1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/f32-dwconv2d-chw/3x3s2p1-sse.c.in
3*4bdc9457SAndroid Build Coastguard Worker // Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker #include <xmmintrin.h>
13*4bdc9457SAndroid Build Coastguard Worker
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/dwconv.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
16*4bdc9457SAndroid Build Coastguard Worker
17*4bdc9457SAndroid Build Coastguard Worker
xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_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*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4(
19*4bdc9457SAndroid Build Coastguard Worker size_t input_height,
20*4bdc9457SAndroid Build Coastguard Worker size_t input_width,
21*4bdc9457SAndroid Build Coastguard Worker const float* input,
22*4bdc9457SAndroid Build Coastguard Worker const float* weights,
23*4bdc9457SAndroid Build Coastguard Worker const float* zero,
24*4bdc9457SAndroid Build Coastguard Worker float* output,
25*4bdc9457SAndroid Build Coastguard Worker uint32_t padding_top,
26*4bdc9457SAndroid Build Coastguard Worker const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27*4bdc9457SAndroid Build Coastguard Worker {
28*4bdc9457SAndroid Build Coastguard Worker assert(input_height != 0);
29*4bdc9457SAndroid Build Coastguard Worker assert(input_width != 0);
30*4bdc9457SAndroid Build Coastguard Worker assert(input_width % sizeof(float) == 0);
31*4bdc9457SAndroid Build Coastguard Worker assert(padding_top >= 0);
32*4bdc9457SAndroid Build Coastguard Worker assert(padding_top <= 1);
33*4bdc9457SAndroid Build Coastguard Worker
34*4bdc9457SAndroid Build Coastguard Worker const __m128 vmask_even = _mm_load_ps((const float*) params->sse.mask_even);
35*4bdc9457SAndroid Build Coastguard Worker const __m128 vmask_odd = _mm_load_ps((const float*) params->sse.mask_odd);
36*4bdc9457SAndroid Build Coastguard Worker const __m128 vmax = _mm_load_ps(params->sse.max);
37*4bdc9457SAndroid Build Coastguard Worker const __m128 vmin = _mm_load_ps(params->sse.min);
38*4bdc9457SAndroid Build Coastguard Worker
39*4bdc9457SAndroid Build Coastguard Worker const __m128 vbias = _mm_load1_ps(weights);
40*4bdc9457SAndroid Build Coastguard Worker const __m128 vk00 = _mm_load1_ps(weights + 1);
41*4bdc9457SAndroid Build Coastguard Worker const __m128 vk01 = _mm_load1_ps(weights + 2);
42*4bdc9457SAndroid Build Coastguard Worker const __m128 vk02 = _mm_load1_ps(weights + 3);
43*4bdc9457SAndroid Build Coastguard Worker const __m128 vk10 = _mm_load1_ps(weights + 4);
44*4bdc9457SAndroid Build Coastguard Worker const __m128 vk11 = _mm_load1_ps(weights + 5);
45*4bdc9457SAndroid Build Coastguard Worker const __m128 vk12 = _mm_load1_ps(weights + 6);
46*4bdc9457SAndroid Build Coastguard Worker const __m128 vk20 = _mm_load1_ps(weights + 7);
47*4bdc9457SAndroid Build Coastguard Worker const __m128 vk21 = _mm_load1_ps(weights + 8);
48*4bdc9457SAndroid Build Coastguard Worker const __m128 vk22 = _mm_load1_ps(weights + 9);
49*4bdc9457SAndroid Build Coastguard Worker
50*4bdc9457SAndroid Build Coastguard Worker const size_t input_decrement = round_down_po2(input_width, 4 /* SIMD output width */ * 2 /* subsampling */ * sizeof(float));
51*4bdc9457SAndroid Build Coastguard Worker
52*4bdc9457SAndroid Build Coastguard Worker const float* i0 = (const float*) ((uintptr_t) input - ((-padding_top) & input_width));
53*4bdc9457SAndroid Build Coastguard Worker const float* i1 = (const float*) ((uintptr_t) i0 + input_width);
54*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(padding_top != 0) {
55*4bdc9457SAndroid Build Coastguard Worker i0 = zero;
56*4bdc9457SAndroid Build Coastguard Worker }
57*4bdc9457SAndroid Build Coastguard Worker const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
