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
17
xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__sse41_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__sse41_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t** restrict a,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(uint8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 2);
44 uint8_t* c0 = c;
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 do {
55 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
56 __m128i vacc1x0123 = vacc0x0123;
57 __m128i vacc2x0123 = vacc0x0123;
58 w = (const void*) ((const int32_t*) w + 4);
59
60 size_t p = ks;
61 do {
62 const uint8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 const uint8_t* restrict a1 = a[1];
67 if XNN_UNPREDICTABLE(a1 != zero) {
68 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
69 }
70 const uint8_t* restrict a2 = a[2];
71 if XNN_UNPREDICTABLE(a2 != zero) {
72 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
73 }
74 a += 3;
75
76 size_t k = kc;
77 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
78 while (k >= 8 * sizeof(uint8_t)) {
79 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
80 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
81 a0 += 8;
82 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
83 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
84 a1 += 8;
85 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
86 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
87 a2 += 8;
88
89 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
90 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
91
92 vacc0x0123 = _mm_add_epi32(vacc0x0123,
93 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
94 vacc1x0123 = _mm_add_epi32(vacc1x0123,
95 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
96 vacc2x0123 = _mm_add_epi32(vacc2x0123,
97 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
98 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
99 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
100
101 vacc0x0123 = _mm_add_epi32(vacc0x0123,
102 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
103 vacc1x0123 = _mm_add_epi32(vacc1x0123,
104 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
105 vacc2x0123 = _mm_add_epi32(vacc2x0123,
106 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
107 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
108 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
109
110 vacc0x0123 = _mm_add_epi32(vacc0x0123,
111 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
112 vacc1x0123 = _mm_add_epi32(vacc1x0123,
113 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
114 vacc2x0123 = _mm_add_epi32(vacc2x0123,
115 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
116 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
117 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
118
119 vacc0x0123 = _mm_add_epi32(vacc0x0123,
120 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
121 vacc1x0123 = _mm_add_epi32(vacc1x0123,
122 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
123 vacc2x0123 = _mm_add_epi32(vacc2x0123,
124 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
125
126 w = (const void*) ((const uint8_t*) w + 32);
127 k -= 8 * sizeof(uint8_t);
128 }
129 if (k != 0) {
130 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
131 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
132 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
133 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
134 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
135 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
136 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
137 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
138 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
139
140 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
141 w = (const void*) ((const uint8_t*) w + 8);
142 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
143
144 vacc0x0123 = _mm_add_epi32(vacc0x0123,
145 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
146 vacc1x0123 = _mm_add_epi32(vacc1x0123,
147 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
148 vacc2x0123 = _mm_add_epi32(vacc2x0123,
149 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
150
151 if (k > 2 * sizeof(uint8_t)) {
152 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
153 w = (const void*) ((const uint8_t*) w + 8);
154 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
155
156 vacc0x0123 = _mm_add_epi32(vacc0x0123,
157 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
158 vacc1x0123 = _mm_add_epi32(vacc1x0123,
159 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
160 vacc2x0123 = _mm_add_epi32(vacc2x0123,
161 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
162
163 if (k > 4 * sizeof(uint8_t)) {
164 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
165 w = (const void*) ((const uint8_t*) w + 8);
166 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
167
168 vacc0x0123 = _mm_add_epi32(vacc0x0123,
169 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
170 vacc1x0123 = _mm_add_epi32(vacc1x0123,
171 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
172 vacc2x0123 = _mm_add_epi32(vacc2x0123,
173 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
174 }
175 }
176 }
177 p -= 3 * sizeof(void*);
178 } while (p != 0);
179
180 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
181 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
182 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
183
184 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
185 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
186 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
187 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
188
189 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
190 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
191 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
192 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
193
194 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
195 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
196 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
197
198 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
199 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
200 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
201
202 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
203
204 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
205
206 if (nc >= 4) {
207 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
208 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
209 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
210 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
211 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
212 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
213
214 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
215
216 nc -= 4;
217 } else {
218 if (nc & 2) {
219 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
220 c2 += 2;
221 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
222 c1 += 2;
223 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
224 c0 += 2;
225 vout = _mm_srli_epi32(vout, 16);
226 }
227 if (nc & 1) {
228 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
229 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
230 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
231 }
232
233 nc = 0;
234 }
235 } while (nc != 0);
236 }
237