexpr unsigned Ho = out_nkhw_desc.GetLength(I2);
constexpr unsigned Wo = out_nkhw_desc.GetLength(I3);
constexpr unsigned K = wei_kcsr_desc.GetLength(I0);
constexpr unsigned C = wei_kcsr_desc.GetLength(I1);
constexpr unsigned Y = wei_kcsr_desc.GetLength(I2);
constexpr unsigned X = wei_kcsr_desc.GetLength(I3);
constexpr unsigned K = wei_kcyx_desc.GetLength(I0);
constexpr unsigned C = wei_kcyx_desc.GetLength(I1);
constexpr unsigned Y = wei_kcyx_desc.GetLength(I2);
constexpr unsigned X = wei_kcyx_desc.GetLength(I3);
// reorder weight
auto wei_csrk_desc = make_ConstantTensorDescriptor(Sequence<C, Y, X, K>{});
ostream_ConstantTensorDescriptor(wei_csrk_desc, std::cout << "wei_csrk_desc: ");
auto wei_cyxk_desc = make_ConstantTensorDescriptor(Sequence<C, Y, X, K>{});
ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
Tensor<T> wei_csrk(make_TensorDescriptor(wei_csrk_desc));
Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
auto f_reorder_kcsr2csrk = [&](auto k, auto c, auto s, auto r) {
wei_csrk(c, s, r, k) = wei_kcsr(k, c, s, r);
auto f_reorder_kcyx2cyxk = [&](auto k, auto c, auto y, auto x) {
wei_cyxk(c, y, x, k) = wei_kcyx(k, c, y, x);
};
make_ParallelTensorFunctor(f_reorder_kcsr2csrk, K, C, Y, X)(
make_ParallelTensorFunctor(f_reorder_kcyx2cyxk, K, C, Y, X)(
std::thread::hardware_concurrency());
// reorder input
@@ -69,11 +69,11 @@ void device_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(InDesc,
std::size_t data_sz = sizeof(T);
DeviceMem in_chwn_device_buf(data_sz * in_chwn.mDesc.GetElementSpace());
DeviceMem wei_csrk_device_buf(data_sz * wei_csrk.mDesc.GetElementSpace());
DeviceMem wei_cyxk_device_buf(data_sz * wei_cyxk.mDesc.GetElementSpace());
DeviceMem out_khwn_device_buf(data_sz * out_khwn.mDesc.GetElementSpace());
in_chwn_device_buf.ToDevice(in_chwn.mData.data());
wei_csrk_device_buf.ToDevice(wei_csrk.mData.data());
wei_cyxk_device_buf.ToDevice(wei_cyxk.mData.data());
out_khwn_device_buf.ToDevice(out_khwn.mData.data());
#if 0
@@ -250,11 +250,11 @@ void device_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(InDesc,
for(unsigned i = 0; i < nrepeat; ++i)
{
float time = launch_kernel(
gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded<GridSize,
gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded<GridSize,
BlockSize,
T,
decltype(in_chwn_desc),
decltype(wei_csrk_desc),
decltype(wei_cyxk_desc),
decltype(out_khwn_desc),
LowerPads,
UpperPads,
@@ -274,7 +274,7 @@ void device_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(InDesc,
dim3(BlockSize),
static_cast<T*>(in_chwn_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_csrk_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_cyxk_device_buf.GetDeviceBuffer()),
static_cast<T*>(out_khwn_device_buf.GetDeviceBuffer()));
printf("Elapsed time : %f ms\n", time);

36
driver/device_implicit_gemm_convolution_2_chwn_csrk_khwn.hpp → driver/device_implicit_gemm_convolution_2_chwn_cyxk_khwn.hpp
View File

@@ -1,13 +1,13 @@
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_buffer.hip.hpp"
#include "gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer.hip.hpp"
template <class T, class InDesc, class WeiDesc, class OutDesc>
void device_implicit_gemm_convolution_2_chwn_csrk_khwn(InDesc,
void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
const Tensor<T>& in_nchw,
WeiDesc,
const Tensor<T>& wei_kcsr,
const Tensor<T>& wei_kcyx,
OutDesc,
Tensor<T>& out_nkhw,
unsigned nrepeat)
@@ -18,7 +18,7 @@ void device_implicit_gemm_convolution_2_chwn_csrk_khwn(InDesc,
constexpr auto I3 = Number<3>{};
constexpr auto in_nchw_desc = InDesc{};
constexpr auto wei_kcsr_desc = WeiDesc{};
constexpr auto wei_kcyx_desc = WeiDesc{};
constexpr auto out_nkhw_desc = OutDesc{};
constexpr unsigned N = in_nchw_desc.GetLength(I0);
@@ -28,10 +28,10 @@ void device_implicit_gemm_convolution_2_chwn_csrk_khwn(InDesc,
constexpr unsigned Ho = out_nkhw_desc.GetLength(I2);
constexpr unsigned Wo = out_nkhw_desc.GetLength(I3);
constexpr unsigned K = wei_kcsr_desc.GetLength(I0);
constexpr unsigned C = wei_kcsr_desc.GetLength(I1);
constexpr unsigned Y = wei_kcsr_desc.GetLength(I2);
constexpr unsigned X = wei_kcsr_desc.GetLength(I3);
constexpr unsigned K = wei_kcyx_desc.GetLength(I0);
constexpr unsigned C = wei_kcyx_desc.GetLength(I1);
constexpr unsigned Y = wei_kcyx_desc.GetLength(I2);
constexpr unsigned X = wei_kcyx_desc.GetLength(I3);
constexpr unsigned BGhostRead = (Y - 1) * Wi + (X - 1);
