| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-vmulc/scalar.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2021 Google LLC |
| 6 |
// |
| 7 |
// This source code is licensed under the BSD-style license found in the |
| 8 |
// LICENSE file in the root directory of this source tree. |
| 9 |
|
| 10 |
#include <assert.h> |
| 11 |
|
| 12 |
#include "xnnpack/math.h" |
| 13 |
#include "xnnpack/vbinary.h" |
| 14 |
|
| 15 |
|
| 16 |
void xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_u4( |
| 17 |
size_t batch, |
| 18 |
const int8_t* input_a, |
| 19 |
const int8_t* input_b, |
| 20 |
int8_t* output, |
| 21 |
const union xnn_qs8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 22 |
{ |
| 23 |
assert(batch != 0); |
| 24 |
assert(batch % sizeof(int8_t) == 0); |
| 25 |
assert(input_a != NULL); |
| 26 |
assert(input_b != NULL); |
| 27 |
assert(output != NULL); |
| 28 |
|
| 29 |
const int32_t va_zero_point = params->scalar.a_zero_point; |
| 30 |
const float vscale = params->scalar.scale; |
| 31 |
const float voutput_min_less_zero_point = (int32_t) params->scalar.output_min - (int32_t) params->scalar.output_zero_point; |
| 32 |
const float voutput_max_less_zero_point = (int32_t) params->scalar.output_max - (int32_t) params->scalar.output_zero_point; |
| 33 |
const float vmagic_bias = 12582912.0f; |
| 34 |
const int32_t vmagic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) params->scalar.output_zero_point; |
| 35 |
|
| 36 |
const int32_t vb = (int32_t) *input_b - params->scalar.b_zero_point; |
| 37 |
for (; batch >= 4 * sizeof(int8_t); batch -= 4 * sizeof(int8_t)) { |
| 38 |
const int32_t va0 = input_a[0] - va_zero_point; |
| 39 |
const int32_t va1 = input_a[1] - va_zero_point; |
| 40 |
const int32_t va2 = input_a[2] - va_zero_point; |
| 41 |
const int32_t va3 = input_a[3] - va_zero_point; |
| 42 |
input_a += 4; |
| 43 |
|
| 44 |
const int32_t vacc0 = va0 * vb; |
| 45 |
const int32_t vacc1 = va1 * vb; |
| 46 |
const int32_t vacc2 = va2 * vb; |
| 47 |
const int32_t vacc3 = va3 * vb; |
| 48 |
|
| 49 |
float vfpacc0 = (float) vacc0 * vscale; |
| 50 |
float vfpacc1 = (float) vacc1 * vscale; |
| 51 |
float vfpacc2 = (float) vacc2 * vscale; |
| 52 |
float vfpacc3 = (float) vacc3 * vscale; |
| 53 |
|
| 54 |
vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point); |
| 55 |
vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point); |
| 56 |
vfpacc2 = math_max_f32(vfpacc2, voutput_min_less_zero_point); |
| 57 |
vfpacc3 = math_max_f32(vfpacc3, voutput_min_less_zero_point); |
| 58 |
|
| 59 |
vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point); |
| 60 |
vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point); |
| 61 |
vfpacc2 = math_min_f32(vfpacc2, voutput_max_less_zero_point); |
| 62 |
vfpacc3 = math_min_f32(vfpacc3, voutput_max_less_zero_point); |
| 63 |
|
| 64 |
vfpacc0 += vmagic_bias; |
| 65 |
vfpacc1 += vmagic_bias; |
| 66 |
vfpacc2 += vmagic_bias; |
| 67 |
vfpacc3 += vmagic_bias; |
| 68 |
|
| 69 |
const int32_t vout0 = (int32_t) float_as_uint32(vfpacc0) - vmagic_bias_less_output_zero_point; |
| 70 |
const int32_t vout1 = (int32_t) float_as_uint32(vfpacc1) - vmagic_bias_less_output_zero_point; |
| 71 |
const int32_t vout2 = (int32_t) float_as_uint32(vfpacc2) - vmagic_bias_less_output_zero_point; |
| 72 |
const int32_t vout3 = (int32_t) float_as_uint32(vfpacc3) - vmagic_bias_less_output_zero_point; |
| 73 |
|
| 74 |
output[0] = (int8_t) vout0; |
| 75 |
output[1] = (int8_t) vout1; |
| 76 |
output[2] = (int8_t) vout2; |
| 77 |
output[3] = (int8_t) vout3; |
| 78 |
output += 4; |
| 79 |
} |
| 80 |
if XNN_UNLIKELY(batch != 0) { |
| 81 |
do { |
| 82 |
const int32_t va = (int32_t) *input_a++ - va_zero_point; |
| 83 |
const int32_t vacc = va * vb; |
| 84 |
|
| 85 |
float vfpacc = (float) vacc * vscale; |
| 86 |
vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point); |
| 87 |
vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point); |
| 88 |
vfpacc += vmagic_bias; |
| 89 |
const int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point; |
| 90 |
*output++ = (int8_t) vout; |
| 91 |
|
| 92 |
batch -= sizeof(int8_t); |
| 93 |
} while (batch != 0); |
| 94 |
} |
| 95 |
} |