Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x4c8-minmax-sse2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_2x4c8__sse2_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
30 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 8 * sizeof(int8_t));
42 const int8_t* a0 = a;
43 float* c0 = c;
44 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50
51 const __m128 vmin = _mm_set1_ps(params->scalar.min);
52 const __m128 vmax = _mm_set1_ps(params->scalar.max);
53 XNN_FORCE_REALIZATION(vmin);
54 XNN_FORCE_REALIZATION(vmax);
55
56
57 do {
58 const __m128i vksum = _mm_load_si128((const __m128i*) w);
59 const __m128i vinput_zero_point01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
60 const __m128i vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(0, 0, 0, 0));
61 const __m128i vinput_zero_point1 = _mm_shuffle_epi32(vinput_zero_point01, _MM_SHUFFLE(2, 2, 2, 2));
62 const __m128i vzero = _mm_setzero_si128();
63 const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
64 const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);
65
66 __m128i vzpprodksumhi0 = _mm_mulhi_epu16(vinput_zero_point0, vksum_lo);
67 const __m128i vzpprodksumlo0 = _mm_mullo_epi16(vinput_zero_point0, vksum_lo);
68 __m128i vzpprodksumhi1 = _mm_mulhi_epu16(vinput_zero_point1, vksum_lo);
69 const __m128i vzpprodksumlo1 = _mm_mullo_epi16(vinput_zero_point1, vksum_lo);
70
71 vzpprodksumhi0 = _mm_add_epi16(vzpprodksumhi0, _mm_mullo_epi16(vinput_zero_point0, vksum_hi));
72 vzpprodksumhi0 = _mm_sub_epi16(vzpprodksumhi0, _mm_and_si128(_mm_srai_epi16(vinput_zero_point0, 15), vksum_lo));
73 vzpprodksumhi1 = _mm_add_epi16(vzpprodksumhi1, _mm_mullo_epi16(vinput_zero_point1, vksum_hi));
74 vzpprodksumhi1 = _mm_sub_epi16(vzpprodksumhi1, _mm_and_si128(_mm_srai_epi16(vinput_zero_point1, 15), vksum_lo));
75
76 vzpprodksumhi0 = _mm_slli_si128(vzpprodksumhi0, 2);
77 vzpprodksumhi1 = _mm_slli_si128(vzpprodksumhi1, 2);
78
79 const __m128i vksumzp0 = _mm_or_si128(vzpprodksumhi0, vzpprodksumlo0);
80 const __m128i vksumzp1 = _mm_or_si128(vzpprodksumhi1, vzpprodksumlo1);
81
82 const __m128i vksum010 = _mm_unpacklo_epi32(vksumzp0, vzero);
83 const __m128i vksum230 = _mm_unpackhi_epi32(vksumzp0, vzero);
84 const __m128i vksum011 = _mm_unpacklo_epi32(vksumzp1, vzero);
85 const __m128i vksum231 = _mm_unpackhi_epi32(vksumzp1, vzero);
86
87 __m128i vacc0x0 = _mm_unpacklo_epi64(vksum010, vzero);
88 __m128i vacc0x1 = _mm_unpackhi_epi64(vksum010, vzero);
89 __m128i vacc1x0 = _mm_unpacklo_epi64(vksum011, vzero);
90 __m128i vacc1x1 = _mm_unpackhi_epi64(vksum011, vzero);
91
92 __m128i vacc0x2 = _mm_unpacklo_epi64(vksum230, vzero);
93 __m128i vacc0x3 = _mm_unpackhi_epi64(vksum230, vzero);
94 __m128i vacc1x2 = _mm_unpacklo_epi64(vksum231, vzero);
95 __m128i vacc1x3 = _mm_unpackhi_epi64(vksum231, vzero);
96 w = (const int32_t*) w + 4;
97
98 size_t k = kc;
99
100
101 while (k >= 8 * sizeof(int8_t)) {
102 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
103 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
104 a0 += 8;
105 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
106 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
107 a1 += 8;
108
109 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
110
111 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
112
113 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
114 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
115 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
116
117 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
118
119 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
120 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
121 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
122
123 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
124
125 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
126 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
127 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
128
129 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
130
131 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
132 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
133
134 w = (const int8_t*) w + 32;
135 k -= 8 * sizeof(int8_t);
136 }
137
138 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
139 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
140 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
141 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
142
143 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
144 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
145
146 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
147 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
148
149 const __m128i vinput_scale01 = _mm_loadu_si128((const __m128i*) &quantization_params[0]);
150 const __m128 vinput_scale0 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(1, 1, 1, 1)));
151 const __m128 vinput_scale1 = _mm_castsi128_ps(_mm_shuffle_epi32(vinput_scale01, _MM_SHUFFLE(3, 3, 3, 3)));
152
153 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
154 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale1);
155
156 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
157 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
158 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
159
160 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
161 w = (const float*) w + 8;
162 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
163 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
164
165 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
166 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
167
168 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
169 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
170
171 if XNN_LIKELY(nc >= 4) {
172 _mm_storeu_ps(c0, vout0x0123);
173 _mm_storeu_ps(c1, vout1x0123);
174
175 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
176 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
177
178 c0 = (float*) ((uintptr_t) c0 + cn_stride);
179 c1 = (float*) ((uintptr_t) c1 + cn_stride);
180
181 nc -= 4;
182 } else {
183 if (nc & 2) {
184 _mm_storel_pi((__m64*) c0, vout0x0123);
185 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
186 c0 += 2;
187 _mm_storel_pi((__m64*) c1, vout1x0123);
188 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
189 c1 += 2;
190 }
191 if (nc & 1) {
192 _mm_store_ss(c0, vout0x0123);
193 _mm_store_ss(c1, vout1x0123);
194 }
195 nc = 0;
196 }
197 } while (nc != 0);
198 }