// Auto-generated file. Do not edit! // Template: src/qs8-igemm/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/math.h" #include "xnnpack/gemm.h" #include "xnnpack/unaligned.h" void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x2__scalar( size_t mr, size_t nc, size_t kc, size_t ks, const int8_t** restrict a, const void* restrict w, float* 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_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 <= 1); assert(nc != 0); assert(kc != 0); assert(ks != 0); assert(ks % (1 * sizeof(void*)) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); float* c0 = c; 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_point = quantization_params->zero_point; int32_t vacc0x0 = vksum0 * vinput_zero_point; int32_t vacc0x1 = vksum1 * vinput_zero_point; w = (const void*) ((const int32_t*) w + 2); size_t p = ks; do { const int8_t* restrict a0 = a[0]; assert(a0 != NULL); if XNN_UNPREDICTABLE(a0 != zero) { a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); } else { a0 = zero_data; } a += 1; size_t k = kc; do { const int32_t va0 = (int32_t) *a0++; const int32_t vb0 = (int32_t) ((const int8_t*) w)[0]; const int32_t vb1 = (int32_t) ((const int8_t*) w)[1]; w = (const void*) ((const int8_t*) w + 2); vacc0x0 += va0 * vb0; vacc0x1 += va0 * vb1; k -= sizeof(int8_t); } while (k != 0); p -= 1 * sizeof(void*); } while (p != 0); float vout0x0 = (float) vacc0x0; float vout0x1 = (float) vacc0x1; const float vinput_scale = quantization_params->inv_scale; vout0x0 *= vinput_scale; vout0x1 *= vinput_scale; const float vfilter_output_scale0 = unaligned_indexed_load_f32(w, 0); vout0x0 *= vfilter_output_scale0; const float vfilter_output_scale1 = unaligned_indexed_load_f32(w, 1); vout0x1 *= vfilter_output_scale1; const float vbias0 = unaligned_indexed_load_f32(w, 2); vout0x0 += vbias0; const float vbias1 = unaligned_indexed_load_f32(w, 3); vout0x1 += vbias1; w = (const float*) w + 4; const float voutput_min = params->scalar.min; vout0x0 = math_max_f32(vout0x0, voutput_min); vout0x1 = math_max_f32(vout0x1, voutput_min); const float voutput_max = params->scalar.max; vout0x0 = math_min_f32(vout0x0, voutput_max); vout0x1 = math_min_f32(vout0x1, voutput_max); if XNN_LIKELY(nc >= 2) { c0[0] = vout0x0; c0[1] = vout0x1; c0 = (float*) ((uintptr_t) c0 + cn_stride); a = (const int8_t**restrict) ((uintptr_t) a - ks); nc -= 2; } else { if (nc & 1) { c0[0] = vout0x0; } nc = 0; } } while (nc != 0); }