diff options
Diffstat (limited to 'src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp')
| -rw-r--r-- | src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp | 418 |
1 files changed, 418 insertions, 0 deletions
diff --git a/src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp b/src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp new file mode 100644 index 0000000000..60fda511e3 --- /dev/null +++ b/src/cpu/kernels/gemm_matrix_mul/generic/neon/fp16.cpp @@ -0,0 +1,418 @@ +/* + * Copyright (c) 2022-2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +#include "src/core/utils/helpers/float_ops.h" +#include "src/cpu/kernels/gemm_matrix_mul/generic/neon/impl.h" + +#include <arm_neon.h> + +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) +{ + const auto width_matrix_b = static_cast<int>(dst->info()->dimension(0)); + const auto in_b_stride = static_cast<int>(rhs->info()->strides_in_bytes()[1] / rhs->info()->element_size()); + const auto num_elems_vec_a = static_cast<int>(lhs->info()->dimension(0)); + + // The implementation computes 32 elements per iteration + 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"); + + Window win_out(window); + win_out.set(Window::DimX, Window::Dimension(0, 1, 1)); + win_out.set(Window::DimY, Window::Dimension(0, 1, 1)); + + Window win_a(window); + win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); + win_a.set(Window::DimY, Window::Dimension(0, 0, 0)); + + 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) + { + win_b = window; + } + win_b.set(Window::DimX, Window::Dimension(0, 1, 1)); + win_b.set(Window::DimY, Window::Dimension(0, 1, 1)); + + Iterator ina(lhs, win_a); + Iterator inb(rhs, win_b); + Iterator out(dst, win_out); + + const bool multiply_alpha = !(helpers::float_ops::is_one(alpha)); + + 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) + { + if (x > width_matrix_b) + { + return; + } + + auto matrix_b = reinterpret_cast<const float16_t *>(inb.ptr()) + x; + + 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 vec_a = reinterpret_cast<const float16_t *>(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; + } + + 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; + } + + // 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); + } + + auto vec_out = reinterpret_cast<float16_t *>(out.ptr()) + x; + + vst1q_f16(vec_out + 0, acc0); + vst1q_f16(vec_out + 8, acc1); + vst1q_f16(vec_out + 16, acc2); + vst1q_f16(vec_out + 24, acc3); + } + + for (; x < window_end_x; ++x) + { + if (x > width_matrix_b) + { + return; + } + + auto matrix_b = reinterpret_cast<const float16_t *>(inb.ptr()) + x; + + float16x4_t vacc = vdup_n_f16(0.f); + + auto vec_a = reinterpret_cast<const float16_t *>(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); + + 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), + }; + + vacc = vadd_f16(vacc, vmul_f16(a0l, b_col)); + + matrix_b += 4 * in_b_stride; + } + + float16_t acc = + vget_lane_f16(vacc, 0) + vget_lane_f16(vacc, 1) + vget_lane_f16(vacc, 2) + vget_lane_f16(vacc, 3); + + for (; vec_a < vec_a_end_addr; ++vec_a) + { + const float16_t a0 = *vec_a; + const float16_t b00 = *matrix_b; + + acc += b00 * a0; + + matrix_b += in_b_stride; + } + + // Multiply by the weight of matrix product (alpha) + if (multiply_alpha) + { + acc *= static_cast<float16_t>(alpha); + } + + auto vec_out = reinterpret_cast<float16_t *>(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) +{ + ARM_COMPUTE_UNUSED(info); + const int out_width = static_cast<int>(dst->info()->dimension(0)); + const int out_height = static_cast<int>(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 + Window win_a(window); + win_a.set(Window::DimX, Window::Dimension(0, 0, 0)); + win_a.set(Window::DimY, Window::Dimension(window.y().start() / 4, std::max(window.y().end() / 4, 1), 1)); + + 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) + { + win_b = window; + } + // Set step_x and step_y for matrix B. Scale by a factor of 8 the X range as the input transposed matrix A has 8 times less the cols of the dst matrix + win_b.set(Window::DimX, Window::Dimension(window.x().start() / 8, window.x().end() / 8, in_b_stride)); + win_b.set(Window::DimY, Window::Dimension(0, 0, 0)); + + Iterator ina(lhs, win_a); + Iterator inb(rhs, win_b); + Iterator out(dst, window); + + const bool multiply_alpha = !(helpers::float_ops::is_one(alpha)); + + const float16x8_t alpha_f16 = vdupq_n_f16(alpha); + + execute_window_loop( + window, + [&](const Coordinates &id) + { + const auto *mtx_a0 = reinterpret_cast<const float16_t *>(ina.ptr()); + const auto *mtx_b0 = reinterpret_cast<const float16_t *>(inb.ptr()); + auto *mtx_out = reinterpret_cast<float16_t *>(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| + |a20 a21 a22 a23 | a24 a25 a26 a27| = | a00 a10 a20 a30 || a01 a11 a21 a31 || a02 a12 a22 a32 || a03 a13 a23 a33 | a40 a50 a60 a70 | ... + |a30 a31 a32 a33 | a34 a35 a36 a37| | a04 a14 a24 a34 || a05 a15 a25 a35 || a06 a15 a26 a36 || a07 a17 a27 a37 | a44 a54 a64 a74 | ... + |a40 a41 a42 a43 | a44 a45 a46 a47| + |a50 a51 a52 a53 | a54 a55 a56 a57| + |a60 a61 a62 a63 | a64 a65 a66 a67| + |a70 a71 a72 a73 | a74 a75 a76 a77| + + After this operation, the dst matrix will have the following shape: [ height * 4, width / 4 ] + + B Matrix has been transposed as shown below + + |b00 b01 b02 b03 b04 b05 b06 b07| + |b10 b11 b12 b13 b14 b15 b16 b17| + |b20 b21 b22 b23 b24 b25 b26 b27| + |b30 b31 b32 b33 b34 b35 b36 b37| + -------------------> + + |b00 b01 b02 b03 b04 b05 b06 b07||b10 b11 b12 b13 b14 b15 b16 b17||b20 b21 b22 b23 b24 b25 b26 b27||b30 b31 b32 b33 b34 b35 b36 b37| + + c.val[0][0] = a00*b00 + a01*b10 + a02*b20 + a03*b30 + c.val[0][1] = a00*b01 + a01*b11 + a02*b21 + a03*b31 + + 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))); + + 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;) + + { + 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))); + + 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 (id.x() < (out_width - 8)) + { + vst1q_f16(mtx_out, c.val[0]); + if (id.y() + 1 < out_height) + { + vst1q_f16(mtx_out + 1 * out_stride, c.val[1]); + if (id.y() + 2 < out_height) + { + 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) + { + *(mtx_out + x) = c.val[0][x]; + if (id.y() + 1 < out_height) + { + *(mtx_out + x + 1 * out_stride) = c.val[1][x]; + if (id.y() + 2 < out_height) + { + *(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); +} + +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); +} +} // namespace cpu +} // namespace arm_compute +#endif //__ARM_FEATURE_FP16_VECTOR_ARITHMETIC |