mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-11 08:50:17 +00:00
added implicit GEMM v3
This commit is contained in:
Chao Liu
@@ -57,7 +57,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(InDesc,
|
||||
wei_cyxk_device_buf.ToDevice(wei_cyxk.mData.data());
|
||||
out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
|
||||
|
||||
#if 1
|
||||
#if 0
|
||||
// for 3x3, 34x34, v1r3, Pascal
|
||||
constexpr index_t BlockSize = 128;
|
||||
|
||||
@@ -127,7 +127,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(InDesc,
|
||||
constexpr index_t WeiBlockCopyDataPerRead_K = 4;
|
||||
|
||||
constexpr index_t OutThreadCopyDataPerWrite_W = 4;
|
||||
#elif 0
|
||||
#elif 1
|
||||
// for 3x3, 34x34, v1r3, Vega 20, WoPerBlock = 16
|
||||
constexpr index_t BlockSize = 256;
|
||||
|
||||
@@ -313,7 +313,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(InDesc,
|
||||
for(index_t i = 0; i < nrepeat; ++i)
|
||||
{
|
||||
constexpr auto gridwise_conv =
|
||||
#if 0
|
||||
#if 1
|
||||
GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
#else
|
||||
GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_nkhw
|
||||
|
||||
@@ -62,7 +62,6 @@ void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
|
||||
constexpr index_t B = (N * Ho * Wo) / (N1 * N2);
|
||||
|
||||
#if 1
|
||||
// for 3x3, 28x28, v3
|
||||
constexpr index_t BlockSize = 256;
|
||||
|
||||
constexpr index_t BPerBlock = 16;
|
||||
|
||||
@@ -83,7 +83,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
constexpr index_t HBlockWork = mod_conv::integer_divide_ceil(Ho, HoPerBlock);
|
||||
constexpr index_t WBlockWork = mod_conv::integer_divide_ceil(Wo, WoPerBlock);
|
||||
|
||||
constexpr auto block_work_desc = make_ConstantTensorDescriptor(
|
||||
constexpr auto block_work_desc = make_ConstantTensorDescriptor_default_rank_packed(
|
||||
Sequence<NBlockWork, KBlockWork, HBlockWork, WBlockWork>{});
|
||||
|
||||
const auto block_work_multi_id =
|
||||
@@ -99,7 +99,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
|
||||
// global tensor view
|
||||
constexpr auto wei_c_k_global_desc =
|
||||
make_ConstantTensorDescriptor(Sequence<C, K>{}, Sequence<Y * X * K, 1>{});
|
||||
make_ConstantTensorDescriptor_default_rank(Sequence<C, K>{}, Sequence<Y * X * K, 1>{});
|
||||
|
||||
// LDS tensor view
|
||||
// be careful of alignment
|
||||
@@ -108,7 +108,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
GemmDataPerReadA,
|
||||
GemmDataPerReadB);
|
||||
|
||||
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_aligned(
|
||||
constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
|
||||
Sequence<CPerBlock, HoPerBlock, WoPerBlock, NPerBlock>{},
|
||||
Number<InBlockReorderDataPerWrite_N>{});
|
||||
|
||||
@@ -117,12 +117,12 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
|
||||
"GemmDataPerReadB alignment requirement is not meet");
|
||||
|
||||
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
|
||||
constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
|
||||
Sequence<CPerBlock, KPerBlock>{},
|
||||
Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
|
||||
|
||||
// tensor view of threadwise output in register
|
||||
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor(
|
||||
constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_default_rank_packed(
|
||||
Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
|
||||
|
||||
// blockwise copy
|
||||
@@ -140,7 +140,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
decltype(map_chwn2nchw),
|
||||
InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW,
|
||||
InBlockReorderDataPerRead_W,
|
||||
InBlockReorderDataPerWrite_N>{};
