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