/* * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. * * 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_depthwise_separable_conv_HWC_q7_nonsquare.c * Description: Q7 depthwise separable convolution function (non-square shape) * * $Date: 17. January 2018 * $Revision: V.1.0.0 * * Target Processor: Cortex-M cores * * -------------------------------------------------------------------- */ #include "arm_math.h" #include "arm_nnfunctions.h" /** * @ingroup groupNN */ /** * @addtogroup NNConv * @{ */ /** * @brief Q7 depthwise separable convolution function (non-square shape) * @param[in] Im_in pointer to input tensor * @param[in] dim_im_in_x input tensor dimention x * @param[in] dim_im_in_y input tensor dimention y * @param[in] ch_im_in number of input tensor channels * @param[in] wt pointer to kernel weights * @param[in] ch_im_out number of filters, i.e., output tensor channels * @param[in] dim_kernel_x filter kernel size x * @param[in] dim_kernel_y filter kernel size y * @param[in] padding_x padding sizes x * @param[in] padding_y padding sizes y * @param[in] stride_x convolution stride x * @param[in] stride_y convolution stride y * @param[in] bias pointer to bias * @param[in] bias_shift amount of left-shift for bias * @param[in] out_shift amount of right-shift for output * @param[in,out] Im_out pointer to output tensor * @param[in] dim_im_out_x output tensor dimension x * @param[in] dim_im_out_y output tensor dimension y * @param[in,out] bufferA pointer to buffer space for input * @param[in,out] bufferB pointer to buffer space for output * @return The function returns either * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. * * This function is the version with full list of optimization tricks, but with * some contraints: * ch_im_in is multiple of 2 * ch_im_out is multiple of 2 */ arm_status arm_depthwise_separable_conv_HWC_q7_nonsquare(const q7_t * Im_in, const uint16_t dim_im_in_x, const uint16_t dim_im_in_y, const uint16_t ch_im_in, const q7_t * wt, const uint16_t ch_im_out, const uint16_t dim_kernel_x, const uint16_t dim_kernel_y, const uint16_t padding_x, const uint16_t padding_y, const uint16_t stride_x, const uint16_t stride_y, const q7_t * bias, const uint16_t bias_shift, const uint16_t out_shift, q7_t * Im_out, const uint16_t dim_im_out_x, const uint16_t dim_im_out_y, q15_t * bufferA, q7_t * bufferB) { #if defined (ARM_MATH_DSP) /* Run the following code for Cortex-M4 and Cortex-M7 */ /* * Implementation: * There are 3 nested loop here: * Inner loop: calculate each output value with MAC instruction over an accumulator * Mid loop: loop over different output channel * Outer loop: loop over different output (x, y) * */ int16_t i_out_y, i_out_x; int16_t i_ker_y, i_ker_x; q7_t *colBuffer = (q7_t *) bufferA; q7_t *pBuffer = colBuffer; const q7_t *pBias = bias; q7_t *pOut = Im_out; uint16_t rowCnt; uint16_t row_shift; /* do some checking here, basically ch_im_in == ch_im_out */ if (ch_im_in != ch_im_out) { return ARM_MATH_SIZE_MISMATCH; } for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++) { for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++) { /* we first do im2col here */ for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; i_ker_y++) { for (i_ker_x = i_out_x * stride_x - padding_x; i_ker_x < i_out_x * stride_x - padding_x + dim_kernel_x; i_ker_x++) { if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_x) { /* arm_fill_q7(0, pBuffer, ch_im_in); */ memset(pBuffer, 0, ch_im_in); } else { /* arm_copy_q7((q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */ memcpy(pBuffer, (q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, ch_im_in); } pBuffer += ch_im_in; } } /* we will do the computation here for each channel */ rowCnt = ch_im_out >> 2; row_shift = 0; pBias = bias; while (rowCnt) { q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); uint16_t colCnt = (dim_kernel_x * dim_kernel_y) >> 1; q7_t *pB = colBuffer + row_shift; const q7_t *pA = wt + row_shift; row_shift += 4; #ifdef USE_INTRINSIC #ifndef ARM_MATH_BIG_ENDIAN while (colCnt) { q31_t inA1, inA2, inB1, inB2, opA, opB; inB1 = *__SIMD32(pB); pB += ch_im_in; opB = *__SIMD32(pB); pB += ch_im_in; inB2 = __PKHTB(opB, inB1, 16); inB1 = __PKHBT(inB1, opB, 16); inA1 = *__SIMD32(pA); pA += ch_im_in; opB = *__SIMD32(pA); pA += ch_im_in; inA2 = __PKHTB(opB, inA1, 16); inA1 = __PKHBT(inA1, opB, 16); opA = __SXTB16(inA1); opB = __SXTB16(inB1); sum = __SMLAD(opA, opB, sum); opA = __SXTB16(__ROR(inA1, 8)); opB = __SXTB16(__ROR(inB1, 8)); sum2 = __SMLAD(opA, opB, sum2); opA = __SXTB16(inA2); opB = __SXTB16(inB2); sum3 = __SMLAD(opA, opB, sum3); opA = __SXTB16(__ROR(inA2, 8)); opB = __SXTB16(__ROR(inB2, 8)); sum4 = __SMLAD(opA, opB, sum4); colCnt--; } #else while (colCnt) { q31_t inA1, inA2, inB1, inB2, opA, opB; inB1 = *__SIMD32(pB); pB += ch_im_in; opB = *__SIMD32(pB); pB += ch_im_in; inB2 = __PKHBT(opB, inB1, 16); inB1 = __PKHTB(inB1, opB, 16); inA1 = *__SIMD32(pA); pA += ch_im_in; opB = *__SIMD32(pA); pA += ch_im_in; inA2 = __PKHBT(opB, inA1, 16); inA1 = __PKHTB(inA1, opB, 16); opA = __SXTB16(inA1); opB = __SXTB16(inB1); sum2 = __SMLAD(opA, opB, sum2); opA = __SXTB16(__ROR(inA1, 8)); opB = __SXTB16(__ROR(inB1, 8)); sum = __SMLAD(opA, opB, sum); opA = __SXTB16(inA2); opB = __SXTB16(inB2); sum4 = __SMLAD(opA, opB, sum4); opA = __SXTB16(__ROR(inA2, 8)); opB = __SXTB16(__ROR(inB2, 8)); sum3 = __SMLAD(opA, opB, sum3); colCnt--; } #endif /* ARM_MATH_BIG_ENDIAN */ #else #ifndef ARM_MATH_BIG_ENDIAN // r0 r1 r2 r3 r4 r5 // inA1, inA2, inB1, inB2, opA, opB asm volatile ("COL_LOOP:\n" "ldr.w r2, [%[pB], #0]\n" "add.w %[pB], %[pB], %[ch_im_in]\n" "ldr.w r5, [%[pB], #0]\n" "add.w %[pB], %[pB], %[ch_im_in]\n" "pkhtb r3, r5, r2, ASR #16\n" "pkhbt r2, r2, r5, LSL #16\n" "ldr.w r0, [%[pA], #0]\n" "add.w %[pA], %[pA], %[ch_im_in]\n" "ldr.w r5, [%[pA], #0]\n" "add.w %[pA], %[pA], %[ch_im_in]\n" "pkhtb r1, r5, r0, ASR #16\n" "pkhbt r0, r0, r5, LSL #16\n" "sxtb16 r4, r0\n" "sxtb16 r5, r2\n" "smlad %[sum], r4, r5, %[sum]\n" "mov.w r4, r0, ror #8\n" "mov.w r5, r2, ror #8\n" "sxtb16 r4, r4\n" "sxtb16 r5, r5\n" "smlad %[sum2], r4, r5, %[sum2]\n" "sxtb16 r4, r1\n" "sxtb16 r5, r3\n" "smlad %[sum3], r4, r5, %[sum3]\n" "mov.w r4, r1, ror #8\n" "mov.w r5, r3, ror #8\n" "sxtb16 r4, r4\n" "sxtb16 r5, r5\n" "smlad %[sum4], r4, r5, %[sum4]\n" "subs %[colCnt], #1\n" "bne COL_LOOP\n":[sum] "+r"(sum),[sum2] "+r"(sum2),[sum3] "+r"(sum3), [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt), [ch_im_in] "r"(ch_im_in):"r0", "r1", "r2", "r3", "r4", "r5"); #else // r0 r1 r2 