From afd38f0c617d6f89b2b4532c6c44f116617e2b6f Mon Sep 17 00:00:00 2001 From: Felix Thomasmathibalan Date: Wed, 27 Sep 2023 17:46:17 +0100 Subject: Apply clang-format on repository Code is formatted as per a revised clang format configuration file(not part of this delivery). Version 14.0.6 is used. Exclusion List: - files with .cl extension - files that are not strictly C/C++ (e.g. Android.bp, Sconscript ...) And the following directories - compute_kernel_writer/validation/ - tests/ - include/ - src/core/NEON/kernels/convolution/ - src/core/NEON/kernels/arm_gemm/ - src/core/NEON/kernels/arm_conv/ - data/ There will be a follow up for formatting of .cl files and the files under tests/ and compute_kernel_writer/validation/. Signed-off-by: Felix Thomasmathibalan Change-Id: Ib7eb1fcf4e7537b9feaefcfc15098a804a3fde0a Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10391 Benchmark: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gunes Bayir --- .../kernels/gemm_matrix_mul/generic/neon/fp16.cpp | 472 +++++------ .../kernels/gemm_matrix_mul/generic/neon/fp32.cpp | 15 +- .../kernels/gemm_matrix_mul/generic/neon/impl.cpp | 898 +++++++++++---------- .../kernels/gemm_matrix_mul/generic/neon/impl.h | 7 +- src/cpu/kernels/gemm_matrix_mul/list.h | 5 +- 5 files changed, 715 insertions(+), 682 deletions(-) (limited to 'src/cpu/kernels/gemm_matrix_mul') diff --git a/src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp b/src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp index 8fd79f9287..60fda511e3 100644 --- a/src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp +++ b/src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp @@ -32,7 +32,8 @@ namespace arm_compute { namespace cpu { -void vector_matrix_multiply_f16(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha) +void vector_matrix_multiply_f16( + const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha) { const auto width_matrix_b = static_cast(dst->info()->dimension(0)); const auto in_b_stride = static_cast(rhs->info()->strides_in_bytes()[1] / rhs->info()->element_size()); @@ -42,7 +43,8 @@ void vector_matrix_multiply_f16(const ITensor *lhs, const ITensor *rhs, ITensor const int window_start_x = 32 * info.thread_id; const int window_step_x = 32 * info.num_threads; const int window_end_x = ceil_to_multiple(width_matrix_b - window_start_x, window_step_x) + window_start_x; - ARM_COMPUTE_ERROR_ON_MSG((window_end_x - window_start_x) % window_step_x, " (window_end_x - window_start_x) must be multiple of window_step_x"); + ARM_COMPUTE_ERROR_ON_MSG((window_end_x - window_start_x) % window_step_x, + " (window_end_x - window_start_x) must be multiple of window_step_x"); Window win_out(window); win_out.set(Window::DimX, Window::Dimension(0, 1, 1)); @@ -55,7 +57,7 @@ void vector_matrix_multiply_f16(const ITensor *lhs, const ITensor *rhs, ITensor Window win_b; // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 // This scenario can happen when the the matrix multiplication is used to perform a convolution operation - if(rhs->info()->num_dimensions() >= 3) + if (rhs->info()->num_dimensions() >= 3) { win_b = window; } @@ -70,169 +72,172 @@ void vector_matrix_multiply_f16(const ITensor *lhs, const ITensor *rhs, ITensor const float16x8_t alpha_f16 = vdupq_n_f16(alpha); - execute_window_loop(win_out, [&](const Coordinates &) - { - int x = window_start_x; - // Here we don't check for x lower equal than (window_end_x - window_step_x) because of - // window_end_x is computed above which may cause out-of-bound writes to the dst. - for(; x < (window_end_x - window_step_x); x += window_step_x) + execute_window_loop( + win_out, + [&](const Coordinates &) { - if(x > width_matrix_b) + int x = window_start_x; + // Here we don't check for x lower equal than (window_end_x - window_step_x) because of + // window_end_x is computed above which may cause out-of-bound writes to the dst. + for (; x < (window_end_x - window_step_x); x += window_step_x) { - return; - } - - auto matrix_b = reinterpret_cast(inb.ptr()) + x; + if (x > width_matrix_b) + { + return; + } - float16x8_t acc0 = vdupq_n_f16(0.f); - float16x8_t acc1 = vdupq_n_f16(0.f); - float16x8_t acc2 = vdupq_n_f16(0.f); - float16x8_t acc3 = vdupq_n_f16(0.f); + auto matrix_b = reinterpret_cast(inb.ptr()) + x; - auto vec_a = reinterpret_cast(ina.ptr()); - const float16_t *vec_a_end_addr = vec_a + num_elems_vec_a; - for(; vec_a <= (vec_a_end_addr - 4);) - { - const float16x4_t a0l = vld1_f16(vec_a); - - float16x8_t b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride); - float16x8_t b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride); - float16x8_t b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride); - float16x8_t b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride); - float16x8_t b10 = vld1q_f16(matrix_b + 0 + 1 * in_b_stride); - float16x8_t b11 = vld1q_f16(matrix_b + 8 + 1 * in_b_stride); - float16x8_t b12 = vld1q_f16(matrix_b + 16 + 1 * in_b_stride); - float16x8_t b13 = vld1q_f16(matrix_b + 24 + 1 * in_b_stride); - - acc0 = vaddq_f16(acc0, vmulq_lane_f16(b00, a0l, 0)); - acc1 = vaddq_f16(acc1, vmulq_lane_f16(b01, a0l, 0)); - acc2 = vaddq_f16(acc2, vmulq_lane_f16(b02, a0l, 0)); - acc3 = vaddq_f16(acc3, vmulq_lane_f16(b03, a0l, 0)); - acc0 = vaddq_f16(acc0, vmulq_lane_f16(b10, a0l, 1)); - acc1 = vaddq_f16(acc1, vmulq_lane_f16(b11, a0l, 1)); - acc2 = vaddq_f16(acc2, vmulq_lane_f16(b12, a0l, 1)); - acc3 = vaddq_f16(acc3, vmulq_lane_f16(b13, a0l, 1)); - - matrix_b += 2 * in_b_stride; - - b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride); - b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride); - b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride); - b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride); - b10 = vld1q_f16(matrix_b + 0 + 1 * in_b_stride); - b11 = vld1q_f16(matrix_b + 8 + 1 * in_b_stride); - b12 = vld1q_f16(matrix_b + 16 + 1 * in_b_stride); - b13 = vld1q_f16(matrix_b + 24 + 1 * in_b_stride); - - acc0 = vaddq_f16(acc0, vmulq_lane_f16(b00, a0l, 2)); - acc1 = vaddq_f16(acc1, vmulq_lane_f16(b01, a0l, 2)); - acc2 = vaddq_f16(acc2, vmulq_lane_f16(b02, a0l, 2)); - acc3 = vaddq_f16(acc3, vmulq_lane_f16(b03, a0l, 2)); - acc0 = vaddq_f16(acc0, vmulq_lane_f16(b10, a0l, 3)); - acc1 = vaddq_f16(acc1, vmulq_lane_f16(b11, a0l, 3)); - acc2 = vaddq_f16(acc2, vmulq_lane_f16(b12, a0l, 3)); - acc3 = vaddq_f16(acc3, vmulq_lane_f16(b13, a0l, 3)); - - vec_a += 4; - matrix_b += 2 * in_b_stride; - } + float16x8_t acc0 = vdupq_n_f16(0.f); + float16x8_t acc1 = vdupq_n_f16(0.f); + float16x8_t acc2 = vdupq_n_f16(0.f); + float16x8_t acc3 = vdupq_n_f16(0.f); - for(; vec_a < vec_a_end_addr; ++vec_a) - { - const float16_t a0 = *vec_a; - const float16x8_t b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride); - const float16x8_t b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride); - const float16x8_t b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride); - const float16x8_t b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride); - - acc0 = vaddq_f16(acc0, vmulq_n_f16(b00, a0)); - acc1 = vaddq_f16(acc1, vmulq_n_f16(b01, a0)); - acc2 = vaddq_f16(acc2, vmulq_n_f16(b02, a0)); - acc3 = vaddq_f16(acc3, vmulq_n_f16(b03, a0)); - - matrix_b += in_b_stride; - } + auto vec_a = reinterpret_cast(ina.ptr()); + const float16_t *vec_a_end_addr = vec_a + num_elems_vec_a; + for (; vec_a <= (vec_a_end_addr - 