Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x4c8-minmax-sse41-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 <smmintrin.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__sse41_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 vinit0 = _mm_mullo_epi32(vksum, vinput_zero_point0);
60 __m128i vacc0x0 = _mm_blend_epi16(vinit0, vzero, 0xFC);
61 __m128i vacc0x1 = _mm_blend_epi16(vinit0, vzero, 0xF3);
62 __m128i vacc0x2 = _mm_blend_epi16(vinit0, vzero, 0xCF);
63 __m128i vacc0x3 = _mm_blend_epi16(vinit0, vzero, 0x3F);
64 w = (const int32_t*) w + 4;
65
66 size_t k = kc;
67
68 while (k >= 16 * sizeof(int8_t)) {
69 const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0);
70 const __m128i vxa0c0 = _mm_cvtepi8_epi16(va0c0);
71 a0 += 8;
72
73 const __m128i vb01c01 = _mm_load_si128((const __m128i*) w);
74 const __m128i vbs01c0 = _mm_slli_epi32(vb01c01, 4);
75 const __m128i vb01c0 = _mm_and_si128(vbs01c0, vmask);
76 const __m128i vsb01c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c0);
77 const __m128i vxb0c0 = _mm_unpacklo_epi8(vb01c0, vsb01c0);
78 const __m128i vxb1c0 = _mm_unpackhi_epi8(vb01c0, vsb01c0);
79
80 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0));
81 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0));
82 const __m128i vb23c01 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
83 const __m128i vbs23c0 = _mm_slli_epi32(vb23c01, 4);
84 const __m128i vb23c0 = _mm_and_si128(vbs23c0, vmask);
85 const __m128i vsb23c0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c0);
86 const __m128i vxb2c0 = _mm_unpacklo_epi8(vb23c0, vsb23c0);
87 const __m128i vxb3c0 = _mm_unpackhi_epi8(vb23c0, vsb23c0);
88
89 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0));
90 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0));
91
92 const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0);
93 const __m128i vxa0c1 = _mm_cvtepi8_epi16(va0c1);
94 a0 += 8;
95
96 const __m128i vb01c1 = _mm_and_si128(vb01c01, vmask);
97 const __m128i vsb01c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01c1);
98 const __m128i vxb0c1 = _mm_unpacklo_epi8(vb01c1, vsb01c1);
99 const __m128i vxb1c1 = _mm_unpackhi_epi8(vb01c1, vsb01c1);
100
101 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1));
102 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1));
103 const __m128i vb23c1 = _mm_and_si128(vb23c01, vmask);
104 const __m128i vsb23c1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23c1);
105 const __m128i vxb2c1 = _mm_unpacklo_epi8(vb23c1, vsb23c1);
106 const __m128i vxb3c1 = _mm_unpackhi_epi8(vb23c1, vsb23c1);
107
108 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1));
109 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1));
110
111
112 w = (const int8_t*) w + 32;
113 k -= 16 * sizeof(int8_t);
114 }
115
116 while (k >= 8 * sizeof(int8_t)) {
117 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
118 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
119 a0 += 8;
120
121 __m128i vb01 = _mm_load_si128((const __m128i*) w);
122 vb01 = _mm_slli_epi32(vb01, 4);
123 vb01 = _mm_and_si128(vb01, vmask);
124
125 const __m128i vxbm1 = _mm_unpackhi_epi8(vb01, vb01);
126 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
127 const __m128i vxb1 = _mm_srai_epi16(vxbm1, 8);
128
129 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
130 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
131 __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
132 vb23 = _mm_slli_epi32(vb23, 4);
133 vb23 = _mm_and_si128(vb23, vmask);
134
135 const __m128i vxbm3 = _mm_unpackhi_epi8(vb23, vb23);
136 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
137 const __m128i vxb3 = _mm_srai_epi16(vxbm3, 8);
138
139 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
140 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
141
142 w = (const int8_t*) w + 32;
143 k -= 8 * sizeof(int8_t);
144 }
145
146 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
147 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
148
149 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
150
151 vacc0x0123 = _mm_srai_epi32(vacc0x0123, 4);
152 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
153
154 const __m128 vinput_scale0 = _mm_load1_ps(&quantization_params[0].inv_scale);
155
156 vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0);
157
158 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
159 vout0x0123 = _mm_mul_ps(vout0x0123, vfilter_output_scale0123);
160
161 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
162 w = (const float*) w + 8;
163 vout0x0123 = _mm_add_ps(vout0x0123, vbias0123);
164
165 vout0x0123 = _mm_max_ps(vout0x0123, vmin);
166
167 vout0x0123 = _mm_min_ps(vout0x0123, vmax);
168
169 if XNN_LIKELY(nc >= 4) {
170 _mm_storeu_ps(c0, vout0x0123);
171
172 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
173
174 c0 = (float*) ((uintptr_t) c0 + cn_stride);
175
176 nc -= 4;
177 } else {
178 if (nc & 2) {
179 _mm_storel_pi((__m64*) c0, vout0x0123);
180 vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123);
181 c0 += 2;
182 }
183 if (nc & 1) {
184 _mm_store_ss(c0, vout0x0123);
185 }
186 nc = 0;
187 }
188 } while (nc != 0);
189 }