| // Auto-generated file. Do not edit! |
| // Template: src/qs8-vmul/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 <assert.h> |
| |
| #include <fp16.h> |
| |
| #include <xnnpack/math.h> |
| #include <xnnpack/vmul.h> |
| |
| |
| void xnn_qs8_vmul_minmax_fp32_ukernel__scalar_x4( |
| size_t n, |
| const int8_t* input_a, |
| const int8_t* input_b, |
| int8_t* output, |
| const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| { |
| const int32_t va_zero_point = params->fp32_scalar.a_zero_point; |
| const int32_t vb_zero_point = params->fp32_scalar.b_zero_point; |
| const float vscale = params->fp32_scalar.scale; |
| const float voutput_min_less_zero_point = params->fp32_scalar.output_min_less_zero_point; |
| const float voutput_max_less_zero_point = params->fp32_scalar.output_max_less_zero_point; |
| const float vmagic_bias = params->fp32_scalar.magic_bias; |
| const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar.magic_bias_less_output_zero_point; |
| |
| for (; n >= 4 * sizeof(int8_t); n -= 4 * sizeof(int8_t)) { |
| const int32_t va0 = input_a[0] - va_zero_point; |
| const int32_t va1 = input_a[1] - va_zero_point; |
| const int32_t va2 = input_a[2] - va_zero_point; |
| const int32_t va3 = input_a[3] - va_zero_point; |
| input_a += 4; |
| |
| const int32_t vb0 = input_b[0] - vb_zero_point; |
| const int32_t vb1 = input_b[1] - vb_zero_point; |
| const int32_t vb2 = input_b[2] - vb_zero_point; |
| const int32_t vb3 = input_b[3] - vb_zero_point; |
| input_b += 4; |
| |
| const int32_t vacc0 = va0 * vb0; |
| const int32_t vacc1 = va1 * vb1; |
| const int32_t vacc2 = va2 * vb2; |
| const int32_t vacc3 = va3 * vb3; |
| |
| float vfpacc0 = (float) vacc0 * vscale; |
| float vfpacc1 = (float) vacc1 * vscale; |
| float vfpacc2 = (float) vacc2 * vscale; |
| float vfpacc3 = (float) vacc3 * vscale; |
| |
| vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point); |
| vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point); |
| vfpacc2 = math_max_f32(vfpacc2, voutput_min_less_zero_point); |
| vfpacc3 = math_max_f32(vfpacc3, voutput_min_less_zero_point); |
| |
| vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point); |
| vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point); |
| vfpacc2 = math_min_f32(vfpacc2, voutput_max_less_zero_point); |
| vfpacc3 = math_min_f32(vfpacc3, voutput_max_less_zero_point); |
| |
| vfpacc0 += vmagic_bias; |
| vfpacc1 += vmagic_bias; |
| vfpacc2 += vmagic_bias; |
| vfpacc3 += vmagic_bias; |
| |
| const int32_t vout0 = (int32_t) fp32_to_bits(vfpacc0) - vmagic_bias_less_output_zero_point; |
| const int32_t vout1 = (int32_t) fp32_to_bits(vfpacc1) - vmagic_bias_less_output_zero_point; |
| const int32_t vout2 = (int32_t) fp32_to_bits(vfpacc2) - vmagic_bias_less_output_zero_point; |
| const int32_t vout3 = (int32_t) fp32_to_bits(vfpacc3) - vmagic_bias_less_output_zero_point; |
| |
| output[0] = (int8_t) vout0; |
| output[1] = (int8_t) vout1; |
| output[2] = (int8_t) vout2; |
| output[3] = (int8_t) vout3; |
| output += 4; |
| } |
| if XNN_UNLIKELY(n != 0) { |
| do { |
| const int32_t va = (int32_t) *input_a++ - va_zero_point; |
| const int32_t vb = (int32_t) *input_b++ - vb_zero_point; |
| const int32_t vacc = va * vb; |
| |
| float vfpacc = (float) vacc * vscale; |
| vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point); |
| vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point); |
| vfpacc += vmagic_bias; |
| const int32_t vout = (int32_t) fp32_to_bits(vfpacc) - vmagic_bias_less_output_zero_point; |
| *output++ = (int8_t) vout; |
| |
| n -= sizeof(int8_t); |
| } while (n != 0); |
| } |
| } |