4);) + { + const float16x4_t a0l = vld1_f16(vec_a); + + float16x8_t b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride); + float16x8_t b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride); + float16x8_t b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride); + float16x8_t b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride); + float16x8_t b10 = vld1q_f16(matrix_b + 0 + 1 * in_b_stride); + float16x8_t b11 = vld1q_f16(matrix_b + 8 + 1 * in_b_stride); + float16x8_t b12 = vld1q_f16(matrix_b + 16 + 1 * in_b_stride); + float16x8_t b13 = vld1q_f16(matrix_b + 24 + 1 * in_b_stride); + + acc0 = vaddq_f16(acc0, vmulq_lane_f16(b00, a0l, 0)); + acc1 = vaddq_f16(acc1, vmulq_lane_f16(b01, a0l, 0)); + acc2 = vaddq_f16(acc2, vmulq_lane_f16(b02, a0l, 0)); + acc3 = vaddq_f16(acc3, vmulq_lane_f16(b03, a0l, 0)); + acc0 = vaddq_f16(acc0, vmulq_lane_f16(b10, a0l, 1)); + acc1 = vaddq_f16(acc1, vmulq_lane_f16(b11, a0l, 1)); + acc2 = vaddq_f16(acc2, vmulq_lane_f16(b12, a0l, 1)); + acc3 = vaddq_f16(acc3, vmulq_lane_f16(b13, a0l, 1)); + + matrix_b += 2 * in_b_stride; + + b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride); + b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride); + b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride); + b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride); + b10 = vld1q_f16(matrix_b + 0 + 1 * in_b_stride); + b11 = vld1q_f16(matrix_b + 8 + 1 * in_b_stride); + b12 = vld1q_f16(matrix_b + 16 + 1 * in_b_stride); + b13 = vld1q_f16(matrix_b + 24 + 1 * in_b_stride); + + acc0 = vaddq_f16(acc0, vmulq_lane_f16(b00, a0l, 2)); + acc1 = vaddq_f16(acc1, vmulq_lane_f16(b01, a0l, 2)); + acc2 = vaddq_f16(acc2, vmulq_lane_f16(b02, a0l, 2)); + acc3 = vaddq_f16(acc3, vmulq_lane_f16(b03, a0l, 2)); + acc0 = vaddq_f16(acc0, vmulq_lane_f16(b10, a0l, 3)); + acc1 = vaddq_f16(acc1, vmulq_lane_f16(b11, a0l, 3)); + acc2 = vaddq_f16(acc2, vmulq_lane_f16(b12, a0l, 3)); + acc3 = vaddq_f16(acc3, vmulq_lane_f16(b13, a0l, 3)); + + vec_a += 4; + matrix_b += 2 * in_b_stride; + } - // Multiply by the weight of matrix product (alpha) - if(multiply_alpha) - { - acc0 = vmulq_f16(acc0, alpha_f16); - acc1 = vmulq_f16(acc1, alpha_f16); - acc2 = vmulq_f16(acc2, alpha_f16); - acc3 = vmulq_f16(acc3, alpha_f16); - } + for (; vec_a < vec_a_end_addr; ++vec_a) + { + const float16_t a0 = *vec_a; + const float16x8_t b00 = vld1q_f16(matrix_b + 0 + 0 * in_b_stride); + const float16x8_t b01 = vld1q_f16(matrix_b + 8 + 0 * in_b_stride); + const float16x8_t b02 = vld1q_f16(matrix_b + 16 + 0 * in_b_stride); + const float16x8_t b03 = vld1q_f16(matrix_b + 24 + 0 * in_b_stride); + + acc0 = vaddq_f16(acc0, vmulq_n_f16(b00, a0)); + acc1 = vaddq_f16(acc1, vmulq_n_f16(b01, a0)); + acc2 = vaddq_f16(acc2, vmulq_n_f16(b02, a0)); + acc3 = vaddq_f16(acc3, vmulq_n_f16(b03, a0)); + + matrix_b += in_b_stride; + } - auto vec_out = reinterpret_cast(out.ptr()) + x; + // Multiply by the weight of matrix product (alpha) + if (multiply_alpha) + { + acc0 = vmulq_f16(acc0, alpha_f16); + acc1 = vmulq_f16(acc1, alpha_f16); + acc2 = vmulq_f16(acc2, alpha_f16); + acc3 = vmulq_f16(acc3, alpha_f16); + } - vst1q_f16(vec_out + 0, acc0); - vst1q_f16(vec_out + 8, acc1); - vst1q_f16(vec_out + 16, acc2); - vst1q_f16(vec_out + 24, acc3); - } + auto vec_out = reinterpret_cast(out.ptr()) + x; - for(; x < window_end_x; ++x) - { - if(x > width_matrix_b) - { - return; + vst1q_f16(vec_out + 0, acc0); + vst1q_f16(vec_out + 8, acc1); + vst1q_f16(vec_out + 16, acc2); + vst1q_f16(vec_out + 24, acc3); } - auto matrix_b = reinterpret_cast(inb.ptr()) + x; + for (; x < window_end_x; ++x) + { + if (x > width_matrix_b) + { + return; + } - float16x4_t vacc = vdup_n_f16(0.f); + auto matrix_b = reinterpret_cast(inb.ptr()) + x; - auto vec_a = reinterpret_cast(ina.ptr()); - const float16_t *vec_a_end_addr = vec_a + num_elems_vec_a; - for(; vec_a <= (vec_a_end_addr - 4); vec_a += 4) - { - const float16x4_t a0l = vld1_f16(vec_a); + float16x4_t vacc = vdup_n_f16(0.f); - const float16x4_t b_col = + auto vec_a = reinterpret_cast(ina.ptr()); + const float16_t *vec_a_end_addr = vec_a + num_elems_vec_a; + for (; vec_a <= (vec_a_end_addr - 4); vec_a += 4) { - *(matrix_b + 0 * in_b_stride), - *(matrix_b + 1 * in_b_stride), - *(matrix_b + 2 * in_b_stride), - *(matrix_b + 3 * in_b_stride), - }; + const float16x4_t a0l = vld1_f16(vec_a); - vacc = vadd_f16(vacc, vmul_f16(a0l, b_col)); + const float16x4_t b_col = { + *(matrix_b + 0 * in_b_stride), + *(matrix_b + 1 * in_b_stride), + *(matrix_b + 2 * in_b_stride), + *(matrix_b + 3 * in_b_stride), + }; - matrix_b += 4 * in_b_stride; - } + vacc = vadd_f16(vacc, vmul_f16(a0l, b_col)); - float16_t acc = vget_lane_f16(vacc, 0) + vget_lane_f16(vacc, 1) + vget_lane_f16(vacc, 2) + vget_lane_f16(vacc, 3); + matrix_b += 4 * in_b_stride; + } - for(; vec_a < vec_a_end_addr; ++vec_a) - { - const float16_t a0 = *vec_a; - const float16_t b00 = *matrix_b; + float16_t acc = + vget_lane_f16(vacc, 0) + vget_lane_f16(vacc, 1) + vget_lane_f16(vacc, 2) + vget_lane_f16(vacc, 3); - acc += b00 * a0; + for (; vec_a < vec_a_end_addr; ++vec_a) + { + const float16_t a0 = *vec_a; + const float16_t b00 = *matrix_b; - matrix_b += in_b_stride; - } + acc += b00 * a0; - // Multiply by the weight of matrix product (alpha) - if(multiply_alpha) - { - acc *= static_cast(alpha); - } + matrix_b += in_b_stride; + } - auto vec_out = reinterpret_cast(out.ptr()) + x; + // Multiply by the weight of matrix product (alpha) + if (multiply_alpha) + { + acc *= static_cast(alpha); + } - *(vec_out) = acc; - } - }, - ina, inb, out); + auto vec_out = reinterpret_cast(out.ptr()) + x; + + *(vec_out) = acc; + } + }, + ina, inb, out); } -void matrix_matrix_multiply_f16(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha) +void matrix_matrix_multiply_f16( + const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha) { ARM_COMPUTE_UNUSED(info); - const int out_width = static_cast(dst->info()->dimension(0)); - const int out_height = static_cast(dst->info()->dimension(1)); - const size_t in_b_stride = rhs->info()->strides_in_bytes()[1] / data_size_from_type(rhs->info()->data_type()); - const size_t out_stride = dst->info()->strides_in_bytes()[1] / data_size_from_type(dst->info()->data_type()); + const int out_width = static_cast(dst->info()->dimension(0)); + const int out_height = static_cast(dst->info()->dimension(1)); + const size_t in_b_stride = rhs->info()->strides_in_bytes()[1] / data_size_from_type(rhs->info()->data_type()); + const size_t out_stride = dst->info()->strides_in_bytes()[1] / data_size_from_type(dst->info()->data_type()); const int num_elems_matrix_b_x = rhs->info()->dimension(0); // Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the dst matrix @@ -243,7 +248,7 @@ void matrix_matrix_multiply_f16(const ITensor *lhs, const ITensor *rhs, ITensor Window win_b; // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 // This scenario can happen when the the matrix multiplication is used to perform a convolution operation - if(rhs->info()->num_dimensions() >= 3) + if (rhs->info()->num_dimensions() >= 3) { win_b = window; } @@ -259,22 +264,16 @@ void matrix_matrix_multiply_f16(const ITensor *lhs, const ITensor *rhs, ITensor const float16x8_t alpha_f16 = vdupq_n_f16(alpha); - execute_window_loop(window, [&](const Coordinates & id) - { - const auto *mtx_a0 = reinterpret_cast(ina.ptr()); - const auto *mtx_b0 = reinterpret_cast(inb.ptr()); - auto *mtx_out = reinterpret_cast(out.ptr()); - float16x8x4_t c = + execute_window_loop( + window, + [&](const Coordinates &id) { - { - vdupq_n_f16(0.f), - vdupq_n_f16(0.f), - vdupq_n_f16(0.f), - vdupq_n_f16(0.f) - } - }; + const auto *mtx_a0 = reinterpret_cast(ina.ptr()); + const auto *mtx_b0 = reinterpret_cast(inb.ptr()); + auto *mtx_out = reinterpret_cast(out.ptr()); + float16x8x4_t c = {{vdupq_n_f16(0.f), vdupq_n_f16(0.f), vdupq_n_f16(0.f), vdupq_n_f16(0.f)}}; - /* + /* This kernel puts the values in a 4x4 block of Matrix A on the same row (Interleaved values) |a00 a01 a02 a03 | a04 a05 a06 a07| |a10 a11 a12 a13 | a14 a15 a16 a17| @@ -302,111 +301,118 @@ void matrix_matrix_multiply_f16(const ITensor *lhs, const ITensor *rhs, ITensor The size of the dst tensor's XY-plane must be the following shape [ width * 8, height / 8 ]. All other dimensions must have the same size. */ - const float16_t *mtx_b0_end_addr = mtx_b0 + num_elems_matrix_b_x; - - for(; mtx_b0 <= (mtx_b0_end_addr - 32);) - - { - const float16x8_t p00 = vld1q_f16(mtx_a0); - const float16x8_t p02 = vld1q_f16(mtx_a0 + 8); - - const float16x8_t q00 = vld1q_f16(mtx_b0); - const float16x8_t q02 = vld1q_f16(mtx_b0 + 8); - const float16x8_t q04 = vld1q_f16(mtx_b0 + 16); - const float16x8_t q06 = vld1q_f16(mtx_b0 + 24); - - c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q00, vgetq_lane_f16(p00, 0))); - c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q00, vgetq_lane_f16(p00, 1))); - c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q00, vgetq_lane_f16(p00, 2))); - c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q00, vgetq_lane_f16(p00, 3))); + const float16_t *mtx_b0_end_addr = mtx_b0 + num_elems_matrix_b_x; - c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q02, vgetq_lane_f16(p00, 4))); - c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q02, vgetq_lane_f16(p00, 5))); - c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q02, vgetq_lane_f16(p00, 6))); - c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q02, vgetq_lane_f16(p00, 7))); + for (; mtx_b0 <= (mtx_b0_end_addr - 32);) - c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q04, vgetq_lane_f16(p02, 0))); - c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q04, vgetq_lane_f16(p02, 1))); - c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q04, vgetq_lane_f16(p02, 2))); - c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q04, vgetq_lane_f16(p02, 3))); - - c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q06, vgetq_lane_f16(p02, 4))); - c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q06, vgetq_lane_f16(p02, 5))); - c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q06, vgetq_lane_f16(p02, 6))); - c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q06, vgetq_lane_f16(p02, 7))); - - mtx_a0 += 16; - mtx_b0 += 32; - } + { + const float16x8_t p00 = vld1q_f16(mtx_a0); + const float16x8_t p02 = vld1q_f16(mtx_a0 + 8); + + const float16x8_t q00 = vld1q_f16(mtx_b0); + const float16x8_t q02 = vld1q_f16(mtx_b0 + 8); + const float16x8_t q04 = vld1q_f16(mtx_b0 + 16); + const float16x8_t q06 = vld1q_f16(mtx_b0 + 24); + + c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q00, vgetq_lane_f16(p00, 0))); + c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q00, vgetq_lane_f16(p00, 1))); + c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q00, vgetq_lane_f16(p00, 2))); + c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q00, vgetq_lane_f16(p00, 3))); + + c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q02, vgetq_lane_f16(p00, 4))); + c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q02, vgetq_lane_f16(p00, 5))); + c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q02, vgetq_lane_f16(p00, 6))); + c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q02, vgetq_lane_f16(p00, 7))); + + c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q04, vgetq_lane_f16(p02, 0))); + c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q04, vgetq_lane_f16(p02, 1))); + c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q04, vgetq_lane_f16(p02, 2))); + c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q04, vgetq_lane_f16(p02, 3))); + + c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q06, vgetq_lane_f16(p02, 4))); + c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q06, vgetq_lane_f16(p02, 5))); + c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q06, vgetq_lane_f16(p02, 6))); + c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q06, vgetq_lane_f16(p02, 7))); + + mtx_a0 += 16; + mtx_b0 += 32; + } - for(; mtx_b0 < mtx_b0_end_addr;) + for (; mtx_b0 < mtx_b0_end_addr;) - { - const float16x4_t p00 = vld1_f16(mtx_a0); - const float16x8_t q00 = vld1q_f16(mtx_b0); + { + const float16x4_t p00 = vld1_f16(mtx_a0); + const float16x8_t q00 = vld1q_f16(mtx_b0); - c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q00, vget_lane_f16(p00, 0))); - c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q00, vget_lane_f16(p00, 1))); - c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q00, vget_lane_f16(p00, 2))); - c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q00, vget_lane_f16(p00, 3))); + c.val[0] = vaddq_f16(c.val[0], vmulq_n_f16(q00, vget_lane_f16(p00, 0))); + c.val[1] = vaddq_f16(c.val[1], vmulq_n_f16(q00, vget_lane_f16(p00, 1))); + c.val[2] = vaddq_f16(c.val[2], vmulq_n_f16(q00, vget_lane_f16(p00, 2))); + c.val[3] = vaddq_f16(c.val[3], vmulq_n_f16(q00, vget_lane_f16(p00, 3))); - mtx_a0 += 4; - mtx_b0 += 8; - } + mtx_a0 += 4; + mtx_b0 += 8; + } - if(multiply_alpha) - { - c.val[0] = vmulq_f16(c.val[0], alpha_f16); - c.val[1] = vmulq_f16(c.val[1], alpha_f16); - c.val[2] = vmulq_f16(c.val[2], alpha_f16); - c.val[3] = vmulq_f16(c.val[3], alpha_f16); - } + if (multiply_alpha) + { + c.val[0] = vmulq_f16(c.val[0], alpha_f16); + c.val[1] = vmulq_f16(c.val[1], alpha_f16); + c.val[2] = vmulq_f16(c.val[2], alpha_f16); + c.val[3] = vmulq_f16(c.val[3], alpha_f16); + } - if(id.x() < (out_width - 8)) - { - vst1q_f16(mtx_out, c.val[0]); - if(id.y() + 1 < out_height) + if (id.x() < (out_width - 8)) { - vst1q_f16(mtx_out + 1 * out_stride, c.val[1]); - if(id.y() + 2 < out_height) + vst1q_f16(mtx_out, c.val[0]); + if (id.y() + 1 < out_height) { - vst1q_f16(mtx_out + 2 * out_stride, c.val[2]); - if(id.y() + 3 < out_height) + vst1q_f16(mtx_out + 1 * out_stride, c.val[1]); + if (id.y() + 2 < out_height) { - vst1q_f16(mtx_out + 3 * out_stride, c.val[3]); + vst1q_f16(mtx_out + 2 * out_stride, c.val[2]); + if (id.y() + 3 < out_height) + { + vst1q_f16(mtx_out + 3 * out_stride, c.val[3]); + } } } } - } - else - { - // Left-over columns - const int columns_left = out_width - id.x(); - for(int x = 0; x < columns_left; ++x) + else { - *(mtx_out + x) = c.val[0][x]; - if(id.y() + 1 < out_height) + // Left-over columns + const int columns_left = out_width - id.x(); + for (int x = 0; x < columns_left; ++x) { - *(mtx_out + x + 1 * out_stride) = c.val[1][x]; - if(id.y() + 2 < out_height) + *(mtx_out + x) = c.val[0][x]; + if (id.y() + 1 < out_height) { - *(mtx_out + x + 2 * out_stride) = c.val[2][x]; - if(id.y() + 3 < out_height) + *(mtx_out + x + 1 * out_stride) = c.val[1][x]; + if (id.y() + 2 < out_height) { - *(mtx_out + x + 3 * out_stride) = c.val[3][x]; + *(mtx_out + x + 2 * out_stride) = c.val[2][x]; + if (id.y() + 3 < out_height) + { + *(mtx_out + x + 3 * out_stride) = c.val[3][x]; + } } } } } - } - }, - ina, inb, out); + }, + ina, inb, out); } -void neon_fp16_gemm_matrix_mul(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha, const bool is_dst_vector) +void neon_fp16_gemm_matrix_mul(const ITensor *lhs, + const ITensor *rhs, + ITensor *dst, + const Window &window, + const ThreadInfo &info, + float alpha, + const bool is_dst_vector) { - return (is_dst_vector) ? vector_matrix_multiply_f16(lhs, rhs, dst, window, info, alpha) : matrix_matrix_multiply_f16(lhs, rhs, dst, window, info, alpha); + return (is_dst_vector) ? vector_matrix_multiply_f16(lhs, rhs, dst, window, info, alpha) + : matrix_matrix_multiply_f16(lhs, rhs, dst, window, info, alpha); } -} // namespce cpu +} // namespace cpu } // namespace arm_compute #endif //__ARM_FEATURE_FP16_VECTOR_ARITHMETIC diff --git a/src/cpu/kernels/gemm_matrix_mul/generic/neon/fp32.cpp b/src/cpu/kernels/gemm_matrix_mul/generic/neon/fp32.cpp index 9c1f6f3c0f..e12a312280 100644 --- a/src/cpu/kernels/gemm_matrix_mul/generic/neon/fp32.cpp +++ b/src/cpu/kernels/gemm_matrix_mul/generic/neon/fp32.cpp @@ -28,9 +28,16 @@ namespace arm_compute { namespace cpu { -void neon_fp32_gemm_matrix_mul(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha, const bool is_dst_vector) +void neon_fp32_gemm_matrix_mul(const ITensor *lhs, + const ITensor *rhs, + ITensor *dst, + const Window &window, + const ThreadInfo &info, + float alpha, + const bool is_dst_vector) { - return (is_dst_vector) ? vector_matrix_multiply_f32(lhs, rhs, dst, window, info, alpha) : matrix_matrix_multiply_f32(lhs, rhs, dst, window, info, alpha); + return (is_dst_vector) ? vector_matrix_multiply_f32(lhs, rhs, dst, window, info, alpha) + : matrix_matrix_multiply_f32(lhs, rhs, dst, window, info, alpha); } -} // namespce cpu -} // namespace arm_compute \ No newline at end of file +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.cpp b/src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.cpp index 0051d3d9dc..404d070a37 100644 --- a/src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.cpp +++ b/src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.cpp @@ -23,6 +23,7 @@ */ #include "src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.h" + #include "src/core/utils/helpers/float_ops.h" #include @@ -31,10 +32,12 @@ namespace arm_compute { namespace cpu { -void vector_matrix_multiply_f32(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha) +void vector_matrix_multiply_f32( + const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha) { - const auto width_matrix_b = static_cast(dst->info()->dimension(0)); - const auto in_b_stride = static_cast(rhs->info()->strides_in_bytes()[1] / data_size_from_type(rhs->info()->data_type())); + const auto width_matrix_b = static_cast(dst->info()->dimension(0)); + const auto in_b_stride = + static_cast(rhs->info()->strides_in_bytes()[1] / data_size_from_type(rhs->info()->data_type())); const auto num_elems_vec_a = static_cast(lhs->info()->dimension(0)); // The implementation computes 16 elements per iteration @@ -54,7 +57,7 @@ void vector_matrix_multiply_f32(const ITensor *lhs, const ITensor *rhs, ITensor Window win_b; // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 // This scenario can happen when the the matrix multiplication is used to perform a convolution operation - if(rhs->info()->num_dimensions() >= 3) + if (rhs->info()->num_dimensions() >= 3) { win_b = window; } @@ -69,209 +72,220 @@ void vector_matrix_multiply_f32(const ITensor *lhs, const ITensor *rhs, ITensor const float32x4_t alpha_f32 = vdupq_n_f32(alpha); - execute_window_loop(win_out, [&](const Coordinates &) - { - int x = window_start_x; - // Here we don't check for x lower equal than (window_end_x - window_step_x) because of - // window_end_x is computed above which may cause out-of-bound writes to the dst. - for(; x < (window_end_x - window_step_x); x += window_step_x) + execute_window_loop( + win_out, + [&](const Coordinates &) { - if(x > width_matrix_b) + int x = window_start_x; + // Here we don't check for x lower equal than (window_end_x - window_step_x) because of + // window_end_x is computed above which may cause out-of-bound writes to the dst. + for (; x < (window_end_x - window_step_x); x += window_step_x) { - return; - } + if (x > width_matrix_b) + { + return; + } - float32x4_t acc0 = vdupq_n_f32(0.f); - float32x4_t acc1 = vdupq_n_f32(0.f); - float32x4_t acc2 = vdupq_n_f32(0.f); - float32x4_t acc3 = vdupq_n_f32(0.f); + float32x4_t acc0 = vdupq_n_f32(0.f); + float32x4_t acc1 = vdupq_n_f32(0.f); + float32x4_t acc2 = vdupq_n_f32(0.f); + float32x4_t acc3 = vdupq_n_f32(0.f); - auto vec_a = reinterpret_cast(ina.ptr()); - auto matrix_b = reinterpret_cast(inb.ptr()) + x; + auto vec_a = reinterpret_cast(ina.ptr()); + auto matrix_b = reinterpret_cast(inb.ptr()) + x; #if __arm__ - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(vec_a))); - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(matrix_b))); - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(matrix_b + in_b_stride))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(vec_a))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(matrix_b))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(matrix_b + in_b_stride))); #endif /* __arm__ */ - auto vec_a_end_addr = vec_a + num_elems_vec_a; - for(; vec_a <= (vec_a_end_addr - 4);) - { - float32x2_t a0l = vld1_f32(vec_a); + auto vec_a_end_addr = vec_a + num_elems_vec_a; + for (; vec_a <= (vec_a_end_addr - 4);) + { + float32x2_t a0l = vld1_f32(vec_a); - float32x4_t b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); - float32x4_t b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); - float32x4_t b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); - float32x4_t b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); + float32x4_t b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); + float32x4_t b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); + float32x4_t b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); + float32x4_t b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); - float32x4_t b10 = vld1q_f32(matrix_b + 0 + 1 * in_b_stride); - float32x4_t b11 = vld1q_f32(matrix_b + 4 + 1 * in_b_stride); - float32x4_t b12 = vld1q_f32(matrix_b + 8 + 1 * in_b_stride); - float32x4_t b13 = vld1q_f32(matrix_b + 12 + 1 * in_b_stride); + float32x4_t b10 = vld1q_f32(matrix_b + 0 + 1 * in_b_stride); + float32x4_t b11 = vld1q_f32(matrix_b + 4 + 1 * in_b_stride); + float32x4_t b12 = vld1q_f32(matrix_b + 8 + 1 * in_b_stride); + float32x4_t b13 = vld1q_f32(matrix_b + 12 + 1 * in_b_stride); #if __arm__ - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(vec_a))); - asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 1 * in_b_stride))); - asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 2 * in_b_stride))); - asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 3 * in_b_stride))); - asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 4 * in_b_stride))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(vec_a))); + asm volatile( + "PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 1 * in_b_stride))); + asm volatile( + "PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 2 * in_b_stride))); + asm volatile( + "PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 3 * in_b_stride))); + asm volatile( + "PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 