Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x4c2-minmax-fp32-sse41-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_3x4c2__sse41_ld128(
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 <= 3);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (3 * 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, 2 * 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
55 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);
56 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
57 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
58 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
59 XNN_FORCE_REALIZATION(voutput_zero_point);
60 XNN_FORCE_REALIZATION(voutput_min);
61
62 do {
63 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
64 __m128i vacc1x0123 = vacc0x0123;
65 __m128i vacc2x0123 = vacc0x0123;
66 w = (const void*) ((const int32_t*) w + 4);
67
68 size_t p = ks;
69 do {
70 const int8_t* restrict a0 = a[0];
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73 }
74 const int8_t* restrict a1 = a[1];
75 if XNN_UNPREDICTABLE(a1 != zero) {
76 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77 }
78 const int8_t* restrict a2 = a[2];
79 if XNN_UNPREDICTABLE(a2 != zero) {
80 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
81 }
82 a += 3;
83
84 size_t k = kc;
85 while (k >= 8 * sizeof(int8_t)) {
86 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
87 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
88 a0 += 8;
89 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
90 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
91 a1 += 8;
92 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
93 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
94 a2 += 8;
95
96 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
97 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
98 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
99
100 vacc0x0123 = _mm_add_epi32(vacc0x0123,
101 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
102 vacc1x0123 = _mm_add_epi32(vacc1x0123,
103 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
104 vacc2x0123 = _mm_add_epi32(vacc2x0123,
105 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
106
107 vacc0x0123 = _mm_add_epi32(vacc0x0123,
108 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
109 vacc1x0123 = _mm_add_epi32(vacc1x0123,
110 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
111 vacc2x0123 = _mm_add_epi32(vacc2x0123,
112 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
113 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
114 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
115 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
116
117 vacc0x0123 = _mm_add_epi32(vacc0x0123,
118 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
119 vacc1x0123 = _mm_add_epi32(vacc1x0123,
120 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
121 vacc2x0123 = _mm_add_epi32(vacc2x0123,
122 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
123
124 vacc0x0123 = _mm_add_epi32(vacc0x0123,
125 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
126 vacc1x0123 = _mm_add_epi32(vacc1x0123,
127 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
128 vacc2x0123 = _mm_add_epi32(vacc2x0123,
129 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
130
131 w = (const void*) ((const int8_t*) w + 32);
132 k -= 8 * sizeof(int8_t);
133 }
134 if (k != 0) {
135 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
136 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
137 a0 = (const int8_t*) ((uintptr_t) a0 + k);
138 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
139 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
140 a1 = (const int8_t*) ((uintptr_t) a1 + k);
141 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
142 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
143 a2 = (const int8_t*) ((uintptr_t) a2 + k);
144
145 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
146 w = (const int8_t*) w + 8;
147 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
148
149 vacc0x0123 = _mm_add_epi32(vacc0x0123,
150 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
151 vacc1x0123 = _mm_add_epi32(vacc1x0123,
152 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
153 vacc2x0123 = _mm_add_epi32(vacc2x0123,
154 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
155
156 if (k > 2 * sizeof(int8_t)) {
157 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
158 w = (const int8_t*) w + 8;
159 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
160
161 vacc0x0123 = _mm_add_epi32(vacc0x0123,
162 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
163 vacc1x0123 = _mm_add_epi32(vacc1x0123,
164 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
165 vacc2x0123 = _mm_add_epi32(vacc2x0123,
166 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
167
168 if (k > 4 * sizeof(int8_t)) {
169 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
170 w = (const int8_t*) w + 8;
171 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
172
173 vacc0x0123 = _mm_add_epi32(vacc0x0123,
174 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
175 vacc1x0123 = _mm_add_epi32(vacc1x0123,
176 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
177 vacc2x0123 = _mm_add_epi32(vacc2x0123,
178 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
179 }
180 }
181 }
182 p -= 3 * sizeof(void*);
183 } while (p != 0);
184
185 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
186 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
187 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
188
189 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
190 w = (const float*) w + 4;
191 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
192 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
193 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
194
195 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
196 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
197 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
198
199 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
200 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
201 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
202
203 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
204 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
205
206
207 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
208
209 vout = _mm_max_epi8(vout, voutput_min);
210
211 if (nc >= 4) {
212 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
213 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
214 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
215 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
216 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
217 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
218
219 a = (const int8_t**restrict) ((uintptr_t) a - ks);
220
221 nc -= 4;
222 } else {
223 if (nc & 2) {
224 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
225 c2 += 2;
226 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
227 c1 += 2;
228 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
229 c0 += 2;
230 vout = _mm_srli_epi32(vout, 16);
231 }
232 if (nc & 1) {
233 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
234 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
235 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
236 }
237
238 nc = 0;
239 }
240 } while (nc != 0);
241 }