| // Auto-generated file. Do not edit! |
| // Template: src/qs8-gemm/MRxNRc4-neondot.c.in |
| // Generator: tools/xngen |
| // |
| // Copyright 2020 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 <assert.h> |
| |
| #include <arm_neon.h> |
| |
| #include <xnnpack/gemm.h> |
| |
| |
| void xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot( |
| size_t mr, |
| size_t nc, |
| size_t kc, |
| const int8_t* restrict a, |
| size_t a_stride, |
| const void* restrict w, |
| int8_t* restrict c, |
| size_t cm_stride, |
| size_t cn_stride, |
| const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN { |
| assert(mr != 0); |
| assert(mr <= 8); |
| assert(nc != 0); |
| assert(kc != 0); |
| |
| const int8_t* a0 = a; |
| int8_t* c0 = c; |
| const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); |
| int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); |
| if XNN_UNPREDICTABLE(mr < 2) { |
| a1 = a0; |
| c1 = c0; |
| } |
| const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); |
| int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride); |
| if XNN_UNPREDICTABLE(mr <= 2) { |
| a2 = a1; |
| c2 = c1; |
| } |
| const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride); |
| int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride); |
| if XNN_UNPREDICTABLE(mr < 4) { |
| a3 = a2; |
| c3 = c2; |
| } |
| const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride); |
| int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride); |
| if XNN_UNPREDICTABLE(mr <= 4) { |
| a4 = a3; |
| c4 = c3; |
| } |
| const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride); |
| int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride); |
| if XNN_UNPREDICTABLE(mr < 6) { |
| a5 = a4; |
| c5 = c4; |
| } |
| const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride); |
| int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride); |
| if XNN_UNPREDICTABLE(mr <= 6) { |
| a6 = a5; |
| c6 = c5; |
| } |
| const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride); |
| int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride); |
| if XNN_UNPREDICTABLE(mr != 8) { |
| a7 = a6; |
| c7 = c6; |
| } |
| |
| // Loop over groups of 8 columns. |
| do { |
| // Initialize accumulators with bias. 8 bias values are loaded from the |
| // weight matrix, at the start of the group of 8 columns. |
| int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| int32x4_t vacc1x0123 = vacc0x0123; |
| int32x4_t vacc1x4567 = vacc0x4567; |
| int32x4_t vacc2x0123 = vacc0x0123; |
| int32x4_t vacc2x4567 = vacc0x4567; |
| int32x4_t vacc3x0123 = vacc0x0123; |
| int32x4_t vacc3x4567 = vacc0x4567; |
| int32x4_t vacc4x0123 = vacc0x0123; |
| int32x4_t vacc4x4567 = vacc0x4567; |
| int32x4_t vacc5x0123 = vacc0x0123; |
| int32x4_t vacc5x4567 = vacc0x4567; |
| int32x4_t vacc6x0123 = vacc0x0123; |
| int32x4_t vacc6x4567 = vacc0x4567; |
| int32x4_t vacc7x0123 = vacc0x0123; |
| int32x4_t vacc7x4567 = vacc0x4567; |
| |
| // Inner accumulation loop along the 8 columns. |
| size_t k = kc; |
| // 2x partial unrolled loop to load 8 bytes at a time. |
| while (k >= 8 * sizeof(int8_t)) { |
| // Load a 8x8 block of activations. |
| const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8; |
| const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8; |
| const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8; |
| const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8; |
| const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8; |
| const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8; |
| const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 8; |
| const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 8; |
| |
| // Load a 8x8 block of weights. |
| const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| |
| // Multiply-accumulate: 8x8 * 8x8 --> 8x8. |
| vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); |
| vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); |
| vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0); |
| vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0); |
| vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0); |
| vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0); |
| vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0); |
| vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0); |
| vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0); |
| vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0); |
| vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0); |
| vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0); |
| vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0); |
| vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0); |
| vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0); |
| vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0); |
| vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); |
| vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); |
| vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); |
| vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1); |
| vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); |
| vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1); |
| vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); |
| vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1); |
| vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); |
| vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1); |
| vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); |
| vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1); |
| vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1); |
| vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb4567x4567, va6x01234567, 1); |
| vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1); |
| vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb4567x4567, va7x01234567, 1); |
| |
| k -= 8 * sizeof(int8_t); |
| } |
| // Handle up to 7 final positions of `k` |
| if XNN_UNLIKELY(k != 0) { |
| // Load a 8x4 block of activations. |
| const int8x8_t va0x01234567 = vld1_s8(a0); a0 += k; |
| const int8x8_t va1x01234567 = vld1_s8(a1); a1 += k; |
| const int8x8_t va2x01234567 = vld1_s8(a2); a2 += k; |
| const int8x8_t va3x01234567 = vld1_s8(a3); a3 += k; |
| const int8x8_t va4x01234567 = vld1_s8(a4); a4 += k; |
| const int8x8_t va5x01234567 = vld1_s8(a5); a5 += k; |
| const int8x8_t va6x01234567 = vld1_s8(a6); a6 += k; |
| const int8x8_t va7x01234567 = vld1_s8(a7); a7 += k; |
| |
| // Load a 4x8 block of weights. |
| const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| |
| // Multiply-accumulate: 8x4 * 4x8 --> 8x8. |
| vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0); |
| vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0); |
| vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0); |
| vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0); |
| vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0); |
| vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0); |
| vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0); |
| vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0); |
| vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0); |
| vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0); |
| vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0); |
| vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0); |
| vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0); |
| vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0); |
| vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0); |
| vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0); |
| |
| if (k > 4) { |
| // Load a 4x8 block of weights. |
| const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); |
| |
| // Multiply-accumulate: 8x4 * 4x8 --> 8x8. |
| vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); |
| vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1); |
| vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); |
| vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1); |
| vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); |
| vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1); |
| vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); |
| vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1); |
| vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); |
| vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1); |
| vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); |
| vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1); |
| vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1); |
| vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb4567x4567, va6x01234567, 1); |
| vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1); |
| vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb4567x4567, va7x01234567, 1); |
| } |
| } |
| // End of accumulation loop. The variable `kc` contains the amount by which |
| // we advanced the `va` pointers, so we rewind by this amount now. |
| a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| a2 = (const int8_t*) ((uintptr_t) a2 - kc); |
| a3 = (const int8_t*) ((uintptr_t) a3 - kc); |
| a4 = (const int8_t*) ((uintptr_t) a4 - kc); |
| a5 = (const int8_t*) ((uintptr_t) a5 - kc); |
| a6 = (const int8_t*) ((uintptr_t) a6 - kc); |
| a7 = (const int8_t*) ((uintptr_t) a7 - kc); |
| |
| // Post-accumulation work |
| |
| const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift); |
| const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0))); |
| |
| const int32x4_t vproduct0x0123 = vqrdmulhq_n_s32(vacc0x0123, params->neon.multiplier); |
| const int32x4_t vproduct0x4567 = vqrdmulhq_n_s32(vacc0x4567, params->neon.multiplier); |
| const int32x4_t vproduct1x0123 = vqrdmulhq_n_s32(vacc1x0123, params->neon.multiplier); |
| const int32x4_t vproduct1x4567 = vqrdmulhq_n_s32(vacc1x4567, params->neon.multiplier); |