4 * in_b_stride))); #endif /* __arm__ */ - acc0 = vmlaq_lane_f32(acc0, b00, a0l, 0); - acc1 = vmlaq_lane_f32(acc1, b01, a0l, 0); - acc2 = vmlaq_lane_f32(acc2, b02, a0l, 0); - acc3 = vmlaq_lane_f32(acc3, b03, a0l, 0); + acc0 = vmlaq_lane_f32(acc0, b00, a0l, 0); + acc1 = vmlaq_lane_f32(acc1, b01, a0l, 0); + acc2 = vmlaq_lane_f32(acc2, b02, a0l, 0); + acc3 = vmlaq_lane_f32(acc3, b03, a0l, 0); - acc0 = vmlaq_lane_f32(acc0, b10, a0l, 1); - acc1 = vmlaq_lane_f32(acc1, b11, a0l, 1); - acc2 = vmlaq_lane_f32(acc2, b12, a0l, 1); - acc3 = vmlaq_lane_f32(acc3, b13, a0l, 1); + acc0 = vmlaq_lane_f32(acc0, b10, a0l, 1); + acc1 = vmlaq_lane_f32(acc1, b11, a0l, 1); + acc2 = vmlaq_lane_f32(acc2, b12, a0l, 1); + acc3 = vmlaq_lane_f32(acc3, b13, a0l, 1); - vec_a += 2; - matrix_b += 2 * in_b_stride; + vec_a += 2; + matrix_b += 2 * in_b_stride; - a0l = vld1_f32(vec_a); + a0l = vld1_f32(vec_a); - b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); - b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); - b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); - b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); + b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); + b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); + b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); + b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); - b10 = vld1q_f32(matrix_b + 0 + 1 * in_b_stride); - b11 = vld1q_f32(matrix_b + 4 + 1 * in_b_stride); - b12 = vld1q_f32(matrix_b + 8 + 1 * in_b_stride); - b13 = vld1q_f32(matrix_b + 12 + 1 * in_b_stride); + b10 = vld1q_f32(matrix_b + 0 + 1 * in_b_stride); + b11 = vld1q_f32(matrix_b + 4 + 1 * in_b_stride); + b12 = vld1q_f32(matrix_b + 8 + 1 * in_b_stride); + b13 = vld1q_f32(matrix_b + 12 + 1 * in_b_stride); - acc0 = vmlaq_lane_f32(acc0, b00, a0l, 0); - acc1 = vmlaq_lane_f32(acc1, b01, a0l, 0); - acc2 = vmlaq_lane_f32(acc2, b02, a0l, 0); - acc3 = vmlaq_lane_f32(acc3, b03, a0l, 0); + acc0 = vmlaq_lane_f32(acc0, b00, a0l, 0); + acc1 = vmlaq_lane_f32(acc1, b01, a0l, 0); + acc2 = vmlaq_lane_f32(acc2, b02, a0l, 0); + acc3 = vmlaq_lane_f32(acc3, b03, a0l, 0); - acc0 = vmlaq_lane_f32(acc0, b10, a0l, 1); - acc1 = vmlaq_lane_f32(acc1, b11, a0l, 1); - acc2 = vmlaq_lane_f32(acc2, b12, a0l, 1); - acc3 = vmlaq_lane_f32(acc3, b13, a0l, 1); + acc0 = vmlaq_lane_f32(acc0, b10, a0l, 1); + acc1 = vmlaq_lane_f32(acc1, b11, a0l, 1); + acc2 = vmlaq_lane_f32(acc2, b12, a0l, 1); + acc3 = vmlaq_lane_f32(acc3, b13, a0l, 1); - vec_a += 2; - matrix_b += 2 * in_b_stride; - } + vec_a += 2; + matrix_b += 2 * in_b_stride; + } - for(; vec_a < vec_a_end_addr; ++vec_a) - { - const float a0 = *vec_a; + for (; vec_a < vec_a_end_addr; ++vec_a) + { + const float a0 = *vec_a; - const float32x4_t b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); - const float32x4_t b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); - const float32x4_t b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); - const float32x4_t b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); + const float32x4_t b00 = vld1q_f32(matrix_b + 0 + 0 * in_b_stride); + const float32x4_t b01 = vld1q_f32(matrix_b + 4 + 0 * in_b_stride); + const float32x4_t b02 = vld1q_f32(matrix_b + 8 + 0 * in_b_stride); + const float32x4_t b03 = vld1q_f32(matrix_b + 12 + 0 * in_b_stride); - acc0 = vmlaq_n_f32(acc0, b00, a0); - acc1 = vmlaq_n_f32(acc1, b01, a0); - acc2 = vmlaq_n_f32(acc2, b02, a0); - acc3 = vmlaq_n_f32(acc3, b03, a0); + acc0 = vmlaq_n_f32(acc0, b00, a0); + acc1 = vmlaq_n_f32(acc1, b01, a0); + acc2 = vmlaq_n_f32(acc2, b02, a0); + acc3 = vmlaq_n_f32(acc3, b03, a0); - matrix_b += in_b_stride; - } + matrix_b += in_b_stride; + } - // Multiply by the weight of matrix product (alpha) - if(multiply_alpha) - { - acc0 = vmulq_f32(acc0, alpha_f32); - acc1 = vmulq_f32(acc1, alpha_f32); - acc2 = vmulq_f32(acc2, alpha_f32); - acc3 = vmulq_f32(acc3, alpha_f32); - } + // Multiply by the weight of matrix product (alpha) + if (multiply_alpha) + { + acc0 = vmulq_f32(acc0, alpha_f32); + acc1 = vmulq_f32(acc1, alpha_f32); + acc2 = vmulq_f32(acc2, alpha_f32); + acc3 = vmulq_f32(acc3, alpha_f32); + } - const auto vec_out = reinterpret_cast(out.ptr()) + x; + const auto vec_out = reinterpret_cast(out.ptr()) + x; - vst1q_f32(vec_out + 0, acc0); - vst1q_f32(vec_out + 4, acc1); - vst1q_f32(vec_out + 8, acc2); - vst1q_f32(vec_out + 12, acc3); - } + vst1q_f32(vec_out + 0, acc0); + vst1q_f32(vec_out + 4, acc1); + vst1q_f32(vec_out + 8, acc2); + vst1q_f32(vec_out + 12, acc3); + } - // Left-over loop - for(; x < window_end_x; ++x) - { - if(x > width_matrix_b) + // Left-over loop + for (; x < window_end_x; ++x) { - return; - } + if (x > width_matrix_b) + { + return; + } - float32x4_t vacc = vdupq_n_f32(0.f); + float32x4_t vacc = vdupq_n_f32(0.f); - auto vec_a = reinterpret_cast(ina.ptr()); - auto matrix_b = reinterpret_cast(inb.ptr()) + x; + auto vec_a = reinterpret_cast(ina.ptr()); + auto matrix_b = reinterpret_cast(inb.ptr()) + x; #if __arm__ - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(vec_a))); - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(matrix_b))); - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(matrix_b + in_b_stride))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(vec_a))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(matrix_b))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(matrix_b + in_b_stride))); #endif /* __arm__ */ - auto vec_a_end_addr = vec_a + num_elems_vec_a; - for(; vec_a <= (vec_a_end_addr - 4); vec_a += 4) - { - const float32x4_t a0l = vld1q_f32(vec_a); - - const float32x4_t b_col = + auto vec_a_end_addr = vec_a + num_elems_vec_a; + for (; vec_a <= (vec_a_end_addr - 4); vec_a += 4) { - *(matrix_b + 0 * in_b_stride), - *(matrix_b + 1 * in_b_stride), - *(matrix_b + 2 * in_b_stride), - *(matrix_b + 3 * in_b_stride), - }; + const float32x4_t a0l = vld1q_f32(vec_a); + + const float32x4_t b_col = { + *(matrix_b + 0 * in_b_stride), + *(matrix_b + 1 * in_b_stride), + *(matrix_b + 2 * in_b_stride), + *(matrix_b + 3 * in_b_stride), + }; #if __arm__ - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(vec_a))); - asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 1 * in_b_stride))); - asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 2 * in_b_stride))); - asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 3 * in_b_stride))); - asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 4 * in_b_stride))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(vec_a))); + asm volatile( + "PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 1 * in_b_stride))); + asm volatile( + "PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 2 * in_b_stride))); + asm volatile( + "PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 3 * in_b_stride))); + asm volatile( + "PLD [%0, #128*1]" ::"r"(reinterpret_cast(matrix_b + 4 * in_b_stride))); #endif /* __arm__ */ - vacc = vmlaq_f32(vacc, b_col, a0l); + vacc = vmlaq_f32(vacc, b_col, a0l); - matrix_b += 4 * in_b_stride; - } + matrix_b += 4 * in_b_stride; + } - float acc = vgetq_lane_f32(vacc, 0) + vgetq_lane_f32(vacc, 1) + vgetq_lane_f32(vacc, 2) + vgetq_lane_f32(vacc, 3); + float acc = vgetq_lane_f32(vacc, 0) + vgetq_lane_f32(vacc, 1) + vgetq_lane_f32(vacc, 2) + + vgetq_lane_f32(vacc, 3); - for(; vec_a < vec_a_end_addr; ++vec_a) - { - const float a0 = *vec_a; + for (; vec_a < vec_a_end_addr; ++vec_a) + { + const float a0 = *vec_a; - const float b00 = *matrix_b; + const float b00 = *matrix_b; - acc += b00 * a0; + acc += b00 * a0; - matrix_b += in_b_stride; - } + matrix_b += in_b_stride; + } - // Multiply by the weight of matrix product (alpha) - if(multiply_alpha) - { - acc *= alpha; - } + // Multiply by the weight of matrix product (alpha) + if (multiply_alpha) + { + acc *= alpha; + } - const auto vec_out = reinterpret_cast(out.ptr()) + x; + const auto vec_out = reinterpret_cast(out.ptr()) + x; - *vec_out = acc; - } - }, - ina, inb, out); + *vec_out = acc; + } + }, + ina, inb, out); } -void matrix_matrix_multiply_f32(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha) +void matrix_matrix_multiply_f32( + const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha) { ARM_COMPUTE_UNUSED(info); - const int out_width = static_cast(dst->info()->dimension(0)); - const int out_height = static_cast(dst->info()->dimension(1)); - const size_t in_b_stride = rhs->info()->strides_in_bytes()[1] / data_size_from_type(rhs->info()->data_type()); - const size_t out_stride1 = dst->info()->strides_in_bytes()[1] / data_size_from_type(dst->info()->data_type()); - const size_t out_stride2 = out_stride1 * 2; - const size_t out_stride3 = out_stride1 * 3; + const int out_width = static_cast(dst->info()->dimension(0)); + const int out_height = static_cast(dst->info()->dimension(1)); + const size_t in_b_stride = rhs->info()->strides_in_bytes()[1] / data_size_from_type(rhs->info()->data_type()); + const size_t out_stride1 = dst->info()->strides_in_bytes()[1] / data_size_from_type(dst->info()->data_type()); + const size_t out_stride2 = out_stride1 * 2; + const size_t out_stride3 = out_stride1 * 3; const int num_elems_matrix_b_x = rhs->info()->dimension(0); // Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the dst matrix @@ -282,7 +296,7 @@ void matrix_matrix_multiply_f32(const ITensor *lhs, const ITensor *rhs, ITensor Window win_b; // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 // This scenario can happen when the the matrix multiplication is used to perform a convolution operation - if(rhs->info()->num_dimensions() >= 3) + if (rhs->info()->num_dimensions() >= 3) { win_b = window; } @@ -302,338 +316,340 @@ void matrix_matrix_multiply_f32(const ITensor *lhs, const ITensor *rhs, ITensor // The implementation assumes that the matrix A and Matrix B have been reshaped respectively with CpuGemmInterleave4x4 and CpuGemmTranspose1xW // The reshaping of the matrices helps to have a cache friendly implementation and helps to avoid the data re-arrangements needed for computing 16x4 elements per iteration // All the values needed for computing a single 4x4 block will be read from consecutive memory positions - execute_window_loop(window, [&](const Coordinates & id) - { - auto mtx_a0 = reinterpret_cast(ina.ptr()); - auto mtx_b0 = reinterpret_cast(inb.ptr()); - auto mtx_b1 = mtx_b0 + in_b_stride; + execute_window_loop( + window, + [&](const Coordinates &id) + { + auto mtx_a0 = reinterpret_cast(ina.ptr()); + auto mtx_b0 = reinterpret_cast(inb.ptr()); + auto mtx_b1 = mtx_b0 + in_b_stride; - float32x4_t acc00 = vdupq_n_f32(0.f); - float32x4_t acc10 = vdupq_n_f32(0.f); - float32x4_t acc20 = vdupq_n_f32(0.f); - float32x4_t acc30 = vdupq_n_f32(0.f); + float32x4_t acc00 = vdupq_n_f32(0.f); + float32x4_t acc10 = vdupq_n_f32(0.f); + float32x4_t acc20 = vdupq_n_f32(0.f); + float32x4_t acc30 = vdupq_n_f32(0.f); - float32x4_t acc01 = vdupq_n_f32(0.f); - float32x4_t acc11 = vdupq_n_f32(0.f); - float32x4_t acc21 = vdupq_n_f32(0.f); - float32x4_t acc31 = vdupq_n_f32(0.f); + float32x4_t acc01 = vdupq_n_f32(0.f); + float32x4_t acc11 = vdupq_n_f32(0.f); + float32x4_t acc21 = vdupq_n_f32(0.f); + float32x4_t acc31 = vdupq_n_f32(0.f); #if __arm__ - asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(mtx_a0))); - asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(mtx_b0))); - asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(mtx_b1))); + asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(mtx_a0))); + asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(mtx_b0))); + asm volatile("PLD [%0, #128*1]" ::"r"(reinterpret_cast(mtx_b1))); #endif /* __arm__ */ - auto mtx_b0_end_addr = mtx_b0 + num_elems_matrix_b_x; - for(; mtx_b0 <= (mtx_b0_end_addr - 32);) - { - float32x4_t a0 = vld1q_dup_f32(mtx_a0 + 0); - float32x4_t a1 = vld1q_dup_f32(mtx_a0 + 1); - float32x4_t a2 = vld1q_dup_f32(mtx_a0 + 2); - float32x4_t a3 = vld1q_dup_f32(mtx_a0 + 3); + auto mtx_b0_end_addr = mtx_b0 + num_elems_matrix_b_x; + for (; mtx_b0 <= (mtx_b0_end_addr - 32);) + { + float32x4_t a0 = vld1q_dup_f32(mtx_a0 + 0); + float32x4_t a1 = vld1q_dup_f32(mtx_a0 + 1); + float32x4_t a2 = vld1q_dup_f32(mtx_a0 + 2); + float32x4_t a3 = vld1q_dup_f32(mtx_a0 + 3); - float32x4_t b00 = vld1q_f32(mtx_b0); - float32x4_t b10 = vld1q_f32(mtx_b1); - float32x4_t b01 = vld1q_f32(mtx_b0 + 4); - float32x4_t b11 = vld1q_f32(mtx_b1 + 4); + float32x4_t b00 = vld1q_f32(mtx_b0); + float32x4_t b10 = vld1q_f32(mtx_b1); + float32x4_t b01 = vld1q_f32(mtx_b0 + 4); + float32x4_t b11 = vld1q_f32(mtx_b1 + 4); #if __arm__ - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_a0))); - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b0))); - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b1))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_a0))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b0))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b1))); #endif /* __arm__ */ - // 4x4 block 0 - acc00 = vmlaq_f32(acc00, b00, a0); - acc10 = vmlaq_f32(acc10, b00, a1); - acc20 = vmlaq_f32(acc20, b00, a2); - acc30 = vmlaq_f32(acc30, b00, a3); - - float32x4_t a4 = vld1q_dup_f32(mtx_a0 + 4); - float32x4_t a5 = vld1q_dup_f32(mtx_a0 + 5); - float32x4_t a6 = vld1q_dup_f32(mtx_a0 + 6); - float32x4_t a7 = vld1q_dup_f32(mtx_a0 + 7); - - // 4x4 block 1 - acc01 = vmlaq_f32(acc01, b10, a0); - acc11 = vmlaq_f32(acc11, b10, a1); - acc21 = vmlaq_f32(acc21, b10, a2); - acc31 = vmlaq_f32(acc31, b10, a3); - - // 4x4 block 0 - acc00 = vmlaq_f32(acc00, b01, a4); - acc10 = vmlaq_f32(acc10, b01, a5); - acc20 = vmlaq_f32(acc20, b01, a6); - acc30 = vmlaq_f32(acc30, b01, a7); - - // 4x4 block 1 - acc01 = vmlaq_f32(acc01, b11, a4); - acc11 = vmlaq_f32(acc11, b11, a5); - acc21 = vmlaq_f32(acc21, b11, a6); - acc31 = vmlaq_f32(acc31, b11, a7); - - mtx_a0 += 8; - mtx_b0 += 8; - mtx_b1 += 8; - - a0 = vld1q_dup_f32(mtx_a0 + 0); - a1 = vld1q_dup_f32(mtx_a0 + 1); - a2 = vld1q_dup_f32(mtx_a0 + 2); - a3 = vld1q_dup_f32(mtx_a0 + 3); - - b00 = vld1q_f32(mtx_b0); - b10 = vld1q_f32(mtx_b1); - b01 = vld1q_f32(mtx_b0 + 4); - b11 = vld1q_f32(mtx_b1 + 4); - - // 4x4 block 0 - acc00 = vmlaq_f32(acc00, b00, a0); - acc10 = vmlaq_f32(acc10, b00, a1); - acc20 = vmlaq_f32(acc20, b00, a2); - acc30 = vmlaq_f32(acc30, b00, a3); - - a4 = vld1q_dup_f32(mtx_a0 + 4); - a5 = vld1q_dup_f32(mtx_a0 + 5); - a6 = vld1q_dup_f32(mtx_a0 + 6); - a7 = vld1q_dup_f32(mtx_a0 + 7); - - // 4x4 block 1 - acc01 = vmlaq_f32(acc01, b10, a0); - acc11 = vmlaq_f32(acc11, b10, a1); - acc21 = vmlaq_f32(acc21, b10, a2); - acc31 = vmlaq_f32(acc31, b10, a3); - - // 4x4 block 0 - acc00 = vmlaq_f32(acc00, b01, a4); - acc10 = vmlaq_f32(acc10, b01, a5); - acc20 = vmlaq_f32(acc20, b01, a6); - acc30 = vmlaq_f32(acc30, b01, a7); - - // 4x4 block 1 - acc01 = vmlaq_f32(acc01, b11, a4); - acc11 = vmlaq_f32(acc11, b11, a5); - acc21 = vmlaq_f32(acc21, b11, a6); - acc31 = vmlaq_f32(acc31, b11, a7); - - mtx_a0 += 8; - mtx_b0 += 8; - mtx_b1 += 8; - - a0 = vld1q_dup_f32(mtx_a0 + 0); - a1 = vld1q_dup_f32(mtx_a0 + 1); - a2 = vld1q_dup_f32(mtx_a0 + 2); - a3 = vld1q_dup_f32(mtx_a0 + 3); - b00 = vld1q_f32(mtx_b0); - b10 = vld1q_f32(mtx_b1); - b01 = vld1q_f32(mtx_b0 + 4); - b11 = vld1q_f32(mtx_b1 + 4); + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b00, a0); + acc10 = vmlaq_f32(acc10, b00, a1); + acc20 = vmlaq_f32(acc20, b00, a2); + acc30 = vmlaq_f32(acc30, b00, a3); + + float32x4_t a4 = vld1q_dup_f32(mtx_a0 + 4); + float32x4_t a5 = vld1q_dup_f32(mtx_a0 + 5); + float32x4_t a6 = vld1q_dup_f32(mtx_a0 + 6); + float32x4_t a7 = vld1q_dup_f32(mtx_a0 + 7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b10, a0); + acc11 = vmlaq_f32(acc11, b10, a1); + acc21 = vmlaq_f32(acc21, b10, a2); + acc31 = vmlaq_f32(acc31, b10, a3); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b01, a4); + acc10 = vmlaq_f32(acc10, b01, a5); + acc20 = vmlaq_f32(acc20, b01, a6); + acc30 = vmlaq_f32(acc30, b01, a7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b11, a4); + acc11 = vmlaq_f32(acc11, b11, a5); + acc21 = vmlaq_f32(acc21, b11, a6); + acc31 = vmlaq_f32(acc31, b11, a7); + + mtx_a0 += 8; + mtx_b0 += 8; + mtx_b1 += 8; + + a0 = vld1q_dup_f32(mtx_a0 + 0); + a1 = vld1q_dup_f32(mtx_a0 + 1); + a2 = vld1q_dup_f32(mtx_a0 + 2); + a3 = vld1q_dup_f32(mtx_a0 + 3); + + b00 = vld1q_f32(mtx_b0); + b10 = vld1q_f32(mtx_b1); + b01 = vld1q_f32(mtx_b0 + 4); + b11 = vld1q_f32(mtx_b1 + 4); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b00, a0); + acc10 = vmlaq_f32(acc10, b00, a1); + acc20 = vmlaq_f32(acc20, b00, a2); + acc30 = vmlaq_f32(acc30, b00, a3); + + a4 = vld1q_dup_f32(mtx_a0 + 4); + a5 = vld1q_dup_f32(mtx_a0 + 5); + a6 = vld1q_dup_f32(mtx_a0 + 6); + a7 = vld1q_dup_f32(mtx_a0 + 7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b10, a0); + acc11 = vmlaq_f32(acc11, b10, a1); + acc21 = vmlaq_f32(acc21, b10, a2); + acc31 = vmlaq_f32(acc31, b10, a3); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b01, a4); + acc10 = vmlaq_f32(acc10, b01, a5); + acc20 = vmlaq_f32(acc20, b01, a6); + acc30 = vmlaq_f32(acc30, b01, a7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b11, a4); + acc11 = vmlaq_f32(acc11, b11, a5); + acc21 = vmlaq_f32(acc21, b11, a6); + acc31 = vmlaq_f32(acc31, b11, a7); + + mtx_a0 += 8; + mtx_b0 += 8; + mtx_b1 += 8; + + a0 = vld1q_dup_f32(mtx_a0 + 0); + a1 = vld1q_dup_f32(mtx_a0 + 1); + a2 = vld1q_dup_f32(mtx_a0 + 2); + a3 = vld1q_dup_f32(mtx_a0 + 3); + b00 = vld1q_f32(mtx_b0); + b10 = vld1q_f32(mtx_b1); + b01 = vld1q_f32(mtx_b0 + 4); + b11 = vld1q_f32(mtx_b1 + 4); #if __arm__ - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_a0))); - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b0))); - asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b1))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_a0))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b0))); + asm volatile("PLD [%0, #128*4]" ::"r"(reinterpret_cast(mtx_b1))); #endif /* __arm__ */ - // 4x4 block 0 - acc00 = vmlaq_f32(acc00, b00, a0); - acc10 = vmlaq_f32(acc10, b00, a1); - acc20 = vmlaq_f32(acc20, b00, a2); - acc30 = vmlaq_f32(acc30, b00, a3); - - a4 = vld1q_dup_f32(mtx_a0 + 4); - a5 = vld1q_dup_f32(mtx_a0 + 5); - a6 = vld1q_dup_f32(mtx_a0 + 6); - a7 = vld1q_dup_f32(mtx_a0 + 7); - - // 4x4 block 1 - acc01 = vmlaq_f32(acc01, b10, a0); - acc11 = vmlaq_f32(acc11, b10, a1); - acc21 = vmlaq_f32(acc21, b10, a2); - acc31 = vmlaq_f32(acc31, b10, a3); - - // 4x4 block 0 - acc00 = vmlaq_f32(acc00, b01, a4); - acc10 = vmlaq_f32(acc10, b01, a5); - acc20 = vmlaq_f32(acc20, b01, a6); - acc30 = vmlaq_f32(acc30, b01, a7); - - // 4x4 block 1 - acc01 = vmlaq_f32(acc01, b11, a4); - acc11 = vmlaq_f32(acc11, b11, a5); - acc21 = vmlaq_f32(acc21, b11, a6); - acc31 = vmlaq_f32(acc31, b11, a7); - - mtx_a0 += 8; - mtx_b0 += 8; - mtx_b1 += 8; - - a0 = vld1q_dup_f32(mtx_a0 + 0); - a1 = vld1q_dup_f32(mtx_a0 + 1); - a2 = vld1q_dup_f32(mtx_a0 + 2); - a3 = vld1q_dup_f32(mtx_a0 + 3); - b00 = vld1q_f32(mtx_b0); - b10 = vld1q_f32(mtx_b1); - b01 = vld1q_f32(mtx_b0 + 4); - b11 = vld1q_f32(mtx_b1 + 4); - - // 4x4 block 0 - acc00 = vmlaq_f32(acc00, b00, a0); - acc10 = vmlaq_f32(acc10, b00, a1); - acc20 = vmlaq_f32(acc20, b00, a2); - acc30 = vmlaq_f32(acc30, b00, a3); - - a4 = vld1q_dup_f32(mtx_a0 + 4); - a5 = vld1q_dup_f32(mtx_a0 + 5); - a6 = vld1q_dup_f32(mtx_a0 + 6); - a7 = vld1q_dup_f32(mtx_a0 + 7); - - // 4x4 block 1 - acc01 = vmlaq_f32(acc01, b10, a0); - acc11 = vmlaq_f32(acc11, b10, a1); - acc21 = vmlaq_f32(acc21, b10, a2); - acc31 = vmlaq_f32(acc31, b10, a3); - - // 4x4 block 0 - acc00 = vmlaq_f32(acc00, b01, a4); - acc10 = vmlaq_f32(acc10, b01, a5); - acc20 = vmlaq_f32(acc20, b01, a6); - acc30 = vmlaq_f32(acc30, b01, a7); - - // 4x4 block 1 - acc01 = vmlaq_f32(acc01, b11, a4); - acc11 = vmlaq_f32(acc11, b11, a5); - acc21 = vmlaq_f32(acc21, b11, a6); - acc31 = vmlaq_f32(acc31, b11, a7); - - mtx_a0 += 8; - mtx_b0 += 8; - mtx_b1 += 8; - } - - for(; mtx_b0 < mtx_b0_end_addr;) - { - float32x4_t a0 = vld1q_dup_f32(mtx_a0 + 0); - float32x4_t a1 = vld1q_dup_f32(mtx_a0 + 1); - float32x4_t a2 = vld1q_dup_f32(mtx_a0 + 2); - float32x4_t a3 = vld1q_dup_f32(mtx_a0 + 3); - float32x4_t b00 = vld1q_f32(mtx_b0); - float32x4_t b10 = vld1q_f32(mtx_b1); + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b00, a0); + acc10 = vmlaq_f32(acc10, b00, a1); + acc20 = vmlaq_f32(acc20, b00, a2); + acc30 = vmlaq_f32(acc30, b00, a3); + + a4 = vld1q_dup_f32(mtx_a0 + 4); + a5 = vld1q_dup_f32(mtx_a0 + 5); + a6 = vld1q_dup_f32(mtx_a0 + 6); + a7 = vld1q_dup_f32(mtx_a0 + 7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b10, a0); + acc11 = vmlaq_f32(acc11, b10, a1); + acc21 = vmlaq_f32(acc21, b10, a2); + acc31 = vmlaq_f32(acc31, b10, a3); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b01, a4); + acc10 = vmlaq_f32(acc10, b01, a5); + acc20 = vmlaq_f32(acc20, b01, a6); + acc30 = vmlaq_f32(acc30, b01, a7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b11, a4); + acc11 = vmlaq_f32(acc11, b11, a5); + acc21 = vmlaq_f32(acc21, b11, a6); + acc31 = vmlaq_f32(acc31, b11, a7); + + mtx_a0 += 8; + mtx_b0 += 8; + mtx_b1 += 8; + + a0 = vld1q_dup_f32(mtx_a0 + 0); + a1 = vld1q_dup_f32(mtx_a0 + 1); + a2 = vld1q_dup_f32(mtx_a0 + 2); + a3 = vld1q_dup_f32(mtx_a0 + 3); + b00 = vld1q_f32(mtx_b0); + b10 = vld1q_f32(mtx_b1); + b01 = vld1q_f32(mtx_b0 + 4); + b11 = vld1q_f32(mtx_b1 + 4); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b00, a0); + acc10 = vmlaq_f32(acc10, b00, a1); + acc20 = vmlaq_f32(acc20, b00, a2); + acc30 = vmlaq_f32(acc30, b00, a3); + + a4 = vld1q_dup_f32(mtx_a0 + 4); + a5 = vld1q_dup_f32(mtx_a0 + 5); + a6 = vld1q_dup_f32(mtx_a0 + 6); + a7 = vld1q_dup_f32(mtx_a0 + 7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b10, a0); + acc11 = vmlaq_f32(acc11, b10, a1); + acc21 = vmlaq_f32(acc21, b10, a2); + acc31 = vmlaq_f32(acc31, b10, a3); + + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b01, a4); + acc10 = vmlaq_f32(acc10, b01, a5); + acc20 = vmlaq_f32(acc20, b01, a6); + acc30 = vmlaq_f32(acc30, b01, a7); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b11, a4); + acc11 = vmlaq_f32(acc11, b11, a5); + acc21 = vmlaq_f32(acc21, b11, a6); + acc31 = vmlaq_f32(acc31, b11, a7); + + mtx_a0 += 8; + mtx_b0 += 8; + mtx_b1 += 