58*4bdc9457SAndroid Build Coastguard Worker
59*4bdc9457SAndroid Build Coastguard Worker float* o0 = output;
60*4bdc9457SAndroid Build Coastguard Worker
61*4bdc9457SAndroid Build Coastguard Worker size_t padded_input_height = input_height + padding_top + 1 /* padding bottom */;
62*4bdc9457SAndroid Build Coastguard Worker size_t output_height = (padded_input_height - 3 /* kernel size */ + 2 /* subsampling */) / 2;
63*4bdc9457SAndroid Build Coastguard Worker do {
64*4bdc9457SAndroid Build Coastguard Worker if XNN_UNPREDICTABLE(padded_input_height < 4) {
65*4bdc9457SAndroid Build Coastguard Worker i2 = zero;
66*4bdc9457SAndroid Build Coastguard Worker }
67*4bdc9457SAndroid Build Coastguard Worker
68*4bdc9457SAndroid Build Coastguard Worker __m128 vi0x7531 = _mm_setzero_ps();
69*4bdc9457SAndroid Build Coastguard Worker __m128 vi1x7531 = _mm_setzero_ps();
70*4bdc9457SAndroid Build Coastguard Worker __m128 vi2x7531 = _mm_setzero_ps();
71*4bdc9457SAndroid Build Coastguard Worker
72*4bdc9457SAndroid Build Coastguard Worker size_t w = input_width;
73*4bdc9457SAndroid Build Coastguard Worker for (; w >= 8 * sizeof(float); w -= 8 * sizeof(float)) {
74*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0x89AB = _mm_loadu_ps(i0);
75*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0xCDEF = _mm_loadu_ps(i0 + 4);
76*4bdc9457SAndroid Build Coastguard Worker i0 += 8;
77*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1x89AB = _mm_loadu_ps(i1);
78*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1xCDEF = _mm_loadu_ps(i1 + 4);
79*4bdc9457SAndroid Build Coastguard Worker i1 += 8;
80*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2x89AB = _mm_loadu_ps(i2);
81*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2xCDEF = _mm_loadu_ps(i2 + 4);
82*4bdc9457SAndroid Build Coastguard Worker i2 += 8;
83*4bdc9457SAndroid Build Coastguard Worker
84*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0x8ACE = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0));
85*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0x9BDF = _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(3, 1, 3, 1));
86*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1x8ACE = _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(2, 0, 2, 0));
87*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1x9BDF = _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1, 3, 1));
88*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2x8ACE = _mm_shuffle_ps(vi2x89AB, vi2xCDEF, _MM_SHUFFLE(2, 0, 2, 0));
89*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2x9BDF = _mm_shuffle_ps(vi2x89AB, vi2xCDEF, _MM_SHUFFLE(3, 1, 3, 1));
90*4bdc9457SAndroid Build Coastguard Worker
91*4bdc9457SAndroid Build Coastguard Worker __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk01));
92*4bdc9457SAndroid Build Coastguard Worker __m128 vo0p1 = _mm_mul_ps(vi1x8ACE, vk11);
93*4bdc9457SAndroid Build Coastguard Worker __m128 vo0p2 = _mm_mul_ps(vi2x8ACE, vk21);
94*4bdc9457SAndroid Build Coastguard Worker
95*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
96*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
97*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2xF9BD = _mm_shuffle_ps(vi2x9BDF, vi2x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
98*4bdc9457SAndroid Build Coastguard Worker
99*4bdc9457SAndroid Build Coastguard Worker __m128 vo0p3 = _mm_mul_ps(vi0x9BDF, vk02);
100*4bdc9457SAndroid Build Coastguard Worker vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk12));
101*4bdc9457SAndroid Build Coastguard Worker vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x9BDF, vk22));
102*4bdc9457SAndroid Build Coastguard Worker
