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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/igemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22
23
xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__xop_ld128(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)])24 void xnn_qu8_igemm_minmax_fp32_ukernel_3x4c2__xop_ld128(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 size_t ks,
29 const uint8_t** restrict a,
30 const void* restrict w,
31 uint8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 size_t a_offset,
35 const uint8_t* zero,
36 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 3);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(ks != 0);
43 assert(ks % (3 * sizeof(void*)) == 0);
44 assert(a_offset % sizeof(uint8_t) == 0);
45 assert(a != NULL);
46 assert(w != NULL);
47 assert(c != NULL);
48
49 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
50 uint8_t* c0 = c;
51 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
52 if XNN_UNPREDICTABLE(mr < 2) {
53 c1 = c0;
54 }
55 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
56 if XNN_UNPREDICTABLE(mr <= 2) {
57 c2 = c1;
58 }
59
60 do {
61 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
62 __m128i vacc1x0123 = vacc0x0123;
63 __m128i vacc2x0123 = vacc0x0123;
64 w = (const void*) ((const int32_t*) w + 4);
65
66 size_t p = ks;
67 do {
68 const uint8_t* restrict a0 = a[0];
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
71 }
72 const uint8_t* restrict a1 = a[1];
73 if XNN_UNPREDICTABLE(a1 != zero) {
74 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
75 }
76 const uint8_t* restrict a2 = a[2];
77 if XNN_UNPREDICTABLE(a2 != zero) {
78 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
79 }
80 a += 3;
81
82 size_t k = kc;
83 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
84 const __m128i vzero = _mm_setzero_si128();
85 while (k >= 8 * sizeof(uint8_t)) {
86 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
87 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
88 a0 += 8;
89 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
90 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
91 a1 += 8;
92 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
93 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
94 a2 += 8;
95
96 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
97 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
98 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
99
100 vacc0x0123 = _mm_maddd_epi16(
101 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
102 vacc1x0123 = _mm_maddd_epi16(
103 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
104 vacc2x0123 = _mm_maddd_epi16(
105 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
106
107 vacc0x0123 = _mm_maddd_epi16(
108 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
109 vacc1x0123 = _mm_maddd_epi16(
110 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
111 vacc2x0123 = _mm_maddd_epi16(
112 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
113 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
114 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
115 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
116
117 vacc0x0123 = _mm_maddd_epi16(
118 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
119 vacc1x0123 = _mm_maddd_epi16(
120 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
121 vacc2x0123 = _mm_maddd_epi16(
122 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
123
124 vacc0x0123 = _mm_maddd_epi16(
125 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
126 vacc1x0123 = _mm_maddd_epi16(
127 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
128 vacc2x0123 = _mm_maddd_epi16(
129 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
130
131 w = (const void*) ((const uint8_t*) w + 32);
132 k -= 8 * sizeof(uint8_t);
133 }
134 if (k != 0) {
135 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
136 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
137 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
138 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
139 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
140 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
141 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
142 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
143 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
144
145 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
146 w = (const void*) ((const uint8_t*) w + 8);
147 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
148
149 vacc0x0123 = _mm_maddd_epi16(
150 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
151 vacc1x0123 = _mm_maddd_epi16(
152 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
153 vacc2x0123 = _mm_maddd_epi16(
154 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
155
156 if (k > 2 * sizeof(uint8_t)) {
157 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
158 w = (const void*) ((const uint8_t*) w + 8);
159 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
160
161 vacc0x0123 = _mm_maddd_epi16(
162 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
163 vacc1x0123 = _mm_maddd_epi16(
164 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
165 vacc2x0123 = _mm_maddd_epi16(
166 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
167
168 if (k > 4 * sizeof(uint8_t)) {
169 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
170 w = (const void*) ((const uint8_t*) w + 8);
171 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
172
173 vacc0x0123 = _mm_maddd_epi16(
174 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
175 vacc1x0123 = _mm_maddd_epi16(
176 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
177 vacc2x0123 = _mm_maddd_epi16(
178 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
179 }
180 }
181 }
182 p -= 3 * sizeof(void*);
183 } while (p != 0);
184
185 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
186 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
187 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
188
189 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
190 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
191 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
192 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
193
194 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
195 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
196 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
197 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
198
199 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
200 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
201 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
202
203 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
204 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
205 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
206
207 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
208
209 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
210
211 if (nc >= 4) {
212 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
213 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
214 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
215 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
216 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
217 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
218
219 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
220
221 nc -= 4;
222 } else {
223 if (nc & 2) {
224 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
225 c2 += 2;
226 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
227 c1 += 2;
228 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
229 c0 += 2;
230 vout = _mm_srli_epi32(vout, 16);
231 }
232 if (nc & 1) {
233 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
234 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
235 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
236 }
237
238 nc = 0;
239 }
240 } while (nc != 0);
241 }
242