// Auto-generated file. Do not edit! // Template: src/qs8-igemm/neon-mlal-lane.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 #include #include "xnnpack/common.h" #include "xnnpack/igemm.h" void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_6x8__neon_mlal_lane( size_t mr, size_t nc, size_t kc, size_t ks, const int8_t** restrict a, const void* restrict w, float* restrict c, size_t cm_stride, size_t cn_stride, size_t a_offset, const int8_t* zero, const int8_t* zero_data, 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 <= 6); assert(nc != 0); assert(kc != 0); assert(ks != 0); assert(ks % (6 * sizeof(void*)) == 0); assert(a_offset % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); float* c0 = c; float* c1 = (float*) ((uintptr_t) c0 + cm_stride); if XNN_UNPREDICTABLE(mr < 2) { c1 = c0; } float* c2 = (float*) ((uintptr_t) c1 + cm_stride); if XNN_UNPREDICTABLE(mr <= 2) { c2 = c1; } float* c3 = (float*) ((uintptr_t) c2 + cm_stride); if XNN_UNPREDICTABLE(mr < 4) { c3 = c2; } float* c4 = (float*) ((uintptr_t) c3 + cm_stride); if XNN_UNPREDICTABLE(mr <= 4) { c4 = c3; } float* c5 = (float*) ((uintptr_t) c4 + cm_stride); if XNN_UNPREDICTABLE(mr != 6) { c5 = c4; } do { int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4; int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point); int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point); int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point); 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; size_t p = ks; do { const int8_t* restrict a0 = a[0]; if XNN_UNPREDICTABLE(a0 != zero) { a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); } else { a0 = zero_data; } const int8_t* restrict a1 = a[1]; if XNN_UNPREDICTABLE(a1 != zero) { a1 = (const int8_t*) ((uintptr_t) a1 + a_offset); } else { a1 = zero_data; } const int8_t* restrict a2 = a[2]; if XNN_UNPREDICTABLE(a2 != zero) { a2 = (const int8_t*) ((uintptr_t) a2 + a_offset); } else { a2 = zero_data; } const int8_t* restrict a3 = a[3]; if XNN_UNPREDICTABLE(a3 != zero) { a3 = (const int8_t*) ((uintptr_t) a3 + a_offset); } else { a3 = zero_data; } const int8_t* restrict a4 = a[4]; if XNN_UNPREDICTABLE(a4 != zero) { a4 = (const int8_t*) ((uintptr_t) a4 + a_offset); } else { a4 = zero_data; } const int8_t* restrict a5 = a[5]; if XNN_UNPREDICTABLE(a5 != zero) { a5 = (const int8_t*) ((uintptr_t) a5 + a_offset); } else { a5 = zero_data; } a += 6; size_t k = kc; while (k >= 8 * sizeof(int8_t)) { const int8x8_t va0 = vld1_s8(a0); a0 += 8; const int16x8_t vxa0 = vmovl_s8(va0); const int8x8_t va1 = vld1_s8(a1); a1 += 8; const int16x8_t vxa1 = vmovl_s8(va1); const int8x8_t va2 = vld1_s8(a2); a2 += 8; const int16x8_t vxa2 = vmovl_s8(va2); const int8x8_t va3 = vld1_s8(a3); a3 += 8; const int16x8_t vxa3 = vmovl_s8(va3); const int8x8_t va4 = vld1_s8(a4); a4 += 8; const int16x8_t vxa4 = vmovl_s8(va4); const int8x8_t va5 = vld1_s8(a5); a5 += 8; const int16x8_t vxa5 = vmovl_s8(va5); const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0); const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1); const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2); const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3); const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0); const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1); const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2); const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa4), 3); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa4), 3); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa5), 3); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa5), 3); k -= 8 * sizeof(int8_t); } if XNN_UNLIKELY(k != 0) { const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k); const int16x8_t vxa0 = vmovl_s8(va0); const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k); const int16x8_t vxa1 = vmovl_s8(va1); const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k); const int16x8_t vxa2 = vmovl_s8(va2); const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k); const