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_switch_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_switch_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);
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*) output;
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 size_t bh = block_height;
48 for (; bh >= 4; bh -= 4) {
49 const v128_t v2_0 = wasm_v128_load(i0);
50 i0 = (uint32_t*) ((uintptr_t) i0 + input_offset);
51 const v128_t v2_1 = wasm_v128_load(i1);
52 i1 = (uint32_t*) ((uintptr_t) i1 + input_offset);
53 const v128_t v2_2 = wasm_v128_load(i2);
54 i2 = (uint32_t*) ((uintptr_t) i2 + input_offset);
55 const v128_t v2_3 = wasm_v128_load(i3);
56 i3 = (uint32_t*) ((uintptr_t) i3 + input_offset);
57
58 const v128_t v1_0 = wasm_v32x4_shuffle(v2_0, v2_2, 0, 4, 1, 5);
59 const v128_t v1_1 = wasm_v32x4_shuffle(v2_0, v2_2, 2, 6, 3, 7);
60 const v128_t v1_2 = wasm_v32x4_shuffle(v2_1, v2_3, 0, 4, 1, 5);
61 const v128_t v1_3 = wasm_v32x4_shuffle(v2_1, v2_3, 2, 6, 3, 7);
62 const v128_t v0_0 = wasm_v32x4_shuffle(v1_0, v1_2, 0, 4, 1, 5);
63 const v128_t v0_1 = wasm_v32x4_shuffle(v1_0, v1_2, 2, 6, 3, 7);
64 const v128_t v0_2 = wasm_v32x4_shuffle(v1_1, v1_3, 0, 4, 1, 5);
65 const v128_t v0_3 = wasm_v32x4_shuffle(v1_1, v1_3, 2, 6, 3, 7);
66
67 uint32_t *oN = (uint32_t*) ((uintptr_t) o + oN_stride);
68 switch (rem) {
69 case 3:
70 wasm_v128_store(oN, v0_3);
71 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
72 case 2:
73 wasm_v128_store(oN, v0_2);
74 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
75 case 1:
76 wasm_v128_store(oN, v0_1);
77 case 0:
78 wasm_v128_store(o, v0_0);
79 o = (uint32_t*) ((uintptr_t) o + tile_hbytes);
80 break;
81 default:
82 XNN_UNREACHABLE;
83 }
84 }
85
86 if (bh != 0) {
87 const v128_t v2_0 = wasm_v128_load(i0);
88 if XNN_UNPREDICTABLE(bh < 2) {
89 i1 = i0;
90 }
91 const v128_t v2_1 = wasm_v128_load(i1);
92 if XNN_UNPREDICTABLE(bh <= 2) {
93 i2 = i0;
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 uint32_t* oN = (uint32_t*) ((uintptr_t) o + oN_stride);
110 switch (rem) {
111 case 3:
112 *((double*) oN) = wasm_f64x2_extract_lane(v0_3, 0);
113 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
114 case 2:
115 *((double*) oN) = wasm_f64x2_extract_lane(v0_2, 0);
116 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
117 case 1:
118 *((double*) oN) = wasm_f64x2_extract_lane(v0_1, 0);
119 case 0:
120 *((double*) o) = wasm_f64x2_extract_lane(v0_0, 0);
121 o += 2;
122 break;
123 default:
124 XNN_UNREACHABLE;
125 }
126 v0_0 = wasm_v64x2_shuffle(v0_0, v0_0, 1, 1);
127 v0_1 = wasm_v64x2_shuffle(v0_1, v0_1, 1, 1);
128 v0_2 = wasm_v64x2_shuffle(v0_2, v0_2, 1, 1);
129 v0_3 = wasm_v64x2_shuffle(v0_3, v0_3, 1, 1);
130 }
131
132 if (bh & 1) {
133 uint32_t* oN = (uint32_t*) ((uintptr_t) o + oN_stride);
134 switch (rem) {
135 case 3:
136 *((float*) oN) = wasm_f32x4_extract_lane(v0_3, 0);
137 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
138 case 2:
139 *((float*) oN) = wasm_f32x4_extract_lane(v0_2, 0);
140 oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
141 case 1:
142 *((float*) oN) = wasm_f32x4_extract_lane(v0_1, 0);
143 case 0:
144 *((float*) o) = wasm_f32x4_extract_lane(v0_0, 0);
145 break;
146 default:
147 XNN_UNREACHABLE;
148 }
149 }
150 }
151
152 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
153 i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
154 i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
155 i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
156 o = (uint32_t*) ((uintptr_t) o + output_reset);
157 block_width = doz(block_width, tile_width);
158 } while (block_width != 0);
159 }
160