added lds doubl buffer for implicit gemm v4 (nchw, kcyx)

driver/device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp

@@ -3,6 +3,7 @@
 #include "device.hpp"
 #include "gridwise_convolution_wrapper.hip.hpp"
 #include "gridwise_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hip.hpp"
+#include "gridwise_convolution_implicit_gemm_v4_lds_double_buffer_nchw_kcyx_nkhw.hip.hpp"
 
 template <class T, class InDesc, class WeiDesc, class OutDesc>
 void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
@@ -86,8 +87,10 @@ void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
     for(index_t i = 0; i < nrepeat; ++i)
     {
         constexpr auto gridwise_conv =
-#if 1
+#if 0
             GridwiseConvolutionImplicitGemm_v4_nchw_kcyx_nkhw
+#else
+            GridwiseConvolutionImplicitGemm_v4_lds_double_buffer_nchw_kcyx_nkhw
 #endif
             <GridSize,
              BlockSize,

driver/driver.hip.cpp

@@ -455,10 +455,10 @@ int main(int argc, char* argv[])
 
     constexpr index_t HPad = 0;
     constexpr index_t WPad = 0;
-#elif 0
+#elif 1
     // 1x1 filter, 28x28 image
-    constexpr index_t N  = 16;
-    constexpr index_t C  = 256;
+    constexpr index_t N  = 128;
+    constexpr index_t C  = 512;
     constexpr index_t HI = 28;
     constexpr index_t WI = 28;
     constexpr index_t K  = 512;

src/include/gridwise_convolution_implicit_gemm_v4_lds_double_buffer_nchw_kcyx_nkhw.hip.hpp

+#pragma once
+#include "common.hip.hpp"
+#include "ConstantTensorDescriptor.hip.hpp"
+#include "ConstantMergedTensorDescriptor.hip.hpp"
+#include "ConstantMatrixDescriptor.hip.hpp"
+#include "blockwise_generic_tensor_slice_op.hip.hpp"
+#include "blockwise_gemm.hip.hpp"
+#include "threadwise_tensor_slice_op.hip.hpp"
+
+// define B = merge(N, Ho, Wo)
+template <index_t GridSize,
+          index_t BlockSize,
+          class Float,
+          class InGlobalDesc,
+          class WeiGlobalDesc,
+          class OutGlobalDesc,
+          index_t BPerBlock,
+          index_t KPerBlock,
+          index_t EPerBlock,
+          index_t N1,
+          index_t N2,
+          index_t GemmMPerThreadSubC,
+          index_t GemmNPerThreadSubC,
+          index_t GemmMLevel0Cluster,
+          index_t GemmNLevel0Cluster,
+          index_t GemmMLevel1Cluster,
+          index_t GemmNLevel1Cluster,
+          index_t GemmKPerThreadLoop,
+          index_t GemmDataPerReadA,
+          index_t GemmDataPerReadB,
+          class InBlockCopySubLengths_E_N1_B_N2,
+          class InBlockCopyClusterLengths_E_N1_B_N2,
+          index_t InBlockCopySrcDataPerRead_B,
+          index_t InBlockCopyDstDataPerWrite_N2,
+          class WeiBlockCopySubLengths_E_K,
+          class WeiBlockCopyClusterLengths_E_K,
+          index_t WeiBlockCopySrcDataPerRead_E,
+          index_t WeiBlockCopyDstDataPerWrite_K>
+struct GridwiseConvolutionImplicitGemm_v4_lds_double_buffer_nchw_kcyx_nkhw
+{
+    __device__ void Run(const Float* const __restrict__ p_in_global,
+                        const Float* const __restrict__ p_wei_global,
+                        Float* const __restrict__ p_out_global) const
+    {
+        // this is a mess
+        // TODO: find more elegent way of specifying (or calculating) performance parameters
+        static_assert(N2 == GemmNPerThreadSubC, "wrong!");
+        static_assert((N1 * N2 * BPerBlock) %
+                              (GemmNPerThreadSubC * GemmNLevel0Cluster * GemmNLevel1Cluster) ==
+                          0,
+                      "wrong!");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+        constexpr auto I2 = Number<2>{};
+        constexpr auto I3 = Number<3>{};
+        constexpr auto I4 = Number<4>{};
+        constexpr auto I5 = Number<5>{};
+        constexpr auto I6 = Number<6>{};
+        constexpr auto I7 = Number<7>{};
+
+        constexpr auto True = integral_constant<bool, true>{};
+
+        constexpr auto in_n_c_h_w_global_desc  = InGlobalDesc{};
+        constexpr auto wei_k_c_y_x_global_desc = WeiGlobalDesc{};
+        constexpr auto out_n_k_h_w_global_desc = OutGlobalDesc{};
+
+        constexpr index_t N  = in_n_c_h_w_global_desc.GetLength(I0);
+        constexpr index_t C  = in_n_c_h_w_global_desc.GetLength(I1);
+        constexpr index_t Hi = in_n_c_h_w_global_desc.GetLength(I2);
