| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-dwconv/unipass-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 |
#include <stddef.h> |
| 12 |
#include <stdint.h> |
| 13 |
|
| 14 |
#include "xnnpack/common.h" |
| 15 |
#include "xnnpack/dwconv.h" |
| 16 |
#include "xnnpack/math.h" |
| 17 |
#include "xnnpack/microparams.h" |
| 18 |
#include "xnnpack/unaligned.h" |
| 19 |
|
| 20 |
void xnn_qs8_dwconv_minmax_fp32_ukernel_9p1c__scalar_fmagic( |
| 21 |
size_t channels, |
| 22 |
size_t output_width, |
| 23 |
const int8_t** input, |
| 24 |
const void* weights, |
| 25 |
int8_t* output, |
| 26 |
intptr_t input_stride, |
| 27 |
size_t output_increment, |
| 28 |
size_t input_offset, |
| 29 |
const int8_t* zero, |
| 30 |
const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 31 |
{ |
| 32 |
assert(channels != 0); |
| 33 |
assert(output_width != 0); |
| 34 |
|
| 35 |
const float vscale = params->fp32_scalar.scale; |
| 36 |
const float voutput_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point; |
| 37 |
const float voutput_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point; |
| 38 |
const float vmagic_bias = 12582912.0f; |
| 39 |
const int32_t vmagic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point; |
| 40 |
do { |
| 41 |
const int8_t* i0 = input[0]; |
| 42 |
assert(i0 != NULL); |
| 43 |
if XNN_UNPREDICTABLE(i0 != zero) { |
| 44 |
i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); |
| 45 |
} |
| 46 |
const int8_t* i1 = input[1]; |
| 47 |
assert(i1 != NULL); |
| 48 |
if XNN_UNPREDICTABLE(i1 != zero) { |
| 49 |
i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); |
| 50 |
} |
| 51 |
const int8_t* i2 = input[2]; |
| 52 |
assert(i2 != NULL); |
| 53 |
if XNN_UNPREDICTABLE(i2 != zero) { |
| 54 |
i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); |
| 55 |
} |
| 56 |
const int8_t* i3 = input[3]; |
| 57 |
assert(i3 != NULL); |
| 58 |
if XNN_UNPREDICTABLE(i3 != zero) { |
| 59 |
i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); |
| 60 |
} |
| 61 |
const int8_t* i4 = input[4]; |
| 62 |
assert(i4 != NULL); |
| 63 |
if XNN_UNPREDICTABLE(i4 != zero) { |
| 64 |
i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); |
| 65 |
} |
| 66 |
const int8_t* i5 = input[5]; |
| 67 |
assert(i5 != NULL); |
| 68 |
if XNN_UNPREDICTABLE(i5 != zero) { |
| 69 |
i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); |
| 70 |
} |
| 71 |
const int8_t* i6 = input[6]; |
| 72 |
assert(i6 != NULL); |
| 73 |
if XNN_UNPREDICTABLE(i6 != zero) { |
| 74 |
i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); |
| 75 |
} |
| 76 |
const int8_t* i7 = input[7]; |
| 77 |
assert(i7 != NULL); |
| 78 |
if XNN_UNPREDICTABLE(i7 != zero) { |
| 79 |
i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); |
| 80 |
} |
| 81 |
const int8_t* i8 = input[8]; |
| 82 |
assert(i8 != NULL); |
| 83 |
if XNN_UNPREDICTABLE(i8 != zero) { |
| 84 |
i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); |
| 85 |
} |
| 86 |
input = (const int8_t**) ((uintptr_t) input + input_stride); |
| 87 |
|
| 88 |
size_t c = channels; |
| 89 |
const void* w = weights; |
| 90 |
do { |
| 91 |
int32_t vacc = unaligned_load_s32(w); |
| 92 |
|
| 93 |
const int32_t vi0 = (int32_t) *i0++; |
| 94 |
const int32_t vk0 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[0]; |
| 95 |
vacc += vi0 * vk0; |
| 96 |
const int32_t vi1 = (int32_t) *i1++; |
| 97 |
const int32_t vk1 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[1]; |
| 98 |
vacc += vi1 * vk1; |
| 99 |
const int32_t vi2 = (int32_t) *i2++; |
| 100 |
const int32_t vk2 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[2]; |
| 101 |
vacc += vi2 * vk2; |
| 102 |
const int32_t vi3 = (int32_t) *i3++; |
| 103 |
const int32_t vk3 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[3]; |
| 104 |
vacc += vi3 * vk3; |
| 105 |
const int32_t vi4 = (int32_t) *i4++; |
| 106 |
const int32_t vk4 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[4]; |
| 107 |
vacc += vi4 * vk4; |
| 108 |
const int32_t vi5 = (int32_t) *i5++; |
| 109 |
const int32_t vk5 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[5]; |
| 110 |
vacc += vi5 * vk5; |
| 111 |
const int32_t vi6 = (int32_t) *i6++; |
| 112 |
const int32_t vk6 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[6]; |
| 113 |
vacc += vi6 * vk6; |
| 114 |
const int32_t vi7 = (int32_t) *i7++; |
| 115 |
const int32_t vk7 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[7]; |
| 116 |
vacc += vi7 * vk7; |
| 117 |
const int32_t vi8 = (int32_t) *i8++; |
| 118 |
const int32_t vk8 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[8]; |
| 119 |
vacc += vi8 * vk8; |
| 120 |
|
| 121 |
w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 9 * sizeof(int8_t)); |
| 122 |
|
| 123 |
float vfpacc = (float) vacc * vscale; |
| 124 |
|
| 125 |
vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point); |
| 126 |
vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point); |
| 127 |
vfpacc += vmagic_bias; |
| 128 |
int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point; |
| 129 |
|
| 130 |
*output++ = (int8_t) vout; |
| 131 |
} while (--c != 0); |
| 132 |
|
| 133 |
output = (int8_t*) ((uintptr_t) output + output_increment); |
| 134 |
} while (--output_width != 0); |
| 135 |
} |