Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x16c4-minmax-avx512vnni.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
20
21 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_1x16c4__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 union xnn_f32_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, 4 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 float* c0 = c;
46
47 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
48 XNN_FORCE_REALIZATION(vsign_mask);
49 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
50 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
51 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
52 // XNN_FORCE_REALIZATION(voutput_min);
53 // XNN_FORCE_REALIZATION(voutput_max);
54 do {
55 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
56 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
57 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
58 w = (const int32_t*) w + 16;
59
60 size_t k = kc;
61 while (k >= 8 * sizeof(int8_t)) {
62 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
63 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
64 a0 += 8;
65
66 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
67 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
68
69 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
70 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
71
72 w = (const int8_t*) w + 128;
73 k -= 8 * sizeof(int8_t);
74 }
75 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
76
77 if (k != 0) {
78 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
79 a0 += 4;
80
81 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
82
83 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
84
85 w = (const int8_t*) w + 64;
86 k -= 4 * sizeof(int8_t);
87 }
88
89 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
90
91 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
92
93 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
94 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
95 w = (const float*) w + 32;
96
97 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
98
99 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
100
101 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
102
103 if XNN_LIKELY(nc >= 16) {
104 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
105
106 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
107
108 c0 = (float*) ((uintptr_t) c0 + cn_stride);
109
110 nc -= 16;
111 } else {
112 // Prepare mask for valid 32-bit elements (depends on nc).
113 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
114 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
115 nc = 0;
116 }
117 } while (nc != 0);
118 }