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 <smmintrin.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__avx_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__avx_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 while (k >= 8 * sizeof(uint8_t)) {
80 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
81 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
82 a0 += 8;
83 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
84 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
85 a1 += 8;
86 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
87 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
88 a2 += 8;
89
90 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
91 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
92
93 vacc0x0123 = _mm_add_epi32(vacc0x0123,
94 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
95 vacc1x0123 = _mm_add_epi32(vacc1x0123,
96 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
97 vacc2x0123 = _mm_add_epi32(vacc2x0123,
98 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
99 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
100 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
101
102 vacc0x0123 = _mm_add_epi32(vacc0x0123,
103 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
104 vacc1x0123 = _mm_add_epi32(vacc1x0123,
105 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
106 vacc2x0123 = _mm_add_epi32(vacc2x0123,
107 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
108 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
109 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), 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 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
118 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
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
127 w = (const void*) ((const uint8_t*) w + 32);
128 k -= 8 * sizeof(uint8_t);
129 }
130 if (k != 0) {
131 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
132 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
133 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
134 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
135 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
136 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
137 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
138 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
139 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
140
141 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
142 w = (const void*) ((const uint8_t*) w + 8);
143 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
144
145 vacc0x0123 = _mm_add_epi32(vacc0x0123,
146 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
147 vacc1x0123 = _mm_add_epi32(vacc1x0123,
148 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
149 vacc2x0123 = _mm_add_epi32(vacc2x0123,
150 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
151
152 if (k > 2 * sizeof(uint8_t)) {
153 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
154 w = (const void*) ((const uint8_t*) w + 8);
155 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
156
157 vacc0x0123 = _mm_add_epi32(vacc0x0123,
158 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
159 vacc1x0123 = _mm_add_epi32(vacc1x0123,
160 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
161 vacc2x0123 = _mm_add_epi32(vacc2x0123,
162 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
163
164 if (k > 4 * sizeof(uint8_t)) {
165 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
166 w = (const void*) ((const uint8_t*) w + 8);
167 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
168
169 vacc0x0123 = _mm_add_epi32(vacc0x0123,
170 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
171 vacc1x0123 = _mm_add_epi32(vacc1x0123,
172 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
173 vacc2x0123 = _mm_add_epi32(vacc2x0123,
174 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
175 }
176 }
177 }
178 p -= 3 * sizeof(void*);
179 } while (p != 0);
180
181 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
182 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
183 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
184
185 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
186 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
187 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
188 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
189
190 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
191 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
192 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
193 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
194
195 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
196 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
197 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
198
199 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
200 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
201 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
202
203 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
204
205 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
206
207 if (nc >= 4) {
208 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
209 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
210 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
211 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
212 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
213 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
214
215 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
216
217 nc -= 4;
218 } else {
219 if (nc & 2) {
220 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
221 c2 += 2;
222 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
223 c1 += 2;
224 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
225 c0 += 2;
226 vout = _mm_srli_epi32(vout, 16);
227 }
228 if (nc & 1) {
229 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
230 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
231 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
232 }
233
234 nc = 0;
235 }
236 } while (nc != 0);
237 }
238