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_3x4c2__sse2_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)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__sse2_ld64(
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 <= 3);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (3 * 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
55 do {
56 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
57 __m128i vacc1x0123 = vacc0x0123;
58 __m128i vacc2x0123 = vacc0x0123;
59 w = (const void*) ((const int32_t*) w + 4);
60
61 size_t p = ks;
62 do {
63 const uint8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 const uint8_t* restrict a1 = a[1];
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 const uint8_t* restrict a2 = a[2];
72 if XNN_UNPREDICTABLE(a2 != zero) {
73 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
74 }
75 a += 3;
76
77 size_t k = kc;
78 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
79 const __m128i vzero = _mm_setzero_si128();
80 while (k >= 8 * sizeof(uint8_t)) {
81 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
82 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
83 a0 += 8;
84 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
85 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
86 a1 += 8;
87 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
88 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
89 a2 += 8;
90
91 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
92 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
93
94 vacc0x0123 = _mm_add_epi32(vacc0x0123,
95 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
96 vacc1x0123 = _mm_add_epi32(vacc1x0123,
97 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
98 vacc2x0123 = _mm_add_epi32(vacc2x0123,
99 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
100 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
101 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
102
103 vacc0x0123 = _mm_add_epi32(vacc0x0123,
104 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
105 vacc1x0123 = _mm_add_epi32(vacc1x0123,
106 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
107 vacc2x0123 = _mm_add_epi32(vacc2x0123,
108 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
109 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
110 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, 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 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
119 const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point);
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
128 w = (const void*) ((const uint8_t*) w + 32);
129 k -= 8 * sizeof(uint8_t);
130 }
131 if (k != 0) {
132 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
133 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
134 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
135 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
136 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
137 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
138 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
139 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
140 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
141
142 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
143 w = (const void*) ((const uint8_t*) w + 8);
144 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
145
146 vacc0x0123 = _mm_add_epi32(vacc0x0123,
147 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
148 vacc1x0123 = _mm_add_epi32(vacc1x0123,
149 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
150 vacc2x0123 = _mm_add_epi32(vacc2x0123,
151 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
152
153 if (k > 2 * sizeof(uint8_t)) {
154 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
155 w = (const void*) ((const uint8_t*) w + 8);
156 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
157
158 vacc0x0123 = _mm_add_epi32(vacc0x0123,
159 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
160 vacc1x0123 = _mm_add_epi32(vacc1x0123,
161 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
162 vacc2x0123 = _mm_add_epi32(vacc2x0123,
163 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
164
165 if (k > 4 * sizeof(uint8_t)) {
166 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
167 w = (const void*) ((const uint8_t*) w + 8);
168 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
169
170 vacc0x0123 = _mm_add_epi32(vacc0x0123,
171 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
172 vacc1x0123 = _mm_add_epi32(vacc1x0123,
173 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
174 vacc2x0123 = _mm_add_epi32(vacc2x0123,
175 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
176 }
177 }
178 }
179 p -= 3 * sizeof(void*);
180 } while (p != 0);
181
182 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
183 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
184 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
185
186 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
187 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
188 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
189 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
190
191 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
192 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
193 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
194 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
195
196 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
197 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
198 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
199
200 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
201 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
202 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
203
204 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
205
206 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
207
208 if (nc >= 4) {
209 unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(2, 2, 2, 2))));
210 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
211 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1))));
212 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
213 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
214 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
215
216 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
217
218 nc -= 4;
219 } else {
220 if (nc & 2) {
221 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
222 c2 += 2;
223 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
224 c1 += 2;
225 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
226 c0 += 2;
227 vout = _mm_srli_epi32(vout, 16);
228 }
229 if (nc & 1) {
230 *c2 = (uint8_t) _mm_extract_epi16(vout, 4);
231 *c1 = (uint8_t) _mm_extract_epi16(vout, 2);
232 *c0 = (uint8_t) _mm_cvtsi128_si32(vout);
233 }
234
235 nc = 0;
236 }
237 } while (nc != 0);
238 }
239