feat(dnn/aarch64): add aarch64 int8X8X16_mk4_k8x8x8 matmul, performance is better

GitOrigin-RevId: b6af21e8e3

dnn/src/aarch64/matrix_mul/algos.cpp

@@ -1310,4 +1310,74 @@ MEGDNN_REG_GEMM_FUNC_FOR_IM2COL_IMPL(AlgoQuint8K8x8x8,
                                      int32_t);
 #endif
 
+/* ===================== Int8x8x16 K8x8x8 algo ===================== */
+namespace {
+void int8x8x16_mk4_8x8x8_kern(const MatrixMulImpl::KernParam& kern_param) {
+    MIDOUT_BEGIN(megdnn_aarch64_matmul_kern,
+                 midout_iv("int8x8x16_mk4_8x8x8_kern"_hash)) {
+        auto M = kern_param.M, N = kern_param.N, K = kern_param.K;
+        auto trA = kern_param.trA, trB = kern_param.trB;
+        auto LDA = kern_param.LDA, LDB = kern_param.LDB, LDC = kern_param.LDC;
+        auto A_type = kern_param.A_type, B_type = kern_param.B_type,
+             C_type = kern_param.C_type;
+        const auto Aptr = kern_param.A<dt_int8>(),
+                   Bptr = kern_param.B<dt_int8>();
+        auto Cptr = kern_param.C<dt_int16>();
+
+        aarch64::matmul::gemm_s8x8x16_mk4_8x8x8 strategy(M, N, K, A_type,
+                                                             B_type, C_type);
+        megdnn::matmul::GemmInterleaved<
+                aarch64::matmul::gemm_s8x8x16_mk4_8x8x8>(M, N, K, trA, trB,
+                                                             strategy)
+                .execute(Aptr, LDA, Bptr, LDB, Cptr, LDC,
+                         kern_param.workspace_ptr);
+    }
+    MIDOUT_END();
+}
+}  // anonymous namespace
+
+bool MatrixMulImpl::AlgoInt8x8x16MK4_K8x8x8::usable(
+        const KernSizeParam& kern_size_param) const {
+    return can_be_treated_as_int8x8x16(kern_size_param) &&
+           kern_size_param.format == param::MatrixMul::Format::MK4 &&
+           kern_size_param.compute_mode == Param::ComputeMode::DEFAULT &&
+           !kern_size_param.trA && !kern_size_param.trB &&
+           kern_size_param.M % 4 == 0 && kern_size_param.K % 4 == 0;
+}
+
+bool MatrixMulImpl::AlgoInt8x8x16MK4_K8x8x8::preferred(
+        const KernSizeParam&) const {
+    return true;
+}
+
+size_t MatrixMulImpl::AlgoInt8x8x16MK4_K8x8x8::get_workspace(
+        const KernSizeParam& kern_size_param) const {
+    MIDOUT_BEGIN(megdnn_aarch64_matmul_kern,
+                 midout_iv("AlgoInt8x8x16_MK4_8x8x8::get_workspace"_hash)) {
+        auto M = kern_size_param.M, N = kern_size_param.N,
+             K = kern_size_param.K;
+        auto trA = kern_size_param.trA, trB = kern_size_param.trB;
+        auto A_type = kern_size_param.A_type, B_type = kern_size_param.B_type,
+             C_type = kern_size_param.C_type;
+        aarch64::matmul::gemm_s8x8x16_mk4_8x8x8 strategy(M, N, K, A_type,
+                                                             B_type, C_type);
+        return megdnn::matmul::GemmInterleaved<
+                       matmul::gemm_s8x8x16_mk4_8x8x8>(M, N, K, trA, trB,
+                                                           strategy)
+                .get_workspace_size();
+    }
+    MIDOUT_END();
+    return 0;
+}
+
+MatrixMulImpl::kern_t MatrixMulImpl::AlgoInt8x8x16MK4_K8x8x8::get_kern(
+        const KernSizeParam&) const {
+    return int8x8x16_mk4_8x8x8_kern;
+}
+
+MEGDNN_REG_GEMM_FUNC_FOR_IM2COL_IMPL(AlgoInt8x8x16MK4_K8x8x8,
+                                     megdnn_aarch64_matmul_kern,
+                                     "AlgoInt8x8x16MK4_K8x8x8Impl"_hash,
+                                     aarch64::matmul::gemm_s8x8x16_mk4_8x8x8, int8_t,
+                                     int16_t);
 // vim: syntax=cpp.doxygen

dnn/src/aarch64/matrix_mul/algos.h

@@ -202,6 +202,22 @@ public:
     MEGDNN_REG_GEMM_FUNC_FOR_IM2COL();
 };
 
