Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-4x4c2s4-minmax-fp32-avx-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2s4-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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 <smmintrin.h>
13
14 #include "xnnpack/igemm.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_4x4c2s4__avx_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 4);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (4 * sizeof(void*)) == 0);
39 assert(a_offset % 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 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 c3 = c2;
57 }
58
59 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
60 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
61 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
62 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
63 XNN_FORCE_REALIZATION(voutput_zero_point);
64 XNN_FORCE_REALIZATION(voutput_min);
65
66
67 do {
68 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
69 __m128i vacc1x0123 = vacc0x0123;
70 __m128i vacc2x0123 = vacc0x0123;
71 __m128i vacc3x0123 = vacc0x0123;
72 w = (const void*) ((const int32_t*) w + 4);
73
74 size_t p = ks;
75 do {
76 const int8_t* restrict a0 = a[0];
77 if XNN_UNPREDICTABLE(a0 != zero) {
78 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
79 }
80 const int8_t* restrict a1 = a[1];
81 if XNN_UNPREDICTABLE(a1 != zero) {
82 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
83 }
84 const int8_t* restrict a2 = a[2];
85 if XNN_UNPREDICTABLE(a2 != zero) {
86 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
87 }
88 const int8_t* restrict a3 = a[3];
89 if XNN_UNPREDICTABLE(a3 != zero) {
90 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
91 }
92 a += 4;
93
94 size_t k = kc;
95 do {
96 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
97 __m128i vxa0 = _mm_cvtepi8_epi16(va0);
98 a0 += 8;
99 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
100 __m128i vxa1 = _mm_cvtepi8_epi16(va1);
101 a1 += 8;
102 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
103 __m128i vxa2 = _mm_cvtepi8_epi16(va2);
104 a2 += 8;
105 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
106 __m128i vxa3 = _mm_cvtepi8_epi16(va3);
107 a3 += 8;
108
109 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
110 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
111
112 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb0));
113 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
114 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb0));
115 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
116 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb0));
117 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
118 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb0));
119 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
120 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
121 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
122
123 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb1));
124 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
125 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb1));
126 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
127 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb1));
128 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
129 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb1));
130 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
131 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
132 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
133
134 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb2));
135 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
136 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb2));
137 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
138 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb2));
139 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
140 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb2));
141 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
142 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
143 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
144
145 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb3));
146 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb3));
147 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb3));
148 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb3));
149
150 w = (const void*) ((const int8_t*) w + 32);
151 k -= 8 * sizeof(int8_t);
152 } while (k != 0);
153 p -= 4 * sizeof(void*);
154 } while (p != 0);
155
156 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
157 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
158 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
159 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
160
161 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
162 w = (const float*) w + 4;
163 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
164 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
165 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
166 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
167
168 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
169 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
170 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
171 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
172
173 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
174 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
175 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
176 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
177
178 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
179 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
180
181
182 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
183
184 vout = _mm_max_epi8(vout, voutput_min);
185
186 if (nc >= 4) {
187 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3));
188 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
189 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
190 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
191 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
192 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
193 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
194 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
195
196 a = (const int8_t**restrict) ((uintptr_t) a - ks);
197
198 nc -= 4;
199 } else {
200 if (nc & 2) {
201 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
202 c3 += 2;
203 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
204 c2 += 2;
205 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
206 c1 += 2;
207 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
208 c0 += 2;
209 vout = _mm_srli_epi32(vout, 16);
210 }
211 if (nc & 1) {
212 *c3 = (int8_t) _mm_extract_epi8(vout, 12);
213 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
214 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
215 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
216 }
217
218 nc = 0;
219 }
220 } while (nc != 0);
221 }