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 <emmintrin.h>
13
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__sse2_ld128(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)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__sse2_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const uint8_t** restrict a,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const uint8_t* zero,
31 const union xnn_qu8_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(uint8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
45 uint8_t* c0 = c;
46 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 c3 = c2;
57 }
58
59 do {
60 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
61 __m128i vacc1x0123 = vacc0x0123;
62 __m128i vacc2x0123 = vacc0x0123;
63 __m128i vacc3x0123 = vacc0x0123;
64 w = (const void*) ((const int32_t*) w + 4);
65
66 size_t p = ks;
67 do {
68 const uint8_t* restrict a0 = a[0];
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
71 }
72 const uint8_t* restrict a1 = a[1];
73 if XNN_UNPREDICTABLE(a1 != zero) {
74 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
75 }
76 const uint8_t* restrict a2 = a[2];
77 if XNN_UNPREDICTABLE(a2 != zero) {
78 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
79 }
80 const uint8_t* restrict a3 = a[3];
81 if XNN_UNPREDICTABLE(a3 != zero) {
82 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
83 }
84 a += 4;
85
86 size_t k = kc;
87 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
88 const __m128i vzero = _mm_setzero_si128();
89 while (k >= 8 * sizeof(uint8_t)) {
90 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
91 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
92 a0 += 8;
93 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
94 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
95 a1 += 8;
96 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
97 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
98 a2 += 8;
99 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
100 const __m128i vxa3 = _mm_unpacklo_epi8(va3, vzero);
101 a3 += 8;
102
103 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
104 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
105 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
106
107 vacc0x0123 = _mm_add_epi32(vacc0x0123,
108 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
109 vacc1x0123 = _mm_add_epi32(vacc1x0123,
110 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
111 vacc2x0123 = _mm_add_epi32(vacc2x0123,
112 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
113 vacc3x0123 = _mm_add_epi32(vacc3x0123,
114 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
115
116 vacc0x0123 = _mm_add_epi32(vacc0x0123,
117 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
118 vacc1x0123 = _mm_add_epi32(vacc1x0123,
119 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
120 vacc2x0123 = _mm_add_epi32(vacc2x0123,
121 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
122 vacc3x0123 = _mm_add_epi32(vacc3x0123,
123 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
124 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
125 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
126 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
127
128 vacc0x0123 = _mm_add_epi32(vacc0x0123,
129 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
130 vacc1x0123 = _mm_add_epi32(vacc1x0123,
131 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
132 vacc2x0123 = _mm_add_epi32(vacc2x0123,
133 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
134 vacc3x0123 = _mm_add_epi32(vacc3x0123,
135 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
136
137 vacc0x0123 = _mm_add_epi32(vacc0x0123,
138 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
139 vacc1x0123 = _mm_add_epi32(vacc1x0123,
140 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
141 vacc2x0123 = _mm_add_epi32(vacc2x0123,
142 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
143 vacc3x0123 = _mm_add_epi32(vacc3x0123,
144 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
145
146 w = (const void*) ((const uint8_t*) w + 32);
147 k -= 8 * sizeof(uint8_t);
148 }
149 if (k != 0) {
150 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
151 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
152 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
153 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
154 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
155 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
156 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
157 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
158 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
159 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
160 const __m128i vxa3 = _mm_unpacklo_epi8(va3, vzero);
161 a3 = (const uint8_t*) ((uintptr_t) a3 + k);
162
163 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
164 w = (const void*) ((const uint8_t*) w + 8);
165 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
166
167 vacc0x0123 = _mm_add_epi32(vacc0x0123,
168 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
169 vacc1x0123 = _mm_add_epi32(vacc1x0123,
170 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
171 vacc2x0123 = _mm_add_epi32(vacc2x0123,
172 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
173 vacc3x0123 = _mm_add_epi32(vacc3x0123,
174 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
175
176 if (k > 2 * sizeof(uint8_t)) {
177 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
178 w = (const void*) ((const uint8_t*) w + 8);
179 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
180
181 vacc0x0123 = _mm_add_epi32(vacc0x0123,
182 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
183 vacc1x0123 = _mm_add_epi32(vacc1x0123,
184 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
185 vacc2x0123 = _mm_add_epi32(vacc2x0123,
186 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
187 vacc3x0123 = _mm_add_epi32(vacc3x0123,
188 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
189
190 if (k > 4 * sizeof(uint8_t)) {
191 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
192 w = (const void*) ((const uint8_t*) w + 8);
193 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
194
195 vacc0x0123 = _mm_add_epi32(vacc0x0123,
196 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
197 vacc1x0123 = _mm_add_epi32(vacc1x0123,
198 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
199 vacc2x0123 = _mm_add_epi32(vacc2x0123,
200 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
201 vacc3x0123 = _mm_add_epi32(vacc3x0123,
202 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
203 }
204 }
205 }
206 p -= 4 * sizeof(void*);
207 } while (p != 0);
208
209 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
210 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
211 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
212 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
213
214 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
215 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
216 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
217 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
218 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
219
220 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
221 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
222 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
223 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
224 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
225
226 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
227 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
228 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
229 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
230
231 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
232 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
233 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
234
235 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);
236
237 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
238
239 if (nc >= 4) {
240 unaligned_store_u32(c3, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(3, 3, 3, 3))));
241 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
242 unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(2, 2, 2, 2))));
243 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
244 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1))));
245 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
246 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
247 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
248
249 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
250
251 nc -= 4;
252 } else {
253 if (nc & 2) {
254 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
255 c3 += 2;
256 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
257 c2 += 2;
258 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
259 c1 += 2;
260 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
261 c0 += 2;
262 vout = _mm_srli_epi32(vout, 16);
263 }
264 if (nc & 1) {
265 *c3 = (uint8_t) _mm_extract_epi16(vout, 6);
266 *c2 = (uint8_t) _mm_extract_epi16(vout, 4);
267 *c1 = (uint8_t) _mm_extract_epi16(vout, 2);
268 *c0 = (uint8_t) _mm_cvtsi128_si32(vout);
269 }
270
271 nc = 0;
272 }
273 } while (nc != 0);
274 }
275