/* * SPDX-FileCopyrightText: Copyright 2023-2024 Arm Limited and/or its affiliates <open-source-office@arm.com> * * SPDX-License-Identifier: Apache-2.0 * * Licensed under the Apache License, Version 2.0 (the License); you may * not use this file except in compliance with the License. * You may obtain a copy of the License at * * www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an AS IS BASIS, WITHOUT * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ /* ---------------------------------------------------------------------- * Project: CMSIS NN Library * Title: arm_nn_vec_mat_mult_t_s4 * Description: s4 vector by matrix (transposed) multiplication * * $Date: 26 April 2024 * $Revision: V.2.0.0 * * Target : Arm(R) M-Profile Architecture * * -------------------------------------------------------------------- */ #include "arm_nnsupportfunctions.h" /** */ /** * @defgroup supportFC Fully Connected * * Support functions for Fully Connected * */ /** * @addtogroup supportFC * @{ */ /* * s4 vector(lhs) by matrix (transposed) multiplication * * Refer header file for details. * */ arm_cmsis_nn_status arm_nn_vec_mat_mult_t_s4(const int8_t *lhs, const int8_t *packed_rhs, const int32_t *bias, int8_t *dst, const int32_t lhs_offset, const int32_t dst_offset, const int32_t dst_multiplier, const int32_t dst_shift, const int32_t rhs_cols, const int32_t rhs_rows, const int32_t activation_min, const int32_t activation_max) { const int32_t row_loop_cnt = rhs_rows / 4; const int rhs_offset = rhs_cols * row_loop_cnt; const int8_t *rhs_ptr = &packed_rhs[0]; #if defined(ARM_MATH_MVEI) const int rhs_cols_offset = rhs_cols % 16; #else const int rhs_cols_offset = rhs_cols; #endif #if defined(ARM_MATH_DSP) const int16_t lhs_offset_s16 = (int16_t)lhs_offset; const uint32_t lhs_offset_s16x2 = PKHBT(lhs_offset_s16, lhs_offset_s16, 16); #endif int32_t spillover0, spillover1; #if defined(ARM_MATH_MVEI) const mve_pred16_t lower_nibble_mask = 21845; // 0101010101010101 const uint8x16_t gather_offset = {0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}; const int32_t col_loop_cnt = rhs_cols >> 5; const int I6_elements_spill = rhs_cols & 0x10; for (int32_t i_row_loop_cnt = 0; i_row_loop_cnt < row_loop_cnt; ++i_row_loop_cnt) { const uint32x4_t scatter_offset = {0, 1, 2 * row_loop_cnt, 2 * row_loop_cnt + 1}; const int8_t *lhs_ptr = &lhs[0]; mve_pred16_t rmdr_mask = vctp8q(rhs_cols_offset); int32_t acc0 = 0; int32_t acc2 = 0; int32_t rhs_sum_0 = 0; int32_t rhs_sum_2 = 0; if (bias) { acc0 += *bias; acc2 += bias[2 * row_loop_cnt]; ++bias; } for (int i = 0; i < col_loop_cnt; i++) { const int8x16x2_t inputx2 = vld2q_s8(lhs_ptr); const int8x16_t ker_0 = vldrbq_s8(rhs_ptr); int8x16_t ker_low_0 = vrshlq_n_s8(ker_0, 4); ker_low_0 = vshrq_n_s8(ker_low_0, 4); int8x16_t ker_high_0 = vshrq_n_s8(ker_0, 4); rhs_sum_0 = vaddvaq_s8(rhs_sum_0, ker_low_0); rhs_sum_0 = vaddvaq_s8(rhs_sum_0, ker_high_0); acc0 = vmladavaq_s8(acc0, ker_low_0, inputx2.val[0]); acc0 = vmladavaq_s8(acc0, ker_high_0, inputx2.val[1]); const int8x16_t ker_1 = vldrbq_s8(&rhs_ptr[rhs_offset]); int8x16_t ker_low_1 = vrshlq_n_s8(ker_1, 4); ker_low_1 = vshrq_n_s8(ker_low_1, 4); int8x16_t ker_high_1 = vshrq_n_s8(ker_1, 4); rhs_sum_2 = vaddvaq_s8(rhs_sum_2, ker_low_1); rhs_sum_2 = vaddvaq_s8(rhs_sum_2, ker_high_1); acc2 = vmladavaq_s8(acc2, ker_low_1, inputx2.val[0]); acc2 = vmladavaq_s8(acc2, ker_high_1, inputx2.val[1]); lhs_ptr += 32; rhs_ptr += 16; } if (I6_elements_spill) { const int8x16_t input = vldrbq_s8(lhs_ptr); int8x16_t ker_0 = vldrbq_gather_offset_s8(rhs_ptr, gather_offset); ker_0 = vrshlq_m_n_s8(ker_0, 4, lower_nibble_mask); ker_0 = vshrq_n_s8(ker_0, 4); rhs_sum_0 = vaddvaq_s8(rhs_sum_0, ker_0); acc0 = vmladavaq_s8(acc0, ker_0, input); int8x16_t ker_1 = vldrbq_gather_offset_s8(&rhs_ptr[rhs_offset], gather_offset); ker_1 = vrshlq_m_n_s8(ker_1, 4, lower_nibble_mask); ker_1 = vshrq_n_s8(ker_1, 4); rhs_sum_2 = vaddvaq_s8(rhs_sum_2, ker_1); acc2 = vmladavaq_s8(acc2, ker_1, input); lhs_ptr += 16; rhs_ptr += 8; } if (rmdr_mask) { const int8x16_t input = vldrbq_z_s8(lhs_ptr, rmdr_mask); int8x16_t ker_0 = vldrbq_gather_offset_z_s8(rhs_ptr, gather_offset, rmdr_mask); ker_0 = vrshlq_m_n_s8(ker_0, 4, lower_nibble_mask); ker_0 = vshrq_n_s8(ker_0, 4); rhs_sum_0 = vaddvaq_s8(rhs_sum_0, ker_0); acc0 = vmladavaq_s8(acc0, ker_0, input); int8x16_t ker_1 = vldrbq_gather_offset_z_s8(&rhs_ptr[rhs_offset], gather_offset, rmdr_mask); ker_1 = vrshlq_m_n_s8(ker_1, 4, lower_nibble_mask); ker_1 = vshrq_n_s8(ker_1, 4); rhs_sum_2 = vaddvaq_s8(rhs_sum_2, ker_1); acc2 = vmladavaq_s8(acc2, ker_1, input); rhs_ptr += rhs_cols_offset >> 1; } if (rhs_cols & 1) { const int32_t rhs_high0 = rhs_ptr[0] >> 4; const int32_t rhs_high1 = rhs_ptr[rhs_offset] >> 4; lhs_ptr = &lhs[0]; const int32_t lhs_high = (int8_t)lhs_ptr[0] + lhs_offset; spillover0 = lhs_high * rhs_high0; spillover1 = lhs_high * rhs_high1; rmdr_mask >>= 1; ++lhs_ptr; ++rhs_ptr; } else { spillover0 = 0; spillover1 = 0; lhs_ptr = &lhs[0]; } int32_t acc1 = spillover0; int32_t acc3 = spillover1; if (bias) { acc1 += *bias; acc3 += bias[2 * row_loop_cnt]; ++bias; } int32_t