+class MatrixMulImpl::AlgoInt8x8x16MK4_K8x8x8 final : public AlgoBase {
+public:
+    bool is_reproducible() const override { return true; }
+    const char* name() const override {
+        return "AARCH64_INT8X8X16_MK4_K8X8X8";
+    }
+    bool usable(const KernSizeParam&) const override;
+    bool preferred(const KernSizeParam&) const override;
+    size_t get_workspace(const KernSizeParam&) const override;
+    kern_t get_kern(const KernSizeParam&) const override;
+    void* type() const override { return sm_arm_common_algo_type; }
+    PackMode packmode() const override { return PackMode::DEFAULT; }
+
+    MEGDNN_REG_GEMM_FUNC_FOR_IM2COL();
+};
+
 class MatrixMulImpl::AlgoInt8x8x16MK4_4x4x8 final : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }

dnn/src/aarch64/matrix_mul/asm/common.h

@@ -2101,6 +2101,62 @@ static inline void transpos_12x4_s8(const int8_t* inptr0, int8_t* outptr) {
                    vreinterpretq_s32_s8(input2), 3);
 }
 
+
+template <typename T>
+static inline void interleave_8x8_mk4_b(const T*& inptr0, const T*& inptr1,
+                                     T*& outptr) {
+
+    static_assert(
+            std::is_same<T, int8_t>::value || std::is_same<T, uint8_t>::value,
+            "transpose_8x4_1_b only support uint8_t and int8_t");
+    asm volatile(
+            "ld1 {v0.4s},  [%[inptr0]], #16\n"
+            "ld1 {v1.4s},  [%[inptr1]], #16\n"
+            "ld1 {v2.4s},  [%[inptr0]], #16\n"
+            "ld1 {v3.4s},  [%[inptr1]], #16\n"
+
+            "zip1 v4.4s, v0.4s, v1.4s \n"
+            "zip2 v5.4s, v0.4s, v1.4s \n"
+
+            "zip1 v6.4s, v2.4s, v3.4s\n"
+            "zip2 v7.4s, v2.4s, v3.4s\n"
+
+            "st1 {v4.4s},[%[outptr]],#16\n"
+            "st1 {v5.4s},[%[outptr]],#16\n"
+            "st1 {v6.4s},[%[outptr]],#16\n"
+            "st1 {v7.4s},[%[outptr]],#16\n"
+
+            : [inptr0] "+r"(inptr0), [inptr1] "+r"(inptr1),
+              [outptr] "+r"(outptr)
+            :
+            : "v0", "v1", "v2", "v3", "v4", "v5","v6","v7","memory");
+}
+
+template <typename T>
+static inline void transpose_8x8_mk4_b(const T*& inptr0, const T*& inptr1,
+                                     T* outptr) {
+
+    static_assert(
+            std::is_same<T, int8_t>::value || std::is_same<T, uint8_t>::value,
+            "transpose_8x4_1_b only support uint8_t and int8_t");
+    asm volatile(
+            "ld4 {v0.8b-v3.8b},  [%[inptr0]], #32\n"
+            "ld4 {v4.8b-v7.8b},  [%[inptr1]], #32\n"
+            "st1 {v0.2s},[%[outptr]],#8\n"
+            "st1 {v1.2s},[%[outptr]],#8\n"
+            "st1 {v2.2s},[%[outptr]],#8\n"
+            "st1 {v3.2s},[%[outptr]],#8\n"
+            "st1 {v4.2s},[%[outptr]],#8\n"
+            "st1 {v5.2s},[%[outptr]],#8\n"
+            "st1 {v6.2s},[%[outptr]],#8\n"
+            "st1 {v7.2s},[%[outptr]],#8\n"
+
+            : [inptr0] "+r"(inptr0), [inptr1] "+r"(inptr1),
+              [outptr] "+r"(outptr)
+            :
+            : "v0", "v1", "v2", "v3", "v4", "v5","v6","v7","memory");
+}
+
 }  // namespace aarch64
 }  // namespace megdnn
 

