Debian Codegraph @

src/f32-ppmm/gen/f32-ppmm-4x8-minmax-sse.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/f32-ppmm/sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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 <xmmintrin.h>
13
14 #include "xnnpack/ppmm.h"
15
16
17 void xnn_f32_ppmm_minmax_ukernel_4x8__sse(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const float* restrict a,
22 const float* restrict w,
23 float* restrict c,
24 size_t cm_stride,
25 size_t cn_stride,
26 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(mr != 0);
29 assert(mr <= 4);
30 assert(nc != 0);
31 assert(kc != 0);
32 assert(kc % sizeof(float) == 0);
33
34 float* c0 = c;
35 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
36 if XNN_UNPREDICTABLE(mr < 2) {
37 c1 = c0;
38 }
39 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
40 if XNN_UNPREDICTABLE(mr <= 2) {
41 c2 = c1;
42 }
43 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 4) {
45 c3 = c2;
46 }
47
48 const __m128 vmin = _mm_set1_ps(params->scalar.min);
49 const __m128 vmax = _mm_set1_ps(params->scalar.max);
50 XNN_FORCE_REALIZATION(vmin);
51 XNN_FORCE_REALIZATION(vmax);
52
53 do {
54 __m128 vacc0x0123 = _mm_load_ps(w);
55 __m128 vacc0x4567 = _mm_load_ps(w + 4);
56 __m128 vacc1x0123 = vacc0x0123;
57 __m128 vacc1x4567 = vacc0x4567;
58 __m128 vacc2x0123 = vacc0x0123;
59 __m128 vacc2x4567 = vacc0x4567;
60 __m128 vacc3x0123 = vacc0x0123;
61 __m128 vacc3x4567 = vacc0x4567;
62 w += 8;
63
64 size_t k = kc;
65 do {
66 const __m128 va0123 = _mm_load_ps(a);
67 a += 4;
68
69 const __m128 vb0123 = _mm_load_ps(w);
70 const __m128 vb4567 = _mm_load_ps(w + 4);
71 w += 8;
72
73 const __m128 va0000 = _mm_shuffle_ps(va0123, va0123, _MM_SHUFFLE(0, 0, 0, 0));
74 const __m128 va1111 = _mm_shuffle_ps(va0123, va0123, _MM_SHUFFLE(1, 1, 1, 1));
75 const __m128 va2222 = _mm_shuffle_ps(va0123, va0123, _MM_SHUFFLE(2, 2, 2, 2));
76 const __m128 va3333 = _mm_shuffle_ps(va0123, va0123, _MM_SHUFFLE(3, 3, 3, 3));
77
78 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0000, vb0123));
79 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1111, vb0123));
80 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2222, vb0123));
81 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3333, vb0123));
82 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0000, vb4567));
83 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1111, vb4567));
84 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2222, vb4567));
85 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3333, vb4567));
86
87 k -= sizeof(float);
88 } while (k != 0);
89
90 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
91 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
92 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
93 vacc3x0123 = _mm_min_ps(vacc3x0123, vmax);
94 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
95 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
96 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
97 vacc3x4567 = _mm_min_ps(vacc3x4567, vmax);
98
99 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
100 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
101 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
102 vacc3x0123 = _mm_max_ps(vacc3x0123, vmin);
103 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
104 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
105 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
106 vacc3x4567 = _mm_max_ps(vacc3x4567, vmin);
107
108 if XNN_LIKELY(nc >= 8) {
109 _mm_storeu_ps(c3, vacc3x0123);
110 _mm_storeu_ps(c3 + 4, vacc3x4567);
111 _mm_storeu_ps(c2, vacc2x0123);
112 _mm_storeu_ps(c2 + 4, vacc2x4567);
113 _mm_storeu_ps(c1, vacc1x0123);
114 _mm_storeu_ps(c1 + 4, vacc1x4567);
115 _mm_storeu_ps(c0, vacc0x0123);
116 _mm_storeu_ps(c0 + 4, vacc0x4567);
117
118 a = (const float*) ((uintptr_t) a - kc * 4);
119
120 c3 = (float*) ((uintptr_t) c3 + cn_stride);
121 c2 = (float*) ((uintptr_t) c2 + cn_stride);
122 c1 = (float*) ((uintptr_t) c1 + cn_stride);
123 c0 = (float*) ((uintptr_t) c0 + cn_stride);
124
125 nc -= 8;
126 } else {
127 if (nc & 4) {
128 _mm_storeu_ps(c3, vacc3x0123);
129 _mm_storeu_ps(c2, vacc2x0123);
130 _mm_storeu_ps(c1, vacc1x0123);
131 _mm_storeu_ps(c0, vacc0x0123);
132
133 vacc3x0123 = vacc3x4567;
134 vacc2x0123 = vacc2x4567;
135 vacc1x0123 = vacc1x4567;
136 vacc0x0123 = vacc0x4567;
137
138 c3 += 4;
139 c2 += 4;
140 c1 += 4;
141 c0 += 4;
142 }
143 if (nc & 2) {
144 _mm_storel_pi((__m64*) c3, vacc3x0123);
145 _mm_storel_pi((__m64*) c2, vacc2x0123);
146 _mm_storel_pi((__m64*) c1, vacc1x0123);
147 _mm_storel_pi((__m64*) c0, vacc0x0123);
148
149 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
150 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
151 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
152 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
153
154 c3 += 2;
155 c2 += 2;
156 c1 += 2;
157 c0 += 2;
158 }
159 if (nc & 1) {
160 _mm_store_ss(c3, vacc3x0123);
161 _mm_store_ss(c2, vacc2x0123);
162 _mm_store_ss(c1, vacc1x0123);
163 _mm_store_ss(c0, vacc0x0123);
164 }
165
166 nc = 0;
167 }
168 } while (nc != 0);
169 }