+        constexpr index_t Wi = in_n_c_h_w_global_desc.GetLength(I3);
+
+        constexpr index_t K  = out_n_k_h_w_global_desc.GetLength(I1);
+        constexpr index_t Ho = out_n_k_h_w_global_desc.GetLength(I2);
+        constexpr index_t Wo = out_n_k_h_w_global_desc.GetLength(I3);
+
+        constexpr index_t Y = wei_k_c_y_x_global_desc.GetLength(I2);
+        constexpr index_t X = wei_k_c_y_x_global_desc.GetLength(I3);
+
+        static_assert(N % (N1 * N2) == 0, "wrong! cannot divice N evenly among thread");
+
+        constexpr index_t N0 = N / (N1 * N2);
+
+        constexpr index_t B = N0 * Ho * Wo;
+
+        constexpr index_t E = C * Y * X;
+
+        // divide block work by [K, B]
+        static_assert(K % KPerBlock == 0 && B % BPerBlock == 0 && E % (2 * EPerBlock) == 0,
+                      "wrong! cannot divide work evenly among block");
+
+        constexpr index_t KBlockWork = K / KPerBlock;
+        constexpr index_t BBlockWork = B / BPerBlock;
+
+        constexpr auto block_work_desc =
+            make_ConstantTensorDescriptor_default_rank_packed(Sequence<KBlockWork, BBlockWork>{});
+
+        const auto block_work_multi_id =
+            block_work_desc.GetMultiIndexFrom1dIndex(get_block_1d_id());
+
+        const index_t k_block_data_on_global = block_work_multi_id[0] * KPerBlock;
+        const index_t b_block_data_on_global = block_work_multi_id[1] * BPerBlock;
+
+        // input tensor
+        //     tensor descriptor in device memory [N0, N1, N2, Ho, Wo]
+        constexpr auto in_n0_n1_n2_h_w_global_desc = in_n_c_h_w_global_desc.Slice(I2, Number<Ho>{})
+                                                         .Slice(I3, Number<Wo>{})
+                                                         .Fold(I0, Number<N1>{}, Number<N2>{})
+                                                         .Extract(Sequence<0, 1, 2, 4, 5>{});
+
+        //     batch descritpor for device memory
+        constexpr auto in_c_y_x_global_desc = in_n_c_h_w_global_desc.Slice(I2, Number<Y>{})
+                                                  .Slice(I3, Number<X>{})
+                                                  .Extract(Sequence<1, 2, 3>{});
+
+        //     merged tensor descriptor in device memory [E, N1, B, N2], src of blockwise copy
+        constexpr auto in_e_n1_b_n2_global_merged_desc = make_ConstantMergedTensorDescriptor(
+            in_c_y_x_global_desc.Embed(in_n0_n1_n2_h_w_global_desc),
+            Sequence<0, 1, 2>{},
+            Sequence<4>{},
+            Sequence<3, 6, 7>{},
+            Sequence<5>{});
+
+#if 0
+        if(get_block_1d_id() == 0 && get_thread_local_1d_id() == 0)
+        {
+            print_ConstantTensorDescriptor(in_n0_n1_n2_h_w_global_desc,
+                                           "in_n0_n1_n2_h_w_global_desc: ");
+            print_ConstantTensorDescriptor(in_c_y_x_global_desc, "in_c_y_x_global_desc: ");
+            print_ConstantMergedTensorDescriptor(in_e_n1_b_n2_global_merged_desc,
+                                                 "in_e_n1_b_n2_global_merged_desc: ");
+        }
+#endif
+
+        //     memory layout descriptor in LDS [E, N1, B, N2], dst of blockwise copy
+        //     be careful of LDS alignment
+        constexpr auto in_e_n1_b_n2_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
+            Sequence<EPerBlock, N1, BPerBlock, N2>{}, Number<InBlockCopyDstDataPerWrite_N2>{});
+
+        //     this check is ad-hoc
+        //     TODO: need to properly implement tensor descriptor with multiple alignment
+        //     requirements
+        static_assert(in_e_n1_b_n2_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
+                      "GemmDataPerReadB alignment requirement is not satisfied");
+
+        // input blockwise copy
+        //     slice a merged tensor, reorder and copy to a normal tensor
+        //     this copy operator already has blockwise offset built-in
+        auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1<
+            BlockSize,
+            Float,
+            decltype(in_e_n1_b_n2_global_merged_desc),
+            decltype(in_e_n1_b_n2_block_desc),
+            decltype(in_e_n1_b_n2_block_desc.GetLengths()),
+            InBlockCopySubLengths_E_N1_B_N2,
+            InBlockCopyClusterLengths_E_N1_B_N2,
+            Sequence<0, 1, 3, 2>, // thread_arrange_order [E, N1, N2, B]