dnn/src/aarch64/matrix_mul/int8x8x16/strategy.cpp

@@ -13,6 +13,7 @@
 #include "src/aarch64/matrix_mul/asm/common.h"
 #include "src/aarch64/matrix_mul/int8x8x16/kernel_4x4x16.h"
 #include "src/aarch64/matrix_mul/int8x8x16/kernel_8x8x8.h"
+#include "src/aarch64/matrix_mul/int8x8x16/kernel_mk4_8x8x8.h"
 #include "src/aarch64/matrix_mul/int8x8x16/kernel_mk4_16x12x4_a53.h"
 #include "src/aarch64/matrix_mul/int8x8x16/kernel_mk4_4x4x8_a72.h"
 #include "src/aarch64/matrix_mul/int8x8x16/strategy.h"
@@ -357,4 +358,81 @@ void gemm_s8x8x16_mk4_4x4_a72::kern(const dt_int8* packA, const dt_int8* packB,
     }
 }
 
+// ===========================gemm_s8x8x16_mk4_8x8x8==================================
+MEGDNN_REG_GEMM_STRATEGY_IMPL(gemm_s8x8x16_mk4_8x8x8);
+
+void gemm_s8x8x16_mk4_8x8x8::pack_A(dt_int8* out, const dt_int8* in,
+                                        int ldin, int y0, int ymax, int k0,
+                                        int kmax, bool) const {
+    matmul_mk4_8x8x8::gemm_s8x8x16_mk4_8x8x8_pack_A(out, in, ldin, y0,
+                                                          ymax, k0, kmax);
+}
+
+void gemm_s8x8x16_mk4_8x8x8::pack_B(dt_int8* out, const dt_int8* in,
+                                        int ldin, int x0, int xmax, int k0,
+                                        int kmax, bool) const {
+    matmul_mk4_8x8x8::gemm_s8x8x16_mk4_8x8x8_pack_B(out, in, ldin, x0,
+                                                          xmax, k0, kmax);
+}
+
+void gemm_s8x8x16_mk4_8x8x8::kern(const dt_int8* packA, const dt_int8* packB,
+                                  size_t M, size_t N, size_t K, dt_int16* C,
+                                  size_t LDC, bool is_first_k, const dt_int16*,
+                                  dt_int16*) const {
+    megdnn_assert(A_dtype.enumv() == B_dtype.enumv() &&
+                  C_dtype.enumv() == DTypeEnum::Int16 &&
+                  A_dtype.enumv() == DTypeEnum::Int8);
+    megdnn_assert(is_first_k == true, "only impl is_first_k");
+    MEGDNN_MARK_USED_VAR(A_dtype);
+    MEGDNN_MARK_USED_VAR(B_dtype);
+    MEGDNN_MARK_USED_VAR(C_dtype);
+    megdnn_assert(M % 4 == 0 && K % 4 == 0, "M and K must be time of 4");
+
+    constexpr size_t pack_size = 4;
+    constexpr size_t pack_m = 8;
+    constexpr size_t pack_n = 8;
+    const size_t remain_n = N % pack_n;
+    size_t remain_m = M % pack_m;
+    K = round_up<size_t>(K, 8);
+    size_t KSIZE8 = K * pack_n;
+    size_t m_idx = 0;
+    for (; m_idx + pack_m <= M; m_idx += pack_m) {
+        int16_t* output = C + (m_idx / pack_size * LDC);
+
+        size_t n_idx = 0;
+        const int8_t* cur_packB = packB;
+        for (; n_idx + pack_n <= N; n_idx += pack_n) {
+            matmul_mk4_8x8x8::kern_8x8(packA, cur_packB, K, output, LDC,
+                                       is_first_k, pack_m, pack_n);
+            output += pack_n * pack_size;
+            cur_packB += KSIZE8;
+        }
+        if (remain_n > 0) {
+            matmul_mk4_8x8x8::kern_8x8_remain(packA, cur_packB, K, output, LDC,
+                                              is_first_k, pack_m, remain_n);
+            output += remain_n * pack_size;
+            cur_packB += KSIZE8;
+        }
+        packA += KSIZE8;
+    }
+
+    if (remain_m == 4) {
+        int16_t* output = C + (m_idx / pack_size * LDC);
+        size_t n_idx = 0;
+        const int8_t* cur_packB = packB;
+        for (; n_idx + pack_n <= N; n_idx += pack_n) {
+            matmul_mk4_8x8x8::kern_4x8(packA, cur_packB, K, output, LDC,
+                                       is_first_k, 4, pack_n);
+            output += pack_n * pack_size;
+            cur_packB += pack_n * K;
+        }
+        if (remain_n > 0) {
+            matmul_mk4_8x8x8::kern_4x8_remain(packA, cur_packB, K, output, LDC,
+                                              is_first_k, 4, remain_n);
+            output += remain_n * pack_size;
+            cur_packB += pack_n * K;
+        }
+    }
+}
+
 // vim: syntax=cpp.doxygen

