| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/x32-packw/scalar.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2023 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 |
|
| 11 |
#include <assert.h> |
| 12 |
#include <stddef.h> |
| 13 |
#include <stdint.h> |
| 14 |
|
| 15 |
#include "xnnpack/math.h" |
| 16 |
#include "xnnpack/packw.h" |
| 17 |
|
| 18 |
|
| 19 |
|
| 20 |
void xnn_x32_packw_gemm_goi_ukernel_x2__scalar_float_u4( |
| 21 |
size_t g, |
| 22 |
size_t nc, |
| 23 |
size_t kc, |
| 24 |
size_t nr, |
| 25 |
size_t kr, |
| 26 |
size_t sr, |
| 27 |
const uint32_t* weights, |
| 28 |
const uint32_t* bias, |
| 29 |
const void* scale, |
| 30 |
uint32_t* packed_weights, |
| 31 |
size_t extra_bytes, |
| 32 |
const void* params) |
| 33 |
{ |
| 34 |
assert(g != 0); |
| 35 |
assert(nc != 0); |
| 36 |
assert(kc != 0); |
| 37 |
assert(nr == 2); |
| 38 |
assert(kr == 1); |
| 39 |
assert(sr == 1); |
| 40 |
assert(weights != NULL); |
| 41 |
assert(packed_weights != NULL); |
| 42 |
|
| 43 |
float* out = (float*) packed_weights; |
| 44 |
const float* b = (const float*) bias; |
| 45 |
|
| 46 |
do { |
| 47 |
// NC main loop multiple of 2 |
| 48 |
const float* w0 = (const float*) weights; |
| 49 |
size_t n = nc; |
| 50 |
for (;n >= 2; n -= 2) { |
| 51 |
if XNN_LIKELY(b != NULL) { |
| 52 |
out[0] = b[0]; |
| 53 |
out[1] = b[1]; |
| 54 |
b += 2; |
| 55 |
} else { |
| 56 |
out[0] = 0; |
| 57 |
out[1] = 0; |
| 58 |
} |
| 59 |
out += 2; |
| 60 |
|
| 61 |
const float* w1 = w0 + kc; |
| 62 |
|
| 63 |
// KC main loop multiple of 2x4 |
| 64 |
size_t k = kc; |
| 65 |
for (; k >= 4; k -= 4) { |
| 66 |
const float v00 = w0[0]; |
| 67 |
const float v01 = w0[1]; |
| 68 |
const float v02 = w0[2]; |
| 69 |
const float v03 = w0[3]; |
| 70 |
w0 += 4; |
| 71 |
const float v10 = w1[0]; |
| 72 |
const float v11 = w1[1]; |
| 73 |
const float v12 = w1[2]; |
| 74 |
const float v13 = w1[3]; |
| 75 |
w1 += 4; |
| 76 |
out[0] = v00; |
| 77 |
out[1] = v10; |
| 78 |
out[2] = v01; |
| 79 |
out[3] = v11; |
| 80 |
out[4] = v02; |
| 81 |
out[5] = v12; |
| 82 |
out[6] = v03; |
| 83 |
out[7] = v13; |
| 84 |
out += 8; |
| 85 |
} |
| 86 |
|
| 87 |
// KC remainder |
| 88 |
for (; k != 0; --k) { |
| 89 |
const float v0 = *w0++; |
| 90 |
out[0] = v0; |
| 91 |
const float v1 = *w1++; |
| 92 |
out[1] = v1; |
| 93 |
out += 2; |
| 94 |
} |
| 95 |
out = (float*) ((uintptr_t) out + extra_bytes); |
| 96 |
w0 = w1; |
| 97 |
} |
| 98 |
|
| 99 |
// NC remainder (1..1) |
| 100 |
if XNN_UNLIKELY(n != 0) { |
| 101 |
if XNN_LIKELY(b != NULL) { |
| 102 |
size_t nb = n; |
| 103 |
do { |
| 104 |
*out++ = *b++; |
| 105 |
} while (--nb != 0); |
| 106 |
} else { |
| 107 |
size_t nb = n; |
| 108 |
do { |
| 109 |
*out++ = 0; |
| 110 |
} while (--nb != 0); |
| 111 |
} |
| 112 |
out += (2 - n); |
| 113 |
|
| 114 |
|
| 115 |
// KC main loop multiple of 2x4 |
| 116 |
size_t k = kc; |
| 117 |
for (; k >= 4; k -= 4) { |
| 118 |
const float v00 = w0[0]; |
| 119 |
const float v01 = w0[1]; |
| 120 |
const float v02 = w0[2]; |
| 121 |
const float v03 = w0[3]; |
| 122 |
w0 += 4; |
| 123 |
out[0] = v00; |
| 124 |
out[2] = v01; |
| 125 |
out[4] = v02; |
| 126 |
out[6] = v03; |
| 127 |
out += 8; |
| 128 |
} |
| 129 |
|
| 130 |
// KC remainder of 1..3 |
| 131 |
for (; k != 0; --k) { |
| 132 |
const float v0 = *w0++; |
| 133 |
out[0] = v0; |
| 134 |
out += 2; |
| 135 |
} |
| 136 |
out = (float*) ((uintptr_t) out + extra_bytes); |
| 137 |
} |
| 138 |
weights += nc * kc; |
| 139 |
} while (--g != 0); |
| 140 |
} |