int16x8_t vxa3 = vmovl_s8(va3); const int8x8_t va4 = vld1_s8(a4); a4 = (const int8_t*) ((uintptr_t) a4 + k); const int16x8_t vxa4 = vmovl_s8(va4); const int8x8_t va5 = vld1_s8(a5); a5 = (const int8_t*) ((uintptr_t) a5 + k); const int16x8_t vxa5 = vmovl_s8(va5); const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0); if (k >= 2 * sizeof(int8_t)) { const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1); if (k > 2 * sizeof(int8_t)) { const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2); if (k >= 4 * sizeof(int8_t)) { const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3); if (k > 4 * sizeof(int8_t)) { const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0); if (k >= 6 * sizeof(int8_t)) { const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1); if (k > 6 * sizeof(int8_t)) { const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8; const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6); vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2); vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2); vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2); vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2); vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2); vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2); vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2); vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2); vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2); vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2); } } } } } } } p -= 6 * sizeof(void*); } while (p != 0); float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123); float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567); float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123); float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567); float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123); float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567); float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123); float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567); float32x4_t vout4x0123 = vcvtq_f32_s32(vacc4x0123); float32x4_t vout4x4567 = vcvtq_f32_s32(vacc4x4567); float32x4_t vout5x0123 = vcvtq_f32_s32(vacc5x0123); float32x4_t vout5x4567 = vcvtq_f32_s32(vacc5x4567); const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale); vout0x0123 = vmulq_f32(vout0x0123, vinput_scale); vout0x4567 = vmulq_f32(vout0x4567, vinput_scale); vout1x0123 = vmulq_f32(vout1x0123, vinput_scale); vout1x4567 = vmulq_f32(vout1x4567, vinput_scale); vout2x0123 = vmulq_f32(vout2x0123, vinput_scale); vout2x4567 = vmulq_f32(vout2x4567, vinput_scale); vout3x0123 = vmulq_f32(vout3x0123, vinput_scale); vout3x4567 = vmulq_f32(vout3x4567, vinput_scale); vout4x0123 = vmulq_f32(vout4x0123, vinput_scale); vout4x4567 = vmulq_f32(vout4x4567, vinput_scale); vout5x0123 = vmulq_f32(vout5x0123, vinput_scale); vout5x4567 = vmulq_f32(vout5x4567, vinput_scale); 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 vbias0123 = vld1q_f32(w); w = (const float*) w + 4; #if XNN_ARCH_ARM64 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123); vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123); vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123); vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123); vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123); vout5x0123 = vfmaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123); #else vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123); vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123); vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123); vout3x0123 = vmlaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123); vout4x0123 = vmlaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123); vout5x0123 = vmlaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123); #endif const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4; #if XNN_ARCH_ARM64 