Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-4x16c4-minmax-avx512vnnigfni.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_qc4w_gemm_minmax_ukernel_4x16c4__avx512vnnigfni(
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_qc4w_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 <= 4);
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 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr != 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64
65 const __m512i vsign_mask = _mm512_set1_epi8(0x80);
66 XNN_FORCE_REALIZATION(vsign_mask);
67 const __m512i vinput_zero_point0 = _mm512_set1_epi32((int) quantization_params[0].zero_point + 128);
68 const __m512i vinput_zero_point1 = _mm512_set1_epi32((int) quantization_params[1].zero_point + 128);
69 const __m512i vinput_zero_point2 = _mm512_set1_epi32((int) quantization_params[2].zero_point + 128);
70 const __m512i vinput_zero_point3 = _mm512_set1_epi32((int) quantization_params[3].zero_point + 128);
71 const __m512 voutput_min = _mm512_set1_ps(params->scalar.min);
72 const __m512 voutput_max = _mm512_set1_ps(params->scalar.max);
73 // XNN_FORCE_REALIZATION(voutput_min);
74 // XNN_FORCE_REALIZATION(voutput_max);
75 const __m512i vmask = _mm512_set1_epi8(0xF0);
76 XNN_FORCE_REALIZATION(vmask);
77 const __m512i vshl4 = _mm512_set1_epi64(0x01020408);
78 XNN_FORCE_REALIZATION(vshl4);
79 do {
80 const __m512i vksum0123456789ABCDEF = _mm512_load_epi32(w);
81 __m512i vacc0x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point0);
82 __m512i vacc1x0x0123456789ABCDEF = _mm512_setzero_epi32();
83 __m512i vacc1x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point1);
84 __m512i vacc1x1x0123456789ABCDEF = _mm512_setzero_epi32();
85 __m512i vacc2x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point2);
86 __m512i vacc1x2x0123456789ABCDEF = _mm512_setzero_epi32();
87 __m512i vacc3x0123456789ABCDEF = _mm512_mullo_epi32(vksum0123456789ABCDEF, vinput_zero_point3);
88 __m512i vacc1x3x0123456789ABCDEF = _mm512_setzero_epi32();
89 w = (const int32_t*) w + 16;
90
91 size_t k = kc;
92 while (k >= 8 * sizeof(int8_t)) {
93 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
94 const __m512i va0x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0 + 4)), vsign_mask);
95 a0 += 8;
96 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
97 const __m512i va1x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1 + 4)), vsign_mask);
98 a1 += 8;
99 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
100 const __m512i va2x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2 + 4)), vsign_mask);
101 a2 += 8;
102 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
103 const __m512i va3x4567 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3 + 4)), vsign_mask);
104 a3 += 8;
105
106 const __m512i vbb0123456789ABCDEFx01234567 = _mm512_load_si512(w);
107 const __m512i vb0123456789ABCDEFx0123 = _mm512_gf2p8affine_epi64_epi8(vbb0123456789ABCDEFx01234567, vshl4, 0);
108 const __m512i vb0123456789ABCDEFx4567 = _mm512_and_si512(vbb0123456789ABCDEFx01234567, vmask);
109
110 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEFx0123);
111 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEFx0123);
112 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEFx0123);
113 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEFx0123);
114 vacc1x0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0x0123456789ABCDEF, va0x4567, vb0123456789ABCDEFx4567);
115 vacc1x1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x1x0123456789ABCDEF, va1x4567, vb0123456789ABCDEFx4567);
116 vacc1x2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x2x0123456789ABCDEF, va2x4567, vb0123456789ABCDEFx4567);
117 vacc1x3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x3x0123456789ABCDEF, va3x4567, vb0123456789ABCDEFx4567);
118
119 w = (const int8_t*) w + 64;
120 k -= 8 * sizeof(int8_t);
121 }
122 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, vacc1x0x0123456789ABCDEF);
123 vacc1x0123456789ABCDEF = _mm512_add_epi32(vacc1x0123456789ABCDEF, vacc1x1x0123456789ABCDEF);
124 vacc2x0123456789ABCDEF = _mm512_add_epi32(vacc2x0123456789ABCDEF, vacc1x2x0123456789ABCDEF);
125 vacc3x0123456789ABCDEF = _mm512_add_epi32(vacc3x0123456789ABCDEF, vacc1x3x0123456789ABCDEF);
126
127 if (k != 0) {
128 const __m512i va0x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a0)), vsign_mask);
129 a0 += 4;
130 const __m512i va1x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a1)), vsign_mask);
131 a1 += 4;
132 const __m512i va2x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a2)), vsign_mask);
133 a2 += 4;
134 const __m512i va3x0123 = _mm512_xor_epi32(_mm512_set1_epi32((int) unaligned_load_u32(a3)), vsign_mask);
135 a3 += 4;
136
137 const __m512i vbb0123456789ABCDEF = _mm512_load_si512(w);
138 const __m512i vb0123456789ABCDEF = _mm512_gf2p8affine_epi64_epi8(vbb0123456789ABCDEF, vshl4, 0);
139
140 vacc0x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc0x0123456789ABCDEF, va0x0123, vb0123456789ABCDEF);
141 vacc1x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc1x0123456789ABCDEF, va1x0123, vb0123456789ABCDEF);
142 vacc2x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc2x0123456789ABCDEF, va2x0123, vb0123456789ABCDEF);
143 vacc3x0123456789ABCDEF = _mm512_dpbusd_epi32(vacc3x0123456789ABCDEF, va3x0123, vb0123456789ABCDEF);
144
145 w = (const int8_t*) w + 64;
146 k -= 4 * sizeof(int8_t);
147 }
148
149 vacc0x0123456789ABCDEF = _mm512_srai_epi32(vacc0x0123456789ABCDEF, 4);
150 vacc1x0123456789ABCDEF = _mm512_srai_epi32(vacc1x0123456789ABCDEF, 4);
151 vacc2x0123456789ABCDEF = _mm512_srai_epi32(vacc2x0123456789ABCDEF, 4);
152 vacc3x0123456789ABCDEF = _mm512_srai_epi32(vacc3x0123456789ABCDEF, 4);
153 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
154 __m512 vscaled1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF);
155 __m512 vscaled2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF);
156 __m512 vscaled3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF);
157
158 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale));
159 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale));
160 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale));
161 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale));
162
163 const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_load_ps((const float*) w);
164 const __m512 vbias0123456789ABCDEF = _mm512_load_ps((const float*) w + 16);
165 w = (const float*) w + 32;
166
167 vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vscaled0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
168 vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vscaled1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
169 vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vscaled2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
170 vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vscaled3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vbias0123456789ABCDEF);
171
172 vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min);
173 vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min);
174 vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min);
175 vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min);
176
177 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max);
178 vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max);
179 vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max);
180 vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max);
181
182 if XNN_LIKELY(nc >= 16) {
183 _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF);
184 _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF);
185 _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF);
186 _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF);
187
188 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
189 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
190 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
191 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
192
193 c0 = (float*) ((uintptr_t) c0 + cn_stride);
194 c1 = (float*) ((uintptr_t) c1 + cn_stride);
195 c2 = (float*) ((uintptr_t) c2 + cn_stride);
196 c3 = (float*) ((uintptr_t) c3 + cn_stride);
197
198 nc -= 16;
199 } else {
200 // Prepare mask for valid 32-bit elements (depends on nc).
201 const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1);
202 _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF);
203 _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF);
204 _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF);
205 _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF);
206 nc = 0;
207 }
208 } while (nc != 0);
209 }