1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c8-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 #ifdef __GNUC__
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
22
xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_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_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qs8_igemm_minmax_ukernel_3x4c8__xop_ld128(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 size_t ks,
28 const int8_t** restrict a,
29 const void* restrict w,
30 int8_t* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 size_t a_offset,
34 const int8_t* zero,
35 const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
36 {
37 assert(mr != 0);
38 assert(mr <= 3);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(ks != 0);
42 assert(ks % (3 * sizeof(void*)) == 0);
43 assert(a_offset % sizeof(int8_t) == 0);
44 assert(a != NULL);
45 assert(w != NULL);
46 assert(c != NULL);
47
48 kc = round_up_po2(kc, 8);
49 int8_t* c0 = c;
50 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 c1 = c0;
53 }
54 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 c2 = c1;
57 }
58
59 do {
60 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
61 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
62 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
63 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
64 __m128i vacc1x0 = vacc0x0;
65 __m128i vacc1x1 = vacc0x1;
66 __m128i vacc1x2 = vacc0x2;
67 __m128i vacc1x3 = vacc0x3;
68 __m128i vacc2x0 = vacc0x0;
69 __m128i vacc2x1 = vacc0x1;
70 __m128i vacc2x2 = vacc0x2;
71 __m128i vacc2x3 = vacc0x3;
72 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
73
74 size_t p = ks;
75 do {
76 const int8_t* restrict a0 = a[0];
77 if XNN_UNPREDICTABLE(a0 != zero) {
78 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
79 }
80 const int8_t* restrict a1 = a[1];
81 if XNN_UNPREDICTABLE(a1 != zero) {
82 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
83 }
84 const int8_t* restrict a2 = a[2];
85 if XNN_UNPREDICTABLE(a2 != zero) {
86 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
87 }
88 a += 3;
89
90 size_t k = 0;
91 while (k < kc) {
92 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
93 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
94 a0 += 8;
95 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
96 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
97 a1 += 8;
98 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
99 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
100 a2 += 8;
101
102 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
103 const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
104 const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
105 const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
106
107 vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
108 vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
109 vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
110 vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
111 vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
112 vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
113 const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
114 const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
115 const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
116 const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
117
118 vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
119 vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
120 vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
121 vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
122 vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
123 vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);
124
125 w = (const void*) ((uintptr_t) w + 32);
126 k += 8 * sizeof(int8_t);
127 }
128 p -= 3 * sizeof(void*);
129 } while (p != 0);
130
131 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
132 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
133 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
134 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
135 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
136 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
137
138 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
139 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
140 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
141
142 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
143 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
144
145 const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
146 const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
147 const __m128i vacc2x1133 = _mm_shuffle_epi32(vacc2x0123, _MM_SHUFFLE(3, 3, 1, 1));
148
149 const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
150 const __m128i vprod1x02 = _mm_add_epi64(_mm_mul_epi32(vacc1x0123, vmultiplier), vrounding);
151 const __m128i vprod2x02 = _mm_add_epi64(_mm_mul_epi32(vacc2x0123, vmultiplier), vrounding);
152
153 const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
154 const __m128i vprod1x13 = _mm_add_epi64(_mm_mul_epi32(vacc1x1133, vmultiplier), vrounding);
155 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding);
156
157 const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
158 const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
159 const __m128i vq31prod1x02 = _mm_srli_epi64(vprod1x02, 31);
160 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13);
161 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31);
162 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13);
163
164 const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
165 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC);
166 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC);
167
168 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
169 const __m128i vrem0x0123 =
170 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
171 const __m128i vrem1x0123 =
172 _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
173 const __m128i vrem2x0123 =
174 _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
175
176 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
177 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
178 vacc0x0123 =
179 _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
180 vacc1x0123 =
181 _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
182 vacc2x0123 =
183 _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
184
185 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
186 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
187 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
188
189 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
190 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
191 vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
192 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max);
193
194 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
195
196 if (nc >= 4) {
197 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
198 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
199 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
200 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
201 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
202 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
203
204 a = (const int8_t**restrict) ((uintptr_t) a - ks);
205
206 nc -= 4;
207 } else {
208 if (nc & 2) {
209 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
210 c2 += 2;
211 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
212 c1 += 2;
213 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
214 c0 += 2;
215 vout = _mm_srli_epi32(vout, 16);
216 }
217 if (nc & 1) {
218 *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
219 *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
220 *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
221 }
222
223 nc = 0;
224 }
225 } while (nc != 0);
226 }
227