// Auto-generated file. Do not edit! // Template: src/qs8-gemm/scalar.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 "xnnpack/gemm.h" #include "xnnpack/math.h" void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_4x4__scalar( size_t mr, size_t nc, size_t kc, const int8_t* restrict a, size_t a_stride, const void* restrict w, float* restrict c, size_t cm_stride, size_t cn_stride, const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) { assert(mr != 0); assert(mr <= 4); assert(nc != 0); assert(kc != 0); const int8_t* a0 = a; float* c0 = c; const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); float* c1 = (float*) ((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); float* c2 = (float*) ((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); float* c3 = (float*) ((uintptr_t) c2 + cm_stride); if XNN_UNPREDICTABLE(mr != 4) { a3 = a2; c3 = c2; } kc = round_up_po2(kc, 2); do { const int32_t vksum0 = ((const int32_t*) w)[0]; const int32_t vksum1 = ((const int32_t*) w)[1]; const int32_t vksum2 = ((const int32_t*) w)[2]; const int32_t vksum3 = ((const int32_t*) w)[3]; const int32_t vinput_zero_point0 = quantization_params[0].zero_point; int32_t vacc0x0 = vksum0 * vinput_zero_point0; int32_t vacc0x1 = vksum1 * vinput_zero_point0; int32_t vacc0x2 = vksum2 * vinput_zero_point0; int32_t vacc0x3 = vksum3 * vinput_zero_point0; const int32_t vinput_zero_point1 = quantization_params[1].zero_point; int32_t vacc1x0 = vksum0 * vinput_zero_point1; int32_t vacc1x1 = vksum1 * vinput_zero_point1; int32_t vacc1x2 = vksum2 * vinput_zero_point1; int32_t vacc1x3 = vksum3 * vinput_zero_point1; const int32_t vinput_zero_point2 = quantization_params[2].zero_point; int32_t vacc2x0 = vksum0 * vinput_zero_point2; int32_t vacc2x1 = vksum1 * vinput_zero_point2; int32_t vacc2x2 = vksum2 * vinput_zero_point2; int32_t vacc2x3 = vksum3 * vinput_zero_point2; const int32_t vinput_zero_point3 = quantization_params[3].zero_point; int32_t vacc3x0 = vksum0 * vinput_zero_point3; int32_t vacc3x1 = vksum1 * vinput_zero_point3; int32_t vacc3x2 = vksum2 * vinput_zero_point3; int32_t vacc3x3 = vksum3 * vinput_zero_point3; w = (const int32_t*) w + 4; size_t k = kc; for (; k >= 2 * sizeof(uint8_t); k -= 2 * sizeof(uint8_t)) { const int32_t va0c0 = (int32_t) a0[0]; const int32_t va0c1 = (int32_t) a0[1]; a0 += 2; const int32_t va1c0 = (int32_t) a1[0]; const int32_t va1c1 = (int32_t) a1[1]; a1 += 2; const int32_t va2c0 = (int32_t) a2[0]; const int32_t va2c1 = (int32_t) a2[1]; a2 += 2; const int32_t va3c0 = (int32_t) a3[0]; const int32_t va3c1 = (int32_t) a3[1]; a3 += 2; const uint8_t vbi0 = ((const uint8_t*) w)[0]; const uint8_t vbi1 = ((const uint8_t*) w)[1]; const uint8_t vbi2 = ((const uint8_t*) w)[2]; const uint8_t vbi3 = ((const uint8_t*) w)[3]; w = (const uint8_t*) w + 4; const int32_t vb0c0 = (int32_t) (int8_t) (vbi0 << 4); const int32_t vb0c1 = (int32_t) (int8_t) (vbi0 & 0xF0); const int32_t vb1c0 = (int32_t) (int8_t) (vbi1 << 4); const int32_t vb1c1 = (int32_t) (int8_t) (vbi1 & 0xF0); const int32_t vb2c0 = (int32_t) (int8_t) (vbi2 << 4); const int32_t vb2c1 = (int32_t) (int8_t) (vbi2 & 0xF0); const int32_t vb3c0 = (int32_t) (int8_t) (vbi3 << 4); const int32_t vb3c1 = (int32_t) (int8_t) (vbi3 & 0xF0); vacc0x0 += va0c0 * vb0c0; vacc0x1 += va0c0 * vb1c0; vacc0x2 += va0c0 * vb2c0; vacc0x3 += va0c0 * vb3c0; vacc1x0 += va1c0 * vb0c0; vacc1x1 += va1c0 * vb1c0; vacc1x2 += va1c0 * vb2c0; vacc1x3 += va1c0 * vb3c0; vacc2x0 += va2c0 * vb0c0; vacc2x1 += va2c0 * vb1c0; vacc2x2 += va2c0 * vb2c0; vacc2x3 += va2c0 * vb3c0; vacc3x0 += va3c0 * vb0c0; vacc3x1 += va3c0 * vb1c0; vacc3x2 += va3c0 * vb2c0; vacc3x3 += va3c0 * vb3c0; vacc0x0 += va0c1 * vb0c1; vacc0x1 += va0c1 * vb1c1; vacc0x2 += va0c1 * vb2c1; vacc0x3 += va0c1 * vb3c1; vacc1x0 += va1c1 * vb0c1; vacc1x1 += va1c1 * vb1c1; vacc1x2 += va1c1 * vb2c1; vacc1x3 += va1c1 * vb3c1; vacc2x0 += va2c1 * vb0c1; vacc2x1 += va2c1 * vb1c1; vacc2x2 += va2c1 * vb2c1; vacc2x3 += va2c1 * vb3c1; vacc3x0 += va3c1 * vb0c1; vacc3x1 += va3c1 * vb1c1; vacc3x2 += va3c1 * vb2c1; vacc3x3 += va3c1 * vb3c1; } float vout0x0 = (float) math_asr_s32(vacc0x0, 4); float vout0x1 = (float) math_asr_s32(vacc0x1, 4); float vout0x2 = (float) math_asr_s32(vacc0x2, 4); float vout0x3 = (float) math_asr_s32(vacc0x3, 4); float vout1x0 = (float) math_asr_s32(vacc1x0, 4); float vout1x1 = (float) math_asr_s32(vacc1x1, 4); float vout1x2 = (float) math_asr_s32(vacc1x2, 4); float vout1x3 = (float) math_asr_s32(vacc1x3, 4); float vout2x0 = (float) math_asr_s32(vacc2x0, 4); float vout2x1 = (float) math_asr_s32(vacc2x1, 4); float vout2x2 = (float) math_asr_s32(vacc2x2, 4); float vout2x3 = (float) math_asr_s32(vacc2x3, 4); float vout3x0 = (float) math_asr_s32(vacc3x0, 4); float vout3x1 = (float) math_asr_s32(vacc3x1, 4); float vout3x2 = (float) math_asr_s32(vacc3x2, 4); float vout3x3 = (float) math_asr_s32(vacc3x3, 4); const float vinput_scale0 = quantization_params[0].inv_scale; vout0x0 *= vinput_scale0; vout0x1 *= vinput_scale0; vout0x2 *= vinput_scale0; vout0x3 *= vinput_scale0; const float vinput_scale1 = quantization_params[1].inv_scale; vout1x0 *= vinput_scale1; vout1x1 *= vinput_scale1; vout1x2 *= vinput_scale1; vout1x3 *= vinput_scale1; const float vinput_scale2 = quantization_params[2].inv_scale; vout2x0 *= vinput_scale2; vout2x1 *= vinput_scale2; vout2x2 *= vinput_scale2; vout2x3 *= vinput_scale2; const float vinput_scale3 = quantization_params[3].inv_scale; vout3x0 *= vinput_scale3; vout3x1 *= vinput_scale3; vout3x2 *= vinput_scale3; vout3x3 *= vinput_scale3; const float vfilter_output_scale0 = ((const float*) w)[0]; vout0x0 *= vfilter_output_scale0; vout1x0 *= vfilter_output_scale0; vout2x0 *= vfilter_output_scale0; vout3x0 *= vfilter_output_scale0; const float vfilter_output_scale1 = ((const float*) w)[1]; vout0x1 *= vfilter_output_scale1; vout1x1 *= vfilter_output_scale1; vout2x1 *= vfilter_output_scale1; vout3x1 *= vfilter_output_scale1; const float vfilter_output_scale2 = ((const float*) w)[2]; vout0x2 *= vfilter_output_scale2; vout1x2 *= vfilter_output_scale2; vout2x2 *= vfilter_output_scale2; vout3x2 *= vfilter_output_scale2; const