@@ -48,14 +48,14 @@ void device_implicit_gemm_convolution_2_chwn_csrk_khwn(InDesc,
Hi,
Wi)(std::thread::hardware_concurrency());
// convert wei_kcsr to wei_csrk
auto wei_csrk_desc = make_ConstantTensorDescriptor(Sequence<C, Y, X, K>{});
ostream_ConstantTensorDescriptor(wei_csrk_desc, std::cout << "wei_csrk_desc: ");
// convert wei_kcyx to wei_cyxk
auto wei_cyxk_desc = make_ConstantTensorDescriptor(Sequence<C, Y, X, K>{});
ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
Tensor<T> wei_csrk(make_TensorDescriptor(wei_csrk_desc));
Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
make_ParallelTensorFunctor(
[&](auto k, auto c, auto s, auto r) { wei_csrk(c, s, r, k) = wei_kcsr(k, c, s, r); },
[&](auto k, auto c, auto y, auto x) { wei_cyxk(c, y, x, k) = wei_kcyx(k, c, y, x); },
K,
C,
Y,
@@ -200,22 +200,22 @@ void device_implicit_gemm_convolution_2_chwn_csrk_khwn(InDesc,
std::size_t data_sz = sizeof(T);
DeviceMem in_chwn_device_buf(data_sz * (in_chwn.mDesc.GetElementSpace() + BGhostRead +
BPerBlock)); // reserve extra space for BGhostRead
DeviceMem wei_csrk_device_buf(data_sz * wei_csrk.mDesc.GetElementSpace());
DeviceMem wei_cyxk_device_buf(data_sz * wei_cyxk.mDesc.GetElementSpace());
DeviceMem out_khwn_device_buf(data_sz * out_khwn.mDesc.GetElementSpace());
in_chwn_device_buf.ToDevice(in_chwn.mData.data());
wei_csrk_device_buf.ToDevice(wei_csrk.mData.data());
wei_cyxk_device_buf.ToDevice(wei_cyxk.mData.data());
out_khwn_device_buf.ToDevice(out_khwn.mData.data());
for(unsigned i = 0; i < nrepeat; ++i)
{
float time =
launch_kernel(gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_buffer<
launch_kernel(gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer<
GridSize,
BlockSize,
T,
decltype(in_chwn_desc),
decltype(wei_csrk_desc),
decltype(wei_cyxk_desc),
decltype(out_khwn_desc),
BPerBlock,
KPerBlock,
@@ -240,7 +240,7 @@ void device_implicit_gemm_convolution_2_chwn_csrk_khwn(InDesc,
dim3(GridSize),
dim3(BlockSize),
static_cast<T*>(in_chwn_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_csrk_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_cyxk_device_buf.GetDeviceBuffer()),
static_cast<T*>(out_khwn_device_buf.GetDeviceBuffer()));
printf("Elapsed time : %f ms\n", time);

120
driver/driver.hip.cpp
View File

@@ -9,9 +9,9 @@
#include "conv_common.hip.hpp"
#include "device_direct_convolution_1.hpp"
#include "device_direct_convolution_2.hpp"
#include "device_implicit_gemm_convolution_1_chwn_csrk_khwn.hpp"
#include "device_implicit_gemm_convolution_1_chwn_csrk_khwn_padded.hpp"
#include "device_implicit_gemm_convolution_2_chwn_csrk_khwn.hpp"
#include "device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp"
#include "device_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded.hpp"
#include "device_implicit_gemm_convolution_2_chwn_cyxk_khwn.hpp"
struct GeneratorTensor_1
{
@@ -108,7 +108,7 @@ auto make_TensorDescriptor(TConstTensorDesc)
template <class T, class LowerPads, class UpperPads>
void host_direct_convolution(
const Tensor<T>& in_nchw, const Tensor<T>& wei_kcsr, Tensor<T>& out, LowerPads, UpperPads)
const Tensor<T>& in_nchw, const Tensor<T>& wei_kcyx, Tensor<T>& out, LowerPads, UpperPads)
{
unsigned h_pad_low = LowerPads{}.Get(Number<0>{});
unsigned w_pad_low = LowerPads{}.Get(Number<1>{});
@@ -118,18 +118,18 @@ void host_direct_convolution(
auto f = [&](auto n, auto k, auto ho, auto wo) {
double v = 0;
for(int c = 0; c < wei_kcsr.mDesc.GetLengths()[1]; ++c)
for(int c = 0; c < wei_kcyx.mDesc.GetLengths()[1]; ++c)
{
for(int y = 0; y < wei_kcsr.mDesc.GetLengths()[2]; ++y)
for(int y = 0; y < wei_kcyx.mDesc.GetLengths()[2]; ++y)
{
int hi = ho + y - h_pad_low;
for(int x = 0; x < wei_kcsr.mDesc.GetLengths()[3]; ++x)
for(int x = 0; x < wei_kcyx.mDesc.GetLengths()[3]; ++x)
{
int wi = wo + x - w_pad_low;
if(hi >= 0 && hi < in_nchw.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in_nchw.mDesc.GetLengths()[3])
{
v += in_nchw(n, c, hi, wi) * wei_kcsr(k, c, y, x);
v += in_nchw(n, c, hi, wi) * wei_kcyx(k, c, y, x);
}
}
}
@@ -148,7 +148,7 @@ void host_direct_convolution(
template <class T, class LowerPads, class UpperPads>
void host_winograd_3x3_convolution(
const Tensor<T>& in_nchw, const Tensor<T>& wei_kcsr, Tensor<T>& out, LowerPads, UpperPads)
const Tensor<T>& in_nchw, const Tensor<T>& wei_kcyx, Tensor<T>& out, LowerPads, UpperPads)
{
constexpr std::size_t HoPerTile = 2;
constexpr std::size_t WoPerTile = 2;
@@ -158,9 +158,9 @@ void host_winograd_3x3_convolution(
