diff options
Diffstat (limited to 'src/cpu/kernels/CpuGemmLowpMatrixMultiplyKernel.cpp')
| -rw-r--r-- | src/cpu/kernels/CpuGemmLowpMatrixMultiplyKernel.cpp | 877 |
1 files changed, 877 insertions, 0 deletions
diff --git a/src/cpu/kernels/CpuGemmLowpMatrixMultiplyKernel.cpp b/src/cpu/kernels/CpuGemmLowpMatrixMultiplyKernel.cpp new file mode 100644 index 0000000000..a3ed2cd171 --- /dev/null +++ b/src/cpu/kernels/CpuGemmLowpMatrixMultiplyKernel.cpp @@ -0,0 +1,877 @@ +/* + * Copyright (c) 2017-2021 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. + */ +#include "src/cpu/kernels/CpuGemmLowpMatrixMultiplyKernel.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/Window.h" + +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" + +#include <arm_neon.h> + +namespace arm_compute +{ +namespace cpu +{ +namespace kernels +{ +namespace +{ +void inline vector_matrix_multiply_u8(Iterator &ina, + Iterator &inb, + Iterator &out, + int width_a, + int width_b, + int width_out, + size_t stride_b, + const Window &window) +{ + execute_window_loop( + window, + [&](const Coordinates &id) + { + if (id.x() > width_b) + { + return; + } + + // Note: Since the input are all positives, we can use uint32_t + // Accumulators for the block 0 + uint32x4x4_t c0 = {{vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0)}}; + + auto vec_a = reinterpret_cast<const uint8_t *>(ina.ptr()); + auto matrix_b = reinterpret_cast<const uint8_t *>(inb.ptr()); + auto vec_a_end_addr = vec_a + width_a; + + // This for loop performs 8 accumulations + for (; vec_a <= (vec_a_end_addr - 8);) + { + const uint8x8_t a00_u8 = vld1_u8(vec_a); + const uint8x16_t b00_u8 = vld1q_u8(matrix_b + 0 * stride_b); + const uint8x16_t b10_u8 = vld1q_u8(matrix_b + 1 * stride_b); + const uint8x16_t b20_u8 = vld1q_u8(matrix_b + 2 * stride_b); + const uint8x16_t b30_u8 = vld1q_u8(matrix_b + 3 * stride_b); + const uint8x16_t b40_u8 = vld1q_u8(matrix_b + 4 * stride_b); + const uint8x16_t b50_u8 = vld1q_u8(matrix_b + 5 * stride_b); + const uint8x16_t b60_u8 = vld1q_u8(matrix_b + 6 * stride_b); + const uint8x16_t b70_u8 = vld1q_u8(matrix_b + 7 * stride_b); + + // Convert a00_u8 to uint16_t and get the lower part + const uint16x4x2_t a00_u16 = {{vget_low_u16(vmovl_u8(a00_u8)), vget_high_u16(vmovl_u8(a00_u8))}}; + + const uint16x4x4_t b00_u16 = { + {vget_low_u16(vmovl_u8(vget_low_u8(b00_u8))), vget_high_u16(vmovl_u8(vget_low_u8(b00_u8))), + vget_low_u16(vmovl_u8(vget_high_u8(b00_u8))), vget_high_u16(vmovl_u8(vget_high_u8(b00_u8)))}}; + + const uint16x4x4_t b10_u16 = { + {vget_low_u16(vmovl_u8(vget_low_u8(b10_u8))), vget_high_u16(vmovl_u8(vget_low_u8(b10_u8))), + vget_low_u16(vmovl_u8(vget_high_u8(b10_u8))), vget_high_u16(vmovl_u8(vget_high_u8(b10_u8)))}}; + + const uint16x4x4_t b20_u16 = { + {vget_low_u16(vmovl_u8(vget_low_u8(b20_u8))), vget_high_u16(vmovl_u8(vget_low_u8(b20_u8))), + vget_low_u16(vmovl_u8(vget_high_u8(b20_u8))), vget_high_u16(vmovl_u8(vget_high_u8(b20_u8)))}}; + + const uint16x4x4_t b30_u16 = { + {vget_low_u16(vmovl_u8(vget_low_u8(b30_u8))), vget_high_u16(vmovl_u8(vget_low_u8(b30_u8))), + vget_low_u16(vmovl_u8(vget_high_u8(b30_u8))), vget_high_u16(vmovl_u8(vget_high_u8(b30_u8)))}}; + + const uint16x4x4_t b40_u16 = { + {vget_low_u16(vmovl_u8(vget_low_u8(b40_u8))), vget_high_u16(vmovl_u8(vget_low_u8(b40_u8))), + vget_low_u16(vmovl_u8(vget_high_u8(b40_u8))), vget_high_u16(vmovl_u8(vget_high_u8(b40_u8)))}}; + + const uint16x4x4_t b50_u16 = { + {vget_low_u16(vmovl_u8(vget_low_u8(b50_u8))), vget_high_u16(vmovl_u8(vget_low_u8(b50_u8))), + vget_low_u16(vmovl_u8(vget_high_u8(b50_u8))), vget_high_u16(vmovl_u8(vget_high_u8(b50_u8)))}}; + + const uint16x4x4_t b60_u16 = { + {vget_low_u16(vmovl_u8(vget_low_u8(b60_u8))), vget_high_u16(vmovl_u8(vget_low_u8(b60_u8))), + vget_low_u16(vmovl_u8(vget_high_u8(b60_u8))), vget_high_u16(vmovl_u8(vget_high_u8(b60_u8)))}}; + + const uint16x4x4_t b70_u16 = { + {vget_low_u16(vmovl_u8(vget_low_u8(b70_u8))), vget_high_u16(vmovl_u8(vget_low_u8(b70_u8))), + vget_low_u16(vmovl_u8(vget_high_u8(b70_u8))), vget_high_u16(vmovl_u8(vget_high_u8(b70_u8)))}}; + + // Accumulate 0: + c0.val[0] = vmlal_lane_u16(c0.val[0], b00_u16.val[0], a00_u16.val[0], 0); + c0.val[1] = vmlal_lane_u16(c0.val[1], b00_u16.val[1], a00_u16.val[0], 0); + c0.val[2] = vmlal_lane_u16(c0.val[2], b00_u16.val[2], a00_u16.val[0], 0); + c0.val[3] = vmlal_lane_u16(c0.val[3], b00_u16.val[3], a00_u16.val[0], 0); + + // Accumulate 1: + c0.val[0] = vmlal_lane_u16(c0.val[0], b10_u16.val[0], a00_u16.val[0], 1); + c0.val[1] = vmlal_lane_u16(c0.val[1], b10_u16.val[1], a00_u16.val[0], 1); + c0.val[2] = vmlal_lane_u16(c0.val[2], b10_u16.val[2], a00_u16.val[0], 1); + c0.val[3] = vmlal_lane_u16(c0.val[3], b10_u16.val[3], a00_u16.val[0], 1); + + // Accumulate 2: + c0.val[0] = vmlal_lane_u16(c0.val[0], b20_u16.val[0], a00_u16.val[0], 2); + c0.val[1] = vmlal_lane_u16(c0.val[1], b20_u16.val[1], a00_u16.val[0], 2); + c0.val[2] = vmlal_lane_u16(c0.val[2], b20_u16.val[2], a00_u16.val[0], 2); + c0.val[3] = vmlal_lane_u16(c0.val[3], b20_u16.val[3], a00_u16.val[0], 2); + + // Accumulate 3: + c0.val[0] = vmlal_lane_u16(c0.val[0], b30_u16.val[0], a00_u16.val[0], 3); + c0.val[1] = vmlal_lane_u16(c0.val[1], b30_u16.val[1], a00_u16.val[0], 3); + c0.val[2] = vmlal_lane_u16(c0.val[2], b30_u16.val[2], a00_u16.val[0], 3); + c0.val[3] = vmlal_lane_u16(c0.val[3], b30_u16.val[3], a00_u16.val[0], 3); + + // Accumulate 4: + c0.val[0] = vmlal_lane_u16(c0.val[0], b40_u16.val[0], a00_u16.val[1], 0); + c0.val[1] = vmlal_lane_u16(c0.val[1], b40_u16.val[1], a00_u16.val[1], 0); + c0.val[2] = vmlal_lane_u16(c0.val[2], b40_u16.val[2], a00_u16.val[1], 0); + c0.val[3] = vmlal_lane_u16(c0.val[3], b40_u16.val[3], a00_u16.val[1], 0); + + // Accumulate 5: + c0.val[0] = vmlal_lane_u16(c0.val[0], b50_u16.val[0], a00_u16.val[1], 1); + c0.val[1] = vmlal_lane_u16(c0.val[1], b50_u16.val[1], a00_u16.val[1], 1); + c0.val[2] = vmlal_lane_u16(c0.val[2], b50_u16.val[2], a00_u16.val[1], 1); + c0.val[3] = vmlal_lane_u16(c0.val[3], b50_u16.val[3], a00_u16.val[1], 1); + + // Accumulate 6: + c0.val[0] = vmlal_lane_u16(c0.val[0], b60_u16.val[0], a00_u16.val[1], 2); + c0.val[1] = vmlal_lane_u16(c0.val[1], b60_u16.val[1], a00_u16.val[1], 2); + c0.val[2] = vmlal_lane_u16(c0.val[2], b60_u16.val[2], a00_u16.val[1], 2); + c0.val[3] = vmlal_lane_u16(c0.val[3], b60_u16.val[3], a00_u16.val[1], 2); + + // Accumulate 7: + c0.val[0] = vmlal_lane_u16(c0.val[0], b70_u16.val[0], a00_u16.val[1], 3); + c0.val[1] = vmlal_lane_u16(c0.val[1], b70_u16.val[1], a00_u16.val[1], 3); + c0.val[2] = vmlal_lane_u16(c0.val[2], b70_u16.val[2], a00_u16.val[1], 3); + c0.val[3] = vmlal_lane_u16(c0.val[3], b70_u16.val[3], a00_u16.val[1], 3); + + vec_a += 8; + matrix_b += 8 * stride_b; + } + + // This for loop performs the left-over accumulations + for (; vec_a < vec_a_end_addr;) + { + const uint8x8_t a00_u8 = vld1_dup_u8(vec_a); + const uint8x16_t b00_u8 = vld1q_u8(matrix_b); + + const uint16x4x4_t b00_u16 = { + {vget_low_u16(vmovl_u8(vget_low_u8(b00_u8))), vget_high_u16(vmovl_u8(vget_low_u8(b00_u8))), + vget_low_u16(vmovl_u8(vget_high_u8(b00_u8))), vget_high_u16(vmovl_u8(vget_high_u8(b00_u8)))}}; + + // Convert a00_u8 to uint16_t and get the lower part + const uint16x4_t a00_u16 = vget_low_u16(vmovl_u8(a00_u8)); + + // Accumulate 0: + c0.val[0] = vmlal_lane_u16(c0.val[0], b00_u16.val[0], a00_u16, 0); + c0.val[1] = vmlal_lane_u16(c0.val[1], b00_u16.val[1], a00_u16, 0); + c0.val[2] = vmlal_lane_u16(c0.val[2], b00_u16.val[2], a00_u16, 0); + c0.val[3] = vmlal_lane_u16(c0.val[3], b00_u16.val[3], a00_u16, 0); + + vec_a += 1; + matrix_b += stride_b; + } + + auto vec_out = reinterpret_cast<int32_t *>(out.ptr()); + if (id.x() < (width_out - 16)) + { + vst1q_s32(vec_out + 0, vreinterpretq_s32_u32(c0.val[0])); + vst1q_s32(vec_out + 4, vreinterpretq_s32_u32(c0.val[1])); + vst1q_s32(vec_out + 8, vreinterpretq_s32_u32(c0.val[2])); + vst1q_s32(vec_out + 12, vreinterpretq_s32_u32(c0.val[3])); + } + else + { + auto left_over = width_out - id.x(); + for (auto k = 0; k < 4 && left_over; ++k) + { + for (auto j = 0; j < 4 && left_over; ++j, --left_over) + { + *(vec_out + k * 4 + j) = c0.val[k][j]; + } + } + } + }, + ina, inb, out); +} + +void inline vector_matrix_multiply_s8(Iterator &ina, + Iterator &inb, + Iterator &out, + int width_a, + int width_b, + int width_out, + size_t stride_b, + const Window &window) +{ + execute_window_loop( + window, + [&](const Coordinates &id) + { + if (id.x() > width_b) + { + return; + } + + // Accumulators for the block 0 + int32x4x4_t c0 = {{vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0)}}; + + auto vec_a = reinterpret_cast<const int8_t *>(ina.ptr()); + auto matrix_b = reinterpret_cast<const int8_t *>(inb.ptr()); + auto vec_a_end_addr = vec_a + width_a; + + // This for loop performs 8 accumulations + for (; vec_a <= (vec_a_end_addr - 8);) + { + const int8x8_t a00_s8 = vld1_s8(vec_a); + const int8x16_t b00_s8 = vld1q_s8(matrix_b + 0 * stride_b); + const int8x16_t b10_s8 = vld1q_s8(matrix_b + 1 * stride_b); + const int8x16_t b20_s8 = vld1q_s8(matrix_b + 2 * stride_b); + const int8x16_t b30_s8 = vld1q_s8(matrix_b + 3 * stride_b); + const int8x16_t b40_s8 = vld1q_s8(matrix_b + 4 * stride_b); + const int8x16_t b50_s8 = vld1q_s8(matrix_b + 5 * stride_b); + const int8x16_t b60_s8 = vld1q_s8(matrix_b + 6 * stride_b); + const int8x16_t b70_s8 = vld1q_s8(matrix_b + 7 * stride_b); + + // Convert a00_s8 to int16_t and get the lower part + const int16x4x2_t a00_s16 = {{vget_low_s16(vmovl_s8(a00_s8)), vget_high_s16(vmovl_s8(a00_s8))}}; + + const int16x4x4_t b00_s16 = { + {vget_low_s16(vmovl_s8(vget_low_s8(b00_s8))), vget_high_s16(vmovl_s8(vget_low_s8(b00_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(b00_s8))), vget_high_s16(vmovl_s8(vget_high_s8(b00_s8)))}}; + + const int16x4x4_t b10_s16 = { + {vget_low_s16(vmovl_s8(vget_low_s8(b10_s8))), vget_high_s16(vmovl_s8(vget_low_s8(b10_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(b10_s8))), vget_high_s16(vmovl_s8(vget_high_s8(b10_s8)))}}; + + const int16x4x4_t b20_s16 = { + {vget_low_s16(vmovl_s8(vget_low_s8(b20_s8))), vget_high_s16(vmovl_s8(vget_low_s8(b20_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(b20_s8))), vget_high_s16(vmovl_s8(vget_high_s8(b20_s8)))}}; + + const int16x4x4_t b30_s16 = { + {vget_low_s16(vmovl_s8(vget_low_s8(b30_s8))), vget_high_s16(vmovl_s8(vget_low_s8(b30_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(b30_s8))), vget_high_s16(vmovl_s8(vget_high_s8(b30_s8)))}}; + + const int16x4x4_t b40_s16 = { + {vget_low_s16(vmovl_s8(vget_low_s8(b40_s8))), vget_high_s16(vmovl_s8(vget_low_s8(b40_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(b40_s8))), vget_high_s16(vmovl_s8(vget_high_s8(b40_s8)))}}; + + const int16x4x4_t b50_s16 = { + {vget_low_s16(vmovl_s8(vget_low_s8(b50_s8))), vget_high_s16(vmovl_s8(vget_low_s8(b50_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(b50_s8))), vget_high_s16(vmovl_s8(vget_high_s8(b50_s8)))}}; + + const int16x4x4_t b60_s16 = { + {vget_low_s16(vmovl_s8(vget_low_s8(b60_s8))), vget_high_s16(vmovl_s8(vget_low_s8(b60_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(b60_s8))), vget_high_s16(vmovl_s8(vget_high_s8(b60_s8)))}}; + + const int16x4x4_t b70_s16 = { + {vget_low_s16(vmovl_s8(vget_low_s8(b70_s8))), vget_high_s16(vmovl_s8(vget_low_s8(b70_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(b70_s8))), vget_high_s16(vmovl_s8(vget_high_s8(b70_s8)))}}; + + // Accumulate 0: + c0.val[0] = vmlal_lane_s16(c0.val[0], b00_s16.val[0], a00_s16.val[0], 0); + c0.val[1] = vmlal_lane_s16(c0.val[1], b00_s16.val[1], a00_s16.val[0], 0); + c0.val[2] = vmlal_lane_s16(c0.val[2], b00_s16.val[2], a00_s16.val[0], 0); + c0.val[3] = vmlal_lane_s16(c0.val[3], b00_s16.val[3], a00_s16.val[0], 