1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
19
xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__sse2(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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__sse2(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_qs8_conv_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(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 2 * sizeof(int8_t));
42 const int8_t* a0 = a;
43 int8_t* c0 = c;
44 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 a2 = a1;
54 c2 = c1;
55 }
56
57 do {
58 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
59 __m128i vacc1x0123 = vacc0x0123;
60 __m128i vacc2x0123 = vacc0x0123;
61 w = (const void*) ((const int32_t*) w + 4);
62
63 size_t k = kc;
64 while (k >= 8 * sizeof(int8_t)) {
65 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
66 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
67 a0 += 8;
68 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
69 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
70 a1 += 8;
71 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
72 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
73 a2 += 8;
74
75 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
76
77 vacc0x0123 = _mm_add_epi32(vacc0x0123,
78 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
79 vacc1x0123 = _mm_add_epi32(vacc1x0123,
80 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
81 vacc2x0123 = _mm_add_epi32(vacc2x0123,
82 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
83 const __m128i vxb1 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 8));
84
85 vacc0x0123 = _mm_add_epi32(vacc0x0123,
86 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
87 vacc1x0123 = _mm_add_epi32(vacc1x0123,
88 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
89 vacc2x0123 = _mm_add_epi32(vacc2x0123,
90 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
91 const __m128i vxb2 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 16));
92
93 vacc0x0123 = _mm_add_epi32(vacc0x0123,
94 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
95 vacc1x0123 = _mm_add_epi32(vacc1x0123,
96 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
97 vacc2x0123 = _mm_add_epi32(vacc2x0123,
98 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
99 const __m128i vxb3 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 24));
100
101 vacc0x0123 = _mm_add_epi32(vacc0x0123,
102 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
103 vacc1x0123 = _mm_add_epi32(vacc1x0123,
104 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
105 vacc2x0123 = _mm_add_epi32(vacc2x0123,
106 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
107
108 w = (const void*) ((const int16_t*) w + 32);
109 k -= 8 * sizeof(int8_t);
110 }
111 if (k != 0) {
112 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
113 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
114 a0 = (const int8_t*) ((uintptr_t) a0 + k);
115 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
116 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
117 a1 = (const int8_t*) ((uintptr_t) a1 + k);
118 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
119 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
120 a2 = (const int8_t*) ((uintptr_t) a2 + k);
121
122 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
123 w = (const void*) ((const int16_t*) w + 8);
124
125 vacc0x0123 = _mm_add_epi32(vacc0x0123,
126 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
127 vacc1x0123 = _mm_add_epi32(vacc1x0123,
128 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
129 vacc2x0123 = _mm_add_epi32(vacc2x0123,
130 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
131
132 if (k > 2 * sizeof(int8_t)) {
133 const __m128i vxb1 = _mm_load_si128((const __m128i*) w);
134 w = (const void*) ((const int16_t*) w + 8);
135
136 vacc0x0123 = _mm_add_epi32(vacc0x0123,
137 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
138 vacc1x0123 = _mm_add_epi32(vacc1x0123,
139 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
140 vacc2x0123 = _mm_add_epi32(vacc2x0123,
141 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
142
143 if (k > 4 * sizeof(int8_t)) {
144 const __m128i vxb2 = _mm_load_si128((const __m128i*) w);
145 w = (const void*) ((const int16_t*) w + 8);
146
147 vacc0x0123 = _mm_add_epi32(vacc0x0123,
148 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
149 vacc1x0123 = _mm_add_epi32(vacc1x0123,
150 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
151 vacc2x0123 = _mm_add_epi32(vacc2x0123,
152 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
153 }
154 }
155 }
156
157 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
158 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
159 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
160
161 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
162 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
163 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
164 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
165
166 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
167 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
168 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
169 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
170
171 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
172 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
173 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
174
175 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
176 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
177 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
178
179 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
180 vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min);
181 vacc22x0123 = _mm_max_epi16(vacc22x0123, voutput_min);
182
183 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
184
185
186 if (nc >= 4) {
187 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
188 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
189 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout));
190 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
191 unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(vout));
192
193 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
194 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
195 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
196
197 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
198 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
199 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
200
201 nc -= 4;
202 } else {
203 if (nc & 2) {
204 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
205 c0 += 2;
206 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
207 c1 += 2;
208 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
209 c2 += 2;
210 vout = _mm_srli_epi32(vout, 16);
211 }
212 if (nc & 1) {
213 *c0 = (int8_t) _mm_cvtsi128_si32(vout);
214 *c1 = (int8_t) _mm_extract_epi16(vout, 2);
215 *c2 = (int8_t) _mm_extract_epi16(vout, 4);
216 }
217
218 nc = 0;
219 }
220 } while (nc != 0);
221 }
222