// Auto-generated file. Do not edit! // Template: src/qs8-igemm/c2-neon-mull-shuffle.c.in // Generator: tools/xngen // // Copyright 2021 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_qc8w_igemm_minmax_ukernel_2x8c2s4__neonfp16arith( 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 <= 2); assert(nc != 0); assert(kc != 0); assert(ks != 0); assert(ks % (2 * sizeof(void*)) == 0); assert(a_offset % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); uint16_t* c0 = (uint16_t*) c; uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride); if XNN_UNPREDICTABLE(mr != 2) { c1 = c0; } const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point); const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale); kc = round_up_po2(kc, 8 * sizeof(int8_t)); do { const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4; int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point); const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4; int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point); int32x4_t vacc1x0123 = vacc0x0123; int32x4_t vacc1x4567 = 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; } a += 2; size_t k = kc; while (k >= 16 * sizeof(int8_t)) { int8x8_t va0x0 = vld1_s8(a0); a0 += 8; int8x8_t va0x1 = vld1_s8(a0); a0 += 8; int8x8_t va1x0 = vld1_s8(a1); a1 += 8; int8x8_t va1x1 = vld1_s8(a1); a1 += 8; const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8; int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0); int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0); const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8; vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1); vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1); vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0); vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0); int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0); int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0); const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8; vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1); vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1); vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0); vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0); va0x0 = vext_s8(va0x0, va0x0, 2); va0x1 = vext_s8(va0x1, va0x1, 2); va1x0 = vext_s8(va1x0, va1x0, 2); va1x1 = vext_s8(va1x1, va1x1, 2); int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0); int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0); const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8; vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1); vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1); vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1); vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1); int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0); int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0); const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8; vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1); vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1); vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1); va0x0 = vext_s8(va0x0, va0x0, 2); va0x1 = vext_s8(va0x1, va0x1, 2); va1x0 = vext_s8(va1x0, va1x0, 2); va1x1 = vext_s8(va1x1, va1x1, 2); int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0); int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0); const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8; vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1); vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1); vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2); vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2); int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0); int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0); const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8; vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1); vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1); vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2); vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2); va0x0 = vext_s8(va0x0, va0x0, 2); va0x1 = vext_s8(va0x1, va0x1, 2); va1x0 = vext_s8(va1x0, va1x0, 2); va1x1 = vext_s8(va1x1, va1x1, 2); int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0); int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0); const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8; vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1); vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1); vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3); vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3); int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0); int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0); const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8; vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1); vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1); vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3); vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3); k -= 16 * sizeof(int8_t); } if (k != 0) { int8x8_t va0x0 = vld1_s8(a0); a0 += 8; int8x8_t va1x0 = vld1_s8(a1); a1 += 8; const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8; const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8; int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0); int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0); vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0); vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0); int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0); int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0); vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0); vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0); va0x0 = vext_s8(va0x0, va0x0, 2); va1x0 = vext_s8(va1x0, va1x0, 2); int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0); int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0); vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1); vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1); int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0); int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0); vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1); va0x0 = vext_s8(va0x0, va0x0, 2); va1x0 = vext_s8(va1x0, va1x0, 2); int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0); int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0); vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2); vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2); int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0); int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0); vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2); vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2); va0x0 = vext_s8(va0x0, va0x0, 2); va1x0 = vext_s8(va1x0, va1x0, 2); int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0); int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0); vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3); vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3); int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0); int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0); vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3); vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3); } p -= 2 * 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); vout0x0123 = vmulq_f32(vout0x0123, vinput_scale); vout0x4567 = vmulq_f32(vout0x4567, vinput_scale); vout1x0123 = vmulq_f32(vout1x0123, vinput_scale); vout1x4567 = vmulq_f32(vout1x4567, 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); #else vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123); vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, 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); #else vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567); vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567); #endif float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567)); float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567)); #if XNN_ARCH_ARM64 const uint16x8x2_t voutput_minmax = vld2q_dup_u16((const uint16_t*) ¶ms->scalar.min); const float16x8_t voutput_min = vreinterpretq_f16_u16(voutput_minmax.val[0]); const float16x8_t voutput_max = vreinterpretq_f16_u16(voutput_minmax.val[1]); #else // vld2_dup is to work around aarch32 clang bug with vld1q_dup const uint16x4x2_t vminmax = vld2_dup_u16((const uint16_t*) ¶ms->scalar.min); const float16x8_t voutput_min = vreinterpretq_f16_u16(vcombine_u16(vminmax.val[0], vminmax.val[0])); const float16x8_t voutput_max = vreinterpretq_f16_u16(vcombine_u16(vminmax.val[1], vminmax.val[1])); #endif vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min); vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min); vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max); vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max); if XNN_LIKELY(nc >= 8) { vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); a = (const int8_t**restrict) ((uintptr_t) a - ks); c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride); c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride); nc -= 8; } else { float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567); float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567); if (nc & 4) { vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4; vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4; vfp16out1x0123 = vget_high_f16(vfp16out1x01234567); vfp16out0x0123 = vget_high_f16(vfp16out0x01234567); } if (nc & 2) { vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2; vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2; vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2); vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2); } if (nc & 1) { vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0); vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0); } nc = 0; } } while (nc != 0); }