std::size_t HI = in_nchw.mDesc.GetLengths()[2];
std::size_t WI = in_nchw.mDesc.GetLengths()[3];
std::size_t K = wei_kcsr.mDesc.GetLengths()[0];
std::size_t Y = wei_kcsr.mDesc.GetLengths()[2];
std::size_t X = wei_kcsr.mDesc.GetLengths()[3];
std::size_t K = wei_kcyx.mDesc.GetLengths()[0];
std::size_t Y = wei_kcyx.mDesc.GetLengths()[2];
std::size_t X = wei_kcyx.mDesc.GetLengths()[3];
std::size_t HO = out.mDesc.GetLengths()[2];
std::size_t WO = out.mDesc.GetLengths()[3];
@@ -259,49 +259,49 @@ void host_winograd_3x3_convolution(
};
auto f_wei_transform = [&](auto k, auto c) {
wei_transform(k, c, 0, 0) = wei_kcsr(k, c, 0, 0);
wei_transform(k, c, 0, 0) = wei_kcyx(k, c, 0, 0);
wei_transform(k, c, 0, 1) =
0.5 * wei_kcsr(k, c, 0, 0) + 0.5 * wei_kcsr(k, c, 0, 1) + 0.5 * wei_kcsr(k, c, 0, 2);
0.5 * wei_kcyx(k, c, 0, 0) + 0.5 * wei_kcyx(k, c, 0, 1) + 0.5 * wei_kcyx(k, c, 0, 2);
wei_transform(k, c, 0, 2) =
0.5 * wei_kcsr(k, c, 0, 0) - 0.5 * wei_kcsr(k, c, 0, 1) + 0.5 * wei_kcsr(k, c, 0, 2);
wei_transform(k, c, 0, 3) = wei_kcsr(k, c, 0, 2);
0.5 * wei_kcyx(k, c, 0, 0) - 0.5 * wei_kcyx(k, c, 0, 1) + 0.5 * wei_kcyx(k, c, 0, 2);
wei_transform(k, c, 0, 3) = wei_kcyx(k, c, 0, 2);
wei_transform(k, c, 1, 0) =
0.5 * wei_kcsr(k, c, 0, 0) + 0.5 * wei_kcsr(k, c, 1, 0) + 0.5 * wei_kcsr(k, c, 2, 0);
wei_transform(k, c, 1, 1) = 0.25 * wei_kcsr(k, c, 0, 0) + 0.25 * wei_kcsr(k, c, 0, 1) +
0.25 * wei_kcsr(k, c, 0, 2) + 0.25 * wei_kcsr(k, c, 1, 0) +
0.25 * wei_kcsr(k, c, 1, 1) + 0.25 * wei_kcsr(k, c, 1, 2) +
0.25 * wei_kcsr(k, c, 2, 0) + 0.25 * wei_kcsr(k, c, 2, 1) +
0.25 * wei_kcsr(k, c, 2, 2);
wei_transform(k, c, 1, 2) = 0.25 * wei_kcsr(k, c, 0, 0) - 0.25 * wei_kcsr(k, c, 0, 1) +
0.25 * wei_kcsr(k, c, 0, 2) + 0.25 * wei_kcsr(k, c, 1, 0) -
0.25 * wei_kcsr(k, c, 1, 1) + 0.25 * wei_kcsr(k, c, 1, 2) +
0.25 * wei_kcsr(k, c, 2, 0) - 0.25 * wei_kcsr(k, c, 2, 1) +
0.25 * wei_kcsr(k, c, 2, 2);
0.5 * wei_kcyx(k, c, 0, 0) + 0.5 * wei_kcyx(k, c, 1, 0) + 0.5 * wei_kcyx(k, c, 2, 0);
wei_transform(k, c, 1, 1) = 0.25 * wei_kcyx(k, c, 0, 0) + 0.25 * wei_kcyx(k, c, 0, 1) +
0.25 * wei_kcyx(k, c, 0, 2) + 0.25 * wei_kcyx(k, c, 1, 0) +
0.25 * wei_kcyx(k, c, 1, 1) + 0.25 * wei_kcyx(k, c, 1, 2) +
0.25 * wei_kcyx(k, c, 2, 0) + 0.25 * wei_kcyx(k, c, 2, 1) +
0.25 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 1, 2) = 0.25 * wei_kcyx(k, c, 0, 0) - 0.25 * wei_kcyx(k, c, 0, 1) +
0.25 * wei_kcyx(k, c, 0, 2) + 0.25 * wei_kcyx(k, c, 1, 0) -
0.25 * wei_kcyx(k, c, 1, 1) + 0.25 * wei_kcyx(k, c, 1, 2) +
0.25 * wei_kcyx(k, c, 2, 0) - 0.25 * wei_kcyx(k, c, 2, 1) +
0.25 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 1, 3) =
0.5 * wei_kcsr(k, c, 0, 2) + 0.5 * wei_kcsr(k, c, 1, 2) + 0.5 * wei_kcsr(k, c, 2, 2);
0.5 * wei_kcyx(k, c, 0, 2) + 0.5 * wei_kcyx(k, c, 1, 2) + 0.5 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 2, 0) =
0.5 * wei_kcsr(k, c, 0, 0) - 0.5 * wei_kcsr(k, c, 1, 0) + 0.5 * wei_kcsr(k, c, 2, 0);
wei_transform(k, c, 2, 1) = 0.25 * wei_kcsr(k, c, 0, 0) + 0.25 * wei_kcsr(k, c, 0, 1) +
0.25 * wei_kcsr(k, c, 0, 2) - 0.25 * wei_kcsr(k, c, 1, 0) -
0.25 * wei_kcsr(k, c, 1, 1) - 0.25 * wei_kcsr(k, c, 1, 2) +
0.25 * wei_kcsr(k, c, 2, 0) + 0.25 * wei_kcsr(k, c, 2, 1) +
0.25 * wei_kcsr(k, c, 2, 2);
wei_transform(k, c, 2, 2) = 0.25 * wei_kcsr(k, c, 0, 0) - 0.25 * wei_kcsr(k, c, 0, 1) +
0.25 * wei_kcsr(k, c, 0, 2) - 0.25 * wei_kcsr(k, c, 1, 0) +
0.25 * wei_kcsr(k, c, 1, 1) - 0.25 * wei_kcsr(k, c, 1, 2) +
0.25 * wei_kcsr(k, c, 2, 0) - 0.25 * wei_kcsr(k, c, 2, 1) +
0.25 * wei_kcsr(k, c, 2, 2);
0.5 * wei_kcyx(k, c, 0, 0) - 0.5 * wei_kcyx(k, c, 1, 0) + 0.5 * wei_kcyx(k, c, 2, 0);
wei_transform(k, c, 2, 1) = 0.25 * wei_kcyx(k, c, 0, 0) + 0.25 * wei_kcyx(k, c, 0, 1) +
0.25 * wei_kcyx(k, c, 0, 2) - 0.25 * wei_kcyx(k, c, 1, 0) -
0.25 * wei_kcyx(k, c, 1, 1) - 0.25 * wei_kcyx(k, c, 1, 2) +
0.25 * wei_kcyx(k, c, 2, 0) + 0.25 * wei_kcyx(k, c, 2, 1) +
0.25 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 2, 2) = 0.25 * wei_kcyx(k, c, 0, 0) - 0.25 * wei_kcyx(k, c, 0, 1) +
0.25 * wei_kcyx(k, c, 0, 2) - 0.25 * wei_kcyx(k, c, 1, 0) +
0.25 * wei_kcyx(k, c, 1, 1) - 0.25 * wei_kcyx(k, c, 1, 2) +
0.25 * wei_kcyx(k, c, 2, 0) - 0.25 * wei_kcyx(k, c, 2, 1) +
0.25 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 2, 3) =
0.5 * wei_kcsr(k, c, 0, 2) - 0.5 * wei_kcsr(k, c, 1, 2) + 0.5 * wei_kcsr(k, c, 2, 2);
0.5 * wei_kcyx(k, c, 0, 2) - 0.5 * wei_kcyx(k, c, 1, 2) + 0.5 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 3, 0) = wei_kcsr(k, c, 2, 0);
