Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x16c4-minmax-avx512vnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx16c4-avx512vnni.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/gemm.h"
16 #include "xnnpack/intrinsics-polyfill.h"
17 #include "xnnpack/math.h"
18 #include "xnnpack/unaligned.h"
19 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x16c4__avx512vnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 size_t ks,
27 const int8_t** restrict a,
28 const void* restrict w,
29 float* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 size_t a_offset,
33 const int8_t* zero,
34 const int8_t* zero_data,
35 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
36 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 1);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(kc % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 kc = round_up_po2(kc, 4 * sizeof(int8_t));
48 float* c0 = c;
49
50 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
51 XNN_FORCE_REALIZATION(vsign_mask);
52 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
53 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
54 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
55 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
56 XNN_FORCE_REALIZATION(vinput_zero_point);
57 // XNN_FORCE_REALIZATION(vinput_inv_scale);
58 // XNN_FORCE_REALIZATION(voutput_min);
59 // XNN_FORCE_REALIZATION(voutput_max);
60 do {
61 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
62 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
63 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
64 w = (const int32_t*) w + 16;
65
66 size_t p = ks;
67 do {
68 const int8_t* restrict a0 = a[0];
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
71 } else {
72 a0 = zero_data;
73 }
74 a += 1;
75
76 size_t k = kc;
77 while (k >= 8 * sizeof(int8_t)) {
78 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
79 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
80 a0 += 8;
81
82 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
83 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
84 xnn_prefetch_to_l1((const int8_t*) w + 896);
85
86 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
87 xnn_prefetch_to_l1((const int8_t*) w + 960);
88 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
89
90 w = (const int8_t*) w + 128;
91 k -= 8 * sizeof(int8_t);
92 }
93
94 if (k != 0) {
95 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
96 a0 += 4;
97
98 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
99 xnn_prefetch_to_l1((const int8_t*) w + 960);
100
101 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
102
103 w = (const int8_t*) w + 64;
104 k -= 4 * sizeof(int8_t);
105 }
106
107 p -= 1 * sizeof(void*);
108 } while (p != 0);
109
110 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
111
112 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
113
114 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
115
116 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
117 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
118 w = (const float*) w + 32;
119
120 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
121
122 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
123
124 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
125
126 if XNN_LIKELY(nc >= 16) {
127 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
128 c0 = (float*) ((uintptr_t) c0 + cn_stride);
129 a = (const int8_t**restrict) ((uintptr_t) a - ks);
130 nc -= 16;
131 } else {
132 // Prepare mask for valid 32-bit elements (depends on nc).
133 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
134 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
135 nc = 0;
136 }
137 } while (nc != 0);
138 }