Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x16c8-minmax-avx512vnni.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c8-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_qb4w_gemm_minmax_ukernel_1x16c8__avx512vnni(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f32_qb4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 float* c0 = c;
46 size_t bl = params->scalar.blocksize;
47 assert(bl != 0);
48 assert(bl <= kc);
49 assert(kc % bl == 0);
50 assert(bl % 32 == 0);
51
52 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
53 XNN_FORCE_REALIZATION(vsign_mask);
54 const __m512 vinput_zero_point0 = _mm512_set1_ps((float) quantization_params[0].zero_point + 128);
55 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
56 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
57 // XNN_FORCE_REALIZATION(voutput_min);
58 // XNN_FORCE_REALIZATION(voutput_max);
59 const __m512i vmask = _mm512_set1_epi8(0xF0);
60 XNN_FORCE_REALIZATION(vmask);
61 do {
62 const __m512 vksum0123456789ABCDEF = _mm512_loadu_ps(w);
63 __m512 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point0);
64 w = (const int32_t*) w + 16;
65
66 for (size_t kb=0; kb < kc; kb+=bl) {
67 __m512i vacc0x01234567 = _mm512_setzero_epi32();
68 __m512i vacc0x89ABCDEF = _mm512_setzero_epi32();
69 __m512i vacc1x0x01234567 = _mm512_setzero_epi32();
70 __m512i vacc1x0x89ABCDEF = _mm512_setzero_epi32();
71 size_t k = bl;
72 while (k >= 16 * sizeof(int8_t)) {
73 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
74 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
75 a0 += 16;
76
77 const __m512i vbb01234567x01234567 = _mm512_loadu_si512(w);
78 const __m512i vbb89ABCDEFx01234567 = _mm512_loadu_si512((const int8_t*) w + 64);
79 const __m512i vbs01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4);
80 const __m512i vbs89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4);
81 const __m512i vb01234567x89ABCDEF = _mm512_and_si512(vbb01234567x01234567, vmask);
82 const __m512i vb89ABCDEFx89ABCDEF = _mm512_and_si512(vbb89ABCDEFx01234567, vmask);
83 const __m512i vb01234567x01234567 = _mm512_and_si512(vbs01234567x01234567, vmask);
84 const __m512i vb89ABCDEFx01234567 = _mm512_and_si512(vbs89ABCDEFx01234567, vmask);
85
86 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
87 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
88 vacc1x0x01234567 = _mm512_dpbusd_epi32(vacc1x0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
89 vacc1x0x89ABCDEF = _mm512_dpbusd_epi32(vacc1x0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
90
91 w = (const int8_t*) w + 128;
92 k -= 16 * sizeof(int8_t);
93 }
94
95 if (k != 0) {
96 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
97 a0 += 8;
98
99 const __m512i vbb01234567x01234567 = _mm512_load_si512(w);
100 const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
101 const __m512i vb01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4);
102 const __m512i vb89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4);
103
104 vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567);
105 vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
106
107 w = (const int8_t*) w + 128;
108 k -= 8 * sizeof(int8_t);
109 }
110 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_castsi512_ps(_mm512_slli_epi32(
111 _mm512_cvtepu16_epi32(_mm256_load_si256((const __m256i*) w)), 16));
112 w = (const uint16_t*) w + 16;
113 vacc0x01234567 = _mm512_add_epi32(vacc0x01234567, vacc1x0x01234567);
114 vacc0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, vacc1x0x89ABCDEF);
115
116 // Add adjacent pairs
117 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
118 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
119 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
120 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
121 __m512 vf0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
122
123 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vf0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled0x0123456789ABCDEF);
124 }
125
126 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
127
128 const __m512 vbias0123456789ABCDEF = _mm512_loadu_ps((const float*) w);
129 w = (const float*) w + 16;
130
131 vscaled0x0123456789ABCDEF = _mm512_add_ps(vscaled0x0123456789ABCDEF, vbias0123456789ABCDEF);
132
133 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
134
135 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
136
137 if XNN_LIKELY(nc >= 16) {
138 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
139
140 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
141
142 c0 = (float*) ((uintptr_t) c0 + cn_stride);
143
144 nc -= 16;
145 } else {
146 // Prepare mask for valid 32-bit elements (depends on nc).
147 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
148 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
149 nc = 0;
150 }
151 } while (nc != 0);
152 }