dnn/src/aarch64/matrix_mul/int8x8x16/strategy.h

@@ -26,6 +26,8 @@ MEGDNN_REG_GEMM_STRATEGY(dt_int8, dt_int16, dt_int16, 4, 4, 8, false, false,
 MEGDNN_REG_GEMM_STRATEGY_WITH_PACK_A_TYPE(dt_int8, dt_int16, dt_int16, dt_int16,
                                           16, 12, 4, false, false,
                                           gemm_s8x8x16_mk4_16x12_a53);
+MEGDNN_REG_GEMM_STRATEGY(dt_int8, dt_int16, dt_int16, 8, 8, 8, false, false,
+                         gemm_s8x8x16_mk4_8x8x8);
 
 }  // namespace matmul
 }  // namespace aarch64

dnn/src/aarch64/matrix_mul/opr_impl.cpp

@@ -39,6 +39,7 @@ class MatrixMulImpl::AlgoPack : NonCopyableObj {
     AlgoInt8x8x16K4x4x16 int8x8x16_k4x4x16;
     AlgoInt8x8x16MK4_16x12x4 int8x8x16_mk4_16x12x4;
     AlgoInt8x8x16MK4_4x4x8 int8x8x16_mk4_4x4x8;
+    AlgoInt8x8x16MK4_K8x8x8 int8x8x16_mk4_k8x8x8;
 
     AlgoInt16x16x32K12x8x1 int16x16x32_k12x8x1;
     AlgoInt16x16x32MK8_8x8 int16x16x32_mk8_8x8;
@@ -73,6 +74,7 @@ public:
 #endif
         all_algos.emplace_back(&int8x8x16_k4x4x16);
         all_algos.emplace_back(&int8x8x16_k8x8x8);
+        all_algos.emplace_back(&int8x8x16_mk4_k8x8x8);
         all_algos.emplace_back(&int8x8x16_mk4_4x4x8);
         all_algos.emplace_back(&int8x8x16_mk4_16x12x4);
 

dnn/src/aarch64/matrix_mul/opr_impl.h

@@ -57,6 +57,7 @@ private:
 #else
     class AlgoQuint8K8x8x8;         // Aarch64 Quint8 Kernel 8x8x8
 #endif
+    class AlgoInt8x8x16MK4_K8x8x8;      // Aarch64 Int4x4x16 Kernel 4x4x16
 
     class AlgoPack;
 };

dnn/test/aarch64/matrix_mul.cpp

@@ -122,6 +122,20 @@ TEST_F(AARCH64, MATRIX_MUL_INT8_MK4) {
                                  std::move(args));
 }
 