+            Sequence<0, 1, 3, 2>, // src_access_order [E, N1, N2, B]
+            Sequence<0, 1, 2, 3>, // dst_access_order [E, N1, B, N2]
+            InBlockCopySrcDataPerRead_B,
+            InBlockCopyDstDataPerWrite_N2>({0, 0, b_block_data_on_global, 0}, {0, 0, 0, 0});
+
+        // weight tensor
+        //     tensor descriptor in device memory, src of blockwise copy
+        constexpr auto wei_e_k_global_desc =
+            wei_k_c_y_x_global_desc.Unfold(I1, I3).ReorderGivenNew2Old(Sequence<1, 0>{});
+
+        //     tensor descriptor in LDS, dst of blockwise copy
+        //     be careful of LDS alignment
+        constexpr auto wei_e_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
+            Sequence<EPerBlock, KPerBlock>{},
+            Number<mod_conv::max(WeiBlockCopyDstDataPerWrite_K, GemmDataPerReadA)>{});
+
+        // operator for blockwise copy of weight into LDS
+        //     slice a tensor, and copy it into another tensor
+        //     this copy operator already have blockwise offset built-in
+        auto blockwise_wei_copy =
+            BlockwiseGenericTensorSliceCopy_v1<BlockSize,
+                                               Float,
+                                               decltype(wei_e_k_global_desc),
+                                               decltype(wei_e_k_block_desc),
+                                               decltype(wei_e_k_block_desc.GetLengths()),
+                                               WeiBlockCopySubLengths_E_K,
+                                               WeiBlockCopyClusterLengths_E_K,
+                                               Sequence<1, 0>, // thread_arrange_order [K, E]
+                                               Sequence<1, 0>, // src_access_order [K, E]
+                                               Sequence<0, 1>, // dst_access_order [E, K]
+                                               WeiBlockCopySrcDataPerRead_E,
+                                               WeiBlockCopyDstDataPerWrite_K>(
+                {0, k_block_data_on_global}, {0, 0});
+
+        // GEMM definition
+        // c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[EPerBlock, KPerBlock] is in LDS
+        //     b_mtx[EPerBlocl, N1 * BPerBlock * N2] is in LDS
+        //     c_mtx[KPerBlock, N1 * BPerBlock * N2] is distributed among threads, and saved in
+        //     register
+        constexpr auto a_e_k_block_mtx_desc = make_ConstantMatrixDescriptor(
+            Number<EPerBlock>{}, Number<KPerBlock>{}, Number<wei_e_k_block_desc.GetStride(I0)>{});
+
+        constexpr auto b_e_n1bn2_block_mtx_desc =
+            make_ConstantMatrixDescriptor(Number<EPerBlock>{},
+                                          Number<N1 * BPerBlock * N2>{},
+                                          Number<in_e_n1_b_n2_block_desc.GetStride(I0)>{});
+
+        // sanity check
+        static_assert(KPerBlock % (GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster) ==
+                          0,
+                      "wrong!");
+
+        constexpr index_t GemmMRepeat =
+            KPerBlock / (GemmMPerThreadSubC * GemmMLevel0Cluster * GemmMLevel1Cluster);
+
+        // c_thread_mtx definition: this is a mess
+        // TODO:: more elegent way of defining c_thread_mtx
+        constexpr auto c_k0k2_n1n2_thread_mtx_desc = make_ConstantMatrixDescriptor(
+            Number<GemmMRepeat * GemmMPerThreadSubC>{}, Number<N1 * N2>{});
+
+        const auto blockwise_gemm = BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2<
+            BlockSize,
+            decltype(a_e_k_block_mtx_desc),
+            decltype(b_e_n1bn2_block_mtx_desc),
+            decltype(c_k0k2_n1n2_thread_mtx_desc),
+            GemmMPerThreadSubC,
+            GemmNPerThreadSubC,
+            GemmMLevel0Cluster,
+            GemmNLevel0Cluster,
+            GemmMLevel1Cluster,
+            GemmNLevel1Cluster,
+            GemmKPerThreadLoop,
+            GemmDataPerReadA,
+            GemmDataPerReadB>{};
+
+        // choose GEMM implementation here
+        const auto run_blockwise_gemm = [&](auto... Xs) {
+#if 1
+            return blockwise_gemm.Run(Xs...);
+#elif 0
+            return blockwise_gemm.Run_asm(Xs...);
+#endif
+        };
+
+        // LDS allocation for input and weight: be careful of alignment