103*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0x7BDF = _mm_move_ss(vi0xF9BD, vi0x7531);
104*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1x7BDF = _mm_move_ss(vi1xF9BD, vi1x7531);
105*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2x7BDF = _mm_move_ss(vi2xF9BD, vi2x7531);
106*4bdc9457SAndroid Build Coastguard Worker
107*4bdc9457SAndroid Build Coastguard Worker vi0x7531 = vi0xF9BD;
108*4bdc9457SAndroid Build Coastguard Worker vi1x7531 = vi1xF9BD;
109*4bdc9457SAndroid Build Coastguard Worker vi2x7531 = vi2xF9BD;
110*4bdc9457SAndroid Build Coastguard Worker
111*4bdc9457SAndroid Build Coastguard Worker vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x7BDF, vk00));
112*4bdc9457SAndroid Build Coastguard Worker vo0p3 = _mm_add_ps(vo0p3, _mm_mul_ps(vi1x7BDF, vk10));
113*4bdc9457SAndroid Build Coastguard Worker vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x7BDF, vk20));
114*4bdc9457SAndroid Build Coastguard Worker
115*4bdc9457SAndroid Build Coastguard Worker vo0p0 = _mm_add_ps(vo0p0, vo0p1);
116*4bdc9457SAndroid Build Coastguard Worker vo0p2 = _mm_add_ps(vo0p2, vo0p3);
117*4bdc9457SAndroid Build Coastguard Worker vo0p0 = _mm_add_ps(vo0p0, vo0p2);
118*4bdc9457SAndroid Build Coastguard Worker
119*4bdc9457SAndroid Build Coastguard Worker __m128 vo0 = _mm_max_ps(vo0p0, vmin);
120*4bdc9457SAndroid Build Coastguard Worker
121*4bdc9457SAndroid Build Coastguard Worker vo0 = _mm_min_ps(vo0, vmax);
122*4bdc9457SAndroid Build Coastguard Worker
123*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_ps(o0, vo0);
124*4bdc9457SAndroid Build Coastguard Worker o0 += 4;
125*4bdc9457SAndroid Build Coastguard Worker }
126*4bdc9457SAndroid Build Coastguard Worker // Potentially process the last block of 0..7 pixels.
127*4bdc9457SAndroid Build Coastguard Worker assert(w < 8 * sizeof(float));
128*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(w != 0) {
129*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0x89AB = _mm_loadu_ps(i0);
130*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0xCDEF = _mm_loadu_ps(i0 + 4);
131*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1x89AB = _mm_loadu_ps(i1);
132*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1xCDEF = _mm_loadu_ps(i1 + 4);
133*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2x89AB = _mm_loadu_ps(i2);
134*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2xCDEF = _mm_loadu_ps(i2 + 4);
135*4bdc9457SAndroid Build Coastguard Worker
136*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0x8ACE = _mm_and_ps(vmask_even, _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(2, 0, 2, 0)));
137*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi0x89AB, vi0xCDEF, _MM_SHUFFLE(3, 1, 3, 1)));
138*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1x8ACE = _mm_and_ps(vmask_even, _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(2, 0, 2, 0)));
139*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1, 3, 1)));
140*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2x8ACE = _mm_and_ps(vmask_even, _mm_shuffle_ps(vi2x89AB, vi2xCDEF, _MM_SHUFFLE(2, 0, 2, 0)));
141*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi2x89AB, vi2xCDEF, _MM_SHUFFLE(3, 1, 3, 1)));
142*4bdc9457SAndroid Build Coastguard Worker
143*4bdc9457SAndroid Build Coastguard Worker __m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x8ACE, vk01));
144*4bdc9457SAndroid Build Coastguard Worker __m128 vo0p1 = _mm_mul_ps(vi1x8ACE, vk11);
145*4bdc9457SAndroid Build Coastguard Worker __m128 vo0p2 = _mm_mul_ps(vi2x8ACE, vk21);
146*4bdc9457SAndroid Build Coastguard Worker
147*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0xF9BD = _mm_shuffle_ps(vi0x9BDF, vi0x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
148*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
