1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
15 #include <xnnpack/math.h>
16
17
18
xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 2);
41 const int8_t* a0 = a;
42 int8_t* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr != 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61
62 do {
63 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
64 __m128i vacc1x0123 = vacc0x0123;
65 __m128i vacc2x0123 = vacc0x0123;
66 __m128i vacc3x0123 = vacc0x0123;
67 w = (const void*) ((const int32_t*) w + 4);
68
69 size_t k = kc;
70 while (k >= 8 * sizeof(int8_t)) {
71 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
72 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
73 a0 += 8;
74 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
75 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
76 a1 += 8;
77 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
78 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
79 a2 += 8;
80 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
81 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
82 a3 += 8;
83
84 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
85 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
86 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
87
88 vacc0x0123 = _mm_add_epi32(vacc0x0123,
89 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
90 vacc1x0123 = _mm_add_epi32(vacc1x0123,
91 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
92 vacc2x0123 = _mm_add_epi32(vacc2x0123,
93 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
94 vacc3x0123 = _mm_add_epi32(vacc3x0123,
95 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
96
97 vacc0x0123 = _mm_add_epi32(vacc0x0123,
98 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
99 vacc1x0123 = _mm_add_epi32(vacc1x0123,
100 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
101 vacc2x0123 = _mm_add_epi32(vacc2x0123,
102 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
103 vacc3x0123 = _mm_add_epi32(vacc3x0123,
104 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
105 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
106 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
107 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
108
109 vacc0x0123 = _mm_add_epi32(vacc0x0123,
110 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
111 vacc1x0123 = _mm_add_epi32(vacc1x0123,
112 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
113 vacc2x0123 = _mm_add_epi32(vacc2x0123,
114 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
115 vacc3x0123 = _mm_add_epi32(vacc3x0123,
116 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
117
118 vacc0x0123 = _mm_add_epi32(vacc0x0123,
119 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
120 vacc1x0123 = _mm_add_epi32(vacc1x0123,
121 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
122 vacc2x0123 = _mm_add_epi32(vacc2x0123,
123 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
124 vacc3x0123 = _mm_add_epi32(vacc3x0123,
125 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
126
127 w = (const void*) ((const int8_t*) w + 32);
128 k -= 8 * sizeof(int8_t);
129 }
130 if (k != 0) {
131 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
132 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
133 a0 = (const int8_t*) ((uintptr_t) a0 + k);
134 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
135 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
136 a1 = (const int8_t*) ((uintptr_t) a1 + k);
137 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
138 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
139 a2 = (const int8_t*) ((uintptr_t) a2 + k);
140 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
141 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
142 a3 = (const int8_t*) ((uintptr_t) a3 + k);
143
144 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
145 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
146 w = (const void*) ((const int8_t*) w + 8);
147
148 vacc0x0123 = _mm_add_epi32(vacc0x0123,
149 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
150 vacc1x0123 = _mm_add_epi32(vacc1x0123,
151 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
152 vacc2x0123 = _mm_add_epi32(vacc2x0123,
153 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
154 vacc3x0123 = _mm_add_epi32(vacc3x0123,
155 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
156
157 if (k > 2 * sizeof(int8_t)) {
158 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
159 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
160 w = (const void*) ((const int8_t*) w + 8);
161
162 vacc0x0123 = _mm_add_epi32(vacc0x0123,
163 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
164 vacc1x0123 = _mm_add_epi32(vacc1x0123,
165 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
166 vacc2x0123 = _mm_add_epi32(vacc2x0123,
167 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
168 vacc3x0123 = _mm_add_epi32(vacc3x0123,
169 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
170
171 if (k > 4 * sizeof(int8_t)) {
172 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
173 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
174 w = (const void*) ((const int8_t*) w + 8);
175
176 vacc0x0123 = _mm_add_epi32(vacc0x0123,
177 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
178 vacc1x0123 = _mm_add_epi32(vacc1x0123,
179 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
180 vacc2x0123 = _mm_add_epi32(vacc2x0123,
181 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
182 vacc3x0123 = _mm_add_epi32(vacc3x0123,
183 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
184 }
185 }
186 }
187
188 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
189 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
190 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
191 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
192
193 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
194 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
195 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
196 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
197 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
198
199 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
200 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
201 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
202 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
203 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
204
205 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
206 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
207 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
208 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
209
210 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
211 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
212 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
213
214
215 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
216
217 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
218
219 if (nc >= 4) {
220 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
221 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
222 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
223 *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
224
225 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
226 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
227 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
228 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
229
230 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
231 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
232 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
233 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
234
235 nc -= 4;
236 } else {
237 if (nc & 2) {
238 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
239 c0 += 2;
240 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
241 c1 += 2;
242 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
243 c2 += 2;
244 *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
245 c3 += 2;
246 vout = _mm_srli_epi32(vout, 16);
247 }
248 if (nc & 1) {
249 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
250 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
251 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
252 *c3 = (int8_t) _mm_extract_epi8(vout, 12);
253 }
254
255 nc = 0;
256 }
257 } while (nc != 0);
258 }
259