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
17
18
xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128(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)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const uint8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 3);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(uint8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 2);
41 const uint8_t* a0 = a;
42 uint8_t* c0 = c;
43 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
44 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
50 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55
56 do {
57 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
58 __m128i vacc1x0123 = vacc0x0123;
59 __m128i vacc2x0123 = vacc0x0123;
60 w = (const void*) ((const int32_t*) w + 4);
61
62 size_t k = kc;
63 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
64 const __m128i vzero = _mm_setzero_si128();
65 while (k >= 8 * sizeof(uint8_t)) {
66 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
67 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
68 a0 += 8;
69 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
70 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
71 a1 += 8;
72 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
73 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
74 a2 += 8;
75
76 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
77 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
78 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, 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
87 vacc0x0123 = _mm_add_epi32(vacc0x0123,
88 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
89 vacc1x0123 = _mm_add_epi32(vacc1x0123,
90 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
91 vacc2x0123 = _mm_add_epi32(vacc2x0123,
92 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
93 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
94 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
95 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), 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
104 vacc0x0123 = _mm_add_epi32(vacc0x0123,
105 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
106 vacc1x0123 = _mm_add_epi32(vacc1x0123,
107 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
108 vacc2x0123 = _mm_add_epi32(vacc2x0123,
109 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
110
111 w = (const void*) ((const uint8_t*) w + 32);
112 k -= 8 * sizeof(uint8_t);
113 }
114 if (k != 0) {
115 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
116 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
117 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
118 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
119 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
120 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
121 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
122 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
123 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
124
125 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
126 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
127 w = (const void*) ((const uint8_t*) w + 8);
128
129 vacc0x0123 = _mm_add_epi32(vacc0x0123,
130 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
131 vacc1x0123 = _mm_add_epi32(vacc1x0123,
132 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
133 vacc2x0123 = _mm_add_epi32(vacc2x0123,
134 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
135
136 if (k > 2 * sizeof(uint8_t)) {
137 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
138 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
139 w = (const void*) ((const uint8_t*) w + 8);
140
141 vacc0x0123 = _mm_add_epi32(vacc0x0123,
142 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
143 vacc1x0123 = _mm_add_epi32(vacc1x0123,
144 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
145 vacc2x0123 = _mm_add_epi32(vacc2x0123,
146 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
147
148 if (k > 4 * sizeof(uint8_t)) {
149 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
150 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
151 w = (const void*) ((const uint8_t*) w + 8);
152
153 vacc0x0123 = _mm_add_epi32(vacc0x0123,
154 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
155 vacc1x0123 = _mm_add_epi32(vacc1x0123,
156 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
157 vacc2x0123 = _mm_add_epi32(vacc2x0123,
158 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
159 }
160 }
161 }
162
163 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
164 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
165 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
166
167 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
168 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
169 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
170 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
171
172 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.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_sse2.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 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
186
187 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
188
189 if (nc >= 4) {
190 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
191 vout = _mm_srli_si128(vout, 4);
192 *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
193 vout = _mm_srli_si128(vout, 4);
194 *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(vout);
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 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
208 c0 += 2;
209 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
210 c1 += 2;
211 *((uint16_t*) 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_cvtsi128_si32(vout);
217 *c1 = (uint8_t) _mm_extract_epi16(vout, 2);
218 *c2 = (uint8_t) _mm_extract_epi16(vout, 4);
219 }
220
221 nc = 0;
222 }
223 } while (nc != 0);
224 }
225