Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-2x4c8-minmax-sse2-ld128.c

raw
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
17
18 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_2x4c8__sse2_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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 2);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (2 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr != 2) {
49 c1 = c0;
50 }
51
52 const __m128 vmin = _mm_set1_ps(params->scalar.min);
53 const __m128 vmax = _mm_set1_ps(params->scalar.max);
54 XNN_FORCE_REALIZATION(vmin);
55 XNN_FORCE_REALIZATION(vmax);
56 __m128i vinput_zero_point = _mm_cvtsi32_si128(*((const int*) &quantization_params->zero_point));
57 vinput_zero_point = _mm_shuffle_epi32(vinput_zero_point, _MM_SHUFFLE(0, 0, 0, 0));
58 const __m128 vinput_scale = _mm_load1_ps(&quantization_params->inv_scale);
59
60
61 do {
62 const __m128i vksum = _mm_load_si128((const __m128i*) w);
63 const __m128i vzero = _mm_setzero_si128();
64 const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
65 const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);
66
67 __m128i vzpprodksumhi0 = _mm_mulhi_epu16(vinput_zero_point, vksum_lo);
68 const __m128i vzpprodksumlo0 = _mm_mullo_epi16(vinput_zero_point, vksum_lo);
69 __m128i vzpprodksumhi1 = _mm_mulhi_epu16(vinput_zero_point, vksum_lo);
70 const __m128i vzpprodksumlo1 = _mm_mullo_epi16(vinput_zero_point, vksum_lo);
71
72 vzpprodksumhi0 = _mm_add_epi16(vzpprodksumhi0, _mm_mullo_epi16(vinput_zero_point, vksum_hi));
73 vzpprodksumhi0 = _mm_sub_epi16(vzpprodksumhi0, _mm_and_si128(_mm_srai_epi16(vinput_zero_point, 15), vksum_lo));
74 vzpprodksumhi1 = _mm_add_epi16(vzpprodksumhi1, _mm_mullo_epi16(vinput_zero_point, vksum_hi));
75 vzpprodksumhi1 = _mm_sub_epi16(vzpprodksumhi1, _mm_and_si128(_mm_srai_epi16(vinput_zero_point, 15), vksum_lo));
76
77 vzpprodksumhi0 = _mm_slli_si128(vzpprodksumhi0, 2);
78 vzpprodksumhi1 = _mm_slli_si128(vzpprodksumhi1, 2);
79
80 const __m128i vksumzp0 = _mm_or_si128(vzpprodksumhi0, vzpprodksumlo0);
81 const __m128i vksumzp1 = _mm_or_si128(vzpprodksumhi1, vzpprodksumlo1);
82
83 const __m128i vksum010 = _mm_unpacklo_epi32(vksumzp0, vzero);
84 const __m128i vksum230 = _mm_unpackhi_epi32(vksumzp0, vzero);
85 const __m128i vksum011 = _mm_unpacklo_epi32(vksumzp1, vzero);
86 const __m128i vksum231 = _mm_unpackhi_epi32(vksumzp1, vzero);
87
88 __m128i vacc0x0 = _mm_unpacklo_epi64(vksum010, vzero);
89 __m128i vacc0x1 = _mm_unpackhi_epi64(vksum010, vzero);
90 __m128i vacc1x0 = _mm_unpacklo_epi64(vksum011, vzero);
91 __m128i vacc1x1 = _mm_unpackhi_epi64(vksum011, vzero);
92
93 __m128i vacc0x2 = _mm_unpacklo_epi64(vksum230, vzero);
94 __m128i vacc0x3 = _mm_unpackhi_epi64(vksum230, vzero);
95 __m128i vacc1x2 = _mm_unpacklo_epi64(vksum231, vzero);
96 __m128i vacc1x3 = _mm_unpackhi_epi64(vksum231, vzero);
97 w = (const int32_t*) w + 4;
98
99 size_t p = ks;
100 do {
101 const int8_t* restrict a0 = a[0];
102 if XNN_UNPREDICTABLE(a0 != zero) {
103 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
104 } else {
105 a0 = zero_data;
106 }
107 const int8_t* restrict a1 = a[1];
108 if XNN_UNPREDICTABLE(a1 != zero) {
109 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
110 } else {
111 a1 = zero_data;
112 }
113 a += 2;
114
115 size_t k = 0;
116 while (k < kc) {
117 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
118 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
119 a0 += 8;
120 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
121 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
122 a1 += 8;
123
124 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
125 const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
126 const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
127 const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
128
129 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
130 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
131 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
132 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
133 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
134 const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
135 const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
136 const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
137
138 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
139 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
140 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
141 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
142
143 w = (const void*) ((const int8_t*) w + 32);
144 k += 8 * sizeof(int8_t);
145 }
146 p -= 2 * sizeof(void*);
147 } while (p != 0);
148
149 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
150 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
151 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
152 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
153
154 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
155 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
156
157 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
158 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
159
160 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale);
161 vout1x0123 = _mm_mul_ps(vout1x0123, vinput_scale);
162
163 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
164 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
165 vout1x0123 = _mm_mul_ps(vout1x0123, vfilter_output_scale0123);
166
167 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
168 w = (const float*) w + 8;
169 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
170 vout1x0123 = _mm_add_ps(vout1x0123, vbias0123);
171
172 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
173 vout1x0123 = _mm_max_ps(vout1x0123, vmin);
174
175 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
176 vout1x0123 = _mm_min_ps(vout1x0123, vmax);
177
178 if XNN_LIKELY(nc >= 4) {
179 _mm_storeu_ps(c1, vout1x0123);
180 _mm_storeu_ps(c0, vout0x0123);
181
182 a = (const int8_t**restrict) ((uintptr_t) a - ks);
183
184 c0 = (float*) ((uintptr_t) c0 + cn_stride);
185 c1 = (float*) ((uintptr_t) c1 + cn_stride);
186
187 nc -= 4;
188 } else {
189 if (nc & 2) {
190 _mm_storel_pi((__m64*) c1, vout1x0123);
191 vout1x0123 = _mm_unpackhi_ps(vout1x0123, vout1x0123);
192 c1 += 2;
193 _mm_storel_pi((__m64*) c0, vout0x0123);
194 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
195 c0 += 2;
196 }
197 if (nc & 1) {
198 _mm_store_ss(c1, vout1x0123);
199 _mm_store_ss(c0, vout0x0123);
200 }
201 nc = 0;
202 }
203 } while (nc != 0);
204 }