Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-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_qb4w_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_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 vb01c01 = _mm_loadu_si128((const __m128i*) w);
80 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
81 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
82 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
83 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
84 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
85
86 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
87 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
88 const __m128i vb23c01 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
89 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
90 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
91 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
92 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
93 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
94
95 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
96 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
97
98 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
99 const __m128i vxa0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c1, va0c1), 8);
100 a0 += 8;
101
102 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
103 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
104 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
105 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
106
107 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
108 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
109 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
110 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
111 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
112 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
113
114 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
115 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
116
117
118 w = (const int8_t*) w + 32;
119 k -= 16 * sizeof(int8_t);
120 }
121
122 while (k >= 8 * sizeof(int8_t)) {
123 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
124 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
125 a0 += 8;
126
127 __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
128 vb01 = _mm_slli_epi32(vb01, 4);
129 vb01 = _mm_and_si128(vb01, vmask);
130
131 const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
132 const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
133 const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
134
135 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
136 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
137 __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
138 vb23 = _mm_slli_epi32(vb23, 4);
139 vb23 = _mm_and_si128(vb23, vmask);
140
141 const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
142 const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
143 const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
144
145 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
146 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
147
148 w = (const int8_t*) w + 32;
149 k -= 8 * sizeof(int8_t);
150 }
151 // accumulate float
152 const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(_mm_setzero_si128(), _mm_loadl_epi64((const __m128i*) w)));
153 w = (const uint16_t*) w + 4;
154
155 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
156 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
157
158 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
159
160 vout0x0123 = _mm_add_ps(vout0x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc0x0123), vfilter_output_scale0123));
161 }
162
163 const __m128 vinput_scale0 = _mm_load1_ps(&quantization_params[0].inv_scale);
164
165 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
166
167
168 const __m128 vbias0123 = _mm_loadu_ps((const float*) w);
169 w = (const float*) w + 4;
170 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
171
172 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
173
174 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
175
176 if XNN_LIKELY(nc >= 4) {
177 _mm_storeu_ps(c0, vout0x0123);
178
179 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
180
181 c0 = (float*) ((uintptr_t) c0 + cn_stride);
182
183 nc -= 4;
184 } else {
185 if (nc & 2) {
186 _mm_storel_pi((__m64*) c0, vout0x0123);
187 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
188 c0 += 2;
189 }
190 if (nc & 1) {
191 _mm_store_ss(c0, vout0x0123);
192 }
193 nc = 0;
194 }
195 } while (nc != 0);
196 }