+TEST_F(AARCH64, MATRIX_MUL_INT8x8x16_MK4) {
+    std::vector<matrix_mul::TestArg> args;
+    for (size_t m : {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17})
+        for (size_t n :
+             {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 24})
+            for (size_t k :
+                 {2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+                  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29})
+                args.emplace_back(m, n, k, 0);
+    matrix_mul::check_matrix_mul(dtype::Int8{}, dtype::Int8{}, dtype::Int16{},
+                                 handle(), "AARCH64_INT8X8X16_MK4_K8X8X8",
+                                 param::MatrixMul::Format::MK4, 1, 1e-3,
+                                 std::move(args));
+}
 TEST_F(AARCH64, MATRIX_MUL_MK4_8x8x16_4x4) {
     matrix_mul::check_matrix_mul(dtype::Int8{}, dtype::Int8{}, dtype::Int16{},
                                  handle(), "AARCH64_INT8X8X16_MK4_4X4X8",
@@ -396,6 +410,71 @@ TEST_F(AARCH64, BENCHMARK_MATRIX_MUL_MK4_8x8x16) {
     run(384, 384, 384);
 }
 
+TEST_F(AARCH64, BENCHMARK_MATRIX_MUL_MK4_8x8x8_8x8x16_vs_4x4x16_8x8x16) {
+    constexpr size_t RUNS = 50;
+    param::MatrixMul param;
+    param.transposeA = false;
+    param.transposeB = false;
+    Benchmarker<MatrixMul> benchmarker(handle());
+    Benchmarker<MatrixMul> benchmarker_mk4(handle());
+    Benchmarker<MatrixMul> benchmarker_mk4_4x4x8(handle());
+    benchmarker.set_times(RUNS)
+            .set_dtype(0, dtype::Int8{})
+            .set_dtype(1, dtype::Int8{})
+            .set_dtype(2, dtype::Int16{})
+            .set_param(param)
+            .set_display(false);
+    benchmarker.set_before_exec_callback(
+            AlgoChecker<MatrixMul>("AARCH64_INT8X8X16_K4X4X16"));
+
+    param.format = MatrixMul::Param::Format::MK4;
+    benchmarker_mk4.set_before_exec_callback(
+            AlgoChecker<MatrixMul>(
+                "AARCH64_INT8X8X16_MK4_K8X8X8"
+                ));
+    benchmarker_mk4.set_times(RUNS)
+            .set_dtype(0, dtype::Int8{})
+            .set_dtype(1, dtype::Int8{})
+            .set_dtype(2, dtype::Int16{})
+            .set_param(param)
+            .set_display(false);
+
+    benchmarker_mk4_4x4x8.set_before_exec_callback(
+            AlgoChecker<MatrixMul>("AARCH64_INT8X8X16_MK4_4X4X8"));
+    benchmarker_mk4_4x4x8.set_times(RUNS)
+            .set_dtype(0, dtype::Int8{})
+            .set_dtype(1, dtype::Int8{})
+            .set_dtype(2, dtype::Int16{})
+            .set_param(param)
+            .set_display(false);
+
+    auto run = [&](size_t M, size_t N, size_t K) {
+        auto default_used = benchmarker.exec({{M, K}, {K, N}, {}}) / RUNS;
+        auto mk_used = benchmarker_mk4.exec(
+                               {{M / 4, K / 4, 4, 4}, {K / 4, N, 4}, {}}) /
+                       RUNS;
+        auto mk4_4x4x8_used =
+                benchmarker_mk4_4x4x8.exec(
+                        {{M / 4, K / 4, 4, 4}, {K / 4, N, 4}, {}}) /
+                RUNS;
+        float computations = 2.f * M * K * N * 1e-6;
+        printf("run: {%zu{M} %zu{K} %zu{N}} normal: %f ms %f Gflops mk4: %f ms "
+               "%f Gflops speedup: %f, mk4_4x4x8 %f Gflops %f ms speedup: %f\n",
+               M, K, N, default_used, computations / default_used, mk_used,
+               computations / mk_used, default_used / mk_used,
+               computations / mk4_4x4x8_used, mk4_4x4x8_used , mk4_4x4x8_used/mk_used);
+    };
+
+    run(384, 384, 384);
+    run(512, 512, 512);
+    run(1024, 1024, 384);
+    run(256, 256, 384);
+    for(int m = 32; m <= 512;m*=2)
+    for(int n = 32; n <= 512;n*=2)
+    for(int k = 32; k < 512;k*=2){
+        run(m,n,k);
+    }
+}
 TEST_F(AARCH64, BENCHMARK_MATRIX_MUL_INT16_4X4X16) {
     constexpr size_t RUNS = 50;
     param::MatrixMul param;