|
||||
InBlockReorderDataPerWrite_N>({0, 0, 0, 0}, {0, 0, 0, 0});
|
||||
|
||||
// blockwise wei copy
|
||||
// format is [CPerBlock, KPerBlock]
|
||||
@@ -150,7 +150,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
decltype(wei_c_k_global_desc),
|
||||
decltype(wei_c_k_block_desc),
|
||||
decltype(wei_c_k_block_desc.GetLengths()),
|
||||
WeiBlockCopyDataPerRead_K>{};
|
||||
WeiBlockCopyDataPerRead_K>({0, 0}, {0, 0});
|
||||
|
||||
// a series of blockwise batched GEMM
|
||||
// C_matrix += transpose(A_matrix) * B_matrix
|
||||
@@ -194,7 +194,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
|
||||
// choose GEMM implementation here
|
||||
const auto run_blockwise_batch_gemm = [&](auto... Xs) {
|
||||
#if 0
|
||||
#if 1
|
||||
return blockwise_batch_gemm.Run(Xs...);
|
||||
#elif 0
|
||||
return blockwise_batch_gemm.Run_asm(Xs...);
|
||||
@@ -249,7 +249,6 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
{
|
||||
for(index_t x = 0; x < X; ++x)
|
||||
{
|
||||
#if 1
|
||||
blockwise_in_copy_reorder.Run(p_in_global_block_offset +
|
||||
in_n_c_h_w_global_desc.GetOffsetFromMultiIndex(0, 0, y, x),
|
||||
p_in_block);
|
||||
@@ -257,23 +256,6 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
blockwise_wei_copy.Run(p_wei_global_block_offset +
|
||||
wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, x, 0),
|
||||
p_wei_block);
|
||||
#else
|
||||
Float p_in_clipboard[blockwise_in_copy_reorder.GetRegisterClipboardSize()];
|
||||
Float p_wei_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
|
||||
|
||||
blockwise_in_copy_reorder.RunLoadRegisterClipboard(
|
||||
p_in_global_block_offset + in_n_c_h_w_global_desc.GetOffsetFromMultiIndex(0, 0, y, x),
|
||||
p_in_clipboard);
|
||||
|
||||
blockwise_wei_copy.RunLoadRegisterClipboard(
|
||||
p_wei_global_block_offset + wei_c_y_x_k_global_desc.GetOffsetFromMultiIndex(0, y, x, 0),
|
||||
p_wei_clipboard);
|
||||
|
||||
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_clipboard, p_wei_block);
|
||||
|
||||
blockwise_in_copy_reorder.RunStoreRegisterClipboard(p_in_clipboard, p_in_block);
|
||||
|
||||
#endif
|
||||
|
||||
__syncthreads();
|
||||
|
||||
@@ -304,24 +286,9 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
p_wei_global_block_offset +=
|
||||
CPerBlock * wei_c_y_x_k_global_desc.GetStride(I0))
|
||||
{
|
||||
#if 0
|
||||
blockwise_in_copy_reorder.Run(p_in_global_block_offset,
|
||||
p_in_block);
|
||||
blockwise_in_copy_reorder.Run(p_in_global_block_offset, p_in_block);
|
||||
|
||||
blockwise_wei_copy.Run(p_wei_global_block_offset,
|
||||
p_wei_block);
|
||||
#else
|
||||
Float p_in_clipboard[blockwise_in_copy_reorder.GetRegisterClipboardSize()];
|
||||
Float p_wei_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
|
||||
|
||||
blockwise_in_copy_reorder.RunLoadRegisterClipboard(p_in_global_block_offset,
|
||||
p_in_clipboard);
|
||||
blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global_block_offset,
|
||||
p_wei_clipboard);
|
||||
|
||||
blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_clipboard, p_wei_block);
|
||||
blockwise_in_copy_reorder.RunStoreRegisterClipboard(p_in_clipboard, p_in_block);
|
||||
#endif
|
||||
blockwise_wei_copy.Run(p_wei_global_block_offset, p_wei_block);
|
||||
|
||||
__syncthreads();
|
||||
|
||||
@@ -342,13 +309,12 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
const index_t wo_thread_data_begin = c_thread_mtx_begin.col / NPerBlock;
|
||||
const index_t n_thread_data_begin = c_thread_mtx_begin.col % NPerBlock;
|
||||
|
||||
static_if<GemmNPerThreadSubC <= NPerBlock>{}([&](auto f_dummy) { // f_dummy do nothing but
|
||||
// perfect forwarding.
