// Auto-generated file. Do not edit! // Template: src/qs8-igemm/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/igemm.h" #include "xnnpack/math.h" void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_1x16c8__neoni8mm( size_t mr, size_t nc, size_t kc, size_t ks, const int8_t** restrict a, const void* restrict w, xnn_float16* 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_f16_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 <= 1); assert(nc != 0); assert(kc != 0); assert(ks != 0); assert(ks % (1 * sizeof(void*)) == 0); assert(a_offset % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8 * sizeof(int8_t)); uint16_t* c0 = (uint16_t*) c; do { const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point); const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0x0123 = vmulq_s32(vksum0123, vinput_zero_point); const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point); const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0x89AB = vmulq_s32(vksum89AB, vinput_zero_point); const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4; const int32x4_t vksumzp0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point); int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp0x0123))); int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp0x0123))); int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp0x4567))); int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp0x4567))); int32x4_t vacc01x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp0x89AB))); int32x4_t vacc01xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp0x89AB))); int32x4_t vacc01xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp0xCDEF))); int32x4_t vacc01xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp0xCDEF))); 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; } a += 1; uint64x2x2_t va01x0123456789ABCDEF; va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0); va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0); // Inner accumulation loop along the 16 columns. size_t k = kc; // 2x partial unrolled loop to load 8 bytes at a time. while (k >= 16 * sizeof(int8_t)) { // Load a 1x16 block of activations. va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16; // Load a 16x16 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 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; // Multiply-accumulate: 1x16 * 16x16 --> 1x16. vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567); vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567); vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567); vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567); vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb89x01234567); vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbABx01234567); vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbCDx01234567); vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbEFx01234567); vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF); vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF); vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF); vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF); vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb89x89ABCDEF); vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbABx89ABCDEF); vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbCDx89ABCDEF); vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbEFx89ABCDEF); k -= 16 * sizeof(int8_t); } // Handle up to 8 final positions of `k` if XNN_UNLIKELY(k != 0) { // Load a 1x8 block of activations. uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8; // Load a 16x16 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; // Multiply-accumulate: 1x4 * 4x16 --> 1x16. 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); } p -= 1 * sizeof(void*); } while (p != 0); int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23)); int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67)); int32x4_t vacc0x89AB = vcombine_s32(vget_low_s32(vacc01x89), vget_low_s32(vacc01xAB)); int32x4_t vacc0xCDEF = vcombine_s32(vget_low_s32(vacc01xCD), vget_low_s32(vacc01xEF)); const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale); 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); vout0x0123 = vmulq_f32(vout0x0123, vinput_scale); vout0x4567 = vmulq_f32(vout0x4567, vinput_scale); vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale); vout0xCDEF = vmulq_f32(vout0xCDEF, 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 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 vbias0123 = vld1q_f32(w); w = (const float*) w + 4; vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123); const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4; vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567); const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4; vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB); const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4; vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF); float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567)); float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF)); const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) ¶ms->scalar.min)); vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min); vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min); const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) ¶ms->scalar.max)); vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max); vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max); if XNN_LIKELY(nc >= 16) { vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF)); c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride); a = (const int8_t**restrict) ((uintptr_t) a - ks); nc -= 16; } else { if (nc & 8) { vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8; vfp16out0x01234567 = vfp16out0x89ABCDEF; } float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567); if (nc & 4) { vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4; vfp16out0x0123 = vget_high_f16(vfp16out0x01234567); } if (nc & 2) { vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2; vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2); } if (nc & 1) { vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0); } nc = 0; } } while (nc != 0); }