ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-24 06:44:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

adding implicit gemm

Browse Source
This commit is contained in:
Chao Liu
2019-01-16 02:01:56 -06:00
parent e7b8705b91
commit 913afaeb5d
6 changed files with 325 additions and 136 deletions
Download Patch File Download Diff File Expand all files Collapse all files

92
driver/conv.cu
View File

@@ -16,21 +16,7 @@ struct GeneratorTensor_1
template <class... Is>
double operator()(Is... is)
{
#if 0
return double(std::rand()) / double(RAND_MAX);
#elif 1
return 1;
#elif 0
std::initializer_list<std::size_t> ls = {static_cast<std::size_t>(is)...};
return std::accumulate(ls.begin(), ls.end(), std::size_t(0));
#else
assert(sizeof...(Is) > 0);
std::initializer_list<std::size_t> ids = {static_cast<std::size_t>(is)...};
std::vector<std::size_t> lens(sizeof...(Is), 100);
std::vector<std::size_t> strides(sizeof...(Is), 1);
std::partial_sum(lens.rbegin(), lens.rbegin() + (sizeof...(Is) - 1), strides.rbegin() + 1);
return std::inner_product(ids.begin(), ids.end(), strides.begin(), std::size_t(0)) + 1;
#endif
}
};
@@ -46,6 +32,25 @@ struct GeneratorTensor_2
}
};
struct GeneratorTensor_3
{
template <class... Is>
double operator()(Is... is)
{
#if 0
std::initializer_list<std::size_t> ls = {static_cast<std::size_t>(is)...};
return std::accumulate(ls.begin(), ls.end(), std::size_t(0));
#elif 1
assert(sizeof...(Is) > 0);
std::initializer_list<std::size_t> ids = {static_cast<std::size_t>(is)...};
std::vector<std::size_t> lens(sizeof...(Is), 100);
std::vector<std::size_t> strides(sizeof...(Is), 1);
std::partial_sum(lens.rbegin(), lens.rbegin() + (sizeof...(Is) - 1), strides.rbegin() + 1);
return std::inner_product(ids.begin(), ids.end(), strides.begin(), std::size_t(0)) + 1;
#endif
}
};
// this is ugly, only for 4d
template <class TConstTensorDesc>
void ostream_ConstantTensorDescriptor(TConstTensorDesc, std::ostream& os = std::cout)
@@ -338,7 +343,7 @@ int main()
constexpr unsigned K = 1;
constexpr unsigned S = 3;
constexpr unsigned R = 3;
#elif 0
#elif 1
constexpr unsigned N = 1;
constexpr unsigned C = 1;
constexpr unsigned HI = 34;
@@ -347,21 +352,21 @@ int main()
constexpr unsigned S = 3;
constexpr unsigned R = 3;
#elif 1
constexpr unsigned N = 64;
constexpr unsigned C = 256;
constexpr unsigned N = 64;
constexpr unsigned C = 256;
constexpr unsigned HI = 34;
constexpr unsigned WI = 34;
constexpr unsigned K = 64;
constexpr unsigned S = 3;
constexpr unsigned R = 3;
constexpr unsigned K = 64;
constexpr unsigned S = 3;
constexpr unsigned R = 3;
#elif 0
constexpr unsigned N = 64;
constexpr unsigned C = 64;
constexpr unsigned N = 64;
constexpr unsigned C = 64;
constexpr unsigned HI = 56;
constexpr unsigned WI = 56;
constexpr unsigned K = 64;
constexpr unsigned S = 3;
constexpr unsigned R = 3;
constexpr unsigned K = 64;
constexpr unsigned S = 3;
constexpr unsigned R = 3;
#elif 0
constexpr unsigned N = 64;
constexpr unsigned C = 64;
@@ -374,34 +379,51 @@ int main()
auto in_nchw_desc = make_ConstantTensorDescriptor(Sequence<N, C, HI, WI>{});
auto wei_kcsr_desc = make_ConstantTensorDescriptor(Sequence<K, C, S, R>{});
auto wei_srck_desc = make_ConstantTensorDescriptor(Sequence<S, R, C, K>{});
auto out_nkhw_desc =
get_convolution_output_default_4d_tensor_descriptor(in_nchw_desc, wei_kcsr_desc);
ostream_ConstantTensorDescriptor(in_nchw_desc, std::cout << "in_nchw_desc: ");
ostream_ConstantTensorDescriptor(wei_kcsr_desc, std::cout << "wei_kcsr_desc: ");
ostream_ConstantTensorDescriptor(wei_srck_desc, std::cout << "wei_srck_desc: ");
ostream_ConstantTensorDescriptor(out_nkhw_desc, std::cout << "out_nkhw_desc: ");
Tensor<float> in_nchw(make_TensorDescriptor(in_nchw_desc));
Tensor<float> wei_kcsr(make_TensorDescriptor(wei_kcsr_desc));
Tensor<float> wei_srck(make_TensorDescriptor(wei_srck_desc));
Tensor<float> out_nkhw_host(make_TensorDescriptor(out_nkhw_desc));
Tensor<float> out_nkhw_device(make_TensorDescriptor(out_nkhw_desc));
#if 0
std::size_t num_thread = std::thread::hardware_concurrency();
#if 0
in_nchw.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei_kcsr.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei_srck.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
#elif 1
std::size_t num_thread = std::thread::hardware_concurrency();
#elif 0
in_nchw.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei_kcsr.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei_srck.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
#elif 0
in_nchw.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei_kcsr.GenerateTensorValue(GeneratorTensor_3{}, num_thread);
#elif 1
in_nchw.GenerateTensorValue(GeneratorTensor_3{}, num_thread);
wei_kcsr.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
#endif
for(int i = 0; i < 40; ++i)
#if 1
auto wei_srck_desc = make_ConstantTensorDescriptor(Sequence<S, R, C, K>{});
Tensor<float> wei_srck(make_TensorDescriptor(wei_srck_desc));
auto f_reorder_kcsr2srck = [&](auto k, auto c, auto s, auto r) {
wei_srck(s, r, c, k) = wei_kcsr(k, c, s, r);
};
make_ParallelTensorFunctor(f_reorder_kcsr2srck, K, C, S, R)(num_thread);
#endif
#if 0
wei_srck.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
out_nkhw_device.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
#endif
for(int i = 0; i < 1; ++i)
{
#if 0
device_direct_convolution_1(in_nchw_desc, in_nchw, wei_kcsr_desc, wei_kcsr, out_nkhw_desc, out_nkhw_device);
@@ -428,7 +450,7 @@ int main()
check_error(out_nkhw_host, out_nkhw_device);
#endif
#if 0
#if 1
LogRange(std::cout << "in_nchw : ", in_nchw.mData, ",") << std::endl;
LogRange(std::cout << "wei_kcsr: ", wei_kcsr.mData, ",") << std::endl;
LogRange(std::cout << "out_nkhw_host : ", out_nkhw_host.mData, ",") << std::endl;

