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_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse2_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__sse2_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 const __m128i vzero = _mm_setzero_si128();
66 while (k >= 8 * sizeof(uint8_t)) {
67 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
68 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
69 a0 += 8;
70 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
71 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
72 a1 += 8;
73 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
74 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
75 a2 += 8;
76
77 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
78 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
79
80 vacc0x0123 = _mm_add_epi32(vacc0x0123,
81 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
82 vacc1x0123 = _mm_add_epi32(vacc1x0123,
83 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
84 vacc2x0123 = _mm_add_epi32(vacc2x0123,
85 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
86 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
87 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
88
89 vacc0x0123 = _mm_add_epi32(vacc0x0123,
90 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
91 vacc1x0123 = _mm_add_epi32(vacc1x0123,
92 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
93 vacc2x0123 = _mm_add_epi32(vacc2x0123,
94 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
95 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
96 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
97
98 vacc0x0123 = _mm_add_epi32(vacc0x0123,
99 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
100 vacc1x0123 = _mm_add_epi32(vacc1x0123,
101 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
102 vacc2x0123 = _mm_add_epi32(vacc2x0123,
103 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
104 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
105 const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point);
106
107 vacc0x0123 = _mm_add_epi32(vacc0x0123,
108 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
109 vacc1x0123 = _mm_add_epi32(vacc1x0123,
110 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
111 vacc2x0123 = _mm_add_epi32(vacc2x0123,
112 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
113
114 w = (const void*) ((const uint8_t*) w + 32);
115 k -= 8 * sizeof(uint8_t);
116 }
117 if (k != 0) {
118 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
119 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
120 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
121 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
122 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
123 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
124 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
125 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
126 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
127
128 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
129 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
130 w = (const void*) ((const uint8_t*) w + 8);
131
132 vacc0x0123 = _mm_add_epi32(vacc0x0123,
133 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
134 vacc1x0123 = _mm_add_epi32(vacc1x0123,
135 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
136 vacc2x0123 = _mm_add_epi32(vacc2x0123,
137 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
138
139 if (k > 2 * sizeof(uint8_t)) {
140 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
141 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
142 w = (const void*) ((const uint8_t*) w + 8);
143
144 vacc0x0123 = _mm_add_epi32(vacc0x0123,
145 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
146 vacc1x0123 = _mm_add_epi32(vacc1x0123,
147 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
148 vacc2x0123 = _mm_add_epi32(vacc2x0123,
149 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
150
151 if (k > 4 * sizeof(uint8_t)) {
152 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
153 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
154 w = (const void*) ((const uint8_t*) w + 8);
155
156 vacc0x0123 = _mm_add_epi32(vacc0x0123,
157 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
158 vacc1x0123 = _mm_add_epi32(vacc1x0123,
159 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
160 vacc2x0123 = _mm_add_epi32(vacc2x0123,
161 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
162 }
163 }
164 }
165
166 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
167 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
168 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
169
170 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
171 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
172 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
173 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
174
175 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
176 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
177 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
178 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
179
180 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
181 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
182 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
183
184 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
185 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
186 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
187
188 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
189
190 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
191
192 if (nc >= 4) {
193 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
194 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
195 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout));
196 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
197 unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(vout));
198
199 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
200 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
201 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
202
203 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
204 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
205 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
206
207 nc -= 4;
208 } else {
209 if (nc & 2) {
210 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
211 c0 += 2;
212 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
213 c1 += 2;
214 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
215 c2 += 2;
216 vout = _mm_srli_epi32(vout, 16);
217 }
218 if (nc & 1) {
219 *c0 = (uint8_t) _mm_cvtsi128_si32(vout);
220 *c1 = (uint8_t) _mm_extract_epi16(vout, 2);
221 *c2 = (uint8_t) _mm_extract_epi16(vout, 4);
222 }
223
224 nc = 0;
225 }
226 } while (nc != 0);
227 }
228