| const int32x4_t vproduct2x0123 = vqrdmulhq_n_s32(vacc2x0123, params->neon.multiplier); |
| const int32x4_t vproduct2x4567 = vqrdmulhq_n_s32(vacc2x4567, params->neon.multiplier); |
| const int32x4_t vproduct3x0123 = vqrdmulhq_n_s32(vacc3x0123, params->neon.multiplier); |
| const int32x4_t vproduct3x4567 = vqrdmulhq_n_s32(vacc3x4567, params->neon.multiplier); |
| const int32x4_t vproduct4x0123 = vqrdmulhq_n_s32(vacc4x0123, params->neon.multiplier); |
| const int32x4_t vproduct4x4567 = vqrdmulhq_n_s32(vacc4x4567, params->neon.multiplier); |
| const int32x4_t vproduct5x0123 = vqrdmulhq_n_s32(vacc5x0123, params->neon.multiplier); |
| const int32x4_t vproduct5x4567 = vqrdmulhq_n_s32(vacc5x4567, params->neon.multiplier); |
| const int32x4_t vproduct6x0123 = vqrdmulhq_n_s32(vacc6x0123, params->neon.multiplier); |
| const int32x4_t vproduct6x4567 = vqrdmulhq_n_s32(vacc6x4567, params->neon.multiplier); |
| const int32x4_t vproduct7x0123 = vqrdmulhq_n_s32(vacc7x0123, params->neon.multiplier); |
| const int32x4_t vproduct7x4567 = vqrdmulhq_n_s32(vacc7x4567, params->neon.multiplier); |
| |
| vacc0x0123 = vsraq_n_s32(vproduct0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31); |
| vacc0x4567 = vsraq_n_s32(vproduct0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31); |
| vacc1x0123 = vsraq_n_s32(vproduct1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31); |
| vacc1x4567 = vsraq_n_s32(vproduct1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31); |
| vacc2x0123 = vsraq_n_s32(vproduct2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31); |
| vacc2x4567 = vsraq_n_s32(vproduct2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31); |
| vacc3x0123 = vsraq_n_s32(vproduct3x0123, vbicq_s32(vacc3x0123, vzero_shift_mask), 31); |
| vacc3x4567 = vsraq_n_s32(vproduct3x4567, vbicq_s32(vacc3x4567, vzero_shift_mask), 31); |
| vacc4x0123 = vsraq_n_s32(vproduct4x0123, vbicq_s32(vacc4x0123, vzero_shift_mask), 31); |
| vacc4x4567 = vsraq_n_s32(vproduct4x4567, vbicq_s32(vacc4x4567, vzero_shift_mask), 31); |
| vacc5x0123 = vsraq_n_s32(vproduct5x0123, vbicq_s32(vacc5x0123, vzero_shift_mask), 31); |
| vacc5x4567 = vsraq_n_s32(vproduct5x4567, vbicq_s32(vacc5x4567, vzero_shift_mask), 31); |
| vacc6x0123 = vsraq_n_s32(vproduct6x0123, vbicq_s32(vacc6x0123, vzero_shift_mask), 31); |
| vacc6x4567 = vsraq_n_s32(vproduct6x4567, vbicq_s32(vacc6x4567, vzero_shift_mask), 31); |
| vacc7x0123 = vsraq_n_s32(vproduct7x0123, vbicq_s32(vacc7x0123, vzero_shift_mask), 31); |
| vacc7x4567 = vsraq_n_s32(vproduct7x4567, vbicq_s32(vacc7x4567, vzero_shift_mask), 31); |
| |
| vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift); |
| vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift); |
| vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift); |
| vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift); |
| vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift); |
| vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift); |
| vacc3x0123 = vrshlq_s32(vacc3x0123, vright_shift); |
| vacc3x4567 = vrshlq_s32(vacc3x4567, vright_shift); |
| vacc4x0123 = vrshlq_s32(vacc4x0123, vright_shift); |
| vacc4x4567 = vrshlq_s32(vacc4x4567, vright_shift); |
| vacc5x0123 = vrshlq_s32(vacc5x0123, vright_shift); |
| vacc5x4567 = vrshlq_s32(vacc5x4567, vright_shift); |
| vacc6x0123 = vrshlq_s32(vacc6x0123, vright_shift); |
| vacc6x4567 = vrshlq_s32(vacc6x4567, vright_shift); |
| vacc7x0123 = vrshlq_s32(vacc7x0123, vright_shift); |
| vacc7x4567 = vrshlq_s32(vacc7x4567, vright_shift); |
| |
| const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point); |
| #if XNN_ARCH_ARM64 |
| const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point); |
| const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point); |
| const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point); |
| const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point); |
| const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point); |
| const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point); |
| const int16x8_t vacc6x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x0123), vacc6x4567), voutput_zero_point); |
| const int16x8_t vacc7x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x0123), vacc7x4567), voutput_zero_point); |
| |
| int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567); |
| int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567); |
| int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567); |
| int8x16_t vout6x01234567_7x01234567 = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc7x01234567); |
| #else |
| const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point); |
| const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point); |
| const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point); |
| const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point); |
| const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point); |
