// Auto-generated file. Do not edit! // Template: src/qs8-gemm/c8-neoni8mm.c.in // Generator: tools/xngen // // Copyright 2023 Google LLC // // This source code is licensed under the BSD-style license found in the // LICENSE file in the root directory of this source tree. #include #include #include "xnnpack/gemm.h" #include "xnnpack/math.h" void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_2x32c8__neoni8mm( size_t mr, size_t nc, size_t kc, const int8_t* restrict a, size_t a_stride, const void* restrict w, float* restrict c, size_t cm_stride, size_t cn_stride, const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 2); assert(nc != 0); assert(kc != 0); assert(kc % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8 * sizeof(int8_t)); const int8_t* a0 = a; float* c0 = c; const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); float* c1 = (float*) ((uintptr_t) c0 + cm_stride); if XNN_UNPREDICTABLE(mr != 2) { a1 = a0; c1 = c0; } // Loop over groups of 32 columns. do { // Initialize accumulators with bias. 32 bias values are loaded from the // weight matrix, at the start of the group of 32 columns. const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point); const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0); const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0); const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0); const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0); const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point01), 0); const int32x4_t vksumzp1x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point01), 0); const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point01), 0); const int32x4_t vksumzp1xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point01), 0); const int32x4_t vksumGHIJ = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0xGHIJ = vmulq_lane_s32(vksumGHIJ, vget_low_s32(vinput_zero_point01), 0); const int32x4_t vksumzp1xGHIJ = vmulq_lane_s32(vksumGHIJ, vget_high_s32(vinput_zero_point01), 0); const int32x4_t vksumKLMN = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0xKLMN = vmulq_lane_s32(vksumKLMN, vget_low_s32(vinput_zero_point01), 0); const int32x4_t vksumzp1xKLMN = vmulq_lane_s32(vksumKLMN, vget_high_s32(vinput_zero_point01), 0); const int32x4_t vksumOPQR = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0xOPQR = vmulq_lane_s32(vksumOPQR, vget_low_s32(vinput_zero_point01), 0); const int32x4_t vksumzp1xOPQR = vmulq_lane_s32(vksumOPQR, vget_high_s32(vinput_zero_point01), 0); const int32x4_t vksumSTUV = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0xSTUV = vmulq_lane_s32(vksumSTUV, vget_low_s32(vinput_zero_point01), 0); const int32x4_t vksumzp1xSTUV = vmulq_lane_s32(vksumSTUV, vget_high_s32(vinput_zero_point01), 0); int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123))); int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123))); int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567))); int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567))); int32x4_t vacc01x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB))); int32x4_t vacc01xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB))); int32x4_t vacc01xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF))); int32x4_t vacc01xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF))); int32x4_t vacc01xGH = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xGHIJ), vreinterpretq_u64_s32(vksumzp1xGHIJ))); int32x4_t vacc01xIJ = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xGHIJ), vreinterpretq_u64_s32(vksumzp1xGHIJ))); int32x4_t vacc01xKL = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xKLMN), vreinterpretq_u64_s32(vksumzp1xKLMN))); int32x4_t vacc01xMN = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xKLMN), vreinterpretq_u64_s32(vksumzp1xKLMN))); int32x4_t vacc01xOP = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xOPQR), vreinterpretq_u64_s32(vksumzp1xOPQR))); int32x4_t