Debian Codegraph @

src/f16-f32acc-rsum/gen/f16-f32acc-rsum-avx512skx-u64-acc2.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_u64_acc2(
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 for (; batch >= 64 * sizeof(uint16_t); batch -= 64 * sizeof(uint16_t)) {
35 const __m512 vt0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) i));
36 const __m512 vt1 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + 16)));
37 const __m512 vt2 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + 32)));
38 const __m512 vt3 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) (i + 48)));
39 i += 64;
40
41 vacc0 = _mm512_add_ps(vacc0, vt0);
42 vacc1 = _mm512_add_ps(vacc1, vt1);
43 vacc0 = _mm512_add_ps(vacc0, vt2);
44 vacc1 = _mm512_add_ps(vacc1, vt3);
45 }
46 vacc0 = _mm512_add_ps(vacc0, vacc1);
47 for (; batch >= 16 * sizeof(uint16_t); batch -= 16 * sizeof(uint16_t)) {
48 const __m512 vt = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i*) i));
49 i += 16;
50
51 vacc0 = _mm512_add_ps(vacc0, vt);
52 }
53 if XNN_UNLIKELY(batch != 0) {
54 assert(batch >= 1 * sizeof(uint16_t));
55 assert(batch <= 15 * sizeof(uint16_t));
56
57 // Prepare mask for valid elements (depends on batch).
58 batch >>= XNN_LOG2_SIZEOF_HALF;
59 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
60
61 const __m512 vt = _mm512_cvtph_ps(_mm256_maskz_loadu_epi16(vmask, i));
62
63 vacc0 = _mm512_add_ps(vacc0, vt);
64 }
65 const __m256 vacc256 = _mm256_add_ps(_mm512_castps512_ps256(vacc0), _mm256_castpd_ps(_mm512_extractf64x4_pd(_mm512_castps_pd(vacc0), 1)));
66 __m128 vacc = _mm_add_ps(_mm256_castps256_ps128(vacc256), _mm256_extractf128_ps(vacc256, 1));
67 vacc = _mm_add_ps(vacc, _mm_movehl_ps(vacc, vacc));
68 vacc = _mm_add_ss(vacc, _mm_movehdup_ps(vacc));
69 vacc = _mm_mul_ss(vacc, _mm_load_ss(&params->scalar.scale));
70
71 float vout = _mm_cvtss_f32(vacc);
72 *output += vout;
73 }