Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x4c8-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 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_1x4c8__sse2_ld64(
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_qb4w_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 size_t bl = params->scalar.blocksize;
43 assert(bl <= round_up_po2(kc, 2));
44 assert(bl != 0);
45 assert(bl % 32 == 0);
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 const int8_t* a0 = a;
48 float* c0 = c;
49
50 const __m128 vmin = _mm_set1_ps(params->scalar.min);
51 const __m128 vmax = _mm_set1_ps(params->scalar.max);
52 XNN_FORCE_REALIZATION(vmin);
53 XNN_FORCE_REALIZATION(vmax);
54
55 const __m128i vmask = _mm_set1_epi8(0xF0);
56 XNN_FORCE_REALIZATION(vmask);
57
58 do {
59 const __m128 vksum = _mm_load_ps((const float*) w);
60 __m128i vinput_zero_point0 = _mm_cvtsi32_si128(*((const int*) &quantization_params[0].zero_point));
61 vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point0, _MM_SHUFFLE(0, 0, 0, 0));
62
63 __m128 vinput_zero_point0_float = _mm_cvtepi32_ps(vinput_zero_point0);
64 __m128 vout0x0123 = _mm_mul_ps(vksum, vinput_zero_point0_float);
65 w = (const int32_t*) w + 4;
66
67 for (size_t kb=0; kb < kc; kb += bl) {
68 __m128i vacc0x0 = _mm_setzero_si128();
69 __m128i vacc0x1 = _mm_setzero_si128();
70 __m128i vacc0x2 = _mm_setzero_si128();
71 __m128i vacc0x3 = _mm_setzero_si128();
72 size_t k = bl;
73
74 while (k >= 16 * sizeof(int8_t)) {
75 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
76 const __m128i vxa0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c0, va0c0), 8);
77 a0 += 8;
78
79 const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w);
80 const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4);
81 const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask);
82 const __m128i vxb0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c0, vb0c0), 8);
83
84 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
85 const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
86 const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4);
87 const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask);
88 const __m128i vxb1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c0, vb1c0), 8);
89
90 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
91 const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
92 const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4);
93 const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask);
94 const __m128i vxb2c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c0, vb2c0), 8);
95
96 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
97 const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
98 const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4);
99 const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask);
100 const __m128i vxb3c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c0, vb3c0), 8);
101
102 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
103
104 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
105 const __m128i vxa0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c1, va0c1), 8);
106 a0 += 8;
107
108 const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask);
109 const __m128i vxb0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c1, vb0c1), 8);
110
111 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
112 const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask);
113 const __m128i vxb1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c1, vb1c1), 8);
114
115 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
116 const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask);
117 const __m128i vxb2c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c1, vb2c1), 8);
118
119 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
120 const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask);
121 const __m128i vxb3c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c1, vb3c1), 8);
122
123 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
124
125
126 w = (const int8_t*) w + 32;
127 k -= 16 * sizeof(int8_t);
128 }
129
130 while (k >= 8 * sizeof(int8_t)) {
131 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
132 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
133 a0 += 8;
134
135 __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
136 vb0 = _mm_slli_epi32(vb0, 4);
137 vb0 = _mm_and_si128(vb0, vmask);
138
139 const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
140
141 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
142 __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
143 vb1 = _mm_slli_epi32(vb1, 4);
144 vb1 = _mm_and_si128(vb1, vmask);
145
146 const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
147
148 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
149 __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
150 vb2 = _mm_slli_epi32(vb2, 4);
151 vb2 = _mm_and_si128(vb2, vmask);
152
153 const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
154
155 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
156 __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
157 vb3 = _mm_slli_epi32(vb3, 4);
158 vb3 = _mm_and_si128(vb3, vmask);
159
160 const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
161
162 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
163
164 w = (const int8_t*) w + 32;
165 k -= 8 * sizeof(int8_t);
166 }
167 // accumulate float
168 const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(_mm_setzero_si128(), _mm_loadl_epi64((const __m128i*) w)));
169 w = (const uint16_t*) w + 4;
170
171 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
172 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
173
174 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
175
176 vout0x0123 = _mm_add_ps(vout0x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc0x0123), vfilter_output_scale0123));
177 }
178
179 const __m128 vinput_scale0 = _mm_load1_ps(&quantization_params[0].inv_scale);
180
181 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
182
183
184 const __m128 vbias0123 = _mm_loadu_ps((const float*) w);
185 w = (const float*) w + 4;
186 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
187
188 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
189
190 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
191
192 if XNN_LIKELY(nc >= 4) {
193 _mm_storeu_ps(c0, vout0x0123);
194
195 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
196
197 c0 = (float*) ((uintptr_t) c0 + cn_stride);
198
199 nc -= 4;
200 } else {
201 if (nc & 2) {
202 _mm_storel_pi((__m64*) c0, vout0x0123);
203 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
204 c0 += 2;
205 }
206 if (nc & 1) {
207 _mm_store_ss(c0, vout0x0123);
208 }
209 nc = 0;
210 }
211 } while (nc != 0);
212 }