1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c8-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/igemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const uint8_t** restrict a,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const uint8_t* zero,
31 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 3);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (3 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(uint8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8);
45 uint8_t* c0 = c;
46 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54
55 do {
56 __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
57 __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
58 __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
59 __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
60 __m128i vacc1x0 = vacc0x0;
61 __m128i vacc1x1 = vacc0x1;
62 __m128i vacc1x2 = vacc0x2;
63 __m128i vacc1x3 = vacc0x3;
64 __m128i vacc2x0 = vacc0x0;
65 __m128i vacc2x1 = vacc0x1;
66 __m128i vacc2x2 = vacc0x2;
67 __m128i vacc2x3 = vacc0x3;
68 w = (const int32_t*) w + 4;
69
70 size_t p = ks;
71 do {
72 const uint8_t* restrict a0 = a[0];
73 if XNN_UNPREDICTABLE(a0 != zero) {
74 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
75 }
76 const uint8_t* restrict a1 = a[1];
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
79 }
80 const uint8_t* restrict a2 = a[2];
81 if XNN_UNPREDICTABLE(a2 != zero) {
82 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
83 }
84 a += 3;
85
86 size_t k = 0;
87 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
88 const __m128i vzero = _mm_setzero_si128();
89 while (k < kc) {
90 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
91 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
92 a0 += 8;
93 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
94 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
95 a1 += 8;
96 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
97 const __m128i vxa2 = _mm_unpacklo_epi8(va2, vzero);
98 a2 += 8;
99
100 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
101 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
102
103 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
104 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
105 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
106 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
107 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
108
109 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
110 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
111 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
112 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
113 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
114
115 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
116 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
117 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
118 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
119 const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point);
120
121 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
122 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
123 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
124
125 w = (const void*) ((const uint8_t*) w + 32);
126 k += 8 * sizeof(uint8_t);
127 }
128 p -= 3 * sizeof(void*);
129 } while (p != 0);
130
131 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
132 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
133 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
134 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
135 const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2));
136 const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3));
137
138 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
139 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
140 __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13));
141
142 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
143 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
144 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
145
146 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
147 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
148 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
149 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
150
151 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
152 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
153 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
154 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
155
156 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
157 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
158 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
159
160 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
161 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
162 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
163
164 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
165
166 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
167
168 if (nc >= 4) {
169 unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(2, 2, 2, 2))));
170 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
171 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1))));
172 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
173 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
174 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
175
176 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
177
178 nc -= 4;
179 } else {
180 if (nc & 2) {
181 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
182 c2 += 2;
183 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
184 c1 += 2;
185 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
186 c0 += 2;
187 vout = _mm_srli_epi32(vout, 16);
188 }
189 if (nc & 1) {
190 *c2 = (uint8_t) _mm_extract_epi16(vout, 4);
191 *c1 = (uint8_t) _mm_extract_epi16(vout, 2);
192 *c0 = (uint8_t) _mm_cvtsi128_si32(vout);
193 }
194
195 nc = 0;
196 }
197 } while (nc != 0);
198 }
199