149*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2xF9BD = _mm_shuffle_ps(vi2x9BDF, vi2x9BDF, _MM_SHUFFLE(2, 1, 0, 3));
150*4bdc9457SAndroid Build Coastguard Worker
151*4bdc9457SAndroid Build Coastguard Worker __m128 vo0p3 = _mm_mul_ps(vi0x9BDF, vk02);
152*4bdc9457SAndroid Build Coastguard Worker vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk12));
153*4bdc9457SAndroid Build Coastguard Worker vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x9BDF, vk22));
154*4bdc9457SAndroid Build Coastguard Worker
155*4bdc9457SAndroid Build Coastguard Worker const __m128 vi0x7BDF = _mm_move_ss(vi0xF9BD, vi0x7531);
156*4bdc9457SAndroid Build Coastguard Worker const __m128 vi1x7BDF = _mm_move_ss(vi1xF9BD, vi1x7531);
157*4bdc9457SAndroid Build Coastguard Worker const __m128 vi2x7BDF = _mm_move_ss(vi2xF9BD, vi2x7531);
158*4bdc9457SAndroid Build Coastguard Worker
159*4bdc9457SAndroid Build Coastguard Worker vi0x7531 = vi0xF9BD;
160*4bdc9457SAndroid Build Coastguard Worker vi1x7531 = vi1xF9BD;
161*4bdc9457SAndroid Build Coastguard Worker vi2x7531 = vi2xF9BD;
162*4bdc9457SAndroid Build Coastguard Worker
163*4bdc9457SAndroid Build Coastguard Worker vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x7BDF, vk00));
164*4bdc9457SAndroid Build Coastguard Worker vo0p3 = _mm_add_ps(vo0p3, _mm_mul_ps(vi1x7BDF, vk10));
165*4bdc9457SAndroid Build Coastguard Worker vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x7BDF, vk20));
166*4bdc9457SAndroid Build Coastguard Worker
167*4bdc9457SAndroid Build Coastguard Worker vo0p0 = _mm_add_ps(vo0p0, vo0p1);
168*4bdc9457SAndroid Build Coastguard Worker vo0p2 = _mm_add_ps(vo0p2, vo0p3);
169*4bdc9457SAndroid Build Coastguard Worker vo0p0 = _mm_add_ps(vo0p0, vo0p2);
170*4bdc9457SAndroid Build Coastguard Worker
171*4bdc9457SAndroid Build Coastguard Worker __m128 vo0 = _mm_max_ps(vo0p0, vmin);
172*4bdc9457SAndroid Build Coastguard Worker
173*4bdc9457SAndroid Build Coastguard Worker vo0 = _mm_min_ps(vo0, vmax);
174*4bdc9457SAndroid Build Coastguard Worker
175*4bdc9457SAndroid Build Coastguard Worker if (w == 7 * sizeof(float)) {
176*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_ps(o0, vo0);
177*4bdc9457SAndroid Build Coastguard Worker o0 += 4;
178*4bdc9457SAndroid Build Coastguard Worker } else {
179*4bdc9457SAndroid Build Coastguard Worker w += 1 * sizeof(float);
180*4bdc9457SAndroid Build Coastguard Worker if (w & (4 * sizeof(float))) {
181*4bdc9457SAndroid Build Coastguard Worker _mm_storel_pi((__m64*) o0, vo0);
182*4bdc9457SAndroid Build Coastguard Worker o0 += 2;
183*4bdc9457SAndroid Build Coastguard Worker
184*4bdc9457SAndroid Build Coastguard Worker vo0 = _mm_movehl_ps(vo0, vo0);
185*4bdc9457SAndroid Build Coastguard Worker }
186*4bdc9457SAndroid Build Coastguard Worker if (w & (2 * sizeof(float))) {
187*4bdc9457SAndroid Build Coastguard Worker _mm_store_ss(o0, vo0);
188*4bdc9457SAndroid Build Coastguard Worker o0 += 1;
189*4bdc9457SAndroid Build Coastguard Worker }
190*4bdc9457SAndroid Build Coastguard Worker }
191*4bdc9457SAndroid Build Coastguard Worker }
192*4bdc9457SAndroid Build Coastguard Worker
193*4bdc9457SAndroid Build Coastguard Worker i0 = (const float*) ((uintptr_t) i2 - input_decrement);
194*4bdc9457SAndroid Build Coastguard Worker i1 = (const float*) ((uintptr_t) i0 + input_width);
195*4bdc9457SAndroid Build Coastguard Worker i2 = (const float*) ((uintptr_t) i1 + input_width);
196*4bdc9457SAndroid Build Coastguard Worker
197*4bdc9457SAndroid Build Coastguard Worker
198*4bdc9457SAndroid Build Coastguard Worker output_height -= 1;
199*4bdc9457SAndroid Build Coastguard Worker padded_input_height -= 2;
200*4bdc9457SAndroid Build Coastguard Worker } while (output_height != 0);
201*4bdc9457SAndroid Build Coastguard Worker }
202