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 <smmintrin.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__sse41_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__sse41_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 while (k < kc) {
89 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
90 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
91 a0 += 8;
92 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
93 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
94 a1 += 8;
95 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
96 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
97 a2 += 8;
98
99 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
100 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
101
102 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
103 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
104 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
105 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
106 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
107
108 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
109 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
110 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
111 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
112 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
113
114 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
115 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
116 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
117 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
118 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
119
120 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
121 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
122 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
123
124 w = (const void*) ((const uint8_t*) w + 32);
125 k += 8 * sizeof(uint8_t);
126 }
127 p -= 3 * sizeof(void*);
128 } while (p != 0);
129
130 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
131 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
132 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
133 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
134 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
135 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
136
137 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
138 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
139 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
140
141 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
142 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
143 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
144
145 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
146 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
147 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
148 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
149
150 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
151 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
152 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
153 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
154
155 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
156 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
157 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
158
159 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
160 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
161 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
162
163 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
164
165 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
166
167 if (nc >= 4) {
168 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
169 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
170 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
171 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
172 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
173 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
174
175 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
176
177 nc -= 4;
178 } else {
179 if (nc & 2) {
180 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
181 c2 += 2;
182 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
183 c1 += 2;
184 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
185 c0 += 2;
186 vout = _mm_srli_epi32(vout, 16);
187 }
188 if (nc & 1) {
189 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
190 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
191 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
192 }
193
194 nc = 0;
195 }
196 } while (nc != 0);
197 }
198