+        constexpr index_t max_align = mod_conv::max(InBlockCopyDstDataPerWrite_N2,
+                                                    WeiBlockCopyDstDataPerWrite_K,
+                                                    GemmDataPerReadA,
+                                                    GemmDataPerReadB);
+
+        constexpr index_t in_block_space =
+            in_e_n1_b_n2_block_desc.GetElementSpace(Number<max_align>{});
+
+        constexpr index_t wei_block_space = wei_e_k_block_desc.GetElementSpace(Number<max_align>{});
+
+        __shared__ Float p_in_block_double[2 * in_block_space];
+        __shared__ Float p_wei_block_double[2 * wei_block_space];
+
+        // register allocation for output
+        Float p_out_thread[c_k0k2_n1n2_thread_mtx_desc.GetElementSpace()];
+
+        // zero out threadwise output
+        threadwise_matrix_set_zero(c_k0k2_n1n2_thread_mtx_desc, p_out_thread);
+
+#if 0
+        if(e == 0 * EPerBlock && get_block_1d_id() == 0)
+        {
+            printf("id %5u %5u: "
+                   "mThreadSrcOffset %u, mThreadDstOffset %u \n",
+                   get_block_1d_id(),
+                   get_thread_local_1d_id(),
+                   blockwise_wei_copy.mThreadSrcOffset,
+                   blockwise_wei_copy.mThreadDstOffset);
+        }
+#endif
+
+        // LDS double buffer: preload data into LDS
+        {
+            Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
+            Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
+
+            blockwise_in_copy.RunLoadRegisterClipboard(p_in_global, p_in_register_clipboard);
+            blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global, p_wei_register_clipboard);
+
+            blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard, p_in_block_double);
+            blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard,
+                                                         p_wei_block_double);
+        }
+
+        // LDS double buffer: main body
+        for(index_t e_block_data_begin = 0; e_block_data_begin + 2 * EPerBlock < E;
+            e_block_data_begin += 2 * EPerBlock)
+        {
+#pragma unroll
+            for(index_t iloop = 0; iloop < 2; ++iloop)
+            {
+                const bool even_loop = (iloop % 2 == 0);
+
+                Float* p_in_block_now =
+                    even_loop ? p_in_block_double : p_in_block_double + in_block_space;
+                Float* p_wei_block_now =
+                    even_loop ? p_wei_block_double : p_wei_block_double + wei_block_space;
+
+                Float* p_in_block_next =
+                    even_loop ? p_in_block_double + in_block_space : p_in_block_double;
+                Float* p_wei_block_next =
+                    even_loop ? p_wei_block_double + wei_block_space : p_wei_block_double;
+
+                Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
+                Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
+
+                blockwise_in_copy.MoveSlicingWindowOnSourceTensor(I0, Number<EPerBlock>{}, True);
+                blockwise_wei_copy.MoveSlicingWindowOnSourceTensor(I0, Number<EPerBlock>{}, True);
+
+                __syncthreads();
+
+                // LDS doubel buffer: load next data from device mem
+                blockwise_in_copy.RunLoadRegisterClipboard(p_in_global, p_in_register_clipboard);
+                blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global, p_wei_register_clipboard);
+
+                // LDS double buffer: GEMM on current data
+                run_blockwise_gemm(p_wei_block_now, p_in_block_now, p_out_thread);
+
+                // LDS double buffer: store next data to LDS
+                blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard,
+                                                            p_in_block_next);
+                blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard,
+                                                             p_wei_block_next);
+            }
+        }
+
+        // LDS double buffer: tail
+        {
+            Float p_in_register_clipboard[blockwise_in_copy.GetRegisterClipboardSize()];
+            Float p_wei_register_clipboard[blockwise_wei_copy.GetRegisterClipboardSize()];
+
+            // even iteration