rhs_sum_1 = 0; int32_t rhs_sum_3 = 0; for (int i = 0; i < col_loop_cnt; i++) { const int8x16x2_t inputx2 = vld2q_s8(lhs_ptr); int8x16_t ker_0 = vldrbq_s8(rhs_ptr); int8x16_t ker_low_0 = vrshlq_n_s8(ker_0, 4); ker_low_0 = vshrq_n_s8(ker_low_0, 4); int8x16_t ker_high_0 = vshrq_n_s8(ker_0, 4); rhs_sum_1 = vaddvaq_s8(rhs_sum_1, ker_low_0); rhs_sum_1 = vaddvaq_s8(rhs_sum_1, ker_high_0); acc1 = vmladavaq_s8(acc1, ker_low_0, inputx2.val[0]); acc1 = vmladavaq_s8(acc1, ker_high_0, inputx2.val[1]); int8x16_t ker_1 = vldrbq_s8(&rhs_ptr[rhs_offset]); int8x16_t ker_low_1 = vrshlq_n_s8(ker_1, 4); ker_low_1 = vshrq_n_s8(ker_low_1, 4); int8x16_t ker_high_1 = vshrq_n_s8(ker_1, 4); rhs_sum_3 = vaddvaq_s8(rhs_sum_3, ker_low_1); rhs_sum_3 = vaddvaq_s8(rhs_sum_3, ker_high_1); acc3 = vmladavaq_s8(acc3, ker_low_1, inputx2.val[0]); acc3 = vmladavaq_s8(acc3, ker_high_1, inputx2.val[1]); lhs_ptr += 32; rhs_ptr += 16; } if (I6_elements_spill) { const int8x16_t input = vldrbq_s8(lhs_ptr); int8x16_t ker_0 = vldrbq_gather_offset_s8(rhs_ptr, gather_offset); ker_0 = vrshlq_m_n_s8(ker_0, 4, lower_nibble_mask); ker_0 = vshrq_n_s8(ker_0, 4); rhs_sum_1 = vaddvaq_s8(rhs_sum_1, ker_0); acc1 = vmladavaq_s8(acc1, ker_0, input); int8x16_t ker_1 = vldrbq_gather_offset_s8(&rhs_ptr[rhs_offset], gather_offset); ker_1 = vrshlq_m_n_s8(ker_1, 4, lower_nibble_mask); ker_1 = vshrq_n_s8(ker_1, 4); rhs_sum_3 = vaddvaq_s8(rhs_sum_3, ker_1); acc3 = vmladavaq_s8(acc3, ker_1, input); lhs_ptr += 16; rhs_ptr += 8; } if (rmdr_mask) { const int8x16_t input = vldrbq_z_s8(lhs_ptr, rmdr_mask); int8x16_t ker_0 = vldrbq_gather_offset_z_s8(rhs_ptr, gather_offset, rmdr_mask); ker_0 = vrshlq_m_n_s8(ker_0, 4, lower_nibble_mask); ker_0 = vshrq_n_s8(ker_0, 4); rhs_sum_1 = vaddvaq_s8(rhs_sum_1, ker_0); acc1 = vmladavaq_s8(acc1, ker_0, input); int8x16_t ker_1 = vldrbq_gather_offset_z_s8(&rhs_ptr[rhs_offset], gather_offset, rmdr_mask); ker_1 = vrshlq_m_n_s8(ker_1, 4, lower_nibble_mask); ker_1 = vshrq_n_s8(ker_1, 4); rhs_sum_3 = vaddvaq_s8(rhs_sum_3, ker_1); acc3 = vmladavaq_s8(acc3, ker_1, input); rhs_ptr += rhs_cols_offset >> 1; } int32x4_t acc = {acc0, acc1, acc2, acc3}; const int32x4_t rhs_sum = {rhs_sum_0, rhs_sum_1, rhs_sum_2, rhs_sum_3}; acc += vdupq_n_s32(lhs_offset) * rhs_sum; acc = arm_requantize_mve(acc, dst_multiplier, dst_shift); acc = vaddq_s32(acc, vdupq_n_s32(dst_offset)); acc = vmaxq_s32(acc, vdupq_n_s32(activation_min)); acc = vminq_s32(acc, vdupq_n_s32(activation_max)); vstrbq_scatter_offset_s32(dst, scatter_offset, acc); dst += 2; } #elif defined(ARM_MATH_DSP) for (int32_t i_row_loop_cnt = 0; i_row_loop_cnt < row_loop_cnt; ++i_row_loop_cnt) { const int8_t *lhs_ptr = &lhs[0]; int32_t res0 = 0; int32_t res1 = 0; if (bias) { res0 += *bias; res1 += bias[2 * row_loop_cnt]; ++bias; } for (int rhs_cols_idx = 0; rhs_cols_idx < (rhs_cols / 4); ++rhs_cols_idx) { int32_t lhs_high, rhs_high0, rhs_low0, lhs_low, rhs_high1, rhs_low1; read_and_pad_s4(rhs_ptr, &rhs_low0, &rhs_high0); read_and_pad_s4((const int8_t *)&rhs_ptr[rhs_offset], &rhs_low1, &rhs_high1); rhs_ptr += 2; lhs_high = arm_nn_read_s8x4_ia((const int8_t **)&lhs_ptr); lhs_low = SXTAB16(lhs_offset_s16x2, lhs_high); lhs_high = SXTAB16_RORn(lhs_offset_s16x2, lhs_high, 8); res0 = SMLAD(lhs_low, rhs_low0, res0); res0 = SMLAD(lhs_high, rhs_high0, res0); res1 = SMLAD(lhs_low, rhs_low1, res1); res1 = SMLAD(lhs_high, rhs_high1, res1); } if (((rhs_cols % 4) == 2) || ((rhs_cols % 4) == 3)) { const int32_t rhs_value0 = rhs_ptr[0]; const int32_t lower0 = (int8_t)(rhs_value0 << 4) >> 4; const int32_t higher0 = rhs_value0 >> 4; const int32_t rhs_value1 = rhs_ptr[rhs_offset]; const int32_t lower1 = (int8_t)(rhs_value1 << 4) >> 4; const int32_t higher1 = rhs_value1 >> 4; const int32_t lhs_value_0 = lhs_ptr[0] + lhs_offset; const int32_t lhs_value_1 = lhs_ptr[1] + lhs_offset; res0 += lhs_value_0 * lower0; res0 += lhs_value_1 * higher0; res1 += lhs_value_0 * lower1; res1 += lhs_value_1 * higher1; ++rhs_ptr; lhs_ptr += 2; } if (rhs_cols % 2 == 1) { const int32_t rhs_low0 = (int8_t)(rhs_ptr[0] << 4) >> 4; const int32_t rhs_high0 = rhs_ptr[0] >> 4; const int32_t rhs_low1 = (int8_t)(rhs_ptr[rhs_offset] << 4) >> 4; const int32_t rhs_high1 = rhs_ptr[rhs_offset] >> 4; const int32_t lhs_low = (int8_t)lhs_ptr[0] + lhs_offset; lhs_ptr = &lhs[0]; const int32_t lhs_high = (int8_t)lhs_ptr[0] + lhs_offset; ++lhs_ptr; res0 += lhs_low * rhs_low0; spillover0 = lhs_high * rhs_high0; res1 += lhs_low * rhs_low1; spillover1 = lhs_high * rhs_high1; ++rhs_ptr; } else { spillover0 = 0; spillover1 = 0; lhs_ptr = &lhs[0]; } // Quantize down res0 = arm_nn_requantize(res0, dst_multiplier, dst_shift); res1 = arm_nn_requantize(res1, dst_multiplier, dst_shift); // Add offset res0 += dst_offset; res1 += dst_offset; // Clamp the result res0 = MAX(res0, activation_min); res0 = MIN(res0, activation_max); res1 = MAX(res1, activation_min); res1 = MIN(res1, activation_max); *dst = (int8_t)res0; *(dst + 2 * row_loop_cnt) = (int8_t)res1; dst++; res0 = spillover0; res1 = spillover1; if (bias) { res0 += *bias; res1 += bias[2 * row_loop_cnt]; ++bias; } for (int rhs_cols_idx = 