| const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point); |
| const int16x8_t vacc6x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x0123), vqmovn_s32(vacc6x4567)), voutput_zero_point); |
| const int16x8_t vacc7x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x0123), vqmovn_s32(vacc7x4567)), voutput_zero_point); |
| |
| int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567)); |
| int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567)); |
| int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x01234567)); |
| int8x16_t vout6x01234567_7x01234567 = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc7x01234567)); |
| #endif |
| const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min); |
| const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max); |
| |
| vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min); |
| vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min); |
| vout4x01234567_5x01234567 = vmaxq_s8(vout4x01234567_5x01234567, voutput_min); |
| vout6x01234567_7x01234567 = vmaxq_s8(vout6x01234567_7x01234567, voutput_min); |
| |
| vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max); |
| vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max); |
| vout4x01234567_5x01234567 = vminq_s8(vout4x01234567_5x01234567, voutput_max); |
| vout6x01234567_7x01234567 = vminq_s8(vout6x01234567_7x01234567, voutput_max); |
| |
| if (nc >= 8) { |
| // Main case where there the 8 columns fit in the destination. |
| vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567)); |
| vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567)); |
| vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567)); |
| vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567)); |
| vst1_s8(c4 + 0, vget_low_s8(vout4x01234567_5x01234567)); |
| vst1_s8(c5 + 0, vget_high_s8(vout4x01234567_5x01234567)); |
| vst1_s8(c6 + 0, vget_low_s8(vout6x01234567_7x01234567)); |
| vst1_s8(c7 + 0, vget_high_s8(vout6x01234567_7x01234567)); |
| |
| // Advance to the next 8 columns. |
| c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); |
| c2 = (int8_t*) ((uintptr_t) c2 + cn_stride); |
| c3 = (int8_t*) ((uintptr_t) c3 + cn_stride); |
| c4 = (int8_t*) ((uintptr_t) c4 + cn_stride); |
| c5 = (int8_t*) ((uintptr_t) c5 + cn_stride); |
| c6 = (int8_t*) ((uintptr_t) c6 + cn_stride); |
| c7 = (int8_t*) ((uintptr_t) c7 + cn_stride); |
| |
| nc -= 8; |
| } else { |
| // Final case where not all of the 8 columns fit in the destination. |
| if (nc & 4) { |
| vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4; |
| vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4; |
| vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4; |
| vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4; |
| vst1q_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4; |
| vst1q_lane_u32(__builtin_assume_aligned(c5, 1), vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4; |
| vst1q_lane_u32(__builtin_assume_aligned(c6, 1), vreinterpretq_u32_s8(vout6x01234567_7x01234567), 0); c6 += 4; |
| vst1q_lane_u32(__builtin_assume_aligned(c7, 1), vreinterpretq_u32_s8(vout6x01234567_7x01234567), 2); c7 += 4; |
| vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4); |
| vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4); |
| vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4); |
| vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 4); |
| } |
| if (nc & 2) { |
| vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2; |
| vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2; |
| vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2; |
| vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2; |
| vst1q_lane_u16(__builtin_assume_aligned(c4, 1), vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2; |
| vst1q_lane_u16(__builtin_assume_aligned(c5, 1), vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2; |
| vst1q_lane_u16(__builtin_assume_aligned(c6, 1), vreinterpretq_u16_s8(vout6x01234567_7x01234567), 0); c6 += 2; |
| vst1q_lane_u16(__builtin_assume_aligned(c7, 1), vreinterpretq_u16_s8(vout6x01234567_7x01234567), 4); c7 += 2; |
| vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2); |
| vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2); |
| vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2); |
| vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 2); |
| } |
| if (nc & 1) { |
| vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0); |
| vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8); |
| vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0); |
| vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8); |
| vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0); |
| vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8); |
| vst1q_lane_s8(c6, vout6x01234567_7x01234567, 0); |
| vst1q_lane_s8(c7, vout6x01234567_7x01234567, 8); |
| } |
| |
| nc = 0; |
| } |
| } while (nc != 0); |
| } |
