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 <emmintrin.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
19
xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse2_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)])20 void xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld128(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const uint8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 uint8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
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 <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(uint8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
42 const uint8_t* a0 = a;
43 uint8_t* c0 = c;
44 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
51 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 a2 = a1;
54 c2 = c1;
55 }
56 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
57 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr != 4) {
59 a3 = a2;
60 c3 = c2;
61 }
62
63 do {
64 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
65 __m128i vacc1x0123 = vacc0x0123;
66 __m128i vacc2x0123 = vacc0x0123;
67 __m128i vacc3x0123 = vacc0x0123;
68 w = (const void*) ((const int32_t*) w + 4);
69
70 size_t k = kc;
71 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
72 const __m128i vzero = _mm_setzero_si128();
73 while (k >= 8 * sizeof(uint8_t)) {
74 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
75 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
76 a0 += 8;
77 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
78 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
79 a1 += 8;
80 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
81 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
82 a2 += 8;
83 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
84 const __m128i vxa3 = _mm_unpacklo_epi8(va3, vzero);
85 a3 += 8;
86
87 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
88 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
89 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
90
91 vacc0x0123 = _mm_add_epi32(vacc0x0123,
92 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
93 vacc1x0123 = _mm_add_epi32(vacc1x0123,
94 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
95 vacc2x0123 = _mm_add_epi32(vacc2x0123,
96 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
97 vacc3x0123 = _mm_add_epi32(vacc3x0123,
98 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
99
100 vacc0x0123 = _mm_add_epi32(vacc0x0123,
101 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
102 vacc1x0123 = _mm_add_epi32(vacc1x0123,
103 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
104 vacc2x0123 = _mm_add_epi32(vacc2x0123,
105 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
106 vacc3x0123 = _mm_add_epi32(vacc3x0123,
107 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
108 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
109 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
110 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
111
112 vacc0x0123 = _mm_add_epi32(vacc0x0123,
113 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
114 vacc1x0123 = _mm_add_epi32(vacc1x0123,
115 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
116 vacc2x0123 = _mm_add_epi32(vacc2x0123,
117 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
118 vacc3x0123 = _mm_add_epi32(vacc3x0123,
119 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
120
121 vacc0x0123 = _mm_add_epi32(vacc0x0123,
122 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
123 vacc1x0123 = _mm_add_epi32(vacc1x0123,
124 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
125 vacc2x0123 = _mm_add_epi32(vacc2x0123,
126 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
127 vacc3x0123 = _mm_add_epi32(vacc3x0123,
128 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
129
130 w = (const void*) ((const uint8_t*) w + 32);
131 k -= 8 * sizeof(uint8_t);
132 }
133 if (k != 0) {
134 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
135 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
136 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
137 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
138 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
139 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
140 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
141 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
142 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
143 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
144 const __m128i vxa3 = _mm_unpacklo_epi8(va3, vzero);
145 a3 = (const uint8_t*) ((uintptr_t) a3 + k);
146
147 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
148 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
149 w = (const void*) ((const uint8_t*) w + 8);
150
151 vacc0x0123 = _mm_add_epi32(vacc0x0123,
152 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
153 vacc1x0123 = _mm_add_epi32(vacc1x0123,
154 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
155 vacc2x0123 = _mm_add_epi32(vacc2x0123,
156 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
157 vacc3x0123 = _mm_add_epi32(vacc3x0123,
158 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
159
160 if (k > 2 * sizeof(uint8_t)) {
161 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
162 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
163 w = (const void*) ((const uint8_t*) w + 8);
164
165 vacc0x0123 = _mm_add_epi32(vacc0x0123,
166 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
167 vacc1x0123 = _mm_add_epi32(vacc1x0123,
168 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
169 vacc2x0123 = _mm_add_epi32(vacc2x0123,
170 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
171 vacc3x0123 = _mm_add_epi32(vacc3x0123,
172 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
173
174 if (k > 4 * sizeof(uint8_t)) {
175 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
176 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
177 w = (const void*) ((const uint8_t*) w + 8);
178
179 vacc0x0123 = _mm_add_epi32(vacc0x0123,
180 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
181 vacc1x0123 = _mm_add_epi32(vacc1x0123,
182 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
183 vacc2x0123 = _mm_add_epi32(vacc2x0123,
184 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
185 vacc3x0123 = _mm_add_epi32(vacc3x0123,
186 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
187 }
188 }
189 }
190
191 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
192 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
193 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
194 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
195
196 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
197 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
198 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
199 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
200 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
201
202 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
203 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
204 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
205 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
206 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
207
208 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
209 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
210 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
211 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
212
213 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
214 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
215 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
216
217 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);
218
219 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
220
221 if (nc >= 4) {
222 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
223 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
224 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout));
225 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
226 unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(vout));
227 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
228 unaligned_store_u32(c3, (uint32_t) _mm_cvtsi128_si32(vout));
229
230 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
231 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
232 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
233 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
234
235 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
236 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
237 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
238 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
239
240 nc -= 4;
241 } else {
242 if (nc & 2) {
243 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
244 c0 += 2;
245 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
246 c1 += 2;
247 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
248 c2 += 2;
249 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
250 c3 += 2;
251 vout = _mm_srli_epi32(vout, 16);
252 }
253 if (nc & 1) {
254 *c0 = (uint8_t) _mm_cvtsi128_si32(vout);
255 *c1 = (uint8_t) _mm_extract_epi16(vout, 2);
256 *c2 = (uint8_t) _mm_extract_epi16(vout, 4);
257 *c3 = (uint8_t) _mm_extract_epi16(vout, 6);
258 }
259
260 nc = 0;
261 }
262 } while (nc != 0);
263 }
264