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_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)])20 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128(
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 vb01 = _mm_loadu_si128((const __m128i*) w);
78 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
79 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
80
81 vacc0x0123 = _mm_add_epi32(vacc0x0123,
82 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
83 vacc1x0123 = _mm_add_epi32(vacc1x0123,
84 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
85 vacc2x0123 = _mm_add_epi32(vacc2x0123,
86 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
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 vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
95 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
96 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, 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
105 vacc0x0123 = _mm_add_epi32(vacc0x0123,
106 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
107 vacc1x0123 = _mm_add_epi32(vacc1x0123,
108 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
109 vacc2x0123 = _mm_add_epi32(vacc2x0123,
110 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
111
112 w = (const void*) ((const uint8_t*) w + 32);
113 k -= 8 * sizeof(uint8_t);
114 }
115 if (k != 0) {
116 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
117 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
118 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
119 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
120 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
121 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
122 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
123 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
124 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
125
126 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
127 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
128 w = (const void*) ((const uint8_t*) w + 8);
129
130 vacc0x0123 = _mm_add_epi32(vacc0x0123,
131 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
132 vacc1x0123 = _mm_add_epi32(vacc1x0123,
133 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
134 vacc2x0123 = _mm_add_epi32(vacc2x0123,
135 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
136
137 if (k > 2 * sizeof(uint8_t)) {
138 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
139 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
140 w = (const void*) ((const uint8_t*) w + 8);
141
142 vacc0x0123 = _mm_add_epi32(vacc0x0123,
143 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
144 vacc1x0123 = _mm_add_epi32(vacc1x0123,
145 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
146 vacc2x0123 = _mm_add_epi32(vacc2x0123,
147 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
148
149 if (k > 4 * sizeof(uint8_t)) {
150 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
151 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
152 w = (const void*) ((const uint8_t*) w + 8);
153
154 vacc0x0123 = _mm_add_epi32(vacc0x0123,
155 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
156 vacc1x0123 = _mm_add_epi32(vacc1x0123,
157 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
158 vacc2x0123 = _mm_add_epi32(vacc2x0123,
159 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
160 }
161 }
162 }
163
164 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
165 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
166 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
167
168 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
169 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
170 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
171 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
172
173 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
174 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
175 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
176 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
177
178 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
179 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
180 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
181
182 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
183 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
184 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
185
186 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
187
188 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
189
190 if (nc >= 4) {
191 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
192 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
193 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout));
194 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
195 unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(vout));
196
197 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
198 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
199 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
200
201 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
202 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
203 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
204
205 nc -= 4;
206 } else {
207 if (nc & 2) {
208 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
209 c0 += 2;
210 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
211 c1 += 2;
212 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
213 c2 += 2;
214 vout = _mm_srli_epi32(vout, 16);
215 }
216 if (nc & 1) {
217 *c0 = (uint8_t) _mm_cvtsi128_si32(vout);
218 *c1 = (uint8_t) _mm_extract_epi16(vout, 2);
219 *c2 = (uint8_t) _mm_extract_epi16(vout, 4);
220 }
221
222 nc = 0;
223 }
224 } while (nc != 0);
225 }
226