Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x16c4-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_qc8w_gemm_minmax_ukernel_4x16c4__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 union xnn_f32_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 <= 4);
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 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 a2 = a1;
57 c2 = c1;
58 }
59 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
61 if XNN_UNPREDICTABLE(mr != 4) {
62 a3 = a2;
63 c3 = c2;
64 }
65
66 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
67 XNN_FORCE_REALIZATION(vsign_mask);
68 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
69 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
70 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
71 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
72 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
73 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
74 // XNN_FORCE_REALIZATION(voutput_min);
75 // XNN_FORCE_REALIZATION(voutput_max);
76 do {
77 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
78 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
79 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
80 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
81 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
82 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
83 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
84 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
85 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
86 w = (const int32_t*) w + 16;
87
88 size_t k = kc;
89 while (k >= 8 * sizeof(int8_t)) {
90 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
91 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
92 a0 += 8;
93 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
94 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
95 a1 += 8;
96 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
97 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
98 a2 += 8;
99 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
100 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
101 a3 += 8;
102
103 const __m512i vb0123456789ABCDEFx0123 = _mm512_load_si512(w);
104 const __m512i vb0123456789ABCDEFx4567 = _mm512_load_si512((const int8_t*) w + 64);
105 xnn_prefetch_to_l1((const int8_t*) w + 896);
106
107 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
108 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
109 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
110 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
111 xnn_prefetch_to_l1((const int8_t*) w + 960);
112 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
113 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
114 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
115 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
116
117 w = (const int8_t*) w + 128;
118 k -= 8 * sizeof(int8_t);
119 }
120 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
121 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
122 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
123 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
124
125 if (k != 0) {
126 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
127 a0 += 4;
128 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
129 a1 += 4;
130 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
131 a2 += 4;
132 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
133 a3 += 4;
134
135 const __m512i vb0123456789ABCDEF = _mm512_load_si512(w);
136 xnn_prefetch_to_l1((const int8_t*) w + 960);
137
138 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
139 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
140 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
141 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
142
143 w = (const int8_t*) w + 64;
144 k -= 4 * sizeof(int8_t);
145 }
146
147 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
148 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
149 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
150 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
151
152 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
153 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
154 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
155 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
156
157 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
158 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
159 w = (const float*) w + 32;
160
161 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
162 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
163 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
164 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
165
166 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
167 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
168 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
169 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
170
171 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
172 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
173 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
174 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
175
176 if XNN_LIKELY(nc >= 16) {
177 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
178 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
179 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
180 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
181
182 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
183 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
184 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
185 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
186
187 c0 = (float*) ((uintptr_t) c0 + cn_stride);
188 c1 = (float*) ((uintptr_t) c1 + cn_stride);
189 c2 = (float*) ((uintptr_t) c2 + cn_stride);
190 c3 = (float*) ((uintptr_t) c3 + cn_stride);
191
192 nc -= 16;
193 } else {
194 // Prepare mask for valid 32-bit elements (depends on nc).
195 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
196 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
197 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
198 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
199 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
200 nc = 0;
201 }
202 } while (nc != 0);
203 }