// 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_f16_qb4w_gemm_minmax_ukernel_2x16c8__neoni8mm( size_t mr, size_t nc, size_t kc, const int8_t* restrict a, size_t a_stride, const void* restrict w, xnn_float16* restrict c, size_t cm_stride, size_t cn_stride, const struct xnn_f16_qb4w_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)); size_t bl = params->scalar.blocksize; assert(bl <= kc); assert(bl != 0); assert(bl % 32 == 0); const int8_t* a0 = a; uint16_t* c0 = (uint16_t*) c; const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride); if XNN_UNPREDICTABLE(mr != 2) { a1 = a0; c1 = c0; } const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0)); // Loop over groups of 16 columns. do { // Initialize accumulators with scaled vksum. 16 scaled vksum values are loaded from the // weight matrix, at the start of the group of 16 columns. const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point)); const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4; float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0); float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0); const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4; float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0); float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0); const float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4; float32x4_t vout0x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point01), 0); float32x4_t vout1x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point01), 0); const float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4; float32x4_t vout0xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point01), 0); float32x4_t vout1xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point01), 0); for (size_t kb=0; kb < kc; kb += bl) { int32x4_t vacc01x01 = vdupq_n_s32(0); int32x4_t vacc01x23 = vdupq_n_s32(0); int32x4_t vacc01x45 = vdupq_n_s32(0); int32x4_t vacc01x67 = vdupq_n_s32(0); int32x4_t vacc01x89 = vdupq_n_s32(0); int32x4_t vacc01xAB = vdupq_n_s32(0); int32x4_t vacc01xCD = vdupq_n_s32(0); int32x4_t vacc01xEF = vdupq_n_s32(0); size_t k = bl; // 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 16x16 block of weights. const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4); const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4); const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4); const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4); const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4); const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4); const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4); const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4); const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask); const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask); const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask); const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask); const int8x16_t vb89x89ABCDEF = vandq_s8(vb89x0123456789ABCDEF, vmask); const int8x16_t vbABx89ABCDEF = vandq_s8(vbABx0123456789ABCDEF, vmask); const int8x16_t vbCDx89ABCDEF = vandq_s8(vbCDx0123456789ABCDEF, vmask); const int8x16_t vbEFx89ABCDEF = vandq_s8(vbEFx0123456789ABCDEF, vmask); // Multiply-accumulate: 2x8 * 8x16 --> 2x16. 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); 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); 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 16x16 block of weights. const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4); const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4); const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4); const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4); const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4); const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4); const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4); const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4); // Multiply-accumulate: 2x4 * 4x16 --> 2x16. 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); } 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))); const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4; const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4; const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4; const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4; float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123); vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123); float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567); vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567); float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB); vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB); float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF); vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF); float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123); vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123); float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567); vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567); float32x4_t vf1x89AB = vcvtq_f32_s32(vacc1x89AB); vout1x89AB = vfmaq_f32(vout1x89AB, vf1x89AB, vfilter_output_scale89AB); float32x4_t vf1xCDEF = vcvtq_f32_s32(vacc1xCDEF); vout1xCDEF = vfmaq_f32(vout1xCDEF, vf1xCDEF, vfilter_output_scaleCDEF); } 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); const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4; vout0x0123 = vaddq_f32(vbias0123, vout0x0123); vout1x0123 = vaddq_f32(vbias0123, vout1x0123); const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4; vout0x4567 = vaddq_f32(vbias4567, vout0x4567); vout1x4567 = vaddq_f32(vbias4567, vout1x4567); const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4; vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB); vout1x89AB = vaddq_f32(vbias89AB, vout1x89AB); const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4; vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF); vout1xCDEF = vaddq_f32(vbiasCDEF, vout1xCDEF); 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)); float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567)); float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF)); 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); vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min); vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, 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); vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max); vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max); if XNN_LIKELY(nc >= 16) { vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF)); vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF)); a0 = (const int8_t*) ((uintptr_t) a0 - kc); a1 = (const int8_t*) ((uintptr_t) a1 - kc); c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride); c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride); nc -= 16; } else { if (nc & 8) { vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8; vfp16out0x01234567 = vfp16out0x89ABCDEF; vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8; vfp16out1x01234567 = vfp16out1x89ABCDEF; } float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567); float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567); if (nc & 4) { vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4; vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4; vfp16out0x0123 = vget_high_f16(vfp16out0x01234567); vfp16out1x0123 = vget_high_f16(vfp16out1x01234567); } if (nc & 2) { vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2; vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2; vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2); vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2); } if (nc & 1) { vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0); vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0); } nc = 0; } } while (nc != 0); }