vacc01xQR = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xOPQR), vreinterpretq_u64_s32(vksumzp1xOPQR))); int32x4_t vacc01xST = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xSTUV), vreinterpretq_u64_s32(vksumzp1xSTUV))); int32x4_t vacc01xUV = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xSTUV), vreinterpretq_u64_s32(vksumzp1xSTUV))); // Inner accumulation loop along the 32 columns. size_t k = kc; // 2x partial unrolled loop to load 8 bytes at a time. uint64x2x2_t va01x0123456789ABCDEF; va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0); va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0); while (k >= 16 * sizeof(int8_t)) { // Load a 2x16 block of activations. va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16; va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16; const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]); const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]); // Load a 16x32 block of weights. const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbGHx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbIJx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbKLx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbMNx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbOPx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbQRx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbSTx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbUVx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb89x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbABx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbCDx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbEFx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbGHx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbIJx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbKLx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbMNx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbOPx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbQRx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbSTx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbUVx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; // Multiply-accumulate: 2x8 * 8x32 --> 2x32. vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567); vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567); vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567); vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567); vacc01x89 = vmmlaq_s32(vacc01x89, va01x01234567, vb89x01234567); vacc01xAB = vmmlaq_s32(vacc01xAB, va01x01234567, vbABx01234567); vacc01xCD = vmmlaq_s32(vacc01xCD, va01x01234567, vbCDx01234567); vacc01xEF = vmmlaq_s32(vacc01xEF, va01x01234567, vbEFx01234567); vacc01xGH = vmmlaq_s32(vacc01xGH, va01x01234567, vbGHx01234567); vacc01xIJ = vmmlaq_s32(vacc01xIJ, va01x01234567, vbIJx01234567); vacc01xKL = vmmlaq_s32(vacc01xKL, va01x01234567, vbKLx01234567); vacc01xMN = vmmlaq_s32(vacc01xMN, va01x01234567, vbMNx01234567); vacc01xOP = vmmlaq_s32(vacc01xOP, va01x01234567, vbOPx01234567); vacc01xQR = vmmlaq_s32(vacc01xQR, va01x01234567, vbQRx01234567); vacc01xST = vmmlaq_s32(vacc01xST, va01x01234567, vbSTx01234567); vacc01xUV = vmmlaq_s32(vacc01xUV, va01x01234567, vbUVx01234567); vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF); vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF); vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF); vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF); vacc01x89 = vmmlaq_s32(vacc01x89, va01x89ABCDEF, vb89x89ABCDEF); vacc01xAB = vmmlaq_s32(vacc01xAB, va01x89ABCDEF, vbABx89ABCDEF); vacc01xCD = vmmlaq_s32(vacc01xCD, va01x89ABCDEF, vbCDx89ABCDEF); vacc01xEF = vmmlaq_s32(vacc01xEF, va01x89ABCDEF, vbEFx89ABCDEF); vacc01xGH = vmmlaq_s32(vacc01xGH, va01x89ABCDEF, vbGHx89ABCDEF); vacc01xIJ = vmmlaq_s32(vacc01xIJ, va01x89ABCDEF, vbIJx89ABCDEF); vacc01xKL = vmmlaq_s32(vacc01xKL, va01x89ABCDEF, vbKLx89ABCDEF); vacc01xMN = vmmlaq_s32(vacc01xMN, va01x89ABCDEF, vbMNx89ABCDEF); vacc01xOP = vmmlaq_s32(vacc01xOP, va01x89ABCDEF, vbOPx89ABCDEF); vacc01xQR = vmmlaq_s32(vacc01xQR, va01x89ABCDEF, vbQRx89ABCDEF); vacc01xST = vmmlaq_s32(vacc01xST, va01x89ABCDEF, vbSTx89ABCDEF); vacc01xUV = vmmlaq_s32(vacc01xUV, va01x89ABCDEF, vbUVx89ABCDEF); k -= 16 * sizeof(int8_t); } // Handle up to 8 final positions of `k` if XNN_UNLIKELY(k != 0) { // Load a 2x8 block of activations. uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8; va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8; // Load a 16x32 block of weights. const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbGHx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbIJx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbKLx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbMNx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbOPx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbQRx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbSTx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbUVx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; // Multiply-accumulate: 2x4 * 4x32 --> 2x32. vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567); vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567); vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567); vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567); vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567); vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567); vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567); vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567); vacc01xGH = vmmlaq_s32(vacc01xGH, vreinterpretq_s8_u64(va01x01234567), vbGHx01234567); vacc01xIJ = vmmlaq_s32(vacc01xIJ, vreinterpretq_s8_u64(va01x01234567), vbIJx01234567); vacc01xKL = vmmlaq_s32(vacc01xKL, vreinterpretq_s8_u64(va01x01234567), vbKLx01234567); vacc01xMN = vmmlaq_s32(vacc01xMN, vreinterpretq_s8_u64(va01x01234567), vbMNx01234567); vacc01xOP = vmmlaq_s32(vacc01xOP, vreinterpretq_s8_u64(va01x01234567), vbOPx01234567); vacc01xQR = vmmlaq_s32(vacc01xQR, vreinterpretq_s8_u64(va01x01234567), vbQRx01234567); vacc01xST = vmmlaq_s32(vacc01xST, vreinterpretq_s8_u64(va01x01234567), vbSTx01234567); vacc01xUV = vmmlaq_s32(vacc01xUV, vreinterpretq_s8_u64(va01x01234567), vbUVx01234567); } int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23))); int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23))); int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67))); int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67))); int32x4_t vacc0x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB))); int32x4_t vacc1x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB))); int32x4_t vacc0xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF))); int32x4_t vacc1xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF))); int32x4_t vacc0xGHIJ = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xGH), vreinterpretq_u64_s32(vacc01xIJ))); int32x4_t vacc1xGHIJ = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xGH), vreinterpretq_u64_s32(vacc01xIJ))); int32x4_t vacc0xKLMN = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xKL), vreinterpretq_u64_s32(vacc01xMN))); int32x4_t vacc1xKLMN = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xKL), vreinterpretq_u64_s32(vacc01xMN))); int32x4_t vacc0xOPQR = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xOP), vreinterpretq_u64_s32(vacc01xQR))); int32x4_t vacc1xOPQR = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xOP), vreinterpretq_u64_s32(vacc01xQR))); int32x4_t vacc0xSTUV = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xST), vreinterpretq_u64_s32(vacc01xUV))); int32x4_t vacc1xSTUV = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xST), vreinterpretq_u64_s32(vacc01xUV))); float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123); float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567); float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB); float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF); float32x4_t vout0xGHIJ = vcvtq_f32_s32(vacc0xGHIJ); float32x4_t vout0xKLMN = vcvtq_f32_s32(vacc0xKLMN); float32x4_t vout0xOPQR = vcvtq_f32_s32(vacc0xOPQR); float32x4_t vout0xSTUV = vcvtq_f32_s32(vacc0xSTUV); float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123); float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567); float32x4_t vout1x89AB = vcvtq_f32_s32(vacc1x89AB); float32x4_t vout1xCDEF = vcvtq_f32_s32(vacc1xCDEF); float32x4_t vout1xGHIJ = vcvtq_f32_s32(vacc1xGHIJ); float32x4_t vout1xKLMN = vcvtq_f32_s32(vacc1xKLMN); float32x4_t vout1xOPQR = vcvtq_f32_s32(vacc1xOPQR); float32x4_t vout1xSTUV = vcvtq_f32_s32(vacc1xSTUV); const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point)); vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1); vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1); vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1); vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1); vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1); vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1); vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1); vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1); vout0xGHIJ = vmulq_lane_f32(vout0xGHIJ, vget_low_f32(vinput_scale01), 1); vout1xGHIJ = vmulq_lane_f32(vout1xGHIJ, vget_high_f32(vinput_scale01), 1); vout0xKLMN = vmulq_lane_f32(vout0xKLMN, vget_low_f32(vinput_scale01), 1); vout1xKLMN = vmulq_lane_f32(vout1xKLMN, vget_high_f32(vinput_scale01), 1); vout0xOPQR = vmulq_lane_f32(vout0xOPQR, vget_low_f32(vinput_scale01), 1); vout1xOPQR = vmulq_lane_f32(vout1xOPQR, vget_high_f32(vinput_scale01), 1); vout0xSTUV = vmulq_lane_f32(vout0xSTUV, vget_low_f32(vinput_scale01), 1); vout1xSTUV = vmulq_lane_f32(vout1xSTUV, vget_high_f32(vinput_scale01), 1); const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4; const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4; const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4; const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4; const float32x4_t vfilter_output_scaleGHIJ = vld1q_f32(w); w = (const float*) w + 4; const float32x4_t vfilter_output_scaleKLMN = vld1q_f32(w); w = (const float*) w + 4; const float32x4_t vfilter_output_scaleOPQR = vld1q_f32(w); w = (const float*) w + 4; const float32x4_t vfilter_output_scaleSTUV = vld1q_f32(w); w = (const float*) w + 4; const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4; vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123); vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123); const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4; vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567); vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567); const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4; vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB); vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB); const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4; vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF); vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF); const float32x4_t vbiasGHIJ = vld1q_f32(w); w = (const float*) w + 4; vout0xGHIJ = vfmaq_f32(vbiasGHIJ, vout0xGHIJ, vfilter_output_scaleGHIJ); vout1xGHIJ = vfmaq_f32(vbiasGHIJ, vout1xGHIJ, vfilter_output_scaleGHIJ); const float32x4_t vbiasKLMN = vld1q_f32(w); w = (const float*) w + 4; vout0xKLMN = vfmaq_f32(vbiasKLMN, vout0xKLMN, vfilter_output_scaleKLMN); vout1xKLMN = vfmaq_f32(vbiasKLMN, vout1xKLMN, vfilter_output_scaleKLMN); const float32x4_t vbiasOPQR = vld1q_f32(w); w = (const float*) w + 4; vout0xOPQR = vfmaq_f32(vbiasOPQR, vout0xOPQR, vfilter_output_scaleOPQR); vout1xOPQR = vfmaq_f32(vbiasOPQR, vout1xOPQR, vfilter_output_scaleOPQR); const