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_mov_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_mov_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) - tile_hbytes;
35
36 const uint32_t* i0 = input;
37 uint32_t* o = (uint32_t*) ((uintptr_t) output - tile_hbytes);
38 const size_t minus_output_stride = -output_stride;
39
40 do {
41 const size_t rem = min(block_width - 1, 3);
42 const size_t oN_stride = rem * output_stride;
43 const size_t oN_offset = oN_stride + tile_hbytes;
44 size_t bh = block_height;
45 for (; bh >= 4; bh -= 4) {
46 const v128_t v2_0 = wasm_v128_load(i0);
47 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
48 const v128_t v2_1 = wasm_v128_load(i0);
49 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
50 const v128_t v2_2 = wasm_v128_load(i0);
51 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
52 const v128_t v2_3 = wasm_v128_load(i0);
53 i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
54
55 const v128_t v1_0 = wasm_v32x4_shuffle(v2_0, v2_2, 0, 4, 1, 5);
56 const v128_t v1_1 = wasm_v32x4_shuffle(v2_0, v2_2, 2, 6, 3, 7);
57 const v128_t v1_2 = wasm_v32x4_shuffle(v2_1, v2_3, 0, 4, 1, 5);
58 const v128_t v1_3 = wasm_v32x4_shuffle(v2_1, v2_3, 2, 6, 3, 7);
59 const v128_t v0_0 = wasm_v32x4_shuffle(v1_0, v1_2, 0, 4, 1, 5);
60 const v128_t v0_1 = wasm_v32x4_shuffle(v1_0, v1_2, 2, 6, 3, 7);
61 const v128_t v0_2 = wasm_v32x4_shuffle(v1_1, v1_3, 0, 4, 1, 5);
62 const v128_t v0_3 = wasm_v32x4_shuffle(v1_1, v1_3, 2, 6, 3, 7);
63
64 o = (uint32_t*) ((uintptr_t) o + oN_offset);
65 wasm_v128_store(o, v0_3);
66 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
67 if XNN_UNPREDICTABLE(block_width > 3) {
68 o = oN;
69 }
70 wasm_v128_store(o, v0_2);
71 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
72 if XNN_UNPREDICTABLE(block_width >= 3) {
73 o = oN;
74 }
75 wasm_v128_store(o, v0_1);
76 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
77 if XNN_UNPREDICTABLE(block_width > 1) {
78 o = oN;
79 }
80 wasm_v128_store(o, v0_0);
81 }
82 o = (uint32_t*) ((uintptr_t) o + tile_hbytes);
83
84 if (bh != 0) {
85 const v128_t v2_0 = wasm_v128_load(i0);
86 const uint32_t *i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
87 if XNN_UNPREDICTABLE(bh < 2) {
88 i1 = i0;
89 }
90 const v128_t v2_1 = wasm_v128_load(i1);
91 const uint32_t *i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
92 if XNN_UNPREDICTABLE(bh <= 2) {
93 i2 = i1;
94 }
95 const v128_t v2_2 = wasm_v128_load(i2);
96 const v128_t v2_3 = wasm_v128_xor(v2_0, v2_0);
97
98 const v128_t v1_0 = wasm_v32x4_shuffle(v2_0, v2_2, 0, 4, 1, 5);
99 const v128_t v1_1 = wasm_v32x4_shuffle(v2_0, v2_2, 2, 6, 3, 7);
100 const v128_t v1_2 = wasm_v32x4_shuffle(v2_1, v2_3, 0, 4, 1, 5);
101 const v128_t v1_3 = wasm_v32x4_shuffle(v2_1, v2_3, 2, 6, 3, 7);
102
103 v128_t v0_0 = wasm_v32x4_shuffle(v1_0, v1_2, 0, 4, 1, 5);
104 v128_t v0_1 = wasm_v32x4_shuffle(v1_0, v1_2, 2, 6, 3, 7);
105 v128_t v0_2 = wasm_v32x4_shuffle(v1_1, v1_3, 0, 4, 1, 5);
106 v128_t v0_3 = wasm_v32x4_shuffle(v1_1, v1_3, 2, 6, 3, 7);
107
108 if (bh & 2) {
109 o = (uint32_t*) ((uintptr_t) o + oN_stride);
110 *((double*) o) = wasm_f64x2_extract_lane(v0_3, 0);
111 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
112 if XNN_UNPREDICTABLE(block_width > 3) {
113 o = oN;
114 }
115 *((double*) o) = wasm_f64x2_extract_lane(v0_2, 0);
116 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
117 if XNN_UNPREDICTABLE(block_width >= 3) {
118 o = oN;
119 }
120 *((double*) o) = wasm_f64x2_extract_lane(v0_1, 0);
121 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
122 if XNN_UNPREDICTABLE(block_width > 1) {
123 o = oN;
124 }
125 *((double*) o) = wasm_f64x2_extract_lane(v0_0, 0);
126 o += 2;
127 v0_0 = wasm_v64x2_shuffle(v0_0, v0_0, 1, 1);
128 v0_1 = wasm_v64x2_shuffle(v0_1, v0_1, 1, 1);
129 v0_2 = wasm_v64x2_shuffle(v0_2, v0_2, 1, 1);
130 v0_3 = wasm_v64x2_shuffle(v0_3, v0_3, 1, 1);
131 }
132
133 if (bh & 1) {
134 o = (uint32_t*) ((uintptr_t) o + oN_stride);
135 *((float*) o) = wasm_f32x4_extract_lane(v0_3, 0);
136 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
137 if XNN_UNPREDICTABLE(block_width > 3) {
138 o = oN;
139 }
140 *((float*) o) = wasm_f32x4_extract_lane(v0_2, 0);
141 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
142 if XNN_UNPREDICTABLE(block_width >= 3) {
143 o = oN;
144 }
145 *((float*) o) = wasm_f32x4_extract_lane(v0_1, 0);
146 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
147 if XNN_UNPREDICTABLE(block_width > 1) {
148 o = oN;
149 }
150 *((float*) o) = wasm_f32x4_extract_lane(v0_0, 0);
151 }
152 }
153
154 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
155 o = (uint32_t*) ((uintptr_t) o + output_reset);
156 block_width = doz(block_width, tile_width);
157 } while (block_width != 0);
158 }
159