Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x16c4-minmax-avx512vnni.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
20
21 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x16c4__avx512vnni(
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, 4 * sizeof(int8_t));
47 float* c0 = c;
48
49 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
50 XNN_FORCE_REALIZATION(vsign_mask);
51 const __m512i vinput_zero_point = _mm512_set1_epi32((int) quantization_params->zero_point + 128);
52 const __m512 vinput_inv_scale = _mm512_set1_ps(quantization_params->inv_scale);
53 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
54 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
55 XNN_FORCE_REALIZATION(vinput_zero_point);
56 // XNN_FORCE_REALIZATION(vinput_inv_scale);
57 // XNN_FORCE_REALIZATION(voutput_min);
58 // XNN_FORCE_REALIZATION(voutput_max);
59 do {
60 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
61 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point);
62 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
63 w = (const int32_t*) w + 16;
64
65 size_t p = ks;
66 do {
67 const int8_t* restrict a0 = a[0];
68 if XNN_UNPREDICTABLE(a0 != zero) {
69 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
70 } else {
71 a0 = zero_data;
72 }
73 a += 1;
74
75 size_t k = kc;
76 while (k >= 8 * sizeof(int8_t)) {
77 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
78 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
79 a0 += 8;
80
81 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
82 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
83
84 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
85 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
86
87 w = (const int8_t*) w + 128;
88 k -= 8 * sizeof(int8_t);
89 }
90
91 if (k != 0) {
92 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
93 a0 += 4;
94
95 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
96
97 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
98
99 w = (const int8_t*) w + 64;
100 k -= 4 * sizeof(int8_t);
101 }
102
103 p -= 1 * sizeof(void*);
104 } while (p != 0);
105
106 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
107
108 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
109
110 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vinput_inv_scale);
111
112 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
113 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
114 w = (const float*) w + 32;
115
116 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
117
118 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
119
120 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
121
122 if XNN_LIKELY(nc >= 16) {
123 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
124 c0 = (float*) ((uintptr_t) c0 + cn_stride);
125 a = (const int8_t**restrict) ((uintptr_t) a - ks);
126 nc -= 16;
127 } else {
128 // Prepare mask for valid 32-bit elements (depends on nc).
129 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
130 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
131 nc = 0;
132 }
133 } while (nc != 0);
134 }