Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x16c8-minmax-avx512skx-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c8-avx512skx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/igemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x16c8__avx512skx_prfm(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const int8_t* zero_data,
34 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
35 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 1);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 float* c0 = c;
48
49 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
50 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point);
51 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
52 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
53 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
54 // XNN_FORCE_REALIZATION(vinput_zero_point);
55 // XNN_FORCE_REALIZATION(vinput_inv_scale);
56 // XNN_FORCE_REALIZATION(voutput_min);
57 // XNN_FORCE_REALIZATION(voutput_max);
58 do {
59 const __m512i vksum0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
60 const __m512i vksum4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 4);
61 const __m512i vksum89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 8);
62 const __m512i vksumCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const int32_t*) w + 12);
63 __m512i vacc0x0123 = _mm512_mullo_epi32(vksum0123, vinput_zero_point);
64 __m512i vacc0x4567 = _mm512_mullo_epi32(vksum4567, vinput_zero_point);
65 __m512i vacc0x89AB = _mm512_mullo_epi32(vksum89AB, vinput_zero_point);
66 __m512i vacc0xCDEF = _mm512_mullo_epi32(vksumCDEF, vinput_zero_point);
67 w = (const int32_t*) w + 16;
68
69 size_t p = ks;
70 do {
71 const int8_t* restrict a0 = a[0];
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
74 } else {
75 a0 = zero_data;
76 }
77 a += 1;
78
79 size_t k = kc;
80
81 while (k >= 8 * sizeof(int8_t)) {
82 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
83 a0 += 8;
84
85 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
86 xnn_prefetch_to_l1((const int8_t*) w + 896);
87
88 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
89 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
90
91 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
92 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
93 xnn_prefetch_to_l1((const int8_t*) w + 960);
94
95 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
96 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
97
98 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
99
100 w = (const int8_t*) w + 128;
101 k -= 8 * sizeof(int8_t);
102 }
103 p -= 1 * sizeof(void*);
104 } while (p != 0);
105
106 // Add 4 adjacent sums
107 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
108 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
109 const __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
110 const __m512i vidx = _mm512_set_epi32(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0);
111 __m512i vacc0x0123456789ABCDEF = _mm512_permutexvar_epi32(vidx, vacc0x084C195D2A6E3B7F);
112
113 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
114
115 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
116
117 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
118 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
119 w = (const float*) w + 32;
120
121 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
122
123 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
124
125 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
126
127 if XNN_LIKELY(nc >= 16) {
128 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
129 c0 = (float*) ((uintptr_t) c0 + cn_stride);
130 a = (const int8_t**restrict) ((uintptr_t) a - ks);
131 nc -= 16;
132 } else {
133 // Prepare mask for valid 32-bit elements (depends on nc).
134 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
135 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
136 nc = 0;
137 }
138 } while (nc != 0);
139 }