Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x4c8-minmax-sse2-ld128.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 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_1x4c8__sse2_ld128(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45
46 const __m128 vmin = _mm_set1_ps(params->scalar.min);
47 const __m128 vmax = _mm_set1_ps(params->scalar.max);
48 XNN_FORCE_REALIZATION(vmin);
49 XNN_FORCE_REALIZATION(vmax);
50
51 const __m128i vmask = _mm_set1_epi8(0xF0);
52 XNN_FORCE_REALIZATION(vmask);
53
54 do {
55 const __m128i vksum = _mm_load_si128((const __m128i*) w);
56 __m128i vinput_zero_point0 = _mm_cvtsi32_si128(*((const int*) &quantization_params[0].zero_point));
57 vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point0, _MM_SHUFFLE(0, 0, 0, 0));
58 const __m128i vzero = _mm_setzero_si128();
59 const __m128i vksum_lo = _mm_srli_epi32(_mm_slli_epi32(vksum, 16), 16);
60 const __m128i vksum_hi = _mm_srli_epi32(vksum, 16);
61
62 __m128i vzpprodksumhi0 = _mm_mulhi_epu16(vinput_zero_point0, vksum_lo);
63 const __m128i vzpprodksumlo0 = _mm_mullo_epi16(vinput_zero_point0, vksum_lo);
64
65 vzpprodksumhi0 = _mm_add_epi16(vzpprodksumhi0, _mm_mullo_epi16(vinput_zero_point0, vksum_hi));
66 vzpprodksumhi0 = _mm_sub_epi16(vzpprodksumhi0, _mm_and_si128(_mm_srai_epi16(vinput_zero_point0, 15), vksum_lo));
67
68 vzpprodksumhi0 = _mm_slli_si128(vzpprodksumhi0, 2);
69
70 const __m128i vksumzp0 = _mm_or_si128(vzpprodksumhi0, vzpprodksumlo0);
71
72 const __m128i vksum010 = _mm_unpacklo_epi32(vksumzp0, vzero);
73 const __m128i vksum230 = _mm_unpackhi_epi32(vksumzp0, vzero);
74
75 __m128i vacc0x0 = _mm_unpacklo_epi64(vksum010, vzero);
76 __m128i vacc0x1 = _mm_unpackhi_epi64(vksum010, vzero);
77
78 __m128i vacc0x2 = _mm_unpacklo_epi64(vksum230, vzero);
79 __m128i vacc0x3 = _mm_unpackhi_epi64(vksum230, vzero);
80 w = (const int32_t*) w + 4;
81
82 size_t k = kc;
83
84 while (k >= 16 * sizeof(int8_t)) {
85 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
86 const __m128i vxa0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c0, va0c0), 8);
87 a0 += 8;
88
89 const __m128i vb01c01 = _mm_load_si128((const __m128i*) w);
90 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
91 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
92 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
93 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
94 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
95
96 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
97 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
98 const __m128i vb23c01 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
99 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
100 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
101 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
102 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
103 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
104
105 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
106 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
107
108 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
109 const __m128i vxa0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c1, va0c1), 8);
110 a0 += 8;
111
112 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
113 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
114 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
115 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
116
117 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
118 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
119 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
120 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
121 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
122 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
123
124 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
125 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
126
127
128 w = (const int8_t*) w + 32;
129 k -= 16 * sizeof(int8_t);
130 }
131
132 while (k >= 8 * sizeof(int8_t)) {
133 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
134 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
135 a0 += 8;
136
137 __m128i vb01 = _mm_load_si128((const __m128i*) w);
138 vb01 = _mm_slli_epi32(vb01, 4);
139 vb01 = _mm_and_si128(vb01, vmask);
140
141 const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
142 const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
143 const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
144
145 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
146 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
147 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
148 vb23 = _mm_slli_epi32(vb23, 4);
149 vb23 = _mm_and_si128(vb23, vmask);
150
151 const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
152 const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
153 const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
154
155 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
156 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
157
158 w = (const int8_t*) w + 32;
159 k -= 8 * sizeof(int8_t);
160 }
161
162 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
163 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
164
165 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
166
167 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
168 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
169
170 const __m128 vinput_scale0 = _mm_load1_ps(&quantization_params[0].inv_scale);
171
172 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
173
174 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
175 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
176
177 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
178 w = (const float*) w + 8;
179 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
180
181 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
182
183 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
184
185 if XNN_LIKELY(nc >= 4) {
186 _mm_storeu_ps(c0, vout0x0123);
187
188 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
189
190 c0 = (float*) ((uintptr_t) c0 + cn_stride);
191
192 nc -= 4;
193 } else {
194 if (nc & 2) {
195 _mm_storel_pi((__m64*) c0, vout0x0123);
196 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
197 c0 += 2;
198 }
199 if (nc & 1) {
200 _mm_store_ss(c0, vout0x0123);
201 }
202 nc = 0;
203 }
204 } while (nc != 0);
205 }