wei_transform(k, c, 3, 0) = wei_kcyx(k, c, 2, 0);
wei_transform(k, c, 3, 1) =
0.5 * wei_kcsr(k, c, 2, 0) + 0.5 * wei_kcsr(k, c, 2, 1) + 0.5 * wei_kcsr(k, c, 2, 2);
0.5 * wei_kcyx(k, c, 2, 0) + 0.5 * wei_kcyx(k, c, 2, 1) + 0.5 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 3, 2) =
0.5 * wei_kcsr(k, c, 2, 0) - 0.5 * wei_kcsr(k, c, 2, 1) + 0.5 * wei_kcsr(k, c, 2, 2);
wei_transform(k, c, 3, 3) = wei_kcsr(k, c, 2, 2);
0.5 * wei_kcyx(k, c, 2, 0) - 0.5 * wei_kcyx(k, c, 2, 1) + 0.5 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 3, 3) = wei_kcyx(k, c, 2, 2);
};
auto f_out_transform = [&](auto n, auto k, auto htile, auto wtile) {
@@ -569,16 +569,16 @@ int main(int argc, char* argv[])
auto upper_pads = Sequence<HPad, WPad>{};
auto in_nchw_desc = make_ConstantTensorDescriptor(Sequence<N, C, HI, WI>{});
auto wei_kcsr_desc = make_ConstantTensorDescriptor(Sequence<K, C, Y, X>{});
auto wei_kcyx_desc = make_ConstantTensorDescriptor(Sequence<K, C, Y, X>{});
auto out_nkhw_desc = get_convolution_with_padding_output_default_4d_tensor_descriptor(
in_nchw_desc, wei_kcsr_desc, lower_pads, upper_pads);
in_nchw_desc, wei_kcyx_desc, lower_pads, upper_pads);
ostream_ConstantTensorDescriptor(in_nchw_desc, std::cout << "in_nchw_desc: ");
ostream_ConstantTensorDescriptor(wei_kcsr_desc, std::cout << "wei_kcsr_desc: ");
ostream_ConstantTensorDescriptor(wei_kcyx_desc, std::cout << "wei_kcyx_desc: ");
ostream_ConstantTensorDescriptor(out_nkhw_desc, std::cout << "out_nkhw_desc: ");
Tensor<half> in_nchw(make_TensorDescriptor(in_nchw_desc));
Tensor<half> wei_kcsr(make_TensorDescriptor(wei_kcsr_desc));
Tensor<half> wei_kcyx(make_TensorDescriptor(wei_kcyx_desc));
Tensor<half> out_nkhw_host(make_TensorDescriptor(out_nkhw_desc));
Tensor<half> out_nkhw_device(make_TensorDescriptor(out_nkhw_desc));
@@ -597,13 +597,13 @@ int main(int argc, char* argv[])
{
#if 0
in_nchw.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei_kcsr.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
#elif 1
in_nchw.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei_kcsr.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
#elif 1
in_nchw.GenerateTensorValue(GeneratorTensor_2{-2, 2}, num_thread);
wei_kcsr.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
#endif
}
@@ -613,17 +613,17 @@ int main(int argc, char* argv[])
#elif 0
device_direct_convolution_2
#elif 1
device_implicit_gemm_convolution_1_chwn_csrk_khwn
device_implicit_gemm_convolution_1_chwn_cyxk_khwn
#elif 0
device_implicit_gemm_convolution_2_chwn_csrk_khwn
device_implicit_gemm_convolution_2_chwn_cyxk_khwn
#endif
(in_nchw_desc, in_nchw, wei_kcsr_desc, wei_kcsr, out_nkhw_desc, out_nkhw_device, nrepeat);
(in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 1
device_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(in_nchw_desc,
device_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded(in_nchw_desc,
in_nchw,
wei_kcsr_desc,
wei_kcsr,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
lower_pads,
@@ -636,18 +636,18 @@ int main(int argc, char* argv[])
#if 0
if(Y == 3 && X == 3)
{
host_winograd_3x3_convolution(in_nchw, wei_kcsr, out_nkhw_host, lower_pads, upper_pads);
host_winograd_3x3_convolution(in_nchw, wei_kcyx, out_nkhw_host, lower_pads, upper_pads);
}
else
{
host_direct_convolution(in_nchw, wei_kcsr, out_nkhw_host, lower_pads, upper_pads);
host_direct_convolution(in_nchw, wei_kcyx, out_nkhw_host, lower_pads, upper_pads);
}
check_error(out_nkhw_host, out_nkhw_device);
#endif
#if 0
LogRange(std::cout << "in_nchw : ", in_nchw.mData, ",") << std::endl;
LogRange(std::cout << "wei_kcsr: ", wei_kcsr.mData, ",") << std::endl;
LogRange(std::cout << "wei_kcyx: ", wei_kcyx.mData, ",") << std::endl;
LogRange(std::cout << "out_nkhw_host : ", out_nkhw_host.mData, ",") << std::endl;
LogRange(std::cout << "out_nkhw_device: ", out_nkhw_device.mData, ",") << std::endl;
#endif

26
src/include/gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn.hip.hpp → src/include/gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn.hip.hpp
View File

@@ -35,7 +35,7 @@ template <unsigned GridSize,
unsigned GemmKPerThreadLoop,
unsigned OutThreadCopyDataPerWrite>
__global__ void
gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn(const Float* const __restrict__ p_in_global,
gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global)
{
@@ -52,7 +52,7 @@ gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn(const Float* const __restric