0); + + // Accumulate 1: + c0.val[0] = vmlal_lane_s16(c0.val[0], b10_s16.val[0], a00_s16.val[0], 1); + c0.val[1] = vmlal_lane_s16(c0.val[1], b10_s16.val[1], a00_s16.val[0], 1); + c0.val[2] = vmlal_lane_s16(c0.val[2], b10_s16.val[2], a00_s16.val[0], 1); + c0.val[3] = vmlal_lane_s16(c0.val[3], b10_s16.val[3], a00_s16.val[0], 1); + + // Accumulate 2: + c0.val[0] = vmlal_lane_s16(c0.val[0], b20_s16.val[0], a00_s16.val[0], 2); + c0.val[1] = vmlal_lane_s16(c0.val[1], b20_s16.val[1], a00_s16.val[0], 2); + c0.val[2] = vmlal_lane_s16(c0.val[2], b20_s16.val[2], a00_s16.val[0], 2); + c0.val[3] = vmlal_lane_s16(c0.val[3], b20_s16.val[3], a00_s16.val[0], 2); + + // Accumulate 3: + c0.val[0] = vmlal_lane_s16(c0.val[0], b30_s16.val[0], a00_s16.val[0], 3); + c0.val[1] = vmlal_lane_s16(c0.val[1], b30_s16.val[1], a00_s16.val[0], 3); + c0.val[2] = vmlal_lane_s16(c0.val[2], b30_s16.val[2], a00_s16.val[0], 3); + c0.val[3] = vmlal_lane_s16(c0.val[3], b30_s16.val[3], a00_s16.val[0], 3); + + // Accumulate 4: + c0.val[0] = vmlal_lane_s16(c0.val[0], b40_s16.val[0], a00_s16.val[1], 0); + c0.val[1] = vmlal_lane_s16(c0.val[1], b40_s16.val[1], a00_s16.val[1], 0); + c0.val[2] = vmlal_lane_s16(c0.val[2], b40_s16.val[2], a00_s16.val[1], 0); + c0.val[3] = vmlal_lane_s16(c0.val[3], b40_s16.val[3], a00_s16.val[1], 0); + + // Accumulate 5: + c0.val[0] = vmlal_lane_s16(c0.val[0], b50_s16.val[0], a00_s16.val[1], 1); + c0.val[1] = vmlal_lane_s16(c0.val[1], b50_s16.val[1], a00_s16.val[1], 1); + c0.val[2] = vmlal_lane_s16(c0.val[2], b50_s16.val[2], a00_s16.val[1], 1); + c0.val[3] = vmlal_lane_s16(c0.val[3], b50_s16.val[3], a00_s16.val[1], 1); + + // Accumulate 6: + c0.val[0] = vmlal_lane_s16(c0.val[0], b60_s16.val[0], a00_s16.val[1], 2); + c0.val[1] = vmlal_lane_s16(c0.val[1], b60_s16.val[1], a00_s16.val[1], 2); + c0.val[2] = vmlal_lane_s16(c0.val[2], b60_s16.val[2], a00_s16.val[1], 2); + c0.val[3] = vmlal_lane_s16(c0.val[3], b60_s16.val[3], a00_s16.val[1], 2); + + // Accumulate 7: + c0.val[0] = vmlal_lane_s16(c0.val[0], b70_s16.val[0], a00_s16.val[1], 3); + c0.val[1] = vmlal_lane_s16(c0.val[1], b70_s16.val[1], a00_s16.val[1], 3); + c0.val[2] = vmlal_lane_s16(c0.val[2], b70_s16.val[2], a00_s16.val[1], 3); + c0.val[3] = vmlal_lane_s16(c0.val[3], b70_s16.val[3], a00_s16.val[1], 3); + + vec_a += 8; + matrix_b += 8 * stride_b; + } + + // This for loop performs the left-over accumulations + for (; vec_a < vec_a_end_addr;) + { + const int8x8_t a00_s8 = vld1_dup_s8(vec_a); + const int8x16_t b00_s8 = vld1q_s8(matrix_b); + + const int16x4x4_t b00_s16 = { + {vget_low_s16(vmovl_s8(vget_low_s8(b00_s8))), vget_high_s16(vmovl_s8(vget_low_s8(b00_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(b00_s8))), vget_high_s16(vmovl_s8(vget_high_s8(b00_s8)))}}; + + // Convert a00_s8 to uint16_t and get the lower part + const int16x4_t a00_s16 = vget_low_s16(vmovl_s8(a00_s8)); + + // Accumulate 0: + c0.val[0] = vmlal_lane_s16(c0.val[0], b00_s16.val[0], a00_s16, 0); + c0.val[1] = vmlal_lane_s16(c0.val[1], b00_s16.val[1], a00_s16, 0); + c0.val[2] = vmlal_lane_s16(c0.val[2], b00_s16.val[2], a00_s16, 0); + c0.val[3] = vmlal_lane_s16(c0.val[3], b00_s16.val[3], a00_s16, 0); + + vec_a += 1; + matrix_b += stride_b; + } + + auto vec_out = reinterpret_cast<int32_t *>(out.ptr()); + if (id.x() < (width_out - 16)) + { + vst1q_s32(vec_out + 0, c0.val[0]); + vst1q_s32(vec_out + 4, c0.val[1]); + vst1q_s32(vec_out + 8, c0.val[2]); + vst1q_s32(vec_out + 12, c0.val[3]); + } + else + { + auto left_over = width_out - id.x(); + for (auto k = 0; k < 4 && left_over; ++k) + { + for (auto j = 0; j < 4 && left_over; ++j, --left_over) + { + *(vec_out + k * 4 + j) = c0.val[k][j]; + } + } + } + }, + ina, inb, out); +} + +void inline matrix_multiply_u8( + Iterator &ina, Iterator &inb, Iterator &out, int width_b, const TensorInfo &out_info, const Window &window) +{ + const auto width_out = static_cast<int>(out_info.dimension(0)); + const