1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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 #include <smmintrin.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse41(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_gemm_xw_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse41(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_gemm_xw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29 {
30 assert(mr != 0);
31 assert(mr <= 3);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 8);
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54
55 do {
56 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
57 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
58 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
59 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
60 __m128i vacc1x0 = vacc0x0;
61 __m128i vacc1x1 = vacc0x1;
62 __m128i vacc1x2 = vacc0x2;
63 __m128i vacc1x3 = vacc0x3;
64 __m128i vacc2x0 = vacc0x0;
65 __m128i vacc2x1 = vacc0x1;
66 __m128i vacc2x2 = vacc0x2;
67 __m128i vacc2x3 = vacc0x3;
68 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
69
70 size_t k = 0;
71 while (k < kc) {
72 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
73 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
74 a0 += 8;
75 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
76 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
77 a1 += 8;
78 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
79 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
80 a2 += 8;
81
82 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
83
84 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
85 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
86 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
87 const __m128i vxb1 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 8 * sizeof(int16_t)));
88
89 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
90 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
91 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
92 const __m128i vxb2 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int16_t)));
93
94 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
95 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
96 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
97 const __m128i vxb3 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 24 * sizeof(int16_t)));
98
99 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
100 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
101 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
102
103 w = (const void*) ((uintptr_t) w + 32 * sizeof(int16_t));
104 k += 8 * sizeof(int8_t);
105 }
106
107 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
108 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
109 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
110 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
111 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
112 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
113
114 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
115 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
116 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
117
118 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
119 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
120
121 const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
122 const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
123 const __m128i vacc2x1133 = _mm_shuffle_epi32(vacc2x0123, _MM_SHUFFLE(3, 3, 1, 1));
124
125 const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
126 const __m128i vprod1x02 = _mm_add_epi64(_mm_mul_epi32(vacc1x0123, vmultiplier), vrounding);
127 const __m128i vprod2x02 = _mm_add_epi64(_mm_mul_epi32(vacc2x0123, vmultiplier), vrounding);
128
129 const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
130 const __m128i vprod1x13 = _mm_add_epi64(_mm_mul_epi32(vacc1x1133, vmultiplier), vrounding);
131 const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding);
132
133 const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
134 const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
135 const __m128i vq31prod1x02 = _mm_srli_epi64(vprod1x02, 31);
136 const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13);
137 const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31);
138 const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13);
139
140 const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
141 const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC);
142 const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC);
143
144 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
145 const __m128i vrem0x0123 =
146 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
147 const __m128i vrem1x0123 =
148 _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
149 const __m128i vrem2x0123 =
150 _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
151
152 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
153 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
154 vacc0x0123 =
155 _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
156 vacc1x0123 =
157 _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
158 vacc2x0123 =
159 _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
160
161 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
162 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
163 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
164
165 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
166 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
167 vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
168 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max);
169
170 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
171
172 if (nc >= 4) {
173 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
174 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
175 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
176
177 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
178 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
179 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
180
181 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
182 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
183 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
184
185 nc -= 4;
186 } else {
187 if (nc & 2) {
188 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
189 c0 += 2;
190 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
191 c1 += 2;
192 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
193 c2 += 2;
194 vout = _mm_srli_epi32(vout, 16);
195 }
196 if (nc & 1) {
197 *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
198 *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
199 *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
200 }
201
202 nc = 0;
203 }
204 } while (nc != 0);
205 }
206