constexpr auto I3 = Number<3>{};
constexpr auto in_chwn_global_desc = InGlobalDesc{};
constexpr auto wei_csrk_global_desc = WeiGlobalDesc{};
constexpr auto wei_cyxk_global_desc = WeiGlobalDesc{};
constexpr auto out_khwn_global_desc = OutGlobalDesc{};
constexpr unsigned C = in_chwn_global_desc.GetLength(I0);
@@ -62,8 +62,8 @@ gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn(const Float* const __restric
constexpr unsigned Wo = out_khwn_global_desc.GetLength(I2);
constexpr unsigned N = out_khwn_global_desc.GetLength(I3);
constexpr unsigned Y = wei_csrk_global_desc.GetLength(I1);
constexpr unsigned X = wei_csrk_global_desc.GetLength(I2);
constexpr unsigned Y = wei_cyxk_global_desc.GetLength(I1);
constexpr unsigned X = wei_cyxk_global_desc.GetLength(I2);
constexpr unsigned HiPerBlock = HoPerBlock + Y - 1;
constexpr unsigned WiPerBlock = WoPerBlock + X - 1;
@@ -100,7 +100,7 @@ gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn(const Float* const __restric
constexpr auto wei_ek_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock * Y * X, KPerBlock>{}, Number<WeiBlockCopyDataPerRead>{});
constexpr auto wei_csrk_block_desc = make_ConstantTensorDescriptor_aligned(
constexpr auto wei_cyxk_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, Y, X, KPerBlock>{}, Number<WeiBlockCopyDataPerRead>{});
// tensor view of threadwise output in register
@@ -133,7 +133,7 @@ gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn(const Float* const __restric
// B_matrix[C,Wo*N] is a sub-matrix of in_block[C,Hi,Wi,N]
// C_matrix[K,Wo*N] is a sub-matrix of out_block[K,Ho,Wo,N]
constexpr auto a_cxk_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_csrk_block_desc.GetStride(I0)>{});
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_cyxk_block_desc.GetStride(I0)>{});
constexpr auto b_cxwn_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
@@ -168,7 +168,7 @@ gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn(const Float* const __restric
in_chwn_block_desc.GetElementSpace(Number<InBlockCopyDataPerRead>{});
constexpr unsigned wei_block_size =
wei_csrk_block_desc.GetElementSpace(Number<WeiBlockCopyDataPerRead>{});
wei_cyxk_block_desc.GetElementSpace(Number<WeiBlockCopyDataPerRead>{});
constexpr unsigned max_align = InBlockCopyDataPerRead > WeiBlockCopyDataPerRead
? InBlockCopyDataPerRead
@@ -188,11 +188,11 @@ gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn(const Float* const __restric
0, hi_block_data_begin, wi_block_data_begin, n_block_data_begin);
const Float* p_wei_global_block_begin =
p_wei_global + wei_csrk_global_desc.Get1dIndex(0, 0, 0, k_block_data_begin);
p_wei_global + wei_cyxk_global_desc.Get1dIndex(0, 0, 0, k_block_data_begin);
for(unsigned c_block_data_begin = 0; c_block_data_begin < C; c_block_data_begin += CPerBlock,
p_in_global_block_begin += CPerBlock * in_chwn_global_desc.GetStride(I0),
p_wei_global_block_begin += CPerBlock * wei_csrk_global_desc.GetStride(I0),
p_wei_global_block_begin += CPerBlock * wei_cyxk_global_desc.GetStride(I0),
__syncthreads())
{
// input: global mem to LDS
@@ -204,12 +204,12 @@ gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn(const Float* const __restric
__syncthreads();
// a series of batched GEMM
for(unsigned s = 0; s < Y; ++s)
for(unsigned y = 0; y < Y; ++y)
{
for(unsigned r = 0; r < X; ++r)
for(unsigned x = 0; x < X; ++x)
{
blockwise_batch_gemm.Run(p_wei_block + wei_csrk_block_desc.Get1dIndex(0, s, r, 0),
p_in_block + in_chwn_block_desc.Get1dIndex(0, s, r, 0),
blockwise_batch_gemm.Run(p_wei_block + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
p_in_block + in_chwn_block_desc.Get1dIndex(0, y, x, 0),
p_out_thread,
[](auto& acc, const auto&& v) { acc += v; });
}

30
src/include/gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded.hip.hpp → src/include/gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded.hip.hpp
View File

@@ -27,7 +27,7 @@ template <unsigned GridSize,
unsigned WoPerThread,
unsigned WeiBlockCopyThreadPerDim0,
unsigned WeiBlockCopyThreadPerDim1>
__global__ void gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(
__global__ void gridwise_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded(
const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global)
@@ -45,7 +45,7 @@ __global__ void gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(
constexpr auto I3 = Number<3>{};
constexpr auto in_chwn_global_desc = InGlobalDesc{};