0; rhs_cols_idx < rhs_cols / 4; ++rhs_cols_idx) { int32_t lhs_high, rhs_high0, rhs_low0, lhs_low, rhs_high1, rhs_low1; read_and_pad_s4(rhs_ptr, &rhs_low0, &rhs_high0); read_and_pad_s4((const int8_t *)&rhs_ptr[rhs_offset], &rhs_low1, &rhs_high1); rhs_ptr += 2; lhs_high = arm_nn_read_s8x4_ia((const int8_t **)&lhs_ptr); lhs_low = SXTAB16(lhs_offset_s16x2, lhs_high); lhs_high = SXTAB16_RORn(lhs_offset_s16x2, lhs_high, 8); res0 = SMLAD(lhs_low, rhs_low0, res0); res0 = SMLAD(lhs_high, rhs_high0, res0); res1 = SMLAD(lhs_low, rhs_low1, res1); res1 = SMLAD(lhs_high, rhs_high1, res1); } if (((rhs_cols % 4) == 2) || ((rhs_cols % 4) == 3)) { const int32_t rhs_value0 = rhs_ptr[0]; const int32_t lower0 = (int8_t)(rhs_value0 << 4) >> 4; const int32_t higher0 = rhs_value0 >> 4; const int32_t rhs_value1 = rhs_ptr[rhs_offset]; const int32_t lower1 = (int8_t)(rhs_value1 << 4) >> 4; const int32_t higher1 = rhs_value1 >> 4; const int32_t lhs_value_0 = lhs_ptr[0] + lhs_offset; const int32_t lhs_value_1 = lhs_ptr[1] + lhs_offset; res0 += lhs_value_0 * lower0; res0 += lhs_value_1 * higher0; res1 += lhs_value_0 * lower1; res1 += lhs_value_1 * higher1; ++rhs_ptr; lhs_ptr += 2; } // Quantize down res0 = arm_nn_requantize(res0, dst_multiplier, dst_shift); res1 = arm_nn_requantize(res1, dst_multiplier, dst_shift); // Add offset res0 += dst_offset; res1 += dst_offset; // Clamp the result res0 = MAX(res0, activation_min); res0 = MIN(res0, activation_max); res1 = MAX(res1, activation_min); res1 = MIN(res1, activation_max); *dst = (int8_t)res0; *(dst + 2 * row_loop_cnt) = (int8_t)res1; dst++; } #else for (int i_row_loop_cnt = 0; i_row_loop_cnt < row_loop_cnt; ++i_row_loop_cnt) { const int8_t *lhs_ptr = &lhs[0]; int32_t res0 = 0; int32_t res1 = 0; if (bias) { res0 += *bias; res1 += bias[2 * row_loop_cnt]; ++bias; } for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols / 2; ++rhs_cols_idx) { const int32_t rhs_low0 = (int8_t)(rhs_ptr[0] << 4) >> 4; const int32_t rhs_high0 = rhs_ptr[0] >> 4; const int32_t rhs_low1 = (int8_t)(rhs_ptr[rhs_offset] << 4) >> 4; const int32_t rhs_high1 = rhs_ptr[rhs_offset] >> 4; const int32_t lhs_low = (int8_t)lhs_ptr[0] + lhs_offset; const int32_t lhs_high = (int8_t)lhs_ptr[1] + lhs_offset; res0 += lhs_low * rhs_low0; res0 += lhs_high * rhs_high0; res1 += lhs_low * rhs_low1; res1 += lhs_high * rhs_high1; ++rhs_ptr; lhs_ptr += 2; } if (rhs_cols % 2 == 1) { const int32_t rhs_low0 = (int8_t)(rhs_ptr[0] << 4) >> 4; const int32_t rhs_high0 = rhs_ptr[0] >> 4; const int32_t rhs_low1 = (int8_t)(rhs_ptr[rhs_offset] << 4) >> 