auto height_out = static_cast<int>(out_info.dimension(1)); + const size_t out_stride = out_info.strides_in_bytes()[1] / out_info.element_size(); + execute_window_loop( + window, + [&](const Coordinates &id) + { + const uint8_t *mtx_a0 = ina.ptr(); + const uint8_t *mtx_b0 = inb.ptr(); + + // Note: Since the input are all positives, we can use uint32_t + // Accumulators for the block 0 + uint32x4x4_t c0 = {{vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0)}}; + + // Accumulators for the block 1 + uint32x4x4_t c1 = {{vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0)}}; + + // Accumulators for the block 2 + uint32x4x4_t c2 = {{vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0)}}; + + // Accumulators for the block 3 + uint32x4x4_t c3 = {{vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0)}}; + + for (int k = 0; k < width_b; k += 16, mtx_a0 += 4, mtx_b0 += 16) + { + const uint8x8_t a00_u8 = vld1_u8(mtx_a0); + const uint8x16_t b00_u8 = vld1q_u8(mtx_b0); + + // Convert a00_u8 to uint16_t and get the lower part + const uint16x4_t a00_u16 = vget_low_u16(vmovl_u8(a00_u8)); + + // Convert b00_s8 to uint16_t + const uint16x4x4_t b00_u16 = { + {vget_low_u16(vmovl_u8(vget_low_u8(b00_u8))), vget_high_u16(vmovl_u8(vget_low_u8(b00_u8))), + vget_low_u16(vmovl_u8(vget_high_u8(b00_u8))), vget_high_u16(vmovl_u8(vget_high_u8(b00_u8)))}}; + + // 4x4 block 0 + c0.val[0] = vmlal_lane_u16(c0.val[0], b00_u16.val[0], a00_u16, 0); + c0.val[1] = vmlal_lane_u16(c0.val[1], b00_u16.val[1], a00_u16, 0); + c0.val[2] = vmlal_lane_u16(c0.val[2], b00_u16.val[2], a00_u16, 0); + c0.val[3] = vmlal_lane_u16(c0.val[3], b00_u16.val[3], a00_u16, 0); + + // 4x4 block 1 + c1.val[0] = vmlal_lane_u16(c1.val[0], b00_u16.val[0], a00_u16, 1); + c1.val[1] = vmlal_lane_u16(c1.val[1], b00_u16.val[1], a00_u16, 1); + c1.val[2] = vmlal_lane_u16(c1.val[2], b00_u16.val[2], a00_u16, 1); + c1.val[3] = vmlal_lane_u16(c1.val[3], b00_u16.val[3], a00_u16, 1); + + // 4x4 block 2 + c2.val[0] = vmlal_lane_u16(c2.val[0], b00_u16.val[0], a00_u16, 2); + c2.val[1] = vmlal_lane_u16(c2.val[1], b00_u16.val[1], a00_u16, 2); + c2.val[2] = vmlal_lane_u16(c2.val[2], b00_u16.val[2], a00_u16, 2); + c2.val[3] = vmlal_lane_u16(c2.val[3], b00_u16.val[3], a00_u16, 2); + + // 4x4 block 3 + c3.val[0] = vmlal_lane_u16(c3.val[0], b00_u16.val[0], a00_u16, 3); + c3.val[1] = vmlal_lane_u16(c3.val[1], b00_u16.val[1], a00_u16, 3); + c3.val[2] = vmlal_lane_u16(c3.val[2], b00_u16.val[2], a00_u16, 3); + c3.val[3] = vmlal_lane_u16(c3.val[3], b00_u16.val[3], a00_u16, 3); + } + + auto mtx_out = reinterpret_cast<int32_t *>(out.ptr()); + + if (id.y() < height_out && id.x() < (width_out - 16)) + { + vst1q_s32(mtx_out + 0 * out_stride + 0, vreinterpretq_s32_u32(c0.val[0])); + vst1q_s32(mtx_out + 0 * out_stride + 4, vreinterpretq_s32_u32(c0.val[1])); + vst1q_s32(mtx_out + 0 * out_stride + 8, vreinterpretq_s32_u32(c0.val[2])); + vst1q_s32(mtx_out + 0 * out_stride + 12, vreinterpretq_s32_u32(c0.val[3])); + if (id.y() + 1 < height_out) + { + vst1q_s32(mtx_out + 1 * out_stride + 0, vreinterpretq_s32_u32(c1.val[0])); + vst1q_s32(mtx_out + 1 * out_stride + 4, vreinterpretq_s32_u32(c1.val[1])); + vst1q_s32(mtx_out + 1 * out_stride + 8, vreinterpretq_s32_u32(c1.val[2])); + vst1q_s32(mtx_out + 1 * out_stride + 12, vreinterpretq_s32_u32(c1.val[3])); + if (id.y() + 2 < height_out) + { + vst1q_s32(mtx_out + 2 * out_stride + 0, vreinterpretq_s32_u32(c2.val[0])); + vst1q_s32(mtx_out + 2 * out_stride + 4, vreinterpretq_s32_u32(c2.val[1])); + vst1q_s32(mtx_out + 2 * out_stride + 8, vreinterpretq_s32_u32(c2.val[2])); + vst1q_s32(mtx_out + 2 * out_stride + 12, vreinterpretq_s32_u32(c2.val[3])); + if (id.y() + 3 < height_out) + { + vst1q_s32(mtx_out + 3 * out_stride + 0, vreinterpretq_s32_u32(c3.val[0])); + vst1q_s32(mtx_out + 3 * out_stride + 4, vreinterpretq_s32_u32(c3.val[1])); + vst1q_s32(mtx_out + 3 * out_stride + 8, vreinterpretq_s32_u32(c3.val[2])); + vst1q_s32(mtx_out + 3 * out_stride + 12, vreinterpretq_s32_u32(c3.val[3])); + } + } + } + } + else + { + const auto left_over_value = width_out - id.x(); + auto left_over = left_over_value; + for (auto k = 0; k < 4 && left_over; ++k) + { + for (auto j = 0; j < 4 && left_over; ++j, --left_over) + { + *(mtx_out + k * 4 + j) = c0.val[k][j]; + } + } + if (id.y() + 1 < height_out) + { + left_over = left_over_value; + for (auto k = 0; k < 4 && left_over; ++k) + { + for (auto j = 0; j < 4 && left_over; ++j, --left_over) + { + *(mtx_out + out_stride + k * 4 + j) = c1.val[k][j]; + } + } + if (id.y() + 2 < height_out) + { + left_over = left_over_value; + for (auto k = 0; k < 4 && left_over; ++k) + { + for (auto j = 0; j < 4 && left_over; ++j, --left_over) + { + *(mtx_out + out_stride * 2 + k * 4 + j) = c2.val[k][j]; + } + } + if (id.y() + 3 < height_out) + { + left_over = left_over_value; + for (auto k = 0; k < 4 && left_over; ++k) + { + for (auto j = 0; j < 4 && left_over; ++j, --left_over) + { + *(mtx_out + out_stride * 3 + k * 4 + j) = c3.val[k][j]; + } + } + } + } + } + } + }, + ina, inb, out); +} + +void inline matrix_multiply_s8( + Iterator &ina, Iterator &inb, Iterator &out, int width_b, const TensorInfo &out_info, const Window &window) +{ + const auto width_out = static_cast<int>(out_info.dimension(0)); + const auto height_out = static_cast<int>(out_info.dimension(1)); + const size_t out_stride = out_info.strides_in_bytes()[1] / out_info.element_size(); + // 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<const int8_t *>(ina.ptr()); + auto *mtx_b0 = reinterpret_cast<const int8_t *>(inb.ptr()); + + // Note: Since the input are all positives, we can use uint32_t + // Accumulators for the block 0 + int32x4x4_t c0 = {{vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0)}}; + + // Accumulators for the block 1 + int32x4x4_t c1 = {{vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0)}}; + + // Accumulators for the block 2 + int32x4x4_t c2 = {{vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0)}}; + + // Accumulators for the block 3 + int32x4x4_t c3 = {{vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0), vdupq_n_s32(0)}}; + + for (int k = 0; k < width_b; k += 16, mtx_a0 += 4, mtx_b0 += 16) + { + const int8x8_t a00_s8 = vld1_s8(mtx_a0); + const int8x16_t b00_s8 = vld1q_s8(mtx_b0); + + // Convert a00_s8 to uint16_t and get the lower part + const int16x4_t a00_s16 = vget_low_s16(vmovl_s8(a00_s8)); + + // Convert b00_s8 to int16_t + const int16x4x4_t b00_s16 = { + {vget_low_s16(vmovl_s8(vget_low_s8(b00_s8))), vget_high_s16(vmovl_s8(vget_low_s8(b00_s8))), + vget_low_s16(vmovl_s8(vget_high_s8(b00_s8))), vget_high_s16(vmovl_s8(vget_high_s8(b00_s8)))}}; + + // 4x4 block 0 + c0.val[0] = vmlal_lane_s16(c0.val[0], b00_s16.val[0], a00_s16, 0); + c0.val[1] = vmlal_lane_s16(c0.val[1], b00_s16.val[1], a00_s16, 0); + c0.val[2] = vmlal_lane_s16(c0.val[2], b00_s16.val[2], a00_s16, 0); + c0.val[3] = vmlal_lane_s16(c0.val[3], b00_s16.val[3], a00_s16, 0); + + // 4x4 block 1 + c1.val[0] = vmlal_lane_s16(c1.val[0], b00_s16.val[0], a00_s16, 1); + c1.val[1] = vmlal_lane_s16(c1.val[1], b00_s16.val[1], a00_s16, 1); + c1.val[2] = vmlal_lane_s16(c1.val[2], b00_s16.val[2], a00_s16, 1); + c1.val[3] = vmlal_lane_s16(c1.val[3], b00_s16.val[3], a00_s16, 1); + + // 4x4 block 2 + c2.val[0] = vmlal_lane_s16(c2.val[0], b00_s16.val[0], a00_s16, 2); + c2.val[1] = vmlal_lane_s16(c2.val[1], b00_s16.val[1], a00_s16, 2); + c2.val[2] = vmlal_lane_s16(c2.val[2], b00_s16.val[2], a00_s16, 2); + c2.val[3] = vmlal_lane_s16(c2.val[3], b00_s16.val[3], a00_s16, 2); + + // 4x4 block 3 + c3.val[0] = vmlal_lane_s16(c3.val[0], b00_s16.val[0], a00_s16, 3); + c3.val[1] = vmlal_lane_s16(c3.val[1], b00_s16.val[1], a00_s16, 3); + c3.val[2] = vmlal_lane_s16(c3.val[2], b00_s16.val[2], a00_s16, 3); + c3.val[3] = vmlal_lane_s16(c3.val[3], b00_s16.val[3], a00_s16, 