constexpr auto wei_csrk_global_desc = WeiGlobalDesc{};
constexpr auto wei_cyxk_global_desc = WeiGlobalDesc{};
constexpr auto out_khwn_global_desc = OutGlobalDesc{};
constexpr unsigned C = in_chwn_global_desc.GetLength(I0);
@@ -55,8 +55,8 @@ __global__ void gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(
constexpr unsigned Wo = out_khwn_global_desc.GetLength(I2);
constexpr unsigned N = out_khwn_global_desc.GetLength(I3);
constexpr unsigned Y = wei_csrk_global_desc.GetLength(I1);
constexpr unsigned X = wei_csrk_global_desc.GetLength(I2);
constexpr unsigned Y = wei_cyxk_global_desc.GetLength(I1);
constexpr unsigned X = wei_cyxk_global_desc.GetLength(I2);
constexpr unsigned HPadLow = LowerPads{}.Get(I0);
constexpr unsigned WPadLow = LowerPads{}.Get(I1);
@@ -92,7 +92,7 @@ __global__ void gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(
constexpr auto in_chwn_block_desc =
make_ConstantTensorDescriptor(Sequence<CPerBlock, HiPerBlock, WiPerBlock, NPerBlock>{});
constexpr auto wei_csrk_block_desc =
constexpr auto wei_cyxk_block_desc =
make_ConstantTensorDescriptor(Sequence<CPerBlock, Y, X, KPerBlock>{});
// flattened (2d) tensor view of wei in LDS
@@ -107,7 +107,7 @@ __global__ void gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_chwn_block_desc, "in_chwn_block_desc");
print_ConstantTensorDescriptor(wei_csrk_block_desc, "wei_csrk_block_desc");
print_ConstantTensorDescriptor(wei_cyxk_block_desc, "wei_cyxk_block_desc");
print_ConstantTensorDescriptor(out_hkwn_thread_desc, "out_hkwn_thread_desc");
}
#endif
@@ -148,9 +148,9 @@ __global__ void gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(
constexpr auto blockwise_wei_copy =
Blockwise4dTensorCopy1<BlockSize,
Float,
decltype(wei_csrk_global_desc),
decltype(wei_csrk_block_desc),
decltype(wei_csrk_block_desc.GetLengths())>{};
decltype(wei_cyxk_global_desc),
decltype(wei_cyxk_block_desc),
decltype(wei_cyxk_block_desc.GetLengths())>{};
#elif 0
// weight: format is [C*Y*X,K]
constexpr auto blockwise_wei_copy =
@@ -177,7 +177,7 @@ __global__ void gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(
// B_matrix[C,Wo*N] is a sub-matrix of in_block[C,Hi,Wi,N]
// C_matrix[K,Wo*N] is a sub-matrix of out_block[Ho,K,Wo,N]
constexpr auto a_cxk_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_csrk_block_desc.GetStride(I0)>{});
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_cyxk_block_desc.GetStride(I0)>{});
constexpr auto b_cxwn_block_mtx_desc =
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
@@ -205,7 +205,7 @@ __global__ void gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(
// LDS
constexpr unsigned in_block_size = in_chwn_block_desc.GetElementSpace();
constexpr unsigned wei_block_size = wei_csrk_block_desc.GetElementSpace();
constexpr unsigned wei_block_size = wei_cyxk_block_desc.GetElementSpace();
__shared__ Float p_in_block[in_block_size];
__shared__ Float p_wei_block[wei_block_size];
@@ -245,14 +245,14 @@ __global__ void gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded(
__syncthreads();
// a series of batched GEMM
for(unsigned s = 0; s < Y; ++s)
for(unsigned y = 0; y < Y; ++y)
{
for(unsigned r = 0; r < X; ++r)
for(unsigned x = 0; x < X; ++x)
{
auto f_accum = [](auto& acc, const auto&& v) { acc += v; };
blockwise_batch_gemm.Run(p_wei_block + wei_csrk_block_desc.Get1dIndex(0, s, r, 0),
p_in_block + in_chwn_block_desc.Get1dIndex(0, s, r, 0),
blockwise_batch_gemm.Run(p_wei_block + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
p_in_block + in_chwn_block_desc.Get1dIndex(0, y, x, 0),
p_out_thread,
f_accum);
}

40
src/include/gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_buffer.hip.hpp → src/include/gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer.hip.hpp
View File

@@ -34,7 +34,7 @@ template <unsigned GridSize,
unsigned WeiBlockCopyThreadPerDim1,
unsigned InBlockCopyDataPerRead,
unsigned WeiBlockCopyDataPerRead>
__global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_buffer(
__global__ void gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer(
const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global)
@@ -45,7 +45,7 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
constexpr auto I3 = Number<3>{};
constexpr auto in_chwn_global_desc = InGlobalDesc{};
constexpr auto wei_csrk_global_desc = WeiGlobalDesc{};
constexpr auto wei_cyxk_global_desc = WeiGlobalDesc{};
constexpr auto out_khwn_global_desc = OutGlobalDesc{};