|
||||
// Using this trick to
|
||||
// make this lambda a generic lambda, so it won't be compiled until
|
||||
// instantiated
|
||||
static_if<GemmNPerThreadSubC <= NPerBlock>{}([&](auto fwd) {
|
||||
// fwd do nothing but perfect forwarding.
|
||||
// Using this trick to make this lambda a generic lambda, so it won't be compiled until
|
||||
// begin instantiated here
|
||||
static_assert(
|
||||
(f_dummy(GemmNPerThreadSubC) <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0),
|
||||
(fwd(GemmNPerThreadSubC) <= NPerBlock && NPerBlock % GemmNPerThreadSubC == 0),
|
||||
"wrong!");
|
||||
|
||||
// output is a 10d tensor
|
||||
@@ -356,38 +322,33 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
constexpr index_t N1 = NPerBlock / N2;
|
||||
|
||||
constexpr index_t W2 =
|
||||
(GemmNLevel0Cluster * GemmNLevel1Cluster) / f_dummy(NPerBlock / GemmNPerThreadSubC);
|
||||
(GemmNLevel0Cluster * GemmNLevel1Cluster) / fwd(NPerBlock / GemmNPerThreadSubC);
|
||||
constexpr index_t W1 = WoPerBlock / W2;
|
||||
|
||||
constexpr index_t K2 = GemmMPerThreadSubC;
|
||||
constexpr index_t K1 = KPerBlock / KPerThread;
|
||||
|
||||
constexpr auto out_10d_global_desc =
|
||||
make_ConstantTensorDescriptor(Sequence<N / f_dummy(N1 * N2),
|
||||
N1,
|
||||
N2,
|
||||
K / (K1 * K2),
|
||||
K1,
|
||||
K2,
|
||||
Ho,
|
||||
Wo / (W1 * W2),
|
||||
W1,
|
||||
W2>{});
|
||||
constexpr auto out_10d_global_desc = fwd(out_n_k_h_w_global_desc)
|
||||
.Fold(I3, Number<W1>{}, Number<W2>{})
|
||||
.Fold(I1, Number<K1>{}, Number<K2>{})
|
||||
.Fold(I0, Number<N1>{}, Number<N2>{});
|
||||
|
||||
constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor(
|
||||
Sequence<KPerThread / K2, 1, K2, HoPerThread, 1, W1, 1, 1, 1, N2>{});
|
||||
constexpr auto out_10d_thread_desc = fwd(out_k_h_w_n_thread_desc)
|
||||
.Fold(I3, Number<1>{}, Number<N2>{})
|
||||
.Fold(I2, Number<W1>{}, Number<1>{})
|
||||
.Fold(I0, Number<1>{}, Number<K2>{});
|
||||
|
||||
#if 0
|
||||
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
|
||||
{
|
||||
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
|
||||
"out_k_h_w_n_thread_desc");
|
||||
print_ConstantTensorDescriptor(out_10d_thread_desc, "out_10d_thread_desc");
|
||||
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
|
||||
{
|
||||
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
|
||||
"a: out_k_h_w_n_thread_desc");
|
||||
print_ConstantTensorDescriptor(out_10d_thread_desc, "a: out_10d_thread_desc");
|
||||
|
||||
print_ConstantTensorDescriptor(out_k_h_w_n_global_desc,
|
||||
"out_k_h_w_n_global_desc");
|
||||
print_ConstantTensorDescriptor(out_10d_global_desc, "out_10d_global_desc");
|
||||
}
|
||||
print_ConstantTensorDescriptor(out_n_k_h_w_global_desc,
|
||||
"a: out_n_k_h_w_global_desc");
|
||||
print_ConstantTensorDescriptor(out_10d_global_desc, "a: out_10d_global_desc");
|
||||
}
|
||||
#endif
|
||||
|
||||
constexpr auto map_out_global2thread = Sequence<7, 8, 9, 0, 1, 2, 3, 4, 5, 6>{};
|
||||