24
driver/device_implicit_gemm_convolution.cuh
View File

@@ -1,5 +1,5 @@
#pragma once
#include "gridwise_implicit_gemm_convolution_nchw_kcsr.cuh"
//#include "gridwise_implicit_gemm_convolution_nchw_kcsr.cuh"
#include "gridwise_implicit_gemm_convolution_nchw_srck.cuh"
template <class T, class InDesc, class WeiDesc, class OutDesc>
@@ -26,20 +26,20 @@ void device_implicit_gemm_convolution(
constexpr auto wei_desc = WeiDesc{};
constexpr auto out_desc = OutDesc{};
#if 0
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 64;
constexpr unsigned CPerBlock = 4;
#if 1
constexpr unsigned NPerBlock = 1;
constexpr unsigned KPerBlock = 1;
constexpr unsigned CPerBlock = 1;
constexpr unsigned HoPerBlock = 2;
constexpr unsigned WoPerBlock = 32;
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 8;
constexpr unsigned CPerThread = 2;
constexpr unsigned HoPerThread = 1;
constexpr unsigned WoPerThread = 4;
constexpr unsigned NPerThread = 1;
constexpr unsigned KPerThread = 1;
constexpr unsigned CPerThread = 1;
constexpr unsigned HoPerThread = 2;
constexpr unsigned WoPerThread = 2;
constexpr unsigned BlockSize = 256;
constexpr unsigned BlockSize = 16;
#elif 1
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 32;
@@ -50,7 +50,7 @@ void device_implicit_gemm_convolution(
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 4;
constexpr unsigned CPerThread = 2;
constexpr unsigned HoPerThread = 1;
constexpr unsigned HoPerThread = 2;
constexpr unsigned WoPerThread = 2;
constexpr unsigned BlockSize = 128;