// 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" #include "xnnpack/unaligned.h" void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_2x2__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 union xnn_f32_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 <= 2); 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; } do { const int32_t vksum0 = unaligned_indexed_load_s32(w, 0); const int32_t vksum1 = unaligned_indexed_load_s32(w, 1); 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; 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; w = (const int32_t*) w + 2; size_t k = kc; do { const int32_t va0 = (int32_t) *a0++; const int32_t va1 = (int32_t) *a1++; const int32_t vb0 = (int32_t) ((const int8_t*) w)[0]; const int32_t vb1 = (int32_t) ((const int8_t*) w)[1]; w = (const int8_t*) w + 2; vacc0x0 += va0 * vb0; vacc0x1 += va0 * vb1; vacc1x0 += va1 * vb0; vacc1x1 += va1 * vb1; k -= sizeof(int8_t); } while (k != 0); float vout0x0 = (float) vacc0x0; float vout0x1 = (float) vacc0x1; float vout1x0 = (float) vacc1x0; float vout1x1 = (float) vacc1x1; const float vinput_scale0 = quantization_params[0].inv_scale; vout0x0 *= vinput_scale0; vout0x1 *= vinput_scale0; const float vinput_scale1 = quantization_params[1].inv_scale; vout1x0 *= vinput_scale1; vout1x1 *= vinput_scale1; const float vfilter_output_scale0 = unaligned_indexed_load_f32(w, 0); vout0x0 *= vfilter_output_scale0; vout1x0 *= vfilter_output_scale0; const float vfilter_output_scale1 = unaligned_indexed_load_f32(w, 1); vout0x1 *= vfilter_output_scale1; vout1x1 *= vfilter_output_scale1; const float vbias0 = unaligned_indexed_load_f32(w, 2); vout0x0 += vbias0; vout1x0 += vbias0; const float vbias1 = unaligned_indexed_load_f32(w, 3); vout0x1 += vbias1; vout1x1 += vbias1; w = (const float*) w + 4; const float voutput_min = params->scalar.min; vout0x0 = math_max_f32(vout0x0, voutput_min); vout1x0 = math_max_f32(vout1x0, voutput_min); vout0x1 = math_max_f32(vout0x1, voutput_min); vout1x1 = math_max_f32(vout1x1, voutput_min); const float voutput_max = params->scalar.max; vout0x0 = math_min_f32(vout0x0, voutput_max); vout1x0 = math_min_f32(vout1x0, voutput_max); vout0x1 = math_min_f32(vout0x1, voutput_max); vout1x1 = math_min_f32(vout1x1, voutput_max); if XNN_LIKELY(nc >= 2) { c0[0] = vout0x0; c0[1] = vout0x1; c1[0] = vout1x0; c1[1] = vout1x1; a0 = (const int8_t*) ((uintptr_t) a0 - kc); a1 = (const int8_t*) ((uintptr_t) a1 - kc); c0 = (float*) ((uintptr_t) c0 + cn_stride); c1 = (float*) ((uintptr_t) c1 + cn_stride); nc -= 2; } else { if (nc & 1) { c0[0] = vout0x0; c1[0] = vout1x0; } nc = 0; } } while (nc != 0); }