@@ -405,8 +366,8 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
out_10d_thread_desc.GetLengths(),
|
||||
map_out_global2thread);
|
||||
// Number<OutThreadCopyDataPerWrite_W>{});
|
||||
}).else_([&](auto f_dummy) {
|
||||
static_assert(f_dummy(GemmNPerThreadSubC) >= NPerBlock && NPerThread == NPerBlock &&
|
||||
}).else_([&](auto fwd) {
|
||||
static_assert(fwd(GemmNPerThreadSubC) >= NPerBlock && NPerThread == NPerBlock &&
|
||||
GemmNPerThreadSubC % NPerThread == 0,
|
||||
"wrong!");
|
||||
|
||||
@@ -415,34 +376,40 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
|
||||
constexpr index_t W3 = GemmNPerThreadSubC / NPerBlock;
|
||||
constexpr index_t W2 = GemmNLevel0Cluster * GemmNLevel1Cluster;
|
||||
constexpr index_t W1 = WoPerBlock / f_dummy(W2 * W3);
|
||||
constexpr index_t W1 = WoPerBlock / fwd(W2 * W3);
|
||||
|
||||
constexpr index_t K2 = GemmMPerThreadSubC;
|
||||
constexpr index_t K1 = KPerBlock / KPerThread;
|
||||
|
||||
constexpr auto out_10d_global_desc = make_ConstantTensorDescriptor(
|
||||
Sequence<N / N1, N1, K / (K1 * K2), K1, K2, Ho, Wo / (W1 * W2 * W3), W1, W2, W3>{});
|
||||
constexpr auto out_10d_global_desc =
|
||||
fwd(out_n_k_h_w_global_desc)
|
||||
.Fold(I3, Number<W1>{}, Number<W2>{}, Number<W3>{})
|
||||
.Fold(I1, Number<K1>{}, Number<K2>{})
|
||||
.Fold(I0, Number<N1>{});
|
||||
|
||||
constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor(
|
||||
Sequence<KPerThread / K2, 1, K2, HoPerThread, 1, W1, 1, W3, 1, N1>{});
|
||||
constexpr auto out_10d_thread_desc =
|
||||
fwd(out_k_h_w_n_thread_desc)
|
||||
.Fold(I3, Number<N1>{})
|
||||
.Fold(I2, Number<W1>{}, Number<1>{}, Number<W3>{})
|
||||
.Fold(I0, Number<1>{}, Number<K2>{});
|
||||
|
||||
#if 0
|
||||
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
|
||||
{
|
||||
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
|
||||
"out_k_h_w_n_thread_desc");
|
||||
print_ConstantTensorDescriptor(out_10d_thread_desc, "out_10d_thread_desc");
|
||||
if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0)
|
||||
{
|
||||
print_ConstantTensorDescriptor(out_k_h_w_n_thread_desc,
|
||||
"b: out_k_h_w_n_thread_desc");
|
||||
print_ConstantTensorDescriptor(out_10d_thread_desc, "b: out_10d_thread_desc");
|
||||
|
||||
print_ConstantTensorDescriptor(out_k_h_w_n_global_desc,
|
||||
"out_k_h_w_n_global_desc");
|
||||
print_ConstantTensorDescriptor(out_10d_global_desc, "out_10d_global_desc");
|
||||
|
||||
}
|
||||
print_ConstantTensorDescriptor(out_n_k_h_w_global_desc,
|
||||
"b: out_n_k_h_w_global_desc");
|
||||
print_ConstantTensorDescriptor(out_10d_global_desc, "b: out_10d_global_desc");
|
||||
}
|
||||
#endif
|
||||
|
||||
constexpr auto map_out_global2thread = Sequence<8, 9, 0, 1, 2, 3, 4, 5, 6, 7>{};
|
||||
|
||||
threadwise_tensor_slice_copy_reorder_given_dst2src_v2(
|
||||
#if 0
|
||||
threadwise_tensor_slice_copy_reorder_given_dst2src_v3(
|
||||
out_10d_thread_desc,
|
||||
p_out_thread,
|
||||
out_10d_global_desc,
|
||||
@@ -453,8 +420,24 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
