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
17
xnn_qc8_igemm_minmax_fp32_ukernel_3x4c2__sse41_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_3x4c2__sse41_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 2);
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 do {
55 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
56 __m128i vacc1x0123 = vacc0x0123;
57 __m128i vacc2x0123 = vacc0x0123;
58 w = (const void*) ((const int32_t*) w + 4);
59
60 size_t p = ks;
61 do {
62 const int8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 const int8_t* restrict a1 = a[1];
67 if XNN_UNPREDICTABLE(a1 != zero) {
68 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
69 }
70 const int8_t* restrict a2 = a[2];
71 if XNN_UNPREDICTABLE(a2 != zero) {
72 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
73 }
74 a += 3;
75
76 size_t k = kc;
77 while (k >= 8 * sizeof(int8_t)) {
78 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
79 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
80 a0 += 8;
81 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
82 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
83 a1 += 8;
84 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
85 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
86 a2 += 8;
87
88 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
89 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
90 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
91
92 vacc0x0123 = _mm_add_epi32(vacc0x0123,
93 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
94 vacc1x0123 = _mm_add_epi32(vacc1x0123,
95 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
96 vacc2x0123 = _mm_add_epi32(vacc2x0123,
97 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _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 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
106 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
107 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
108
109 vacc0x0123 = _mm_add_epi32(vacc0x0123,
110 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
111 vacc1x0123 = _mm_add_epi32(vacc1x0123,
112 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
113 vacc2x0123 = _mm_add_epi32(vacc2x0123,
114 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
115
116 vacc0x0123 = _mm_add_epi32(vacc0x0123,
117 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
118 vacc1x0123 = _mm_add_epi32(vacc1x0123,
119 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
120 vacc2x0123 = _mm_add_epi32(vacc2x0123,
121 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
122
123 w = (const void*) ((const int8_t*) w + 32);
124 k -= 8 * sizeof(int8_t);
125 }
126 if (k != 0) {
127 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
128 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
129 a0 = (const int8_t*) ((uintptr_t) a0 + k);
130 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
131 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
132 a1 = (const int8_t*) ((uintptr_t) a1 + k);
133 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
134 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
135 a2 = (const int8_t*) ((uintptr_t) a2 + k);
136
137 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
138 w = (const void*) ((const int8_t*) w + 8);
139 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
140
141 vacc0x0123 = _mm_add_epi32(vacc0x0123,
142 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
143 vacc1x0123 = _mm_add_epi32(vacc1x0123,
144 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
145 vacc2x0123 = _mm_add_epi32(vacc2x0123,
146 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
147
148 if (k > 2 * sizeof(int8_t)) {
149 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
150 w = (const void*) ((const int8_t*) w + 8);
151 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
152
153 vacc0x0123 = _mm_add_epi32(vacc0x0123,
154 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
155 vacc1x0123 = _mm_add_epi32(vacc1x0123,
156 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
157 vacc2x0123 = _mm_add_epi32(vacc2x0123,
158 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
159
160 if (k > 4 * sizeof(int8_t)) {
161 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
162 w = (const void*) ((const int8_t*) w + 8);
163 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
164
165 vacc0x0123 = _mm_add_epi32(vacc0x0123,
166 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
167 vacc1x0123 = _mm_add_epi32(vacc1x0123,
168 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
169 vacc2x0123 = _mm_add_epi32(vacc2x0123,
170 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
171 }
172 }
173 }
174 p -= 3 * sizeof(void*);
175 } while (p != 0);
176
177 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
178 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
179 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
180
181 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
182 w = (const void*) ((const float*) w + 4);
183 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
184 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
185 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
186
187 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
188 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
189 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
190 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
191
192 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
193 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
194 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
195
196 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
197 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
198 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
199
200
201 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
202
203 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
204
205 if (nc >= 4) {
206 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
207 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
208 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
209 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
210 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
211 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
212
213 a = (const int8_t**restrict) ((uintptr_t) a - ks);
214
215 nc -= 4;
216 } else {
217 if (nc & 2) {
218 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
219 c2 += 2;
220 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
221 c1 += 2;
222 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
223 c0 += 2;
224 vout = _mm_srli_epi32(vout, 16);
225 }
226 if (nc & 1) {
227 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
228 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
229 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
230 }
231
232 nc = 0;
233 }
234 } while (nc != 0);
235 }
236