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567); vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567); vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567); vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567); vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567); vout5x4567 = vfmaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567); #else vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567); vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567); vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567); vout3x4567 = vmlaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567); vout4x4567 = vmlaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567); vout5x4567 = vmlaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567); #endif const float32x4_t voutput_min = vld1q_dup_f32(¶ms->scalar.min); vout0x0123 = vmaxq_f32(vout0x0123, voutput_min); vout0x4567 = vmaxq_f32(vout0x4567, voutput_min); vout1x0123 = vmaxq_f32(vout1x0123, voutput_min); vout1x4567 = vmaxq_f32(vout1x4567, voutput_min); vout2x0123 = vmaxq_f32(vout2x0123, voutput_min); vout2x4567 = vmaxq_f32(vout2x4567, voutput_min); vout3x0123 = vmaxq_f32(vout3x0123, voutput_min); vout3x4567 = vmaxq_f32(vout3x4567, voutput_min); vout4x0123 = vmaxq_f32(vout4x0123, voutput_min); vout4x4567 = vmaxq_f32(vout4x4567, voutput_min); vout5x0123 = vmaxq_f32(vout5x0123, voutput_min); vout5x4567 = vmaxq_f32(vout5x4567, 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); vout1x0123 = vminq_f32(vout1x0123, voutput_max); vout1x4567 = vminq_f32(vout1x4567, voutput_max); vout2x0123 = vminq_f32(vout2x0123, voutput_max); vout2x4567 = vminq_f32(vout2x4567, voutput_max); vout3x0123 = vminq_f32(vout3x0123, voutput_max); vout3x4567 = vminq_f32(vout3x4567, voutput_max); vout4x0123 = vminq_f32(vout4x0123, voutput_max); vout4x4567 = vminq_f32(vout4x4567, voutput_max); vout5x0123 = vminq_f32(vout5x0123, voutput_max); vout5x4567 = vminq_f32(vout5x4567, voutput_max); if XNN_LIKELY(nc >= 8) { vst1q_f32(&c5[0], vout5x0123); vst1q_f32(&c5[4], vout5x4567); vst1q_f32(&c4[0], vout4x0123); vst1q_f32(&c4[4], vout4x4567); vst1q_f32(&c3[0], vout3x0123); vst1q_f32(&c3[4], vout3x4567); vst1q_f32(&c2[0], vout2x0123); vst1q_f32(&c2[4], vout2x4567); vst1q_f32(&c1[0], vout1x0123); vst1q_f32(&c1[4], vout1x4567); vst1q_f32(&c0[0], vout0x0123); vst1q_f32(&c0[4], vout0x4567); a = (const int8_t**restrict) ((uintptr_t) a - ks); c0 = (float*) ((uintptr_t) c0 + cn_stride); c1 = (float*) ((uintptr_t) c1 + cn_stride); c2 = (float*) ((uintptr_t) c2 + cn_stride); c3 = (float*) ((uintptr_t) c3 + cn_stride); c4 = (float*) ((uintptr_t) c4 + cn_stride); c5 = (float*) ((uintptr_t) c5 + cn_stride); nc -= 8; } else { if (nc & 4) { vst1q_f32(c5, vout5x0123); c5 += 4; vout5x0123 = vout5x4567; vst1q_f32(c4, vout4x0123); c4 += 4; vout4x0123 = vout4x4567; vst1q_f32(c3, vout3x0123); c3 += 4; vout3x0123 = vout3x4567; vst1q_f32(c2, vout2x0123); c2 += 4; vout2x0123 = vout2x4567; vst1q_f32(c1, vout1x0123); c1 += 4; vout1x0123 = vout1x4567; vst1q_f32(c0, vout0x0123); c0 += 4; vout0x0123 = vout0x4567; } float32x2_t vout5x01 = vget_low_f32(vout5x0123); float32x2_t vout4x01 = vget_low_f32(vout4x0123); float32x2_t vout3x01 = vget_low_f32(vout3x0123); float32x2_t vout2x01 = vget_low_f32(vout2x0123); float32x2_t vout1x01 = vget_low_f32(vout1x0123); float32x2_t vout0x01 = vget_low_f32(vout0x0123); if (nc & 2) { vst1_f32(c5, vout5x01); c5 += 2; vst1_f32(c4, vout4x01); c4 += 2; vst1_f32(c3, vout3x01); c3 += 2; vst1_f32(c2, vout2x01); c2 += 2; vst1_f32(c1, vout1x01); c1 += 2; vst1_f32(c0, vout0x01); c0 += 2; vout5x01 = vget_high_f32(vout5x0123); vout4x01 = vget_high_f32(vout4x0123); vout3x01 = vget_high_f32(vout3x0123); vout2x01 = vget_high_f32(vout2x0123); vout1x01 = vget_high_f32(vout1x0123); vout0x01 = vget_high_f32(vout0x0123); } if (nc & 1) { vst1_lane_f32(c5, vout5x01, 0); vst1_lane_f32(c4, vout4x01, 0); vst1_lane_f32(c3, vout3x01, 0); vst1_lane_f32(c2, vout2x01, 0); vst1_lane_f32(c1, vout1x01, 0); vst1_lane_f32(c0, vout0x01, 0); } nc = 0; } } while (nc != 0); }