1 // Auto-generated file. Do not edit!
2 // Template: src/x32-transposec/wasmsimd.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 <wasm_simd128.h>
11
12 #include <assert.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/transpose.h>
17
xnn_x32_transposec_ukernel__4x4_reuse_multi_wasmsimd(const uint32_t * input,uint32_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)18 void xnn_x32_transposec_ukernel__4x4_reuse_multi_wasmsimd(
19 const uint32_t* input,
20 uint32_t* output,
21 size_t input_stride,
22 size_t output_stride,
23 size_t block_width,
24 size_t block_height) XNN_OOB_READS
25 {
26 assert(output_stride >= block_height * sizeof(uint32_t));
27 assert(input_stride >= block_width * sizeof(uint32_t));
28
29 const size_t tile_height = 4;
30 const size_t tile_width = 4;
31 const size_t tile_hbytes = tile_height * sizeof(uint32_t);
32 const size_t tile_wbytes = tile_width * sizeof(uint32_t);
33 const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
34 const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t);
35
36 const uint32_t* i0 = input;
37 uint32_t* o0 = (uint32_t*) output;
38 uint32_t* o1 = (uint32_t*) ((uintptr_t) o0 + output_stride);
39 uint32_t* o2 = (uint32_t*) ((uintptr_t) o1 + output_stride);
40 uint32_t* o3 = (uint32_t*) ((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 const v128_t v2_0 = wasm_v128_load(i0);
55 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
56 const v128_t v2_1 = wasm_v128_load(i0);
57 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
58 const v128_t v2_2 = wasm_v128_load(i0);
59 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
60 const v128_t v2_3 = wasm_v128_load(i0);
61 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
62
63 const v128_t v1_0 = wasm_v32x4_shuffle(v2_0, v2_2, 0, 4, 1, 5);
64 const v128_t v1_1 = wasm_v32x4_shuffle(v2_0, v2_2, 2, 6, 3, 7);
65 const v128_t v1_2 = wasm_v32x4_shuffle(v2_1, v2_3, 0, 4, 1, 5);
66 const v128_t v1_3 = wasm_v32x4_shuffle(v2_1, v2_3, 2, 6, 3, 7);
67 const v128_t v0_0 = wasm_v32x4_shuffle(v1_0, v1_2, 0, 4, 1, 5);
68 const v128_t v0_1 = wasm_v32x4_shuffle(v1_0, v1_2, 2, 6, 3, 7);
69 const v128_t v0_2 = wasm_v32x4_shuffle(v1_1, v1_3, 0, 4, 1, 5);
70 const v128_t v0_3 = wasm_v32x4_shuffle(v1_1, v1_3, 2, 6, 3, 7);
71
72 wasm_v128_store(o3, v0_3);
73 o3 = (uint32_t*) ((uintptr_t) o3 + tile_hbytes);
74 wasm_v128_store(o2, v0_2);
75 o2 = (uint32_t*) ((uintptr_t) o2 + tile_hbytes);
76 wasm_v128_store(o1, v0_1);
77 o1 = (uint32_t*) ((uintptr_t) o1 + tile_hbytes);
78 wasm_v128_store(o0, v0_0);
79 o0 = (uint32_t*) ((uintptr_t) o0 + tile_hbytes);
80 }
81
82 if (bh != 0) {
83 const v128_t v2_0 = wasm_v128_load(i0);
84 const uint32_t *i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
85 if XNN_UNPREDICTABLE(bh < 2) {
86 i1 = i0;
87 }
88 const v128_t v2_1 = wasm_v128_load(i1);
89 const uint32_t *i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
90 if XNN_UNPREDICTABLE(bh <= 2) {
91 i2 = i1;
92 }
93 const v128_t v2_2 = wasm_v128_load(i2);
94 const v128_t v2_3 = wasm_v128_xor(v2_0, v2_0);
95
96 const v128_t v1_0 = wasm_v32x4_shuffle(v2_0, v2_2, 0, 4, 1, 5);
97 const v128_t v1_1 = wasm_v32x4_shuffle(v2_0, v2_2, 2, 6, 3, 7);
98 const v128_t v1_2 = wasm_v32x4_shuffle(v2_1, v2_3, 0, 4, 1, 5);
99 const v128_t v1_3 = wasm_v32x4_shuffle(v2_1, v2_3, 2, 6, 3, 7);
100
101 v128_t v0_0 = wasm_v32x4_shuffle(v1_0, v1_2, 0, 4, 1, 5);
102 v128_t v0_1 = wasm_v32x4_shuffle(v1_0, v1_2, 2, 6, 3, 7);
103 v128_t v0_2 = wasm_v32x4_shuffle(v1_1, v1_3, 0, 4, 1, 5);
104 v128_t v0_3 = wasm_v32x4_shuffle(v1_1, v1_3, 2, 6, 3, 7);
105
106 if (bh & 2) {
107 *((double*) o3) = wasm_f64x2_extract_lane(v0_3, 0);
108 o3 += 2;
109 *((double*) o2) = wasm_f64x2_extract_lane(v0_2, 0);
110 o2 += 2;
111 *((double*) o1) = wasm_f64x2_extract_lane(v0_1, 0);
112 o1 += 2;
113 *((double*) o0) = wasm_f64x2_extract_lane(v0_0, 0);
114 o0 += 2;
115 v0_0 = wasm_v64x2_shuffle(v0_0, v0_0, 1, 1);
116 v0_1 = wasm_v64x2_shuffle(v0_1, v0_1, 1, 1);
117 v0_2 = wasm_v64x2_shuffle(v0_2, v0_2, 1, 1);
118 v0_3 = wasm_v64x2_shuffle(v0_3, v0_3, 1, 1);
119 }
120
121 if (bh & 1) {
122 *((float*) o3) = wasm_f32x4_extract_lane(v0_3, 0);
123 *((float*) o2) = wasm_f32x4_extract_lane(v0_2, 0);
124 *((float*) o1) = wasm_f32x4_extract_lane(v0_1, 0);
125 *((float*) o0) = wasm_f32x4_extract_lane(v0_0, 0);
126 }
127 }
128
129 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
130 o0 = (uint32_t*) ((uintptr_t) o0 + output_reset);
131 o1 = (uint32_t*) ((uintptr_t) o1 + output_reset);
132 o2 = (uint32_t*) ((uintptr_t) o2 + output_reset);
133 o3 = (uint32_t*) ((uintptr_t) o3 + output_reset);
134 block_width = doz(block_width, tile_width);
135 } while (block_width != 0);
136 }
137