1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2s4-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__sse2_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2s4__sse2_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_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(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54
55 do {
56 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
57 __m128i vacc1x0123 = vacc0x0123;
58 __m128i vacc2x0123 = vacc0x0123;
59 w = (const void*) ((const int32_t*) w + 4);
60
61 size_t p = ks;
62 do {
63 const int8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 const int8_t* restrict a1 = a[1];
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 const int8_t* restrict a2 = a[2];
72 if XNN_UNPREDICTABLE(a2 != zero) {
73 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
74 }
75 a += 3;
76
77 size_t k = kc;
78 do {
79 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
80 __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
81 a0 += 8;
82 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
83 __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
84 a1 += 8;
85 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
86 __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
87 a2 += 8;
88
89 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
90 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
91
92 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb0));
93 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
94 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb0));
95 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
96 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb0));
97 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
98 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
99 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
100
101 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb1));
102 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
103 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb1));
104 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
105 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb1));
106 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
107 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
108 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
109
110 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb2));
111 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
112 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb2));
113 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
114 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb2));
115 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
116 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
117 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
118
119 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb3));
120 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb3));
121 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb3));
122
123 w = (const void*) ((const int8_t*) w + 32);
124 k -= 8 * sizeof(int8_t);
125 } while (k != 0);
126 p -= 3 * sizeof(void*);
127 } while (p != 0);
128
129 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
130 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
131 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
132
133 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
134 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
135 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
136 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
137
138 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
139 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
140 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
141 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
142
143 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
144 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
145 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
146
147 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
148 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
149 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
150
151 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
152 vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min);
153 vacc22x0123 = _mm_max_epi16(vacc22x0123, voutput_min);
154
155 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
156
157
158 if (nc >= 4) {
159 unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(2, 2, 2, 2))));
160 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
161 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1))));
162 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
163 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
164 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
165
166 a = (const int8_t**restrict) ((uintptr_t) a - ks);
167
168 nc -= 4;
169 } else {
170 if (nc & 2) {
171 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
172 c2 += 2;
173 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
174 c1 += 2;
175 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
176 c0 += 2;
177 vout = _mm_srli_epi32(vout, 16);
178 }
179 if (nc & 1) {
180 *c2 = (int8_t) _mm_extract_epi16(vout, 4);
181 *c1 = (int8_t) _mm_extract_epi16(vout, 2);
182 *c0 = (int8_t) _mm_cvtsi128_si32(vout);
183 }
184
185 nc = 0;
186 }
187 } while (nc != 0);
188 }
189