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
17
xnn_qc8_igemm_minmax_fp32_ukernel_3x4c8__sse41_ld128(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_igemm_minmax_fp32_ukernel_3x4c8__sse41_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_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(ks != 0);
37 assert(ks % (3 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8);
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 do {
55 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
56 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
57 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
58 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
59 __m128i vacc1x0 = vacc0x0;
60 __m128i vacc1x1 = vacc0x1;
61 __m128i vacc1x2 = vacc0x2;
62 __m128i vacc1x3 = vacc0x3;
63 __m128i vacc2x0 = vacc0x0;
64 __m128i vacc2x1 = vacc0x1;
65 __m128i vacc2x2 = vacc0x2;
66 __m128i vacc2x3 = vacc0x3;
67 w = (const void*) ((const int32_t*) w + 4);
68
69 size_t p = ks;
70 do {
71 const int8_t* restrict a0 = a[0];
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
74 }
75 const int8_t* restrict a1 = a[1];
76 if XNN_UNPREDICTABLE(a1 != zero) {
77 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
78 }
79 const int8_t* restrict a2 = a[2];
80 if XNN_UNPREDICTABLE(a2 != zero) {
81 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
82 }
83 a += 3;
84
85 size_t k = 0;
86 while (k < kc) {
87 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
88 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
89 a0 += 8;
90 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
91 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
92 a1 += 8;
93 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
94 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
95 a2 += 8;
96
97 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
98 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
99 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
100
101 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
102 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
103 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
104 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
105 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
106 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
107 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
108 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
109 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
110
111 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
112 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
113 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
114 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
115 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
116 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
117
118 w = (const void*) ((const int8_t*) w + 32);
119 k += 8 * sizeof(int8_t);
120 }
121 p -= 3 * sizeof(void*);
122 } while (p != 0);
123
124 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
125 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
126 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
127 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
128 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
129 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
130
131 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
132 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
133 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
134
135 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
136 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
137 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
138
139 const __m128 vscale0123 = _mm_load_ps((const float*) w);
140 w = (const void*) ((const float*) w + 4);
141 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
142 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
143 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
144
145 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
146 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
147 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
148 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
149
150 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
151 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
152 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
153
154 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
155 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
156 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
157
158
159 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
160
161 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
162
163 if (nc >= 4) {
164 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
165 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
166 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
167 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
168 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
169 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
170
171 a = (const int8_t**restrict) ((uintptr_t) a - ks);
172
173 nc -= 4;
174 } else {
175 if (nc & 2) {
176 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
177 c2 += 2;
178 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
179 c1 += 2;
180 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
181 c0 += 2;
182 vout = _mm_srli_epi32(vout, 16);
183 }
184 if (nc & 1) {
185 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
186 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
187 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
188 }
189
190 nc = 0;
191 }
192 } while (nc != 0);
193 }
194