ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 09:16:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

adding implicit gemm

Browse Source
This commit is contained in:
Chao Liu
2019-01-16 02:44:10 -06:00
parent 913afaeb5d
commit 89ee259752
6 changed files with 44 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/include/blockwise_tensor_op.cuh
View File

@@ -135,7 +135,7 @@ __device__ void blockwise_4d_tensor_pointwise_operation_binary_reorder_by_get_ds
const unsigned bindex = dst_desc.Get1dIndex(did[IR0], did[IR1], did[IR2], did[IR3]);
#if 1
#if 0
printf("did %u %u %u %u, did_IR %u %u %u %u, index %u %u\n",
did[0],
did[1],

3
src/include/gemm.cuh
View File

@@ -145,9 +145,6 @@ struct blockwise_1d_strided_batched_gemm_block_a_block_b_thread_c
static_assert(BlockSize == BThreadWork * MThreadWork * NThreadWork,
"wrong! wrong BlockSize");
// printf("%u %u, %u %u\n", get_block_1d_id(), get_thread_local_1d_id(), MThreadWork,
// NThreadWork);
if(DistributeThreadAlongColumnFirst)
{
// num of operations can be reduced

35
src/include/gridwise_implicit_gemm_convolution_nchw_kcsr.cuh
View File

@@ -82,25 +82,24 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc,
make_ConstantTensorDescriptor(Sequence<KPerBlock, CPerBlock, S, R>{});
// tensor view of reordered blockwise input and weight in LDS
constexpr auto reorder_chwn_from_nchw = Sequence<1, 2, 3, 0>{};
constexpr auto in_chwn_block_desc = make_ConstantTensorDescriptor(
in_nchw_block_desc.GetLengths().ReorderByGetNewFromOld(reorder_chwn_from_nchw));
constexpr auto reorder_srck_from_kcsr = Sequence<2, 3, 1, 0>{};
constexpr auto wei_srck_block_desc = make_ConstantTensorDescriptor(
wei_kcsr_block_desc.GetLengths().ReorderByGetNewFromOld(reorder_srck_from_kcsr));
constexpr auto reorder_chwn_from_nchw = Sequence<1, 2, 3, 0>{};
constexpr auto in_chwn_block_desc = make_ConstantTensorDescriptor(
in_nchw_block_desc.GetLengths().ReorderByGetNewFromOld(reorder_chwn_from_nchw));
// tensor view of threadwise output in register
constexpr auto out_hkwn_thread_desc =
make_ConstantTensorDescriptor(Sequence<HoPerThread, KPerThread, WoPerThread, NPerThread>{});
#if 0
#if 1
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_nchw_block_desc, "in_nchw_block_desc");
print_ConstantTensorDescriptor(in_chwn_block_desc, "in_chwn_block_desc");
print_ConstantTensorDescriptor(wei_kcsr_block_desc, "wei_kcsr_block_desc");
print_ConstantTensorDescriptor(wei_srck_block_desc, "wei_srck_block_desc");
print_ConstantTensorDescriptor(out_hkwn_thread_desc, "out_hkwn_thread_desc");
@@ -124,8 +123,6 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc,
const auto c_kxwn_thread_mtx_desc = make_ConstantMatrixDescriptor(
Number<KPerThread>{}, Number<WoPerThread * NPerThread>{}); // constexpr doesn't compile
auto f_accum = [](auto& c, auto& ab) { c += ab; };
const auto blockwise_batch_gemm =
blockwise_1d_strided_batched_gemm_block_a_block_b_thread_c<BlockSize,
decltype(a_cxk_block_mtx_desc),
@@ -137,11 +134,11 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc,
0,
in_chwn_block_desc.GetStride(I1),
out_hkwn_thread_desc.GetStride(
I1),
I0),
HoPerBlock,
HoPerThread,
CPerThread,
decltype(f_accum)>{};
true>{};
// LDS
constexpr unsigned in_block_size = in_chwn_block_desc.GetElementSpace();
@@ -176,7 +173,6 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc,
#if 1
// weight: global mem to LDS,
// convert 4d-tensor wei[K,C,S,R] to matrix wei_matrix[S*R*C,K]
blockwise_4d_tensor_copy_reorder_by_get_dst_from_src<BlockSize>(
wei_kcsr_global_desc,
p_wei_global +
@@ -189,24 +185,29 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc,
__syncthreads();
#if 1
// a series of batched GEMM
for(unsigned s = 0; s < S; ++s)
{
for(unsigned r = 0; r < R; ++r)
{
auto f_accum = [](auto& c, const auto&& ab) { c += ab; };
blockwise_batch_gemm.run(p_wei_block + wei_srck_block_desc.Get1dIndex(s, r, 0, 0),
p_in_block + in_chwn_block_desc.Get1dIndex(0, 0, r, 0),
p_out_thread);
p_out_thread,
f_accum);
}
}
#endif
}
const auto matrix_c_index =
blockwise_batch_gemm.CalculateThreadMatrixCIndex(get_thread_local_1d_id());
const unsigned ho_thread_data_begin = matrix_c_index.batch_begin;
const unsigned k_thread_data_begin = matrix_c_index.col_begin;
const unsigned wo_thread_data_begin = matrix_c_index.row_begin / NPerThread;
const unsigned k_thread_data_begin = matrix_c_index.row_begin;
const unsigned wo_thread_data_begin = matrix_c_index.col_begin / NPerThread;
// output: register to global mem,
// convert out_thread[Ho,K,Wo,N] to out_global[N,K,Ho,Wo]
@@ -222,4 +223,10 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_kcsr(InGlobalDesc,
wo_block_data_begin + wo_thread_data_begin),
out_hkwn_thread_desc.GetLengths(),
reorder_nkhw_from_hkwn);
// printf("%f %f %f %f\n", p_out_thread[0], p_out_thread[1], p_out_thread[2], p_out_thread[3]);
// printf("%u %u, %u %u %u\n", get_block_1d_id(), get_thread_local_1d_id(),
// matrix_c_index.batch_begin, matrix_c_index.row_begin, matrix_c_index.col_begin); printf("%u
// %u, %u %u %u\n", get_block_1d_id(), get_thread_local_1d_id(), ho_thread_data_begin,
// k_thread_data_begin, wo_thread_data_begin);
}

4
src/include/gridwise_implicit_gemm_convolution_nchw_srck.cuh
View File

@@ -90,7 +90,7 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_srck(InGlobalDesc,
constexpr auto out_hkwn_thread_desc =
make_ConstantTensorDescriptor(Sequence<HoPerThread, KPerThread, WoPerThread, NPerThread>{});
#if 1
#if 0
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
{
print_ConstantTensorDescriptor(in_nchw_block_desc, "in_nchw_block_desc");
@@ -189,7 +189,7 @@ __global__ void gridwise_implicit_gemm_convolution_nchw_srck(InGlobalDesc,
auto f_accum = [](auto& c, const auto&& ab) { c += ab; };
blockwise_batch_gemm.run(p_wei_block + wei_srck_block_desc.Get1dIndex(s, r, 0, 0),
p_in_block + in_chwn_block_desc.Get1dIndex(0, 0, r, 0),
p_in_block + in_chwn_block_desc.Get1dIndex(0, s, r, 0),
p_out_thread,
f_accum);
}