1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vmulcaddc/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/math.h>
15 #include <xnnpack/vmulcaddc.h>
16
17
xnn_f32_vmulcaddc_minmax_ukernel_c8__sse_2x(size_t rows,size_t channels,const float * restrict input,size_t input_stride,const float * restrict weights,float * restrict output,size_t output_stride,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vmulcaddc_minmax_ukernel_c8__sse_2x(
19 size_t rows,
20 size_t channels,
21 const float*restrict input,
22 size_t input_stride,
23 const float*restrict weights,
24 float*restrict output,
25 size_t output_stride,
26 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
27 {
28 assert(rows != 0);
29 assert(channels != 0);
30 assert(channels % sizeof(float) == 0);
31
32 const float* i0 = input;
33 float* o0 = output;
34 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
35 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
36 if XNN_UNPREDICTABLE(rows < 2) {
37 i1 = i0;
38 o1 = o0;
39 }
40
41 const size_t input_increment = input_stride * 2 - channels;
42 const size_t output_increment = output_stride * 2 - channels;
43
44 const __m128 vmin = _mm_load_ps(params->sse.min);
45 const __m128 vmax = _mm_load_ps(params->sse.max);
46 do {
47 const float* w = weights;
48 size_t c = channels;
49 for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
50 const __m128 vscale0123 = _mm_load_ps(w);
51 const __m128 vscale4567 = _mm_load_ps(w + 4);
52
53 __m128 vacc0x0123 = _mm_loadu_ps(i0);
54 __m128 vacc0x4567 = _mm_loadu_ps(i0 + 4);
55 i0 += 8;
56 __m128 vacc1x0123 = _mm_loadu_ps(i1);
57 __m128 vacc1x4567 = _mm_loadu_ps(i1 + 4);
58 i1 += 8;
59
60 vacc0x0123 = _mm_mul_ps(vacc0x0123, vscale0123);
61 vacc0x4567 = _mm_mul_ps(vacc0x4567, vscale4567);
62 vacc1x0123 = _mm_mul_ps(vacc1x0123, vscale0123);
63 vacc1x4567 = _mm_mul_ps(vacc1x4567, vscale4567);
64
65 const __m128 vbias0123 = _mm_load_ps(w + 8);
66 const __m128 vbias4567 = _mm_load_ps(w + 12);
67
68 vacc0x0123 = _mm_add_ps(vacc0x0123, vbias0123);
69 vacc0x4567 = _mm_add_ps(vacc0x4567, vbias4567);
70 vacc1x0123 = _mm_add_ps(vacc1x0123, vbias0123);
71 vacc1x4567 = _mm_add_ps(vacc1x4567, vbias4567);
72
73 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
74 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
75 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
76 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
77
78 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
79 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
80 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
81 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
82
83 _mm_storeu_ps(o0, vacc0x0123);
84 _mm_storeu_ps(o0 + 4, vacc0x4567);
85 o0 += 8;
86 _mm_storeu_ps(o1, vacc1x0123);
87 _mm_storeu_ps(o1 + 4, vacc1x4567);
88 o1 += 8;
89
90 w += 16;
91 }
92 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
93 const __m128 vscale0123 = _mm_load_ps(w);
94
95 __m128 vacc0x0123 = _mm_loadu_ps(i0);
96 i0 += 4;
97 __m128 vacc1x0123 = _mm_loadu_ps(i1);
98 i1 += 4;
99
100 vacc0x0123 = _mm_mul_ps(vacc0x0123, vscale0123);
101 vacc1x0123 = _mm_mul_ps(vacc1x0123, vscale0123);
102
103 const __m128 vbias0123 = _mm_load_ps(w + 8);
104
105 vacc0x0123 = _mm_add_ps(vacc0x0123, vbias0123);
106 vacc1x0123 = _mm_add_ps(vacc1x0123, vbias0123);
107
108 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
109 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
110
111 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
112 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
113
114 _mm_storeu_ps(o0, vacc0x0123);
115 o0 += 4;
116 _mm_storeu_ps(o1, vacc1x0123);
117 o1 += 4;
118
119 w += 4;
120 }
121 if XNN_UNLIKELY(c != 0) {
122 const __m128 vscale0123 = _mm_load_ps(w);
123
124 __m128 vacc0x0123 = _mm_loadu_ps(i0);
125 i0 = (const float*) ((uintptr_t) i0 + c);
126 __m128 vacc1x0123 = _mm_loadu_ps(i1);
127 i1 = (const float*) ((uintptr_t) i1 + c);
128
129 vacc0x0123 = _mm_mul_ps(vacc0x0123, vscale0123);
130 vacc1x0123 = _mm_mul_ps(vacc1x0123, vscale0123);
131
132 const __m128 vbias0123 = _mm_load_ps(w + 8);
133
134 vacc0x0123 = _mm_add_ps(vacc0x0123, vbias0123);
135 vacc1x0123 = _mm_add_ps(vacc1x0123, vbias0123);
136
137 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
138 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
139
140 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
141 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
142
143 if (c & (2 * sizeof(float))) {
144 _mm_storel_pi((__m64*) o0, vacc0x0123);
145 _mm_storel_pi((__m64*) o1, vacc1x0123);
146
147 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
148 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
149
150 o0 += 2;
151 o1 += 2;
152 }
153 if (c & (1 * sizeof(float))) {
154 _mm_store_ss(o0, vacc0x0123);
155 _mm_store_ss(o1, vacc1x0123);
156
157 o0 += 1;
158 o1 += 1;
159 }
160 }
161 i0 = (const float*) ((uintptr_t) i0 + input_increment);
162 o0 = (float*) ((uintptr_t) o0 + output_increment);
163 i1 = (const float*) ((uintptr_t) i1 + input_increment);
164 o1 = (float*) ((uintptr_t) o1 + output_increment);
165 if XNN_UNPREDICTABLE(rows < 4) {
166 i1 = i0;
167 o1 = o0;
168 }
169 rows = doz(rows, 2);
170 } while (rows != 0);
171 }
172