Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x16c4-minmax-avx512vnni-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_qc4w_gemm_minmax_ukernel_1x16c4__avx512vnni_prfm(
23 size_t mr,
24 size_t nc,
25 size_t kc,
26 const int8_t* restrict a,
27 size_t a_stride,
28 const void* restrict w,
29 float* restrict c,
30 size_t cm_stride,
31 size_t cn_stride,
32 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_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 const int8_t* a0 = a;
46 float* c0 = c;
47
48 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
49 XNN_FORCE_REALIZATION(vsign_mask);
50 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
51 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
52 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
53 // XNN_FORCE_REALIZATION(voutput_min);
54 // XNN_FORCE_REALIZATION(voutput_max);
55 const __m512i vmask = _mm512_set1_epi8(0xF0);
56 XNN_FORCE_REALIZATION(vmask);
57 do {
58 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
59 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
60 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
61 w = (const int32_t*) w + 16;
62
63 size_t k = kc;
64 while (k >= 8 * sizeof(int8_t)) {
65 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
66 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
67 a0 += 8;
68
69 const __m512i vbb0123456789ABCDEFx01234567 = _mm512_load_si512(w);
70 const __m512i vbs0123456789ABCDEFx0123 = _mm512_slli_epi32(vbb0123456789ABCDEFx01234567, 4);
71 const __m512i vb0123456789ABCDEFx4567 = _mm512_and_si512(vbb0123456789ABCDEFx01234567, vmask);
72 const __m512i vb0123456789ABCDEFx0123 = _mm512_and_si512(vbs0123456789ABCDEFx0123, vmask);
73
74 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
75 xnn_prefetch_to_l1((const int8_t*) w + 960);
76 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
77
78 w = (const int8_t*) w + 64;
79 k -= 8 * sizeof(int8_t);
80 }
81 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
82
83 if (k != 0) {
84 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
85 a0 += 4;
86
87 const __m512i vbb0123456789ABCDEF = _mm512_load_si512(w);
88 const __m512i vb0123456789ABCDEF = _mm512_slli_epi32(vbb0123456789ABCDEF, 4);
89 xnn_prefetch_to_l1((const int8_t*) w + 960);
90
91 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
92
93 w = (const int8_t*) w + 64;
94 k -= 4 * sizeof(int8_t);
95 }
96
97 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
98 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
99
100 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
101
102 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
103 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
104 w = (const float*) w + 32;
105
106 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
107
108 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
109
110 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
111
112 if XNN_LIKELY(nc >= 16) {
113 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
114
115 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
116
117 c0 = (float*) ((uintptr_t) c0 + cn_stride);
118
119 nc -= 16;
120 } else {
121 // Prepare mask for valid 32-bit elements (depends on nc).
122 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
123 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
124 nc = 0;
125 }
126 } while (nc != 0);
127 }