Debian Codegraph @

src/f16-f32acc-rsum/gen/f16-f32acc-rsum-avx512skx-u128-acc4.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/f16-f32acc-rsum/avx512skx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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__avx512skx_u128_acc4(
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 assert(batch != 0);
27 assert(batch % sizeof(uint16_t) == 0);
28 assert(input != NULL);
29 assert(output != NULL);
30
31 const uint16_t* i = (const uint16_t*) input;
32 __m512 vacc0 = _mm512_setzero_ps();
33 __m512 vacc1 = _mm512_setzero_ps();
34 __m512 vacc2 = _mm512_setzero_ps();
35 __m512 vacc3 = _mm512_setzero_ps();
36 for (; batch >= 128 * sizeof(uint16_t); batch -= 128 * sizeof(uint16_t)) {
37 const __m512 vt0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) i));
38 const __m512 vt1 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + 16)));
39 const __m512 vt2 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + 32)));
40 const __m512 vt3 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + 48)));
41 const __m512 vt4 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + 64)));
42 const __m512 vt5 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + 80)));
43 const __m512 vt6 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + 96)));
44 const __m512 vt7 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + 112)));
45 i += 128;
46
47 vacc0 = _mm512_add_ps(vacc0, vt0);
48 vacc1 = _mm512_add_ps(vacc1, vt1);
49 vacc2 = _mm512_add_ps(vacc2, vt2);
50 vacc3 = _mm512_add_ps(vacc3, vt3);
51 vacc0 = _mm512_add_ps(vacc0, vt4);
52 vacc1 = _mm512_add_ps(vacc1, vt5);
53 vacc2 = _mm512_add_ps(vacc2, vt6);
54 vacc3 = _mm512_add_ps(vacc3, vt7);
55 }
56 vacc0 = _mm512_add_ps(vacc0, vacc1);
57 vacc2 = _mm512_add_ps(vacc2, vacc3);
58 vacc0 = _mm512_add_ps(vacc0, vacc2);
59 for (; batch >= 16 * sizeof(uint16_t); batch -= 16 * sizeof(uint16_t)) {
60 const __m512 vt = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) i));
61 i += 16;
62
63 vacc0 = _mm512_add_ps(vacc0, vt);
64 }
65 if XNN_UNLIKELY(batch != 0) {
66 assert(batch >= 1 * sizeof(uint16_t));
67 assert(batch <= 15 * sizeof(uint16_t));
68
69 // Prepare mask for valid elements (depends on batch).
70 batch >>= XNN_LOG2_SIZEOF_HALF;
71 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
72
73 const __m512 vt = _mm512_cvtph_ps(_mm256_maskz_loadu_epi16(vmask, i));
74
75 vacc0 = _mm512_add_ps(vacc0, vt);
76 }
77 const __m256 vacc256 = _mm256_add_ps(_mm512_castps512_ps256(vacc0), _mm256_castpd_ps(_mm512_extractf64x4_pd(_mm512_castps_pd(vacc0), 1)));
78 __m128 vacc = _mm_add_ps(_mm256_castps256_ps128(vacc256), _mm256_extractf128_ps(vacc256, 1));
79 vacc = _mm_add_ps(vacc, _mm_movehl_ps(vacc, vacc));
80 vacc = _mm_add_ss(vacc, _mm_movehdup_ps(vacc));
81 vacc = _mm_mul_ss(vacc, _mm_load_ss(&params->scalar.scale));
82
83 float vout = _mm_cvtss_f32(vacc);
84 *output += vout;
85 }