+            blockwise_in_copy.MoveSlicingWindowOnSourceTensor(I0, Number<EPerBlock>{}, True);
+            blockwise_wei_copy.MoveSlicingWindowOnSourceTensor(I0, Number<EPerBlock>{}, True);
+
+            __syncthreads();
+
+            // LDS doubel buffer: load next data from device mem
+            blockwise_in_copy.RunLoadRegisterClipboard(p_in_global, p_in_register_clipboard);
+            blockwise_wei_copy.RunLoadRegisterClipboard(p_wei_global, p_wei_register_clipboard);
+
+            // LDS double buffer: GEMM on current data
+            run_blockwise_gemm(p_wei_block_double, p_in_block_double, p_out_thread);
+
+            // LDS double buffer: store next data to LDS
+            blockwise_in_copy.RunStoreRegisterClipboard(p_in_register_clipboard,
+                                                        p_in_block_double + in_block_space);
+            blockwise_wei_copy.RunStoreRegisterClipboard(p_wei_register_clipboard,
+                                                         p_wei_block_double + wei_block_space);
+
+            // odd iteration
+            __syncthreads();
+
+            // LDS double buffer: GEMM on current data
+            run_blockwise_gemm(p_wei_block_double + wei_block_space,
+                               p_in_block_double + in_block_space,
+                               p_out_thread);
+        }
+
+        // copy output: register to global memory
+        {
+            constexpr index_t K2 = GemmMPerThreadSubC;
+            constexpr index_t K1 = GemmMLevel0Cluster * GemmMLevel1Cluster;
+            constexpr index_t K0 = K / (K1 * K2);
+
+            // define tensor descriptor for threadwise copy
+            //     output memory layout descriptor in register
+            constexpr auto out_k0_k1_k2_n1_n0_h_w_n2_thread_mem_desc =
+                make_ConstantTensorDescriptor_default_rank_packed(
+                    Sequence<KPerBlock / (K1 * K2), 1, K2, N1, 1, 1, 1, N2>{});
+
+            //     output tensor descriptor in register, src of threadwise copy
+            constexpr auto out_n0_n1_n2_k0_k1_k2_h_w_thread_desc =
+                out_k0_k1_k2_n1_n0_h_w_n2_thread_mem_desc.ReorderGivenNew2Old(
+                    Sequence<4, 3, 7, 0, 1, 2, 5, 6>{});
+
+            //     output memory layout descriptor in device memory, dst of threadwise copy
+            constexpr auto out_n0_n1_n2_k0_k1_k2_h_w_global_mem_desc =
+                out_n_k_h_w_global_desc.Fold(I1, Number<K1>{}, Number<K2>{})
+                    .Fold(I0, Number<N1>{}, Number<N2>{});
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.GetBeginOfThreadMatrixC(get_thread_local_1d_id());
+
+            const index_t k_thread_data_on_global =
+                k_block_data_on_global + c_thread_mtx_on_block.row;
+
+            const index_t b_thread_data_on_global =
+                b_block_data_on_global + c_thread_mtx_on_block.col / N2;
+
+            //     output merged global tensor descriptor, for calculating origin of thread tensor
+            //     in global memory
+            constexpr auto out_k_n1_b_n2_global_merged_desc = make_ConstantMergedTensorDescriptor(
+                out_n0_n1_n2_k0_k1_k2_h_w_global_mem_desc.Unfold(I3, I5),
+                Sequence<3>{},
+                Sequence<1>{},
+                Sequence<0, 4, 5>{},
+                Sequence<2>{});
+
+            //     origin of dst in device memory
+            Float* p_out_thread_on_global =
+                p_out_global +
+                out_k_n1_b_n2_global_merged_desc.GetOffsetFromMultiIndex(
+                    k_thread_data_on_global, 0, b_thread_data_on_global, 0);
+
+            threadwise_generic_tensor_slice_copy(out_n0_n1_n2_k0_k1_k2_h_w_thread_desc,
+                                                 p_out_thread,
+                                                 {0, 0, 0, 0, 0, 0, 0, 0},
+                                                 out_n0_n1_n2_k0_k1_k2_h_w_global_mem_desc,
+                                                 p_out_thread_on_global,
+                                                 {0, 0, 0, 0, 0, 0, 0, 0},
+                                                 out_n0_n1_n2_k0_k1_k2_h_w_thread_desc.GetLengths(),
+                                                 arithmetic_sequence_gen<0, 8, 1>::SeqType{});
+        }
+    }
+};