1 // Copyright 2021 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <xmmintrin.h>
7
8 #include <assert.h>
9
10 #include <xnnpack/common.h>
11 #include <xnnpack/math.h>
12 #include <xnnpack/transpose.h>
13
xnn_x32_transposec_ukernel__4x4_sse(const uint32_t * input,uint32_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)14 void xnn_x32_transposec_ukernel__4x4_sse(
15 const uint32_t* input,
16 uint32_t* output,
17 size_t input_stride,
18 size_t output_stride,
19 size_t block_width,
20 size_t block_height) XNN_OOB_READS
21 {
22 assert(output_stride >= block_height * sizeof(uint32_t));
23 assert(input_stride >= block_width * sizeof(uint32_t));
24
25 const size_t tile_height = 4;
26 const size_t tile_width = 4;
27 const size_t tile_wbytes = tile_width * sizeof(float);
28 const size_t input_vreset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
29 const size_t output_vreset = tile_height * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t);
30 const size_t input_offset = tile_height * input_stride;
31
32 const float* i0 = (const float*) input;
33 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
34 const float* i2 = (const float*) ((uintptr_t) i1 + input_stride);
35 const float* i3 = (const float*) ((uintptr_t) i2 + input_stride);
36
37 float* o0 = (float*) output;
38 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
39 float* o2 = (float*) ((uintptr_t) o1 + output_stride);
40 float* o3 = (float*) ((uintptr_t) o2 + output_stride);
41
42 do {
43 if XNN_UNPREDICTABLE(block_width < 2) {
44 o1 = o0;
45 }
46 if XNN_UNPREDICTABLE(block_width <= 2) {
47 o2 = o0;
48 }
49 if XNN_UNPREDICTABLE(block_width < 4) {
50 o3 = o0;
51 }
52 size_t bh = block_height;
53 for (; bh >= 4; bh -= 4) {
54 __m128 v0 = _mm_loadu_ps(i0);
55 i0 = (const float*) ((uintptr_t) i0 + input_offset);
56 __m128 v1 = _mm_loadu_ps(i1);
57 i1 = (const float*) ((uintptr_t) i1 + input_offset);
58 __m128 v2 = _mm_loadu_ps(i2);
59 i2 = (const float*) ((uintptr_t) i2 + input_offset);
60 __m128 v3 = _mm_loadu_ps(i3);
61 i3 = (const float*) ((uintptr_t) i3 + input_offset);
62
63 _MM_TRANSPOSE4_PS(v0, v1, v2, v3);
64
65 _mm_storeu_ps(o3, v3);
66 o3 = (float*) ((uintptr_t) o3 + tile_wbytes);
67 _mm_storeu_ps(o2, v2);
68 o2 = (float*) ((uintptr_t) o2 + tile_wbytes);
69 _mm_storeu_ps(o1, v1);
70 o1 = (float*) ((uintptr_t) o1 + tile_wbytes);
71 _mm_storeu_ps(o0, v0);
72 o0 = (float*) ((uintptr_t) o0 + tile_wbytes);
73 }
74
75 if (bh != 0) {
76 if XNN_UNPREDICTABLE(bh <= 2) {
77 i2 = i0;
78 }
79 if XNN_UNPREDICTABLE(bh < 2) {
80 i1 = i0;
81 }
82 __m128 v0 = _mm_loadu_ps(i0);
83 __m128 v1 = _mm_loadu_ps(i1);
84 __m128 v2 = _mm_loadu_ps(i2);
85 __m128 v3 = _mm_setzero_ps();
86
87 _MM_TRANSPOSE4_PS(v0, v1, v2, v3);
88
89 if (bh & 2) {
90 _mm_storel_pi((__m64*) o3, v3);
91 o3 += 2;
92 _mm_storel_pi((__m64*) o2, v2);
93 o2 += 2;
94 _mm_storel_pi((__m64*) o1, v1);
95 o1 += 2;
96 _mm_storel_pi((__m64*) o0, v0);
97 o0 += 2;
98 v0 = _mm_movehl_ps(v0, v0);
99 v1 = _mm_movehl_ps(v1, v1);
100 v2 = _mm_movehl_ps(v2, v2);
101 v3 = _mm_movehl_ps(v3, v3);
102 }
103 if (bh & 1) {
104 _mm_store_ss(o3, v3);
105 _mm_store_ss(o2, v2);
106 _mm_store_ss(o1, v1);
107 _mm_store_ss(o0, v0);
108 }
109 }
110 i0 = (const float*) ((uintptr_t) i0 + input_vreset);
111 i1 = (const float*) ((uintptr_t) i0 + input_stride);
112 i2 = (const float*) ((uintptr_t) i1 + input_stride);
113 i3 = (const float*) ((uintptr_t) i2 + input_stride);
114 o0 = (float*) ((uintptr_t) o0 + output_vreset);
115 o1 = (float*) ((uintptr_t) o1 + output_vreset);
116 o2 = (float*) ((uintptr_t) o2 + output_vreset);
117 o3 = (float*) ((uintptr_t) o3 + output_vreset);
118 block_width = doz(block_width, tile_width);
119 } while (block_width != 0);
120 }
121