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_multi_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_multi_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 input_offset = tile_height * input_stride;
35 const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t) - tile_hbytes;
36
37 const uint32_t* i0 = input;
38 const uint32_t* i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
39 const uint32_t* i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
40 const uint32_t* i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
41 uint32_t* o = (uint32_t*) ((uintptr_t) output - tile_hbytes);
42 const size_t minus_output_stride = -output_stride;
43
44 do {
45 const size_t rem = min(block_width - 1, 3);
46 const size_t oN_stride = rem * output_stride;
47 const size_t oN_offset = oN_stride + tile_hbytes;
48 size_t bh = block_height;
49 for (; bh >= 4; bh -= 4) {
50 const v128_t v2_0 = wasm_v128_load(i0);
51 i0 = (uint32_t*) ((uintptr_t) i0 + input_offset);
52 const v128_t v2_1 = wasm_v128_load(i1);
53 i1 = (uint32_t*) ((uintptr_t) i1 + input_offset);
54 const v128_t v2_2 = wasm_v128_load(i2);
55 i2 = (uint32_t*) ((uintptr_t) i2 + input_offset);
56 const v128_t v2_3 = wasm_v128_load(i3);
57 i3 = (uint32_t*) ((uintptr_t) i3 + input_offset);
58
59 const v128_t v1_0 = wasm_v32x4_shuffle(v2_0, v2_2, 0, 4, 1, 5);
60 const v128_t v1_1 = wasm_v32x4_shuffle(v2_0, v2_2, 2, 6, 3, 7);
61 const v128_t v1_2 = wasm_v32x4_shuffle(v2_1, v2_3, 0, 4, 1, 5);
62 const v128_t v1_3 = wasm_v32x4_shuffle(v2_1, v2_3, 2, 6, 3, 7);
63 const v128_t v0_0 = wasm_v32x4_shuffle(v1_0, v1_2, 0, 4, 1, 5);
64 const v128_t v0_1 = wasm_v32x4_shuffle(v1_0, v1_2, 2, 6, 3, 7);
65 const v128_t v0_2 = wasm_v32x4_shuffle(v1_1, v1_3, 0, 4, 1, 5);
66 const v128_t v0_3 = wasm_v32x4_shuffle(v1_1, v1_3, 2, 6, 3, 7);
67
68 o = (uint32_t*) ((uintptr_t) o + oN_offset);
69 wasm_v128_store(o, v0_3);
70 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
71 if XNN_UNPREDICTABLE(block_width > 3) {
72 o = oN;
73 }
74 wasm_v128_store(o, v0_2);
75 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
76 if XNN_UNPREDICTABLE(block_width >= 3) {
77 o = oN;
78 }
79 wasm_v128_store(o, v0_1);
80 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
81 if XNN_UNPREDICTABLE(block_width > 1) {
82 o = oN;
83 }
84 wasm_v128_store(o, v0_0);
85 }
86 o = (uint32_t*) ((uintptr_t) o + tile_hbytes);
87
88 if (bh != 0) {
89 const v128_t v2_0 = wasm_v128_load(i0);
90 if XNN_UNPREDICTABLE(bh < 2) {
91 i1 = i0;
92 }
93 const v128_t v2_1 = wasm_v128_load(i1);
94 if XNN_UNPREDICTABLE(bh <= 2) {
95 i2 = i0;
96 }
97 const v128_t v2_2 = wasm_v128_load(i2);
98 const v128_t v2_3 = wasm_v128_xor(v2_0, v2_0);
99
100 const v128_t v1_0 = wasm_v32x4_shuffle(v2_0, v2_2, 0, 4, 1, 5);
101 const v128_t v1_1 = wasm_v32x4_shuffle(v2_0, v2_2, 2, 6, 3, 7);
102 const v128_t v1_2 = wasm_v32x4_shuffle(v2_1, v2_3, 0, 4, 1, 5);
103 const v128_t v1_3 = wasm_v32x4_shuffle(v2_1, v2_3, 2, 6, 3, 7);
104
105 v128_t v0_0 = wasm_v32x4_shuffle(v1_0, v1_2, 0, 4, 1, 5);
106 v128_t v0_1 = wasm_v32x4_shuffle(v1_0, v1_2, 2, 6, 3, 7);
107 v128_t v0_2 = wasm_v32x4_shuffle(v1_1, v1_3, 0, 4, 1, 5);
108 v128_t v0_3 = wasm_v32x4_shuffle(v1_1, v1_3, 2, 6, 3, 7);
109
110 if (bh & 2) {
111 o = (uint32_t*) ((uintptr_t) o + oN_stride);
112 *((double*) o) = wasm_f64x2_extract_lane(v0_3, 0);
113 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
114 if XNN_UNPREDICTABLE(block_width > 3) {
115 o = oN;
116 }
117 *((double*) o) = wasm_f64x2_extract_lane(v0_2, 0);
118 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
119 if XNN_UNPREDICTABLE(block_width >= 3) {
120 o = oN;
121 }
122 *((double*) o) = wasm_f64x2_extract_lane(v0_1, 0);
123 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
124 if XNN_UNPREDICTABLE(block_width > 1) {
125 o = oN;
126 }
127 *((double*) o) = wasm_f64x2_extract_lane(v0_0, 0);
128 o += 2;
129 v0_0 = wasm_v64x2_shuffle(v0_0, v0_0, 1, 1);
130 v0_1 = wasm_v64x2_shuffle(v0_1, v0_1, 1, 1);
131 v0_2 = wasm_v64x2_shuffle(v0_2, v0_2, 1, 1);
132 v0_3 = wasm_v64x2_shuffle(v0_3, v0_3, 1, 1);
133 }
134
135 if (bh & 1) {
136 o = (uint32_t*) ((uintptr_t) o + oN_stride);
137 *((float*) o) = wasm_f32x4_extract_lane(v0_3, 0);
138 uint32_t *oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
139 if XNN_UNPREDICTABLE(block_width > 3) {
140 o = oN;
141 }
142 *((float*) o) = wasm_f32x4_extract_lane(v0_2, 0);
143 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
144 if XNN_UNPREDICTABLE(block_width >= 3) {
145 o = oN;
146 }
147 *((float*) o) = wasm_f32x4_extract_lane(v0_1, 0);
148 oN = (uint32_t*) ((uintptr_t) o + minus_output_stride);
149 if XNN_UNPREDICTABLE(block_width > 1) {
150 o = oN;
151 }
152 *((float*) o) = wasm_f32x4_extract_lane(v0_0, 0);
153 }
154 }
155
156 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
157 i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
158 i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
159 i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
160 o = (uint32_t*) ((uintptr_t) o + output_reset);
161 block_width = doz(block_width, tile_width);
162 } while (block_width != 0);
163 }
164