float vfilter_output_scale3 = ((const float*) w)[3]; vout0x3 *= vfilter_output_scale3; vout1x3 *= vfilter_output_scale3; vout2x3 *= vfilter_output_scale3; vout3x3 *= vfilter_output_scale3; const float vbias0 = ((const float*) w)[4]; vout0x0 += vbias0; vout1x0 += vbias0; vout2x0 += vbias0; vout3x0 += vbias0; const float vbias1 = ((const float*) w)[5]; vout0x1 += vbias1; vout1x1 += vbias1; vout2x1 += vbias1; vout3x1 += vbias1; const float vbias2 = ((const float*) w)[6]; vout0x2 += vbias2; vout1x2 += vbias2; vout2x2 += vbias2; vout3x2 += vbias2; const float vbias3 = ((const float*) w)[7]; vout0x3 += vbias3; vout1x3 += vbias3; vout2x3 += vbias3; vout3x3 += vbias3; w = (const float*) w + 8; const float voutput_min = params->scalar.min; vout0x0 = math_max_f32(vout0x0, voutput_min); vout1x0 = math_max_f32(vout1x0, voutput_min); vout2x0 = math_max_f32(vout2x0, voutput_min); vout3x0 = math_max_f32(vout3x0, voutput_min); vout0x1 = math_max_f32(vout0x1, voutput_min); vout1x1 = math_max_f32(vout1x1, voutput_min); vout2x1 = math_max_f32(vout2x1, voutput_min); vout3x1 = math_max_f32(vout3x1, voutput_min); vout0x2 = math_max_f32(vout0x2, voutput_min); vout1x2 = math_max_f32(vout1x2, voutput_min); vout2x2 = math_max_f32(vout2x2, voutput_min); vout3x2 = math_max_f32(vout3x2, voutput_min); vout0x3 = math_max_f32(vout0x3, voutput_min); vout1x3 = math_max_f32(vout1x3, voutput_min); vout2x3 = math_max_f32(vout2x3, voutput_min); vout3x3 = math_max_f32(vout3x3, voutput_min); const float voutput_max = params->scalar.max; vout0x0 = math_min_f32(vout0x0, voutput_max); vout1x0 = math_min_f32(vout1x0, voutput_max); vout2x0 = math_min_f32(vout2x0, voutput_max); vout3x0 = math_min_f32(vout3x0, voutput_max); vout0x1 = math_min_f32(vout0x1, voutput_max); vout1x1 = math_min_f32(vout1x1, voutput_max); vout2x1 = math_min_f32(vout2x1, voutput_max); vout3x1 = math_min_f32(vout3x1, voutput_max); vout0x2 = math_min_f32(vout0x2, voutput_max); vout1x2 = math_min_f32(vout1x2, voutput_max); vout2x2 = math_min_f32(vout2x2, voutput_max); vout3x2 = math_min_f32(vout3x2, voutput_max); vout0x3 = math_min_f32(vout0x3, voutput_max); vout1x3 = math_min_f32(vout1x3, voutput_max); vout2x3 = math_min_f32(vout2x3, voutput_max); vout3x3 = math_min_f32(vout3x3, voutput_max); if XNN_LIKELY(nc >= 4) { c0[0] = vout0x0; c0[1] = vout0x1; c0[2] = vout0x2; c0[3] = vout0x3; c1[0] = vout1x0; c1[1] = vout1x1; c1[2] = vout1x2; c1[3] = vout1x3; c2[0] = vout2x0; c2[1] = vout2x1; c2[2] = vout2x2; c2[3] = vout2x3; c3[0] = vout3x0; c3[1] = vout3x1; c3[2] = vout3x2; c3[3] = vout3x3; 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); 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); nc -= 4; } else { if (nc & 2) { c0[0] = vout0x0; c0[1] = vout0x1; vout0x0 = vout0x2; c0 += 2; c1[0] = vout1x0; c1[1] = vout1x1; vout1x0 = vout1x2; c1 += 2; c2[0] = vout2x0; c2[1] = vout2x1; vout2x0 = vout2x2; c2 += 2; c3[0] = vout3x0; c3[1] = vout3x1; vout3x0 = vout3x2; c3 += 2; } if (nc & 1) { c0[0] = vout0x0; c1[0] = vout1x0; c2[0] = vout2x0; c3[0] = vout3x0; } nc = 0; } } while (nc != 0); }