Debian Codegraph @

src/f16-f32acc-rsum/gen/f16-f32acc-rsum-f16c-u24-acc3.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/f16-f32acc-rsum/f16c.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 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/reduce.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_f16_f32acc_rsum_ukernel__f16c_u24_acc3(
21 size_t batch,
22 const xnn_float16* input,
23 float* output,
24 const struct xnn_f16_f32acc_scale_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
27
28 assert(batch != 0);
29 assert(batch % sizeof(uint16_t) == 0);
30 assert(input != NULL);
31 assert(output != NULL);
32
33 const uint16_t* i = (const uint16_t*) input;
34 __m256 vacc0 = _mm256_setzero_ps();
35 __m256 vacc1 = _mm256_setzero_ps();
36 __m256 vacc2 = _mm256_setzero_ps();
37 for (; batch >= 24 * sizeof(uint16_t); batch -= 24 * sizeof(uint16_t)) {
38 const __m256 vt0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
39 const __m256 vt1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 8)));
40 const __m256 vt2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 16)));
41 i += 24;
42
43 vacc0 = _mm256_add_ps(vacc0, vt0);
44 vacc1 = _mm256_add_ps(vacc1, vt1);
45 vacc2 = _mm256_add_ps(vacc2, vt2);
46 }
47 vacc0 = _mm256_add_ps(vacc0, vacc1);
48 vacc0 = _mm256_add_ps(vacc0, vacc2);
49 for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
50 const __m256 vt = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
51 i += 8;
52
53 vacc0 = _mm256_add_ps(vacc0, vt);
54 }
55 if XNN_UNLIKELY(batch != 0) {
56 assert(batch >= 1 * sizeof(uint16_t));
57 assert(batch <= 7 * sizeof(uint16_t));
58 const __m128i vmask = _mm_loadu_si128((const __m128i*) ((uintptr_t) &mask_table[7] - batch));
59 const __m128i vh = _mm_castps_si128(_mm_maskload_ps((const float*) i, vmask));
60 const __m256 vt = _mm256_cvtph_ps(vh);
61 vacc0 = _mm256_add_ps(vacc0, vt);
62 i = (const void*) ((uintptr_t) i + batch);
63 if (batch & (1 * sizeof(uint16_t))) {
64 const __m128i vh = _mm_insert_epi16(_mm_setzero_si128(), (int) unaligned_load_u16(i - 1), 0);
65 const __m256 vt = _mm256_zextps128_ps256(_mm_cvtph_ps(vh));
66 vacc0 = _mm256_add_ps(vacc0, vt);
67 }
68 }
69 __m128 vacc = _mm_add_ps(_mm256_castps256_ps128(vacc0), _mm256_extractf128_ps(vacc0, 1));
70 vacc = _mm_add_ps(vacc, _mm_movehl_ps(vacc, vacc));
71 vacc = _mm_add_ss(vacc, _mm_movehdup_ps(vacc));
72 vacc = _mm_mul_ss(vacc, _mm_load_ss(&params->scalar.scale));
73
74 float vout = _mm_cvtss_f32(vacc);
75 *output += vout;
76 }