r3 r4 r5 // inA1, inA2, inB1, inB2, opA, opB asm volatile ("COL_LOOP:\n" "ldr.w r2, [%[pB], #0]\n" "add.w %[pB], %[pB], %[ch_im_in]\n" "ldr.w r5, [%[pB], #0]\n" "add.w %[pB], %[pB], %[ch_im_in]\n" "pkhbt r3, r5, r2, LSL #16\n" "pkhtb r2, r2, r5, ASR #16\n" "ldr.w r0, [%[pA], #0]\n" "add.w %[pA], %[pA], %[ch_im_in]\n" "ldr.w r5, [%[pA], #0]\n" "add.w %[pA], %[pA], %[ch_im_in]\n" "pkhbt r1, r5, r0, LSL #16\n" "pkhtb r0, r0, r5, ASR #16\n" "sxtb16 r4, r0\n" "sxtb16 r5, r2\n" "smlad %[sum2], r4, r5, %[sum2]\n" "mov.w r4, r0, ror #8\n" "mov.w r5, r2, ror #8\n" "sxtb16 r4, r4\n" "sxtb16 r5, r5\n" "smlad %[sum], r4, r5, %[sum]\n" "sxtb16 r4, r1\n" "sxtb16 r5, r3\n" "smlad %[sum4], r4, r5, %[sum4]\n" "mov.w r4, r1, ror #8\n" "mov.w r5, r3, ror #8\n" "sxtb16 r4, r4\n" "sxtb16 r5, r5\n" "smlad %[sum3], r4, r5, %[sum3]\n" "subs %[colCnt], #1\n" "bne COL_LOOP\n":[sum] "+r"(sum),[sum2] "+r"(sum2),[sum3] "+r"(sum3), [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt), [ch_im_in] "r"(ch_im_in):"r0", "r1", "r2", "r3", "r4", "r5"); #endif /*ARM_MATH_BIG_ENDIAN */ #endif /* USE_INTRINSIC */ colCnt = (dim_kernel_x * dim_kernel_y) & 0x1; while (colCnt) { union arm_nnword inA, inB; inA.word = *__SIMD32(pA); pA += ch_im_in; inB.word = *__SIMD32(pB); pB += ch_im_in; sum += inA.bytes[0] * inB.bytes[0]; sum2 += inA.bytes[1] * inB.bytes[1]; sum3 += inA.bytes[2] * inB.bytes[2]; sum4 += inA.bytes[3] * inB.bytes[3]; colCnt--; } *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); *pOut++ = (q7_t) __SSAT((sum2 >> out_shift), 8); *pOut++ = (q7_t) __SSAT((sum3 >> out_shift), 8); *pOut++ = (q7_t) __SSAT((sum4 >> out_shift), 8); rowCnt--; } rowCnt = ch_im_out & 0x3; while (rowCnt) { q7_t *pB = colBuffer + row_shift; const q7_t *pA = wt + row_shift; q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); uint16_t colCnt = (dim_kernel_x * dim_kernel_y); row_shift += 1; while (colCnt) { q7_t A1 = *pA; q7_t B1 = *pB; pA += ch_im_in; pB += ch_im_in; sum += A1 * B1; colCnt--; } *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); rowCnt--; } // clear counter and pointers pBuffer = colBuffer; } } #else /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ int i_out_y, i_out_x, i_ch_out; int i_ker_y, i_ker_x; /* do some checking here, basically ch_im_in == ch_im_out */ if (ch_im_in != ch_im_out) { return ARM_MATH_SIZE_MISMATCH; } for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++) { for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++) { for (i_ch_out = 0; i_ch_out < ch_im_out; i_ch_out++) { // for each output int conv_out = ((q31_t)(bias[i_ch_out]) << bias_shift) + NN_ROUND(out_shift); for (i_ker_y = 0; i_ker_y < dim_kernel_y; i_ker_y++) { for (i_ker_x = 0; i_ker_x < dim_kernel_x; i_ker_x++) { int in_row = stride_y * i_out_y + i_ker_y - padding_y; int in_col = stride_x * i_out_x + i_ker_x - padding_x; if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x) { conv_out += Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + i_ch_out] * wt[(i_ker_y * dim_kernel_x + i_ker_x) * ch_im_out + i_ch_out]; } } } Im_out[(i_out_y * dim_im_out_x + i_out_x) * ch_im_out + i_ch_out] = (q7_t) __SSAT((conv_out >> out_shift), 8); } } } #endif /* ARM_MATH_DSP */ /* Return to application */ return ARM_MATH_SUCCESS; } /** * @} end of NNConv group */