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 <smmintrin.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
19
xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld128(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld128(
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_qc8_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_cvtepi8_epi16(va0);
67 a0 += 8;
68 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
69 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
70 a1 += 8;
71 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
72 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
73 a2 += 8;
74
75 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
76 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
77 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
78
79 vacc0x0123 = _mm_add_epi32(vacc0x0123,
80 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
81 vacc1x0123 = _mm_add_epi32(vacc1x0123,
82 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
83 vacc2x0123 = _mm_add_epi32(vacc2x0123,
84 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
85
86 vacc0x0123 = _mm_add_epi32(vacc0x0123,
87 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
88 vacc1x0123 = _mm_add_epi32(vacc1x0123,
89 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
90 vacc2x0123 = _mm_add_epi32(vacc2x0123,
91 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
92 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
93 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
94 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
95
96 vacc0x0123 = _mm_add_epi32(vacc0x0123,
97 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
98 vacc1x0123 = _mm_add_epi32(vacc1x0123,
99 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
100 vacc2x0123 = _mm_add_epi32(vacc2x0123,
101 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
102
103 vacc0x0123 = _mm_add_epi32(vacc0x0123,
104 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
105 vacc1x0123 = _mm_add_epi32(vacc1x0123,
106 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
107 vacc2x0123 = _mm_add_epi32(vacc2x0123,
108 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
109
110 w = (const void*) ((const int8_t*) w + 32);
111 k -= 8 * sizeof(int8_t);
112 }
113 if (k != 0) {
114 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
115 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
116 a0 = (const int8_t*) ((uintptr_t) a0 + k);
117 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
118 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
119 a1 = (const int8_t*) ((uintptr_t) a1 + k);
120 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
121 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
122 a2 = (const int8_t*) ((uintptr_t) a2 + k);
123
124 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
125 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
126 w = (const void*) ((const int8_t*) w + 8);
127
128 vacc0x0123 = _mm_add_epi32(vacc0x0123,
129 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
130 vacc1x0123 = _mm_add_epi32(vacc1x0123,
131 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
132 vacc2x0123 = _mm_add_epi32(vacc2x0123,
133 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
134
135 if (k > 2 * sizeof(int8_t)) {
136 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
137 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
138 w = (const void*) ((const int8_t*) w + 8);
139
140 vacc0x0123 = _mm_add_epi32(vacc0x0123,
141 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
142 vacc1x0123 = _mm_add_epi32(vacc1x0123,
143 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
144 vacc2x0123 = _mm_add_epi32(vacc2x0123,
145 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
146
147 if (k > 4 * sizeof(int8_t)) {
148 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
149 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
150 w = (const void*) ((const int8_t*) w + 8);
151
152 vacc0x0123 = _mm_add_epi32(vacc0x0123,
153 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
154 vacc1x0123 = _mm_add_epi32(vacc1x0123,
155 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
156 vacc2x0123 = _mm_add_epi32(vacc2x0123,
157 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
158 }
159 }
160 }
161
162 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
163 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
164 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
165
166 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
167 w = (const void*) ((const float*) w + 4);
168 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
169 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
170 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
171
172 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
173 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
174 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
175 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
176
177 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
178 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
179 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
180
181 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
182 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
183 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
184
185
186 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
187
188 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
189
190 if (nc >= 4) {
191 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
192 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
193 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
194
195 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
196 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
197 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
198
199 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
200 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
201 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
202
203 nc -= 4;
204 } else {
205 if (nc & 2) {
206 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
207 c0 += 2;
208 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
209 c1 += 2;
210 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
211 c2 += 2;
212 vout = _mm_srli_epi32(vout, 16);
213 }
214 if (nc & 1) {
215 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
216 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
217 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
218 }
219
220 nc = 0;
221 }
222 } while (nc != 0);
223 }
224