float32x4_t vbiasSTUV = vld1q_f32(w); w = (const float*) w + 4; vout0xSTUV = vfmaq_f32(vbiasSTUV, vout0xSTUV, vfilter_output_scaleSTUV); vout1xSTUV = vfmaq_f32(vbiasSTUV, vout1xSTUV, vfilter_output_scaleSTUV); const float32x4_t voutput_min = vld1q_dup_f32(¶ms->scalar.min); vout0x0123 = vmaxq_f32(vout0x0123, voutput_min); vout0x4567 = vmaxq_f32(vout0x4567, voutput_min); vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min); vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min); vout0xGHIJ = vmaxq_f32(vout0xGHIJ, voutput_min); vout0xKLMN = vmaxq_f32(vout0xKLMN, voutput_min); vout0xOPQR = vmaxq_f32(vout0xOPQR, voutput_min); vout0xSTUV = vmaxq_f32(vout0xSTUV, voutput_min); vout1x0123 = vmaxq_f32(vout1x0123, voutput_min); vout1x4567 = vmaxq_f32(vout1x4567, voutput_min); vout1x89AB = vmaxq_f32(vout1x89AB, voutput_min); vout1xCDEF = vmaxq_f32(vout1xCDEF, voutput_min); vout1xGHIJ = vmaxq_f32(vout1xGHIJ, voutput_min); vout1xKLMN = vmaxq_f32(vout1xKLMN, voutput_min); vout1xOPQR = vmaxq_f32(vout1xOPQR, voutput_min); vout1xSTUV = vmaxq_f32(vout1xSTUV, voutput_min); const float32x4_t voutput_max = vld1q_dup_f32(¶ms->scalar.max); vout0x0123 = vminq_f32(vout0x0123, voutput_max); vout0x4567 = vminq_f32(vout0x4567, voutput_max); vout0x89AB = vminq_f32(vout0x89AB, voutput_max); vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max); vout0xGHIJ = vminq_f32(vout0xGHIJ, voutput_max); vout0xKLMN = vminq_f32(vout0xKLMN, voutput_max); vout0xOPQR = vminq_f32(vout0xOPQR, voutput_max); vout0xSTUV = vminq_f32(vout0xSTUV, voutput_max); vout1x0123 = vminq_f32(vout1x0123, voutput_max); vout1x4567 = vminq_f32(vout1x4567, voutput_max); vout1x89AB = vminq_f32(vout1x89AB, voutput_max); vout1xCDEF = vminq_f32(vout1xCDEF, voutput_max); vout1xGHIJ = vminq_f32(vout1xGHIJ, voutput_max); vout1xKLMN = vminq_f32(vout1xKLMN, voutput_max); vout1xOPQR = vminq_f32(vout1xOPQR, voutput_max); vout1xSTUV = vminq_f32(vout1xSTUV, voutput_max); if XNN_LIKELY(nc >= 32) { vst1q_f32(c0, vout0x0123); vst1q_f32(c0 + 4, vout0x4567); vst1q_f32(c0 + 8, vout0x89AB); vst1q_f32(c0 + 12, vout0xCDEF); vst1q_f32(c0 + 16, vout0xGHIJ); vst1q_f32(c0 + 20, vout0xKLMN); vst1q_f32(c0 + 24, vout0xOPQR); vst1q_f32(c0 + 28, vout0xSTUV); vst1q_f32(c1, vout1x0123); vst1q_f32(c1 + 4, vout1x4567); vst1q_f32(c1 + 8, vout1x89AB); vst1q_f32(c1 + 12, vout1xCDEF); vst1q_f32(c1 + 16, vout1xGHIJ); vst1q_f32(c1 + 20, vout1xKLMN); vst1q_f32(c1 + 24, vout1xOPQR); vst1q_f32(c1 + 28, vout1xSTUV); a0 = (const int8_t*) ((uintptr_t) a0 - kc); a1 = (const int8_t*) ((uintptr_t) a1 - kc); c0 = (float*) ((uintptr_t) c0 + cn_stride); c1 = (float*) ((uintptr_t) c1 + cn_stride); nc -= 32; } else { if (nc & 16) { vst1q_f32(c0, vout0x0123); c0 += 4; vout0x0123 = vout0xGHIJ; vst1q_f32(c1, vout1x0123); c1 += 4; vout1x0123 = vout1xGHIJ; vst1q_f32(c0, vout0x4567); c0 += 4; vout0x4567 = vout0xKLMN; vst1q_f32(c1, vout1x4567); c1 += 4; vout1x4567 = vout1xKLMN; vst1q_f32(c0, vout0x89AB); c0 += 4; vout0x89AB = vout0xOPQR; vst1q_f32(c1, vout1x89AB); c1 += 4; vout1x89AB = vout1xOPQR; vst1q_f32(c0, vout0xCDEF); c0 += 4; vout0xCDEF = vout0xSTUV; vst1q_f32(c1, vout1xCDEF); c1 += 4; vout1xCDEF = vout1xSTUV; } if (nc & 8) { vst1q_f32(c0, vout0x0123); c0 += 4; vout0x0123 = vout0x89AB; vst1q_f32(c1, vout1x0123); c1 += 4; vout1x0123 = vout1x89AB; vst1q_f32(c0, vout0x4567); c0 += 4; vout0x4567 = vout0xCDEF; vst1q_f32(c1, vout1x4567); c1 += 4; vout1x4567 = vout1xCDEF; } if (nc & 4) { vst1q_f32(c0, vout0x0123); c0 += 4; vout0x0123 = vout0x4567; vst1q_f32(c1, vout1x0123); c1 += 4; vout1x0123 = vout1x4567; } float32x2_t vout0x01 = vget_low_f32(vout0x0123); float32x2_t vout1x01 = vget_low_f32(vout1x0123); if (nc & 2) { vst1_f32(c0, vout0x01); c0 += 2; vst1_f32(c1, vout1x01); c1 += 2; vout0x01 = vget_high_f32(vout0x0123); vout1x01 = vget_high_f32(vout1x0123); } if (nc & 1) { vst1_lane_f32(c0, vout0x01, 0); vst1_lane_f32(c1, vout1x01, 0); } nc = 0; } } while (nc != 0); }