1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-dwconv2d-chw/3x3p1-ssse3.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 <tmmintrin.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4(size_t input_height,size_t input_width,const float * input,const float * weights,const float * zero,float * output,uint32_t padding_top,const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4(
19     size_t input_height,
20     size_t input_width,
21     const float* input,
22     const float* weights,
23     const float* zero,
24     float* output,
25     uint32_t padding_top,
26     const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28   assert(input_height != 0);
29   assert(input_width != 0);
30   assert(input_width % sizeof(float) == 0);
31   assert(padding_top == 1);
32 
33   const __m128 vmask = _mm_load_ps((const float*) params->sse.mask);
34   const __m128 vmax = _mm_load_ps(params->sse.max);
35   const __m128 vmin = _mm_load_ps(params->sse.min);
36 
37   const __m128 vbias = _mm_load1_ps(weights);
38   const __m128 vk00 = _mm_load1_ps(weights + 1);
39   const __m128 vk01 = _mm_load1_ps(weights + 2);
40   const __m128 vk02 = _mm_load1_ps(weights + 3);
41   const __m128 vk10 = _mm_load1_ps(weights + 4);
42   const __m128 vk11 = _mm_load1_ps(weights + 5);
43   const __m128 vk12 = _mm_load1_ps(weights + 6);
44   const __m128 vk20 = _mm_load1_ps(weights + 7);
45   const __m128 vk21 = _mm_load1_ps(weights + 8);
46   const __m128 vk22 = _mm_load1_ps(weights + 9);
47 
48   const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
49 
50   const float* i0 = zero;
51   const float* i1 = input;
52   const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
53 
54   float* o0 = output;
55 
56   size_t output_height = input_height;
57   do {
58     if XNN_UNPREDICTABLE(output_height < 2) {
59       i2 = zero;
60     }
61 
62     __m128 vi0x0123 = _mm_setzero_ps();
63     __m128 vi1x0123 = _mm_setzero_ps();
64     __m128 vi2x0123 = _mm_setzero_ps();
65 
66     __m128 vi0x4567 = _mm_loadu_ps(i0);
67     i0 += 4;
68     __m128 vi1x4567 = _mm_loadu_ps(i1);
69     i1 += 4;
70     __m128 vi2x4567 = _mm_loadu_ps(i2);
71     i2 += 4;
72 
73     size_t w = input_width;
74     for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
75       const __m128 vi0x89AB = _mm_loadu_ps(i0);
76       i0 += 4;
77       const __m128 vi1x89AB = _mm_loadu_ps(i1);
78       i1 += 4;
79       const __m128 vi2x89AB = _mm_loadu_ps(i2);
80       i2 += 4;
81 
82       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
83       __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
84       __m128 vo0p2 = _mm_mul_ps(vi2x4567, vk21);
85 
86       const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
87       const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
88       const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
89 
90       __m128 vo0p3 = _mm_mul_ps(vi0x3456, vk00);
91       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
92       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
93 
94       vi0x0123 = vi0x4567;
95       vi1x0123 = vi1x4567;
96       vi2x0123 = vi2x4567;
97 
98       const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si128(vi0x4567), 4));
99       const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si128(vi1x4567), 4));
100       const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si128(vi2x4567), 4));
101 
102       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x5678, vk02));
103       vo0p3 = _mm_add_ps(vo0p3, _mm_mul_ps(vi1x5678, vk12));
104       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
105 
106       vi0x4567 = vi0x89AB;
107       vi1x4567 = vi1x89AB;
108       vi2x4567 = vi2x89AB;
109 
110       vo0p0 = _mm_add_ps(vo0p0, vo0p1);
111       vo0p2 = _mm_add_ps(vo0p2, vo0p3);
112       vo0p0 = _mm_add_ps(vo0p0, vo0p2);
113 
114       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
115 
116       vo0 = _mm_min_ps(vo0, vmax);
117 
118       _mm_storeu_ps(o0, vo0);
119       o0 += 4;
120     }
121     // Always process the last block of 1..4 pixels.
122     assert(w >= 1 * sizeof(float));
123     assert(w <= 4 * sizeof(float));
124     {
125       vi0x4567 = _mm_and_ps(vmask, vi0x4567);
126       vi1x4567 = _mm_and_ps(vmask, vi1x4567);
127       vi2x4567 = _mm_and_ps(vmask, vi2x4567);
128 
129       __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
130       __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
131       __m128 vo0p2 = _mm_mul_ps(vi2x4567, vk21);
132 
133       const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
134       const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
135       const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
136 
137       __m128 vo0p3 = _mm_mul_ps(vi0x3456, vk00);
138       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
139       vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
140 
141       const __m128i vzero = _mm_setzero_si128();
142       const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4));
143       const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi1x4567), 4));
144       const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi2x4567), 4));
145 
146       vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x5678, vk02));
147       vo0p3 = _mm_add_ps(vo0p3, _mm_mul_ps(vi1x5678, vk12));
148       vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
149 
150       vo0p0 = _mm_add_ps(vo0p0, vo0p1);
151       vo0p2 = _mm_add_ps(vo0p2, vo0p3);
152       vo0p0 = _mm_add_ps(vo0p0, vo0p2);
153 
154       __m128 vo0 = _mm_max_ps(vo0p0, vmin);
155 
156       vo0 = _mm_min_ps(vo0, vmax);
157 
158       if XNN_LIKELY(w == 4 * sizeof(float)) {
159         _mm_storeu_ps(o0, vo0);
160         o0 += 4;
161       } else {
162         if (w & (2 * sizeof(float))) {
163           _mm_storel_pi((__m64*) o0, vo0);
164           o0 += 2;
165 
166           vo0 = _mm_movehl_ps(vo0, vo0);
167         }
168         if (w & (1 * sizeof(float))) {
169           _mm_store_ss(o0, vo0);
170           o0 += 1;
171         }
172       }
173     }
174 
175     i0 = (const float*) ((uintptr_t) i1 - input_decrement);
176     i1 = (const float*) ((uintptr_t) i2 - input_decrement);
177     i2 = (const float*) ((uintptr_t) i1 + input_width);
178 
179 
180   } while (--output_height != 0);
181 }
182