|
||||
ho_block_data_begin + ho_thread_data_begin,
|
||||
wo_block_data_begin + wo_thread_data_begin),
|
||||
out_10d_thread_desc.GetLengths(),
|
||||
map_out_global2thread);
|
||||
// Number<OutThreadCopyDataPerWrite_W>{});
|
||||
map_out_global2thread,
|
||||
Number<OutThreadCopyDataPerWrite_W>{});
|
||||
#else
|
||||
threadwise_tensor_slice_copy_generic(
|
||||
out_10d_thread_desc.ReorderGivenNew2Old(map_out_global2thread),
|
||||
p_out_thread,
|
||||
make_zero_array<index_t, 10>(),
|
||||
out_10d_global_desc,
|
||||
p_out_global +
|
||||
out_n_k_h_w_global_desc.GetOffsetFromMultiIndex(
|
||||
n_block_data_begin + n_thread_data_begin,
|
||||
k_block_data_begin + k_thread_data_begin,
|
||||
ho_block_data_begin + ho_thread_data_begin,
|
||||
wo_block_data_begin + wo_thread_data_begin),
|
||||
make_zero_array<index_t, 10>(),
|
||||
out_10d_thread_desc.GetLengths().ReorderGivenNew2Old(map_out_global2thread),
|
||||
arithmetic_sequence_gen<0, 10, 1>::SeqType{});
|
||||
#endif
|
||||
});
|
||||
}
|
||||
};
|
||||
|
||||
@@ -151,6 +151,7 @@ struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw
|
||||
// slice a tensor, and copy it into another tensor
|
||||
// this copy operator already have blockwise offset built-in
|
||||
const auto blockwise_wei_copy =
|
||||
#if 0
|
||||
BlockwiseTensorSliceCopy_generic_v1<BlockSize,
|
||||
Float,
|
||||
decltype(wei_c_k_global_desc),
|
||||
@@ -164,15 +165,26 @@ struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw
|
||||
WeiBlockCopyDataPerAccess_K,
|
||||
WeiBlockCopyDataPerAccess_K>(
|
||||
{0, k_block_data_on_global}, {0, 0});
|
||||
#else
|
||||
Blockwise2dTensorCopy3<BlockSize,
|
||||
Float,
|
||||
decltype(wei_c_k_global_desc),
|
||||
decltype(wei_c_k_block_desc),
|
||||
decltype(wei_c_k_block_desc.GetLengths()),
|
||||
WeiBlockCopyDataPerAccess_K>({0, k_block_data_on_global},
|
||||
{0, 0});
|
||||
#endif
|
||||
|
||||
// GEMM definition
|
||||
// c_mtx += transpose(a_mtx) * b_mtx
|
||||
// a_mtx[CPerBlock, KPerBlock] is in LDS
|
||||
// b_mtx[CPerBlocl, N1 * BPerBlock * N2] is in LDS
|
||||
// c_mtx[KPerBlock, N1 * BPerBlock * N2] is distributed among threads, and saved in
|
||||
// register
|
||||
constexpr auto a_c_k_block_mtx_desc = make_ConstantMatrixDescriptor(
|
||||
Number<CPerBlock>{}, Number<KPerBlock>{}, Number<wei_c_k_block_desc.GetStride(I0)>{});
|
||||
// GEMM definition
|
||||
// c_mtx += transpose(a_mtx) * b_mtx
|
||||
// a_mtx[CPerBlock, KPerBlock] is in LDS
|
||||
// b_mtx[CPerBlocl, N1 * BPerBlock * N2] is in LDS
|
||||
// c_mtx[KPerBlock, N1 * BPerBlock * N2] is distributed among threads, and saved in
|
||||
// register
|
||||
constexpr auto a_c_k_block_mtx_desc =
|
||||
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
|
||||
Number<KPerBlock>{},
|
||||
Number<wei_c_k_block_desc.GetStride(I0)>{});
|
||||
|
||||
constexpr auto b_c_n1bn2_block_mtx_desc =
|
||||
make_ConstantMatrixDescriptor(Number<CPerBlock>{},
|
||||
|
||||
Reference in New Issue
Block a user