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_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld64(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld64(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const uint8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 uint8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_qu8_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(uint8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
42 const uint8_t* a0 = a;
43 uint8_t* c0 = c;
44 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
51 uint8_t* c2 = (uint8_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 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
65 while (k >= 8 * sizeof(uint8_t)) {
66 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
67 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
68 a0 += 8;
69 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
70 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
71 a1 += 8;
72 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
73 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
74 a2 += 8;
75
76 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
77 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
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 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
86 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
87
88 vacc0x0123 = _mm_add_epi32(vacc0x0123,
89 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
90 vacc1x0123 = _mm_add_epi32(vacc1x0123,
91 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
92 vacc2x0123 = _mm_add_epi32(vacc2x0123,
93 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
94 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
95 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
96
97 vacc0x0123 = _mm_add_epi32(vacc0x0123,
98 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
99 vacc1x0123 = _mm_add_epi32(vacc1x0123,
100 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
101 vacc2x0123 = _mm_add_epi32(vacc2x0123,
102 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
103 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
104 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
105
106 vacc0x0123 = _mm_add_epi32(vacc0x0123,
107 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
108 vacc1x0123 = _mm_add_epi32(vacc1x0123,
109 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
110 vacc2x0123 = _mm_add_epi32(vacc2x0123,
111 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
112
113 w = (const void*) ((const uint8_t*) w + 32);
114 k -= 8 * sizeof(uint8_t);
115 }
116 if (k != 0) {
117 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
118 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
119 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
120 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
121 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
122 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
123 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
124 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
125 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
126
127 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
128 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
129 w = (const void*) ((const uint8_t*) w + 8);
130
131 vacc0x0123 = _mm_add_epi32(vacc0x0123,
132 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
133 vacc1x0123 = _mm_add_epi32(vacc1x0123,
134 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
135 vacc2x0123 = _mm_add_epi32(vacc2x0123,
136 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
137
138 if (k > 2 * sizeof(uint8_t)) {
139 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
140 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
141 w = (const void*) ((const uint8_t*) w + 8);
142
143 vacc0x0123 = _mm_add_epi32(vacc0x0123,
144 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
145 vacc1x0123 = _mm_add_epi32(vacc1x0123,
146 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
147 vacc2x0123 = _mm_add_epi32(vacc2x0123,
148 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
149
150 if (k > 4 * sizeof(uint8_t)) {
151 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
152 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
153 w = (const void*) ((const uint8_t*) w + 8);
154
155 vacc0x0123 = _mm_add_epi32(vacc0x0123,
156 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
157 vacc1x0123 = _mm_add_epi32(vacc1x0123,
158 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
159 vacc2x0123 = _mm_add_epi32(vacc2x0123,
160 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
161 }
162 }
163 }
164
165 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
166 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
167 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
168
169 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
170 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
171 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
172 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
173
174 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
175 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
176 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
177 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
178
179 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
180 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
181 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
182
183 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
184 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
185 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
186
187 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
188
189 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
190
191 if (nc >= 4) {
192 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
193 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
194 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
195
196 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
197 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
198 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
199
200 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
201 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
202 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
203
204 nc -= 4;
205 } else {
206 if (nc & 2) {
207 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
208 c0 += 2;
209 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
210 c1 += 2;
211 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
212 c2 += 2;
213 vout = _mm_srli_epi32(vout, 16);
214 }
215 if (nc & 1) {
216 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
217 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
218 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
219 }
220
221 nc = 0;
222 }
223 } while (nc != 0);
224 }
225