Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-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_qs8_qc8w_igemm_minmax_fp32_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 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 1);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(kc % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 4 * sizeof(int8_t));
45 int8_t* c0 = c;
46
47 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
48 XNN_FORCE_REALIZATION(vsign_mask);
49 const __m512 voutput_max_less_zero_point = _mm512_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
50 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
51 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
52 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
53 // XNN_FORCE_REALIZATION(voutput_zero_point);
54 // XNN_FORCE_REALIZATION(voutput_min);
55 do {
56 __m512i vacc0x0123456789ABCDEF = _mm512_load_epi32(w);
57 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
58 w = (const int32_t*) w + 16;
59
60 size_t p = ks;
61 do {
62 const int8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 a += 1;
67
68 size_t k = kc;
69 while (k >= 8 * sizeof(int8_t)) {
70 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
71 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
72 a0 += 8;
73
74 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
75 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
76
77 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
78 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
79
80 w = (const int8_t*) w + 128;
81 k -= 8 * sizeof(int8_t);
82 }
83
84 if (k != 0) {
85 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
86 a0 += 4;
87
88 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
89
90 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
91
92 w = (const int8_t*) w + 64;
93 k -= 4 * sizeof(int8_t);
94 }
95
96 p -= 1 * sizeof(void*);
97 } while (p != 0);
98
99 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
100
101 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
102
103 const __m512 vscale012345678ABCDEF = _mm512_load_ps(w);
104 w = (const float*) w + 16;
105 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale012345678ABCDEF);
106
107 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
108
109 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
110
111 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
112
113 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
114
115 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
116
117 if (nc >= 16) {
118 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
119 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
120 a = (const int8_t**restrict) ((uintptr_t) a - ks);
121 nc -= 16;
122 } else {
123 // Prepare mask for valid 8-bit elements (depends on nc).
124 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - UINT32_C(1));
125 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
126 nc = 0;
127 }
128 } while (nc != 0);
129 }