3); + } + auto mtx_out = reinterpret_cast<int32_t *>(out.ptr()); + if (id.y() < height_out && id.x() < (width_out - 16)) + { + vst1q_s32(mtx_out + 0 * out_stride + 0, c0.val[0]); + vst1q_s32(mtx_out + 0 * out_stride + 4, c0.val[1]); + vst1q_s32(mtx_out + 0 * out_stride + 8, c0.val[2]); + vst1q_s32(mtx_out + 0 * out_stride + 12, c0.val[3]); + if (id.y() + 1 < height_out) + { + vst1q_s32(mtx_out + 1 * out_stride + 0, c1.val[0]); + vst1q_s32(mtx_out + 1 * out_stride + 4, c1.val[1]); + vst1q_s32(mtx_out + 1 * out_stride + 8, c1.val[2]); + vst1q_s32(mtx_out + 1 * out_stride + 12, c1.val[3]); + if (id.y() + 2 < height_out) + { + vst1q_s32(mtx_out + 2 * out_stride + 0, c2.val[0]); + vst1q_s32(mtx_out + 2 * out_stride + 4, c2.val[1]); + vst1q_s32(mtx_out + 2 * out_stride + 8, c2.val[2]); + vst1q_s32(mtx_out + 2 * out_stride + 12, c2.val[3]); + if (id.y() + 3 < height_out) + { + vst1q_s32(mtx_out + 3 * out_stride + 0, c3.val[0]); + vst1q_s32(mtx_out + 3 * out_stride + 4, c3.val[1]); + vst1q_s32(mtx_out + 3 * out_stride + 8, c3.val[2]); + vst1q_s32(mtx_out + 3 * out_stride + 12, c3.val[3]); + } + } + } + } + else if (id.y() < height_out) + { + const auto left_over_value = width_out - id.x(); + auto left_over = left_over_value; + for (auto k = 0; k < 4 && left_over; ++k) + { + for (auto j = 0; j < 4 && left_over; ++j, --left_over) + { + *(mtx_out + k * 4 + j) = c0.val[k][j]; + } + } + if (id.y() + 1 < height_out) + { + left_over = left_over_value; + for (auto k = 0; k < 4 && left_over; ++k) + { + for (auto j = 0; j < 4 && left_over; ++j, --left_over) + { + *(mtx_out + out_stride + k * 4 + j) = c1.val[k][j]; + } + } + if (id.y() + 2 < height_out) + { + left_over = left_over_value; + for (auto k = 0; k < 4 && left_over; ++k) + { + for (auto j = 0; j < 4 && left_over; ++j, --left_over) + { + *(mtx_out + out_stride * 2 + k * 4 + j) = c2.val[k][j]; + } + } + if (id.y() + 3 < height_out) + { + left_over = left_over_value; + for (auto k = 0; k < 4 && left_over; ++k) + { + for (auto j = 0; j < 4 && left_over; ++j, --left_over) + { + *(mtx_out + out_stride * 3 + k * 4 + j) = c3.val[k][j]; + } + } + } + } + } + } + }, + ina, inb, out); +} + +Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::S8, DataType::U8); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src1, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::QSYMM8, DataType::QSYMM8_PER_CHANNEL, DataType::S8, + DataType::U8); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::S32); + + TensorShape in0_shape = src0->tensor_shape(); + TensorShape in1_shape = src1->tensor_shape(); + TensorShape out_shape = dst->tensor_shape(); + + // Check vector-by-matrix case + if (out_shape[1] == 1) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(in0_shape[0] != in1_shape[1], + "The number of input0's columns must be equal to input1's rows"); + } + else + { + in0_shape.collapse(2); + in1_shape.collapse(2); + out_shape.collapse(2); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(in0_shape[2] != out_shape[2], + "Output tensor must have the same number of batches of input0 tensor"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG( + in1_shape[2] != 1 && in0_shape[2] != in1_shape[2], + "Input1 tensor must have the same number of batches of input0 or the number of batches must be set to 1"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(in1_shape[0] % 16, "Input1's width must be a multiple of 16"); + } + + return Status{}; +} +} // namespace + +void CpuGemmLowpMatrixMultiplyKernel::configure(const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst) +{ + ARM_COMPUTE_UNUSED(src0); + ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, dst)); + + TensorShape in1_shape = src1->tensor_shape(); + in1_shape.collapse(2); + + _slide_matrix_b = in1_shape[2] != 1; + + constexpr unsigned int num_elems_processed_per_iteration_x = 16; + constexpr unsigned int num_elems_processed_per_iteration_y = 4; + + Window win; + // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication + if ((dst->dimension(1) == 1)) + { + // Configure kernel window + win = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x)); + } + else + { + win = + calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); + } + + ICpuKernel::configure(win); +} + +Status +CpuGemmLowpMatrixMultiplyKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, dst)); + return Status{}; +} + +void CpuGemmLowpMatrixMultiplyKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); + + auto src0 = tensors.get_const_tensor(TensorType::ACL_SRC_0); + auto src1 = tensors.get_const_tensor(TensorType::ACL_SRC_1); + auto dst = tensors.get_tensor(TensorType::ACL_DST); + + // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication path + if ((dst->info()->dimension(1) == 1)) + { + const auto width_matrix_a = static_cast<int>(src0->info()->dimension(0)); + const auto width_matrix_b = static_cast<int>(src1->info()->dimension(0)); + const auto width_out = static_cast<int>(dst->info()->dimension(0)); + const auto in_b_stride = + static_cast<int>(src1->info()->strides_in_bytes()[1] / data_size_from_type(src1->info()->data_type())); + + // The implementation computes 16 elements per iteration + const int window_start_x = 16 * info.thread_id; + const int window_step_x = 16 * info.num_threads; + // Make sure (window_end_x - window_start_x) is a multiple of window_step_x + const int window_end_x = ceil_to_multiple(width_matrix_b - window_start_x, window_step_x) + window_start_x; + + Window win_out(window); + win_out.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); + 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 (src1->info()->num_dimensions() >= 3) + { + win_b = window; + } + win_b.set(Window::DimX, Window::Dimension(window_start_x, window_end_x, window_step_x)); + win_b.set(Window::DimY, Window::Dimension(0, 1, 1)); + + Iterator ina(src0, win_a); + Iterator inb(src1, win_b); + Iterator out(dst, win_out); + + switch (src0->info()->data_type()) + { + case DataType::S8: + case DataType::QASYMM8_SIGNED: + { + vector_matrix_multiply_s8(ina, inb, out, width_matrix_a, width_matrix_b, width_out, in_b_stride, + window); + break; + } + case DataType::U8: + case DataType::QASYMM8: + { + vector_matrix_multiply_u8(ina, inb, out, width_matrix_a, width_matrix_b, width_out, in_b_stride, + window); + break; + } + default: + { + ARM_COMPUTE_ERROR("Not supported"); + break; + } + } + } + else + { + const size_t in_b_stride = src1->info()->strides_in_bytes()[1]; + const int width_b = src1->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 output 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, window.y().end() / 4, 1)); + + // Set step_x and step_y for matrix B. Scale by a factor of 16 the X range as the input transposed matrix A has 16 times less the columns of the output matrix + 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 (_slide_matrix_b) + { + win_b = window; + } + win_b.set(Window::DimX, Window::Dimension(window.x().start() / 16, window.x().end() / 16, in_b_stride)); + win_b.set(Window::DimY, Window::Dimension(0, 0, 0)); + + // The step x and step y for the output matrix has been already set using in configure() + Iterator ina(src0, win_a); + Iterator inb(src1, win_b); + Iterator out(dst, window); + + switch (src0->info()->data_type()) + { + case DataType::S8: + case DataType::QASYMM8_SIGNED: + { + matrix_multiply_s8(ina, inb, out, width_b, *dst->info(), window); + break; + } + case DataType::U8: + case DataType::QASYMM8: + { + matrix_multiply_u8(ina, inb, out, width_b, *dst->info(), window); + break; + } + default: + { + ARM_COMPUTE_ERROR("Not supported"); + break; + } + } + } +} + +const char *CpuGemmLowpMatrixMultiplyKernel::name() const +{ + return "CpuGemmLowpMatrixMultiplyKernel"; +} +} // namespace kernels +} // namespace cpu +} // namespace arm_compute |