diff --git a/driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp b/driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp index 3a68099e2e..335655d5a3 100644 --- a/driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp +++ b/driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp @@ -4,6 +4,7 @@ #include "gridwise_convolution_wrapper.hip.hpp" #include "gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hip.hpp" #include "gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn_lds_double_buffer.hip.hpp" +#include "gridwise_convolution_implicit_gemm_v1r2_chwn_cyxk_khwn.hip.hpp" template void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc, @@ -178,33 +179,12 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc, constexpr index_t HoPerThread = 1; constexpr index_t WoPerThread = 1; - constexpr index_t BlockSize = 128; -#elif 0 - // for 7x7, 38x38 - constexpr index_t NPerBlock = 8; - constexpr index_t KPerBlock = 64; - constexpr index_t CPerBlock = 1; - constexpr index_t HoPerBlock = 4; - constexpr index_t WoPerBlock = 4; - - constexpr index_t NPerThread = 4; - constexpr index_t KPerThread = 16; - constexpr index_t CPerThread = 1; - constexpr index_t HoPerThread = 1; - constexpr index_t WoPerThread = 1; - - constexpr index_t WeiBlockCopyThreadPerDim0 = 4; - constexpr index_t WeiBlockCopyThreadPerDim1 = 32; - - constexpr index_t InBlockCopyDataPerRead = 4; // not used, yet - constexpr index_t WeiBlockCopyDataPerRead = 4; - constexpr index_t BlockSize = 128; #elif 1 - // for 3x3, 56x56 - constexpr index_t NPerBlock = 32; - constexpr index_t KPerBlock = 64; - constexpr index_t CPerBlock = 4; + // for 7x7, 38x38 + constexpr index_t NPerBlock = 16; + constexpr index_t KPerBlock = 128; + constexpr index_t CPerBlock = 8; constexpr index_t HoPerBlock = 2; constexpr index_t WoPerBlock = 2; @@ -216,19 +196,50 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc, constexpr index_t GemmMPerThreadSubC = 4; constexpr index_t GemmNPerThreadSubC = 4; constexpr index_t GemmMLevel0Cluster = 4; - constexpr index_t GemmNLevel0Cluster = 4; - constexpr index_t GemmMLevel1Cluster = 2; + constexpr index_t GemmNLevel0Cluster = 2; + constexpr index_t GemmMLevel1Cluster = 4; constexpr index_t GemmNLevel1Cluster = 2; constexpr index_t GemmKPerThreadLoop = 1; - constexpr index_t InBlockCopy_ThreadPerDimC = 1; + constexpr index_t InBlockCopy_ThreadPerDimC = 2; constexpr index_t InBlockCopy_ThreadPerDimH = 4; constexpr index_t InBlockCopy_ThreadPerDimW = 4; - constexpr index_t InBlockCopy_ThreadPerDimN = 8; + constexpr index_t InBlockCopy_ThreadPerDimN = 4; constexpr index_t InBlockCopyDataPerRead = 4; constexpr index_t WeiBlockCopyDataPerRead = 4; - constexpr index_t OutThreadCopyDataPerWrite = 2; + constexpr index_t OutThreadCopyDataPerWrite = 4; + + constexpr index_t BlockSize = 128; +#elif 1 + // for 3x3, 56x56 + constexpr index_t NPerBlock = 16; + constexpr index_t KPerBlock = 128; + constexpr index_t CPerBlock = 8; + constexpr index_t HoPerBlock = 2; + constexpr index_t WoPerBlock = 2; + + constexpr index_t NPerThread = 4; + constexpr index_t KPerThread = 8; + constexpr index_t HoPerThread = 1; + constexpr index_t WoPerThread = 2; + + constexpr index_t GemmMPerThreadSubC = 4; + constexpr index_t GemmNPerThreadSubC = 4; + constexpr index_t GemmMLevel0Cluster = 4; + constexpr index_t GemmNLevel0Cluster = 2; + constexpr index_t GemmMLevel1Cluster = 4; + constexpr index_t GemmNLevel1Cluster = 2; + constexpr index_t