8; + } + + for (; mtx_b0 < mtx_b0_end_addr;) + { + float32x4_t a0 = vld1q_dup_f32(mtx_a0 + 0); + float32x4_t a1 = vld1q_dup_f32(mtx_a0 + 1); + float32x4_t a2 = vld1q_dup_f32(mtx_a0 + 2); + float32x4_t a3 = vld1q_dup_f32(mtx_a0 + 3); + float32x4_t b00 = vld1q_f32(mtx_b0); + float32x4_t b10 = vld1q_f32(mtx_b1); #if __arm__ - asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_a0))); - asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_b0))); - asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_b1))); + asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_a0))); + asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_b0))); + asm volatile("PLD [%0, #128*2]" ::"r"(reinterpret_cast(mtx_b1))); #endif /* __arm__ */ - // 4x4 block 0 - acc00 = vmlaq_f32(acc00, b00, a0); - acc10 = vmlaq_f32(acc10, b00, a1); - acc20 = vmlaq_f32(acc20, b00, a2); - acc30 = vmlaq_f32(acc30, b00, a3); - - // 4x4 block 1 - acc01 = vmlaq_f32(acc01, b10, a0); - acc11 = vmlaq_f32(acc11, b10, a1); - acc21 = vmlaq_f32(acc21, b10, a2); - acc31 = vmlaq_f32(acc31, b10, a3); - - mtx_a0 += 4; - mtx_b0 += 4; - mtx_b1 += 4; - } - - // Multiply by the weight of matrix product (alpha) - if(multiply_alpha) - { - acc00 = vmulq_f32(acc00, alpha_f32); - acc10 = vmulq_f32(acc10, alpha_f32); - acc20 = vmulq_f32(acc20, alpha_f32); - acc30 = vmulq_f32(acc30, alpha_f32); - acc01 = vmulq_f32(acc01, alpha_f32); - acc11 = vmulq_f32(acc11, alpha_f32); - acc21 = vmulq_f32(acc21, alpha_f32); - acc31 = vmulq_f32(acc31, alpha_f32); - } - - const auto mtx_out0 = reinterpret_cast(out.ptr()); - const auto mtx_out1 = mtx_out0 + 4; - - if(id.x() < (out_width - 8)) - { - vst1q_f32(mtx_out0, acc00); - vst1q_f32(mtx_out1, acc01); - if(id.y() + 1 < out_height) + // 4x4 block 0 + acc00 = vmlaq_f32(acc00, b00, a0); + acc10 = vmlaq_f32(acc10, b00, a1); + acc20 = vmlaq_f32(acc20, b00, a2); + acc30 = vmlaq_f32(acc30, b00, a3); + + // 4x4 block 1 + acc01 = vmlaq_f32(acc01, b10, a0); + acc11 = vmlaq_f32(acc11, b10, a1); + acc21 = vmlaq_f32(acc21, b10, a2); + acc31 = vmlaq_f32(acc31, b10, a3); + + mtx_a0 += 4; + mtx_b0 += 4; + mtx_b1 += 4; + } + + // Multiply by the weight of matrix product (alpha) + if (multiply_alpha) + { + acc00 = vmulq_f32(acc00, alpha_f32); + acc10 = vmulq_f32(acc10, alpha_f32); + acc20 = vmulq_f32(acc20, alpha_f32); + acc30 = vmulq_f32(acc30, alpha_f32); + acc01 = vmulq_f32(acc01, alpha_f32); + acc11 = vmulq_f32(acc11, alpha_f32); + acc21 = vmulq_f32(acc21, alpha_f32); + acc31 = vmulq_f32(acc31, alpha_f32); + } + + const auto mtx_out0 = reinterpret_cast(out.ptr()); + const auto mtx_out1 = mtx_out0 + 4; + + if (id.x() < (out_width - 8)) { - vst1q_f32(mtx_out0 + out_stride1, acc10); - vst1q_f32(mtx_out1 + out_stride1, acc11); - if(id.y() + 2 < out_height) + vst1q_f32(mtx_out0, acc00); + vst1q_f32(mtx_out1, acc01); + if (id.y() + 1 < out_height) { - vst1q_f32(mtx_out0 + out_stride2, acc20); - vst1q_f32(mtx_out1 + out_stride2, acc21); - if(id.y() + 3 < out_height) + vst1q_f32(mtx_out0 + out_stride1, acc10); + vst1q_f32(mtx_out1 + out_stride1, acc11); + if (id.y() + 2 < out_height) { - vst1q_f32(mtx_out0 + out_stride3, acc30); - vst1q_f32(mtx_out1 + out_stride3, acc31); + vst1q_f32(mtx_out0 + out_stride2, acc20); + vst1q_f32(mtx_out1 + out_stride2, acc21); + if (id.y() + 3 < out_height) + { + vst1q_f32(mtx_out0 + out_stride3, acc30); + vst1q_f32(mtx_out1 + out_stride3, acc31); + } } } } - } - else if(id.x() < (out_width - 4)) - { - vst1q_f32(mtx_out0, acc00); - if(id.y() + 1 < out_height) + else if (id.x() < (out_width - 4)) { - vst1q_f32(mtx_out0 + out_stride1, acc10); - if(id.y() + 2 < out_height) + vst1q_f32(mtx_out0, acc00); + if (id.y() + 1 < out_height) { - vst1q_f32(mtx_out0 + out_stride2, acc20); - if(id.y() + 3 < out_height) + vst1q_f32(mtx_out0 + out_stride1, acc10); + if (id.y() + 2 < out_height) { - vst1q_f32(mtx_out0 + out_stride3, acc30); + vst1q_f32(mtx_out0 + out_stride2, acc20); + if (id.y() + 3 < out_height) + { + vst1q_f32(mtx_out0 + out_stride3, acc30); + } } } - } - // Left-over columns - const int columns_left = out_width - id.x() - 4; - for(auto x = 0; x < columns_left; ++x) - { - *(mtx_out1 + x) = acc01[x]; - if(id.y() + 1 < out_height) + // Left-over columns + const int columns_left = out_width - id.x() - 4; + for (auto x = 0; x < columns_left; ++x) { - *(mtx_out1 + x + out_stride1) = acc11[x]; - if(id.y() + 2 < out_height) + *(mtx_out1 + x) = acc01[x]; + if (id.y() + 1 < out_height) { - *(mtx_out1 + x + out_stride2) = acc21[x]; - if(id.y() + 3 < out_height) + *(mtx_out1 + x + out_stride1) = acc11[x]; + if (id.y() + 2 < out_height) { - *(mtx_out1 + x + out_stride3) = acc31[x]; + *(mtx_out1 + x + out_stride2) = acc21[x]; + if (id.y() + 3 < out_height) + { + *(mtx_out1 + x + out_stride3) = acc31[x]; + } } } } } - } - else - { - // Left-over columns - const int columns_left = out_width - id.x(); - for(int x = 0; x < columns_left; ++x) + else { - *(mtx_out0 + x) = acc00[x]; - if(id.y() + 1 < out_height) + // Left-over columns + const int columns_left = out_width - id.x(); + for (int x = 0; x < columns_left; ++x) { - *(mtx_out0 + x + out_stride1) = acc10[x]; - if(id.y() + 2 < out_height) + *(mtx_out0 + x) = acc00[x]; + if (id.y() + 1 < out_height) { - *(mtx_out0 + x + out_stride2) = acc20[x]; - if(id.y() + 3 < out_height) + *(mtx_out0 + x + out_stride1) = acc10[x]; + if (id.y() + 2 < out_height) { - *(mtx_out0 + x + out_stride3) = acc30[x]; + *(mtx_out0 + x + out_stride2) = acc20[x]; + if (id.y() + 3 < out_height) + { + *(mtx_out0 + x + out_stride3) = acc30[x]; + } } } } } - } - }, - ina, inb, out); + }, + ina, inb, out); } } // namespace cpu diff --git a/src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.h b/src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.h index f9f1f247ac..74ea4c2b17 100644 --- a/src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.h +++ b/src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.h @@ -24,15 +24,18 @@ #ifndef SRC_CORE_KERNELS_GEMMMATRIXMUL_IMPL_H #define SRC_CORE_KERNELS_GEMMMATRIXMUL_IMPL_H #include "arm_compute/core/Helpers.h" + #include "src/core/CPP/Validate.h" namespace arm_compute { namespace cpu { -void vector_matrix_multiply_f32(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha); +void vector_matrix_multiply_f32( + const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha); -void matrix_matrix_multiply_f32(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha); +void matrix_matrix_multiply_f32( + const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha); } // namespace cpu } // namespace arm_compute diff --git a/src/cpu/kernels/gemm_matrix_mul/list.h b/src/cpu/kernels/gemm_matrix_mul/list.h index 9cdb58ae06..15b23b1d81 100644 --- a/src/cpu/kernels/gemm_matrix_mul/list.h +++ b/src/cpu/kernels/gemm_matrix_mul/list.h @@ -27,8 +27,9 @@ namespace arm_compute { namespace cpu { -#define DECLARE_GEMMMATRIXMUL_KERNEL(func_name) \ - void func_name(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, float alpha, const bool is_dst_vector) +#define DECLARE_GEMMMATRIXMUL_KERNEL(func_name) \ + void func_name(const ITensor *lhs, const ITensor *rhs, ITensor *dst, const Window &window, const ThreadInfo &info, \ + float alpha, const bool is_dst_vector) DECLARE_GEMMMATRIXMUL_KERNEL(neon_fp32_gemm_matrix_mul); DECLARE_GEMMMATRIXMUL_KERNEL(neon_fp16_gemm_matrix_mul); #undef DECLARE_GEMMMATRIXMUL_KERNEL -- cgit v1.2.1