constexpr unsigned C = in_chwn_global_desc.GetLength(I0);
@@ -57,8 +57,8 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
constexpr unsigned Ho = out_khwn_global_desc.GetLength(I1);
constexpr unsigned Wo = out_khwn_global_desc.GetLength(I2);
constexpr unsigned Y = wei_csrk_global_desc.GetLength(I1);
constexpr unsigned X = wei_csrk_global_desc.GetLength(I2);
constexpr unsigned Y = wei_cyxk_global_desc.GetLength(I1);
constexpr unsigned X = wei_cyxk_global_desc.GetLength(I2);
constexpr unsigned B = N * Hi * Wi;
constexpr unsigned BGhostRead = (Y - 1) * Wi + (X - 1);
@@ -85,7 +85,7 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
constexpr auto wei_ek_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock * Y * X, KPerBlock>{}, Number<WeiBlockCopyDataPerRead>{});
constexpr auto wei_csrk_block_desc = make_ConstantTensorDescriptor_aligned(
constexpr auto wei_cyxk_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<CPerBlock, Y, X, KPerBlock>{}, Number<WeiBlockCopyDataPerRead>{});
// tensor view of threadwise output in register
@@ -96,14 +96,14 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_chwn_global_desc, "in_chwn_global_desc");
print_ConstantTensorDescriptor(wei_csrk_global_desc, "wei_csrk_global_desc");
print_ConstantTensorDescriptor(wei_cyxk_global_desc, "wei_cyxk_global_desc");
print_ConstantTensorDescriptor(out_khwn_global_desc, "out_khwn_global_desc");
print_ConstantTensorDescriptor(in_cb_global_desc, "in_cb_global_desc");
print_ConstantTensorDescriptor(wei_ek_global_desc, "wei_ek_global_desc");
print_ConstantTensorDescriptor(in_cb_block_desc, "in_cb_block_desc");
print_ConstantTensorDescriptor(wei_csrk_block_desc, "wei_csrk_block_desc");
print_ConstantTensorDescriptor(wei_cyxk_block_desc, "wei_cyxk_block_desc");
print_ConstantTensorDescriptor(wei_ek_block_desc, "wei_ek_block_desc");
print_ConstantTensorDescriptor(out_kb_thread_desc, "out_kb_thread_desc");
@@ -170,7 +170,7 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
// b_mtx[C,B] is a subset of in_block[C,B + BGhostRead]
// c_mtx[K,B] is out_block[K,B]
constexpr auto a_cxk_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_csrk_block_desc.GetStride(I0)>{});
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_cyxk_block_desc.GetStride(I0)>{});
constexpr auto b_cxb_block_mtx_desc = make_ConstantMatrixDescriptor(
Number<CPerBlock>{}, Number<BPerBlock>{}, Number<in_cb_block_desc.GetStride(I0)>{});
@@ -210,7 +210,7 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
in_cb_block_desc.GetElementSpace(Number<InBlockCopyDataPerRead>{});
constexpr unsigned wei_block_size =
wei_csrk_block_desc.GetElementSpace(Number<WeiBlockCopyDataPerRead>{});
wei_cyxk_block_desc.GetElementSpace(Number<WeiBlockCopyDataPerRead>{});
constexpr unsigned max_align = InBlockCopyDataPerRead > WeiBlockCopyDataPerRead
? InBlockCopyDataPerRead
@@ -227,14 +227,14 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
p_in_global + in_cb_global_desc.Get1dIndex(0, b_block_data_begin);
const Float* p_wei_global_block_offset =
p_wei_global + wei_csrk_global_desc.Get1dIndex(0, 0, 0, k_block_data_begin);
p_wei_global + wei_cyxk_global_desc.Get1dIndex(0, 0, 0, k_block_data_begin);
// preload data into LDS
blockwise_in_copy.Run(p_in_global_block_offset, p_in_block_0);
blockwise_wei_copy.Run(p_wei_global_block_offset, p_wei_block_0);
p_in_global_block_offset += CPerBlock * in_cb_global_desc.GetStride(I0);
p_wei_global_block_offset += CPerBlock * wei_csrk_global_desc.GetStride(I0);
p_wei_global_block_offset += CPerBlock * wei_cyxk_global_desc.GetStride(I0);
// register
Float p_out_thread[out_kb_thread_desc.GetElementSpace()];
@@ -247,7 +247,7 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
for(unsigned c_block_data_begin = 0; c_block_data_begin + CPerBlock < C;
c_block_data_begin += CPerBlock,
p_in_global_block_offset += CPerBlock * in_cb_global_desc.GetStride(I0),
p_wei_global_block_offset += CPerBlock * wei_csrk_global_desc.GetStride(I0),
p_wei_global_block_offset += CPerBlock * wei_cyxk_global_desc.GetStride(I0),
even_loop = !even_loop)
{
Float* p_in_block_now = even_loop ? p_in_block_0 : p_in_block_1;
@@ -275,9 +275,9 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
// compute on current data
// a series of GEMM
for(unsigned s = 0; s < Y; ++s)
for(unsigned y = 0; y < Y; ++y)
{
for(unsigned r = 0; r < X; ++r)
for(unsigned x = 0; x < X; ++x)
{
auto f_accum = [](auto& acc, const auto&& v) { acc += v; };
#if 1
