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
17
xnn_qc8_igemm_minmax_fp32_ukernel_3x4c2__sse2_ld64(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__sse2_ld64(
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_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
80 a0 += 8;
81 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
82 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
83 a1 += 8;
84 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
85 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
86 a2 += 8;
87
88 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
89 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
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 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
98 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
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 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
107 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 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 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
116 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
117
118 vacc0x0123 = _mm_add_epi32(vacc0x0123,
119 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
120 vacc1x0123 = _mm_add_epi32(vacc1x0123,
121 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
122 vacc2x0123 = _mm_add_epi32(vacc2x0123,
123 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
124
125 w = (const void*) ((const int8_t*) w + 32);
126 k -= 8 * sizeof(int8_t);
127 }
128 if (k != 0) {
129 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
130 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
131 a0 = (const int8_t*) ((uintptr_t) a0 + k);
132 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
133 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
134 a1 = (const int8_t*) ((uintptr_t) a1 + k);
135 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
136 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
137 a2 = (const int8_t*) ((uintptr_t) a2 + k);
138
139 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
140 w = (const void*) ((const int8_t*) w + 8);
141 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
142
143 vacc0x0123 = _mm_add_epi32(vacc0x0123,
144 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
145 vacc1x0123 = _mm_add_epi32(vacc1x0123,
146 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
147 vacc2x0123 = _mm_add_epi32(vacc2x0123,
148 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
149
150 if (k > 2 * sizeof(int8_t)) {
151 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
152 w = (const void*) ((const int8_t*) w + 8);
153 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
154
155 vacc0x0123 = _mm_add_epi32(vacc0x0123,
156 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
157 vacc1x0123 = _mm_add_epi32(vacc1x0123,
158 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
159 vacc2x0123 = _mm_add_epi32(vacc2x0123,
160 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
161
162 if (k > 4 * sizeof(int8_t)) {
163 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
164 w = (const void*) ((const int8_t*) w + 8);
165 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
166
167 vacc0x0123 = _mm_add_epi32(vacc0x0123,
168 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
169 vacc1x0123 = _mm_add_epi32(vacc1x0123,
170 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
171 vacc2x0123 = _mm_add_epi32(vacc2x0123,
172 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
173 }
174 }
175 }
176 p -= 3 * sizeof(void*);
177 } while (p != 0);
178
179 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
180 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
181 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
182
183 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
184 w = (const void*) ((const float*) w + 4);
185 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
186 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
187 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
188
189 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point);
190 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
191 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
192 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
193
194 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
195 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
196 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
197
198 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
199 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
200 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
201
202 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
203 vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min);
204 vacc22x0123 = _mm_max_epi16(vacc22x0123, voutput_min);
205
206 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
207
208
209 if (nc >= 4) {
210 *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(2, 2, 2, 2)));
211 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
212 *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1)));
213 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
214 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
215 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
216
217 a = (const int8_t**restrict) ((uintptr_t) a - ks);
218
219 nc -= 4;
220 } else {
221 if (nc & 2) {
222 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
223 c2 += 2;
224 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
225 c1 += 2;
226 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
227 c0 += 2;
228 vout = _mm_srli_epi32(vout, 16);
229 }
230 if (nc & 1) {
231 *c2 = (int8_t) _mm_extract_epi16(vout, 4);
232 *c1 = (int8_t) _mm_extract_epi16(vout, 2);
233 *c0 = (int8_t) _mm_cvtsi128_si32(vout);
234 }
235
236 nc = 0;
237 }
238 } while (nc != 0);
239 }
240