GemmKPerThreadLoop = 1; + + constexpr index_t InBlockCopy_ThreadPerDimC = 2; + constexpr index_t InBlockCopy_ThreadPerDimH = 4; + constexpr index_t InBlockCopy_ThreadPerDimW = 4; + constexpr index_t InBlockCopy_ThreadPerDimN = 4; + constexpr index_t InBlockCopyDataPerRead = 4; + + constexpr index_t WeiBlockCopyDataPerRead = 4; + constexpr index_t OutThreadCopyDataPerWrite = 4; constexpr index_t BlockSize = 128; #elif 0 @@ -264,7 +275,7 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc, constexpr index_t OutThreadCopyDataPerWrite = 2; constexpr index_t BlockSize = 128; -#elif 0 +#elif 1 // for 1x1, 14x14, Pascal constexpr index_t NPerBlock = 16; constexpr index_t KPerBlock = 128; @@ -306,9 +317,11 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc, for(index_t i = 0; i < nrepeat; ++i) { constexpr auto gridwise_conv = -#if 1 +#if 0 GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn -#else +#elif 1 + GridwiseConvolutionImplicitGemm_v1r2_chwn_cyxk_khwn +#elif 0 GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn_lds_double_buffer #endif {}, - Number{}); + Number{}); constexpr auto wei_ek_block_desc = make_ConstantTensorDescriptor_aligned( - Sequence{}, Number{}); + Sequence{}, Number{}); constexpr auto wei_cyxk_block_desc = make_ConstantTensorDescriptor_aligned( - Sequence{}, Number{}); + Sequence{}, Number{}); // tensor view of threadwise output in register constexpr auto out_khwn_thread_desc = make_ConstantTensorDescriptor( @@ -147,7 +150,7 @@ struct GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn constexpr auto c_kxwn_thread_mtx_desc = make_ConstantMatrixDescriptor(Number{}, Number{}, - Number{}); + Number{}); const auto blockwise_batch_gemm = BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2< @@ -169,9 +172,6 @@ struct GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn HoPerThread>{}; // LDS: be careful of alignment - constexpr index_t max_align = - mod_conv::max(index_t(4), InBlockCopyDataPerRead, WeiBlockCopyDataPerRead); - constexpr index_t in_block_space = in_chwn_block_desc.GetElementSpace(Number{}); constexpr index_t wei_block_space = diff --git a/src/include/gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn_lds_double_buffer.hip.hpp b/src/include/gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn_lds_double_buffer.hip.hpp index ac15d60c01..322a7007c6 100644 --- a/src/include/gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn_lds_double_buffer.hip.hpp +++ b/src/include/gridwise_convolution_implicit_gemm_v1_chwn_cyxk_khwn_lds_double_buffer.hip.hpp @@ -98,15 +98,18 @@ struct GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn_lds_double_buffer // tensor view of blockwise input and weight in LDS // be careful of alignment + constexpr index_t max_align = + mod_conv::max(index_t(4), InBlockCopyDataPerRead, WeiBlockCopyDataPerRead); + constexpr auto in_chwn_block_desc = make_ConstantTensorDescriptor_aligned( Sequence{}, - Number{}); + Number{}); constexpr auto wei_ek_block_desc = make_ConstantTensorDescriptor_aligned( - Sequence{}, Number{}); + Sequence{}, Number{}); constexpr auto wei_cyxk_block_desc = make_ConstantTensorDescriptor_aligned( - Sequence{}, Number{}); + Sequence{}, Number{}); // tensor view of threadwise output in register constexpr auto out_khwn_thread_desc = make_ConstantTensorDescriptor( @@ -150,7 +153,7 @@ struct GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn_lds_double_buffer constexpr auto c_kxwn_thread_mtx_desc = make_ConstantMatrixDescriptor(Number{}, Number{}, - Number{}); + Number{}); const auto blockwise_batch_gemm = BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2< @@ -172,9 +175,6 @@ struct GridwiseConvolutionImplicitGemm_v1_chwn_cyxk_khwn_lds_double_buffer HoPerThread>{}; // LDS: be careful of alignment - constexpr index_t max_align = - mod_conv::max(index_t(4), InBlockCopyDataPerRead, WeiBlockCopyDataPerRead); - constexpr index_t in_block_space = in_chwn_block_desc.GetElementSpace(Number{}); constexpr index_t wei_block_space = diff --git a/src/include/gridwise_convolution_implicit_gemm_v1r2_chwn_cyxk_khwn.hip.hpp b/src/include/gridwise_convolution_implicit_gemm_v1r2_chwn_cyxk_khwn.hip.hpp new file mode 100644 index 0000000000..a7d01ec4ca --- /dev/null +++ b/src/include/gridwise_convolution_implicit_gemm_v1r2_chwn_cyxk_khwn.hip.hpp @@ -0,0 +1,335 @@ +#pragma once +#include "common.hip.hpp" +#include "ConstantTensorDescriptor.hip.hpp" +#include "ConstantMatrixDescriptor.hip.hpp" +#include "blockwise_4d_tensor_op.hip.hpp" +#include "blockwise_2d_tensor_op.hip.hpp" +#include "threadwise_nd_tensor_op.hip.hpp" +#include "threadwise_4d_tensor_op.hip.hpp" +#include "blockwise_batched_gemm.hip.hpp" + +template +struct GridwiseConvolutionImplicitGemm_v1r2_chwn_cyxk_khwn +{ + __device__ void Run(const Float* const __restrict__ p_in_global, + const Float* const __restrict__ p_wei_global, + Float* const __restrict__ p_out_global) const + { + // be careful of this assertion + static_assert(NPerThread <= NPerBlock && NPerBlock % NPerThread == 0, + "wrong! should satisfy: NPerThread <= NPerBlock && NPerBlock % NPerThread == 0"); + + constexpr auto I0 = Number<0>{}; + constexpr auto I1 = Number<1>{}; + constexpr auto I2 = Number<2>{}; + constexpr auto I3 = Number<3>{}; + + constexpr auto in_chwn_global_desc = InGlobalDesc{}; + constexpr auto wei_cyxk_global_desc = WeiGlobalDesc{}; + constexpr auto out_khwn_global_desc = OutGlobalDesc{}; + + constexpr index_t C = in_chwn_global_desc.GetLength(I0); + + constexpr index_t K = out_khwn_global_desc.GetLength(I0); + constexpr index_t Ho = out_khwn_global_desc.GetLength(I1); + constexpr index_t Wo = out_khwn_global_desc.GetLength(I2); + constexpr index_t N = out_khwn_global_desc.GetLength(I3); + + constexpr index_t Y = wei_cyxk_global_desc.GetLength(I1); + constexpr index_t X = wei_cyxk_global_desc.GetLength(I2); + + constexpr index_t HiPerBlock = HoPerBlock + Y - 1; + constexpr index_t WiPerBlock = WoPerBlock + X - 1; + + // divide block work: [K, Ho, Wo, N] + static_assert(N % NPerBlock == 0 && K % KPerBlock == 0 && C % CPerBlock == 0 && Ho % HoPerBlock == 0 && Wo % WoPerBlock == 0, + "wrong! cannot evenly divide work for workgroup "); + + constexpr index_t KBlockWork = (K + KPerBlock - 1) / KPerBlock; + constexpr index_t HBlockWork = (Ho + HoPerBlock - 1) / HoPerBlock; + constexpr index_t WBlockWork = (Wo + WoPerBlock - 1) / WoPerBlock; + constexpr index_t NBlockWork = (N + NPerBlock - 1) / NPerBlock; + + const index_t k_block_work_id = get_block_1d_id() / (HBlockWork * WBlockWork * NBlockWork); + index_t itmp = get_block_1d_id() - k_block_work_id * (HBlockWork * WBlockWork * NBlockWork); + const index_t h_block_work_id = itmp / (WBlockWork * NBlockWork); + itmp -= h_block_work_id * (WBlockWork * NBlockWork); + const index_t w_block_work_id = itmp / NBlockWork; + const index_t n_block_work_id = itmp - w_block_work_id * NBlockWork; + + const index_t k_block_data_begin = k_block_work_id * KPerBlock; + const index_t ho_block_data_begin = h_block_work_id * HoPerBlock; + const index_t wo_block_data_begin = w_block_work_id * WoPerBlock; + const index_t n_block_data_begin = n_block_work_id * NPerBlock; + + const index_t hi_block_data_begin = ho_block_data_begin; + const index_t wi_block_data_begin = wo_block_data_begin; + + // 2d tensor view of gridwise weight + constexpr auto wei_ck_global_desc = make_ConstantTensorDescriptor(Sequence{}, Sequence{}); + + // tensor view of blockwise input and weight in LDS + // be careful of alignment + constexpr index_t max_align = + mod_conv::max(index_t(4), InBlockCopyDataPerRead, WeiBlockCopyDataPerRead); + + constexpr auto in_chwn_block_desc = make_ConstantTensorDescriptor_aligned( + Sequence{}, Number{}); + + constexpr auto wei_ck_block_desc = make_ConstantTensorDescriptor_aligned( + Sequence{}, Number{}); + + // tensor view of threadwise output in register + constexpr auto out_khwn_thread_desc = make_ConstantTensorDescriptor( + Sequence{}); + + // blockwise copy + // input: format is [C, Hi, Wi, N] + const auto blockwise_in_copy = + Blockwise4dTensorCopy3{}; + + // blockwise wei copy + // format is [CPerBlock, KPerBlock] + const auto blockwise_wei_copy = +#if 0//debug + Blockwise2dTensorCopy1{}; +#else + Blockwise2dTensorCopy3{}; +#endif + + // a series of blockwise batched GEMM + // C_matrix += transpose(A_matrix) * B_matrix + // A_matrix and B_matrix saved in LDS, C_matrix saved in register + // A_matrix[C,K] is a sub-matrix of wei_block[C,K] + // 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{}, Number{}, Number{}); + + constexpr auto b_cxwn_block_mtx_desc = + make_ConstantMatrixDescriptor(Number{}, + Number{}, + Number{}); + + constexpr auto c_kxwn_thread_mtx_desc = + make_ConstantMatrixDescriptor(Number{}, + Number{}, + Number{}); + + const auto blockwise_batch_gemm = + BlockwiseBatchGemmBlockABlockBThreadCTransANormalBNormalC_V2< + BlockSize, + decltype(a_cxk_block_mtx_desc), + decltype(b_cxwn_block_mtx_desc), + decltype(c_kxwn_thread_mtx_desc), + 0, + in_chwn_block_desc.GetStride(I1), + out_khwn_thread_desc.GetStride(I1), + HoPerBlock, + GemmMPerThreadSubC, + GemmNPerThreadSubC, + GemmMLevel0Cluster, + GemmNLevel0Cluster, + GemmMLevel1Cluster, + GemmNLevel1Cluster, + GemmKPerThreadLoop, + HoPerThread>{}; + + // LDS: be careful of alignment + constexpr index_t in_block_space = in_chwn_block_desc.GetElementSpace(Number{}); + constexpr index_t wei_block_space = wei_ck_block_desc.GetElementSpace(Number{}); + + __shared__ Float p_in_block[in_block_space]; + __shared__ Float p_wei_block[wei_block_space]; + + // register + Float p_out_thread[out_khwn_thread_desc.GetElementSpace()]; + +#if 1 + 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_cyxk_global_desc, "wei_cyxk_global_desc"); + print_ConstantTensorDescriptor(wei_ck_global_desc, "wei_ck_global_desc"); + + print_ConstantTensorDescriptor(in_chwn_block_desc, "in_chwn_block_desc"); + print_ConstantTensorDescriptor(wei_ck_block_desc, "wei_ck_block_desc"); + + printf("in_block_space %u, wei_block_space %u\n", in_block_space, wei_block_space); + } +#endif + + // set threadwise output tensor to 0 + threadwise_4d_tensor_set_zero(out_khwn_thread_desc, p_out_thread); + + const Float* p_in_global_block_offset = + p_in_global + + in_chwn_global_desc.Get1dIndex( + 0, hi_block_data_begin, wi_block_data_begin, n_block_data_begin); + + const Float* p_wei_global_block_offset = + p_wei_global + wei_cyxk_global_desc.Get1dIndex(0, 0, 0, k_block_data_begin); + + for(index_t c_block_data_begin = 0; c_block_data_begin < C; c_block_data_begin += CPerBlock, + p_in_global_block_offset += CPerBlock * in_chwn_global_desc.GetStride(I0), + p_wei_global_block_offset += CPerBlock * wei_cyxk_global_desc.GetStride(I0)) + { + // input: global mem to LDS + blockwise_in_copy.Run(p_in_global_block_offset, p_in_block); + + for(index_t y = 0; y < Y; ++y) + { + for(index_t x = 0; x < X; ++x) + { + // weight: global mem to LDS + blockwise_wei_copy.Run(p_wei_global_block_offset + + wei_cyxk_global_desc.Get1dIndex(0, y, x, 0), + p_wei_block); + + __syncthreads(); + + blockwise_batch_gemm.Run(p_wei_block, + p_in_block + in_chwn_block_desc.Get1dIndex(0, y, x, 0), + p_out_thread); + + __syncthreads(); + } + } + } + +// output: register to global mem, +#if 0 + const auto c_thread_mtx_begin = + blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id()); + + for(index_t k = 0; k < out_khwn_thread_desc.GetLength(I0); ++k) + { + for(index_t ho = 0; ho < out_khwn_thread_desc.GetLength(I1); ++ho) + { + for(index_t wo = 0; wo < out_khwn_thread_desc.GetLength(I2); ++wo) + { + for(index_t n = 0; n < out_khwn_thread_desc.GetLength(I3); ++n) + { + const index_t b = out_khwn_thread_desc.Get1dIndex(0, 0, wo, n); + + const auto c_thread_mtx_distance = + blockwise_batch_gemm.GetDistanceFromBeginOfThreadMatrixC(ho, k, b); + + const index_t ho_thread = + c_thread_mtx_begin.batch + c_thread_mtx_distance.batch; + const index_t k_thread = c_thread_mtx_begin.row + c_thread_mtx_distance.row; + const index_t b_thread = c_thread_mtx_begin.col + c_thread_mtx_distance.col; + + const index_t wo_thread = b_thread / NPerBlock; + const index_t n_thread = b_thread % NPerBlock; + + p_out_global[out_khwn_global_desc.Get1dIndex(k_block_data_begin + k_thread, + ho_block_data_begin + ho_thread, + wo_block_data_begin + wo_thread, + n_block_data_begin + n_thread)] = + p_out_thread[out_khwn_thread_desc.Get1dIndex(k, ho, wo, n)]; + } + } + } + } +#elif 1 + const auto c_thread_mtx_begin = + blockwise_batch_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id()); + + const index_t k_thread_data_begin = c_thread_mtx_begin.row; + const index_t ho_thread_data_begin = c_thread_mtx_begin.batch; + 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 * wo_thread_data_begin; + + // output is a 10d tensor + constexpr index_t N2 = GemmNPerThreadSubC; + constexpr index_t N1 = NPerBlock / N2; + + constexpr index_t W2 = + (GemmNLevel0Cluster * GemmNLevel1Cluster) / (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{}); + + constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor( + Sequence{}); + +#if 0 + if(get_thread_local_1d_id() == 0 && get_block_1d_id() == 0) + { + print_ConstantTensorDescriptor(out_khwn_thread_desc, "out_khwn_thread_desc"); + print_ConstantTensorDescriptor(out_10d_thread_desc, "out_10d_thread_desc"); + + print_ConstantTensorDescriptor(out_khwn_global_desc, "out_khwn_global_desc"); + print_ConstantTensorDescriptor(out_10d_global_desc, "out_10d_global_desc"); + } +#endif + + threadwise_10d_tensor_copy( + out_10d_thread_desc, + p_out_thread, + out_10d_global_desc, + p_out_global + + out_khwn_global_desc.Get1dIndex(k_block_data_begin + k_thread_data_begin, + ho_block_data_begin + ho_thread_data_begin, + wo_block_data_begin + wo_thread_data_begin, + n_block_data_begin + n_thread_data_begin), + out_10d_thread_desc.GetLengths(), + Number{}); +#endif + } +};