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_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const uint8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qu8_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(uint8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 2);
41 const uint8_t* a0 = a;
42 uint8_t* c0 = c;
43 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
44 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
50 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
56 uint8_t* c3 = (uint8_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 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
71 const __m128i vzero = _mm_setzero_si128();
72 while (k >= 8 * sizeof(uint8_t)) {
73 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
74 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
75 a0 += 8;
76 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
77 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
78 a1 += 8;
79 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
80 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
81 a2 += 8;
82 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
83 const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
84 a3 += 8;
85
86 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
87 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
88 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
89
90 vacc0x0123 = _mm_add_epi32(vacc0x0123,
91 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
92 vacc1x0123 = _mm_add_epi32(vacc1x0123,
93 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
94 vacc2x0123 = _mm_add_epi32(vacc2x0123,
95 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
96 vacc3x0123 = _mm_add_epi32(vacc3x0123,
97 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
98
99 vacc0x0123 = _mm_add_epi32(vacc0x0123,
100 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
101 vacc1x0123 = _mm_add_epi32(vacc1x0123,
102 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
103 vacc2x0123 = _mm_add_epi32(vacc2x0123,
104 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
105 vacc3x0123 = _mm_add_epi32(vacc3x0123,
106 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
107 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
108 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
109 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
110
111 vacc0x0123 = _mm_add_epi32(vacc0x0123,
112 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
113 vacc1x0123 = _mm_add_epi32(vacc1x0123,
114 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
115 vacc2x0123 = _mm_add_epi32(vacc2x0123,
116 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
117 vacc3x0123 = _mm_add_epi32(vacc3x0123,
118 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
119
120 vacc0x0123 = _mm_add_epi32(vacc0x0123,
121 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
122 vacc1x0123 = _mm_add_epi32(vacc1x0123,
123 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
124 vacc2x0123 = _mm_add_epi32(vacc2x0123,
125 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
126 vacc3x0123 = _mm_add_epi32(vacc3x0123,
127 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
128
129 w = (const void*) ((const uint8_t*) w + 32);
130 k -= 8 * sizeof(uint8_t);
131 }
132 if (k != 0) {
133 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
134 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
135 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
136 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
137 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
138 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
139 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
140 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
141 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
142 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
143 const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
144 a3 = (const uint8_t*) ((uintptr_t) a3 + k);
145
146 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
147 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
148 w = (const void*) ((const uint8_t*) w + 8);
149
150 vacc0x0123 = _mm_add_epi32(vacc0x0123,
151 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
152 vacc1x0123 = _mm_add_epi32(vacc1x0123,
153 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
154 vacc2x0123 = _mm_add_epi32(vacc2x0123,
155 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
156 vacc3x0123 = _mm_add_epi32(vacc3x0123,
157 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
158
159 if (k > 2 * sizeof(uint8_t)) {
160 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
161 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
162 w = (const void*) ((const uint8_t*) w + 8);
163
164 vacc0x0123 = _mm_add_epi32(vacc0x0123,
165 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
166 vacc1x0123 = _mm_add_epi32(vacc1x0123,
167 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
168 vacc2x0123 = _mm_add_epi32(vacc2x0123,
169 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
170 vacc3x0123 = _mm_add_epi32(vacc3x0123,
171 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
172
173 if (k > 4 * sizeof(uint8_t)) {
174 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
175 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
176 w = (const void*) ((const uint8_t*) w + 8);
177
178 vacc0x0123 = _mm_add_epi32(vacc0x0123,
179 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
180 vacc1x0123 = _mm_add_epi32(vacc1x0123,
181 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
182 vacc2x0123 = _mm_add_epi32(vacc2x0123,
183 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
184 vacc3x0123 = _mm_add_epi32(vacc3x0123,
185 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
186 }
187 }
188 }
189
190 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
191 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
192 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
193 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
194
195 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
196 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
197 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
198 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
199 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
200
201 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
202 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
203 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
204 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
205 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
206
207 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
208 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
209 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
210 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
211
212 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
213 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
214 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
215
216 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);
217
218 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
219
220 if (nc >= 4) {
221 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
222 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
223 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
224 *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
225
226 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
227 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
228 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
229 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
230
231 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
232 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
233 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
234 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
235
236 nc -= 4;
237 } else {
238 if (nc & 2) {
239 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
240 c0 += 2;
241 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
242 c1 += 2;
243 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
244 c2 += 2;
245 *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
246 c3 += 2;
247 vout = _mm_srli_epi32(vout, 16);
248 }
249 if (nc & 1) {
250 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
251 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
252 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
253 *c3 = (uint8_t) _mm_extract_epi8(vout, 12);
254 }
255
256 nc = 0;
257 }
258 } while (nc != 0);
259 }
260