Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x16c8-minmax-avx512skx-madd-prfm.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 #include "xnnpack/prefetch.h"
20
21
22 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_1x16c8__avx512skx_madd_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, 8 * 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(0x0F);
56 XNN_FORCE_REALIZATION(vmask);
57 do {
58 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
59 __m512i vsum0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
60 __m512i vacc0x01234567 = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 0));
61 __m512i vacc0x89ABCDEF = _mm512_cvtepu32_epi64(_mm512_extracti64x4_epi64(vsum0x0123456789ABCDEF, 1));
62 __m512i vacc1x0x01234567 = _mm512_setzero_epi32();
63 __m512i vacc1x0x89ABCDEF = _mm512_setzero_epi32();
64 w = (const int32_t*) w + 16;
65
66 size_t k = kc;
67 while (k >= 16 * sizeof(int8_t)) {
68 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
69 const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
70 a0 += 16;
71
72 const __m512i vbb01234567x01234567 = _mm512_load_si512(w);
73 const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
74 const __m512i vbs01234567x89ABCDEF = _mm512_srli_epi32(vbb01234567x01234567, 4);
75 const __m512i vbs89ABCDEFx89ABCDEF = _mm512_srli_epi32(vbb89ABCDEFx01234567, 4);
76 const __m512i vb01234567x01234567 = _mm512_and_si512(vbb01234567x01234567, vmask);
77 const __m512i vb89ABCDEFx01234567 = _mm512_and_si512(vbb89ABCDEFx01234567, vmask);
78 const __m512i vb01234567x89ABCDEF = _mm512_and_si512(vbs01234567x89ABCDEF, vmask);
79 const __m512i vb89ABCDEFx89ABCDEF = _mm512_and_si512(vbs89ABCDEFx89ABCDEF, vmask);
80
81 vacc0x01234567 = _mm512_dpbusd_epi32_madd(vacc0x01234567, va0x01234567, vb01234567x01234567);
82 vacc0x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
83 xnn_prefetch_to_l1((const int8_t*) w + 896);
84 xnn_prefetch_to_l1((const int8_t*) w + 960);
85 vacc1x0x01234567 = _mm512_dpbusd_epi32_madd(vacc1x0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF);
86 vacc1x0x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc1x0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF);
87
88 w = (const int8_t*) w + 128;
89 k -= 16 * sizeof(int8_t);
90 }
91
92 if (k != 0) {
93 const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask);
94 a0 += 8;
95
96 const __m512i vbb01234567x01234567 = _mm512_load_si512(w);
97 const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64);
98 const __m512i vb01234567x01234567 = _mm512_and_si512(vbb01234567x01234567, vmask);
99 const __m512i vb89ABCDEFx01234567 = _mm512_and_si512(vbb89ABCDEFx01234567, vmask);
100
101 vacc0x01234567 = _mm512_dpbusd_epi32_madd(vacc0x01234567, va0x01234567, vb01234567x01234567);
102 vacc0x89ABCDEF = _mm512_dpbusd_epi32_madd(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567);
103 xnn_prefetch_to_l1((const int8_t*) w + 896);
104 xnn_prefetch_to_l1((const int8_t*) w + 960);
105
106 w = (const int8_t*) w + 128;
107 k -= 8 * sizeof(int8_t);
108 }
109 vacc0x01234567 = _mm512_add_epi32(vacc0x01234567, vacc1x0x01234567);
110 vacc0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, vacc1x0x89ABCDEF);
111
112 // Add adjacent pairs
113 const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
114 const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32));
115 const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32));
116 __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF);
117
118 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
119
120 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
121
122 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
123 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
124 w = (const float*) w + 32;
125
126 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
127
128 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
129
130 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
131
132 if XNN_LIKELY(nc >= 16) {
133 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
134
135 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
136
137 c0 = (float*) ((uintptr_t) c0 + cn_stride);
138
139 nc -= 16;
140 } else {
141 // Prepare mask for valid 32-bit elements (depends on nc).
142 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
143 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
144 nc = 0;
145 }
146 } while (nc != 0);
147 }