@@ -285,8 +285,8 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
#else
blockwise_gemm.Run_RegisterDoubleBuffer
#endif
(p_wei_block_now + wei_csrk_block_desc.Get1dIndex(0, s, r, 0),
p_in_block_now + s * Wi + r,
(p_wei_block_now + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
p_in_block_now + y * Wi + x,
p_out_thread,
f_accum);
}
@@ -305,9 +305,9 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
__syncthreads();
for(unsigned s = 0; s < Y; ++s)
for(unsigned y = 0; y < Y; ++y)
{
for(unsigned r = 0; r < X; ++r)
for(unsigned x = 0; x < X; ++x)
{
auto f_accum = [](auto& acc, const auto&& v) { acc += v; };
#if 0
@@ -315,8 +315,8 @@ __global__ void gridwise_implicit_gemm_convolution_2_chwn_csrk_khwn_lds_double_b
#else
blockwise_gemm.Run_RegisterDoubleBuffer
#endif
(p_wei_block_now + wei_csrk_block_desc.Get1dIndex(0, s, r, 0),
p_in_block_now + s * Wi + r,
(p_wei_block_now + wei_cyxk_block_desc.Get1dIndex(0, y, x, 0),
p_in_block_now + y * Wi + x,
p_out_thread,
f_accum);
}

6
src/include/tensor.hpp
View File

@@ -8,16 +8,16 @@
#include <iostream>
template <class Range>
std::ostream& LogRange(std::ostream& os, Range&& r, std::string delim)
std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim)
{
bool first = true;
for(auto&& x : r)
for(auto&& v : range)
{
if(first)
first = false;
else
os << delim;
os << x;
os << v;
}
return os;
}

30
src/include/threadwise_direct_convolution.hip.hpp
View File

@@ -38,16 +38,16 @@ __device__ void threadwise_direct_convolution_1(InDesc,
{
for(unsigned c = 0; c < wei_desc.GetLength(I1); ++c)
{
for(unsigned s = 0; s < wei_desc.GetLength(I2); ++s)
for(unsigned y = 0; y < wei_desc.GetLength(I2); ++y)
{
for(unsigned r = 0; r < wei_desc.GetLength(I3); ++r)
for(unsigned x = 0; x < wei_desc.GetLength(I3); ++x)
{
const unsigned hi = ho + s;
const unsigned wi = wo + r;
const unsigned hi = ho + y;
const unsigned wi = wo + x;
const unsigned in_index = in_desc.Get1dIndex(n, c, hi, wi);
const unsigned wei_index = wei_desc.Get1dIndex(k, c, s, r);
const unsigned wei_index = wei_desc.Get1dIndex(k, c, y, x);
const unsigned out_index = out_desc.Get1dIndex(n, k, ho, wo);
@@ -153,18 +153,18 @@ __device__ void threadwise_direct_convolution_3(InDesc,
#if 0
// this verison reused old input data in register, and read new data from LDS
// loop over vertical direction
for(unsigned s = 0; s < wei_desc.GetLength(I2); ++s)
for(unsigned y = 0; y < wei_desc.GetLength(I2); ++y)
{
// read first input
threadwise_4d_tensor_copy(in_desc,
p_in + in_desc.Get1dIndex(0, 0, s, 0),
p_in + in_desc.Get1dIndex(0, 0, y, 0),
in_reg_desc,
p_in_reg,
in_reg_desc.GetLengths());
// read first 1x1 weight
threadwise_4d_tensor_copy(wei_desc,
p_wei + wei_desc.Get1dIndex(0, 0, s, 0),
p_wei + wei_desc.Get1dIndex(0, 0, y, 0),
wei_reg_desc,
p_wei_reg,
wei_reg_desc.GetLengths());
@@ -174,11 +174,11 @@ __device__ void threadwise_direct_convolution_3(InDesc,
in_reg_desc, p_in_reg, wei_reg_desc, p_wei_reg, out_desc, p_out);
// loop over horizontal direction
for(unsigned r = 1; r < wei_desc.GetLength(I3); ++r)
for(unsigned x = 1; x < wei_desc.GetLength(I3); ++x)
{
// read new weight
threadwise_4d_tensor_copy(wei_desc,
p_wei + wei_desc.Get1dIndex(0, 0, s, r),
p_wei + wei_desc.Get1dIndex(0, 0, y, x),
wei_reg_desc,
p_wei_reg,
wei_reg_desc.GetLengths());
@@ -189,7 +189,7 @@ __device__ void threadwise_direct_convolution_3(InDesc,
// read new input
threadwise_4d_tensor_copy(
in_desc,
p_in + in_desc.Get1dIndex(0, 0, s, r + in_reg_desc.GetLength(I3) - 1),
p_in + in_desc.Get1dIndex(0, 0, y, x + in_reg_desc.GetLength(I3) - 1),
in_reg_desc,
p_in_reg +
in_reg_desc.Get1dIndex(0, 0, 0, in_reg_desc.GetLength(I3) - in_w_new_read),
@@ -203,21 +203,21 @@ __device__ void threadwise_direct_convolution_3(InDesc,
#elif 1
// this version read all input from LDS when filter moves
// loop over vertical direction
for(unsigned s = 0; s < wei_desc.GetLength(I2); ++s)
for(unsigned y = 0; y < wei_desc.GetLength(I2); ++y)
{
// loop over horizontal direction
for(unsigned r = 0; r < wei_desc.GetLength(I3); ++r)
for(unsigned x = 0; x < wei_desc.GetLength(I3); ++x)
{
// read new weight
threadwise_4d_tensor_copy(wei_desc,
p_wei + wei_desc.Get1dIndex(0, 0, s, r),
p_wei + wei_desc.Get1dIndex(0, 0, y, x),
wei_reg_desc,
p_wei_reg,
wei_reg_desc.GetLengths());
// read new input
threadwise_4d_tensor_copy(in_desc,
p_in + in_desc.Get1dIndex(0, 0, s, r),
p_in + in_desc.Get1dIndex(0, 0, y, x),
in_reg_desc,
p_in_reg,
in_reg_desc.GetLengths());