4; const int32_t rhs_high1 = rhs_ptr[rhs_offset] >> 4; const int32_t lhs_low = (int8_t)lhs_ptr[0] + lhs_offset; lhs_ptr = &lhs[0]; const int32_t lhs_high = (int8_t)lhs_ptr[0] + lhs_offset; ++lhs_ptr; res0 += lhs_low * rhs_low0; spillover0 = lhs_high * rhs_high0; res1 += lhs_low * rhs_low1; spillover1 = lhs_high * rhs_high1; ++rhs_ptr; } else { spillover0 = 0; spillover1 = 0; lhs_ptr = &lhs[0]; } // Quantize down res0 = arm_nn_requantize(res0, dst_multiplier, dst_shift); res1 = arm_nn_requantize(res1, dst_multiplier, dst_shift); // Add offset res0 += dst_offset; res1 += dst_offset; // Clamp the result res0 = MAX(res0, activation_min); res0 = MIN(res0, activation_max); res1 = MAX(res1, activation_min); res1 = MIN(res1, activation_max); *dst = (int8_t)res0; *(dst + 2 * row_loop_cnt) = (int8_t)res1; dst++; res0 = spillover0; res1 = spillover1; if (bias) { res0 += *bias; res1 += bias[2 * row_loop_cnt]; ++bias; } for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols / 2; ++rhs_cols_idx) { const int32_t rhs_low0 = (int8_t)(rhs_ptr[0] << 4) >> 4; const int32_t rhs_high0 = rhs_ptr[0] >> 4; const int32_t rhs_low1 = (int8_t)(rhs_ptr[rhs_offset] << 4) >> 4; const int32_t rhs_high1 = rhs_ptr[rhs_offset] >> 4; const int32_t lhs_low = (int8_t)lhs_ptr[0] + lhs_offset; const int32_t lhs_high = (int8_t)lhs_ptr[1] + lhs_offset; res0 += lhs_low * rhs_low0; res0 += lhs_high * rhs_high0; res1 += lhs_low * rhs_low1; res1 += lhs_high * rhs_high1; ++rhs_ptr; lhs_ptr += 2; } // Quantize down res0 = arm_nn_requantize(res0, dst_multiplier, dst_shift); res1 = arm_nn_requantize(res1, dst_multiplier, dst_shift); // Add offset res0 += dst_offset; res1 += dst_offset; // Clamp the result res0 = MAX(res0, activation_min); res0 = MIN(res0, activation_max); res1 = MAX(res1, activation_min); res1 = MIN(res1, activation_max); *dst = (int8_t)res0; *(dst + 2 * row_loop_cnt) = (int8_t)res1; dst++; } #endif const int8_t *lhs_ptr = &lhs[0]; spillover0 = 0; for (int32_t i_row_loop_cnt = 0; i_row_loop_cnt < rhs_rows % 4; ++i_row_loop_cnt) { int32_t res0 = spillover0; if (bias) { res0 += bias[2 * row_loop_cnt]; ++bias; } #if defined(ARM_MATH_MVEI) int32_t rhs_sum_0 = 0; for (int i = 0; i < col_loop_cnt; i++) { const int8x16x2_t inputx2 = vld2q_s8(lhs_ptr); const int8x16_t ker_0 = vldrbq_s8(&rhs_ptr[rhs_offset]); int8x16_t ker_low_0 = vrshlq_n_s8(ker_0, 4); ker_low_0 = vshrq_n_s8(ker_low_0, 4); int8x16_t ker_high_0 = vshrq_n_s8(ker_0, 4); rhs_sum_0 = vaddvaq_s8(rhs_sum_0, ker_low_0); rhs_sum_0 = vaddvaq_s8(rhs_sum_0, ker_high_0); res0 = vmladavaq_s8(res0, ker_low_0, inputx2.val[0]); res0 = vmladavaq_s8(res0, ker_high_0, inputx2.val[1]); lhs_ptr += 32; rhs_ptr += 16; } if (I6_elements_spill) { const int8x16_t input = vldrbq_s8(lhs_ptr); int8x16_t ker_0 = vldrbq_gather_offset_s8(&rhs_ptr[rhs_offset], gather_offset); ker_0 = vrshlq_m_n_s8(ker_0, 4, lower_nibble_mask); ker_0 = vshrq_n_s8(ker_0, 4); rhs_sum_0 = vaddvaq_s8(rhs_sum_0, ker_0); res0 = vmladavaq_s8(res0, ker_0, input); lhs_ptr += 16; rhs_ptr += 8; } res0 += lhs_offset * rhs_sum_0; #endif #if defined(ARM_MATH_DSP) for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols_offset / 4; ++rhs_cols_idx) { int32_t lhs_high, rhs_high0, rhs_low0, lhs_low; read_and_pad_s4((const int8_t *)&rhs_ptr[rhs_offset], &rhs_high0, &rhs_low0); rhs_ptr += 2; lhs_high = arm_nn_read_s8x4_ia((const int8_t **)&lhs_ptr); lhs_low = SXTAB16(lhs_offset_s16x2, lhs_high); lhs_high = SXTAB16_RORn(lhs_offset_s16x2, lhs_high, 8); res0 = SMLAD(lhs_low, rhs_high0, res0); res0 = SMLAD(lhs_high, rhs_low0, res0); } if ((rhs_cols % 4) == 2 || (rhs_cols % 4 == 3)) { const int32_t rhs_value0 = rhs_ptr[rhs_offset]; const int32_t lower0 = (int8_t)(rhs_value0 << 4) >> 4; const int32_t higher0 = rhs_value0 >> 4; const int32_t lhs_value_0 = lhs_ptr[0] + lhs_offset; const int32_t lhs_value_1 = lhs_ptr[1] + lhs_offset; res0 += lhs_value_0 * lower0; res0 += lhs_value_1 * higher0; ++rhs_ptr; lhs_ptr += 2; } #else for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols_offset / 2; ++rhs_cols_idx) { const int32_t rhs_low0 = (int8_t)(rhs_ptr[rhs_offset] << 4) >> 4; const int32_t rhs_high0 = rhs_ptr[rhs_offset] >> 4; const int32_t lhs_low = (int8_t)lhs_ptr[0] + lhs_offset; const int32_t lhs_high = (int8_t)lhs_ptr[1] + lhs_offset; res0 += lhs_low * rhs_low0; res0 += lhs_high * rhs_high0; ++rhs_ptr; lhs_ptr += 2; } #endif if ((rhs_cols % 2 == 1) && (i_row_loop_cnt % 2 == 0)) { const int32_t rhs_low0 = (int8_t)(rhs_ptr[rhs_offset] << 4) >> 4; const int32_t rhs_high0 = rhs_ptr[rhs_offset] >> 4; const int32_t lhs_low = (int8_t)lhs_ptr[0] + lhs_offset; lhs_ptr = &lhs[0]; const int32_t lhs_high = (int8_t)lhs_ptr[0] + lhs_offset; ++lhs_ptr; res0 += lhs_low * rhs_low0; spillover0 = lhs_high * rhs_high0; ++rhs_ptr; } else { spillover0 = 0; lhs_ptr = &lhs[0]; } // Quantize down res0 = arm_nn_requantize(res0, dst_multiplier, dst_shift); // Add offset res0 += dst_offset; // Clamp the result res0 = MAX(res0, activation_min); res0 = MIN(res0, activation_max); *(dst + 2 * row_loop_cnt) = (int8_t)res0; dst++; } return ARM_CMSIS_NN_SUCCESS; } /** * @} end of Doxygen group */