1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
#include "ops_windowing.h"
#include "math_utils.h"
#include "types.h"
#include <float.h>
#include <math.h>
#include <stdbool.h>
#include <stdint.h>
#include <string.h>
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#define MIN(a, b) ((a) < (b) ? (a) : (b))
// ============================================================================
// WINDOWING OPERATION IMPLEMENTATION MACROS
// ============================================================================
//
// These macros generate sliding window reduction operations for various types.
//
// Metadata layout (same for all operations):
// - metadata[0]: output_size (total number of elements in output)
// - metadata[1]: num_dims (number of dimensions)
// - metadata[2..2+num_dims]: input_shape
// - metadata[2+num_dims..2+2*num_dims]: input_strides
// - metadata[2+2*num_dims]: input_offset (starting offset in input)
// - metadata[3+2*num_dims..3+3*num_dims]: window_shape (size of window in each dimension)
// - metadata[3+3*num_dims..3+4*num_dims]: strides (step size in each dimension)
// - metadata[3+4*num_dims..3+4*num_dims+2*num_dims]: padding (before and after for each dimension)
// - metadata[3+6*num_dims..]: output_shape
//
// Algorithm:
// 1. For each output element, compute its coordinates
// 2. For each position in the window:
// - Compute corresponding input coordinates with stride and padding
// - Check if position is within bounds (considering padding)
// - Apply reduction operation on valid values
// 3. Out-of-bounds values are treated according to operation:
// - max: -infinity, min: +infinity, sum/mean: 0
/**
* @brief Macro to implement windowing reduction operations (max, min, sum)
*
* @param IN_TYPE C type of input elements
* @param OUT_TYPE C type of output elements
* @param TYPE_SUFFIX Suffix for function name
* @param INIT_VAL Initial value for accumulator
* @param ACCUMULATE Statement to accumulate values (uses 'acc' and 'val')
*/
#define REDUCE_WINDOW_OP(IN_TYPE, OUT_TYPE, TYPE_SUFFIX, INIT_VAL, ACCUMULATE) \
void reduce_window_##TYPE_SUFFIX(const void *input_ptr, void *output_ptr, \
const size_t *metadata) { \
const IN_TYPE *input = (const IN_TYPE *)input_ptr; \
OUT_TYPE *output = (OUT_TYPE *)output_ptr; \
\
const size_t num_els = metadata[0]; \
const size_t num_dims = metadata[1]; \
const size_t *input_shape = metadata + 2; \
const size_t *input_strides = metadata + 2 + num_dims; \
const size_t input_offset = metadata[2 + 2 * num_dims]; \
const size_t *window_shape = metadata + 2 + 2 * num_dims + 1; \
const size_t *strides = metadata + 2 + 3 * num_dims + 1; \
const size_t *padding = metadata + 2 + 4 * num_dims + 1; \
const size_t *output_shape = metadata + 2 + 6 * num_dims + 1; \
\
for (size_t out_idx = 0; out_idx < num_els; out_idx++) { \
size_t out_coords[16]; \
size_t tmp = out_idx; \
for (int i = (int)num_dims - 1; i >= 0; i--) { \
out_coords[i] = tmp % output_shape[i]; \
tmp /= output_shape[i]; \
} \
\
OUT_TYPE acc = INIT_VAL; \
\
size_t window_size = 1; \
for (size_t i = 0; i < num_dims; i++) { \
window_size *= window_shape[i]; \
} \
\
for (size_t win_idx = 0; win_idx < window_size; win_idx++) { \
size_t window_coords[16]; \
size_t tmp_win = win_idx; \
for (int i = (int)num_dims - 1; i >= 0; i--) { \
window_coords[i] = tmp_win % window_shape[i]; \
tmp_win /= window_shape[i]; \
} \
\
bool in_bounds = true; \
size_t input_coords[16]; \
for (size_t i = 0; i < num_dims; i++) { \
size_t padded_pos = out_coords[i] * strides[i] + window_coords[i]; \
size_t pad_before = padding[i * 2]; \
if (padded_pos < pad_before) { \
in_bounds = false; \
break; \
} \
size_t input_pos = padded_pos - pad_before; \
if (input_pos >= input_shape[i]) { \
in_bounds = false; \
break; \
} \
input_coords[i] = input_pos; \
} \
\
IN_TYPE val; \
if (in_bounds) { \
size_t idx = input_offset; \
for (size_t i = 0; i < num_dims; i++) { \
idx += input_coords[i] * input_strides[i]; \
} \
val = input[idx]; \
} else { \
val = (IN_TYPE)INIT_VAL; \
} \
\
ACCUMULATE; \
} \
\
output[out_idx] = acc; \
} \
}
/**
* @brief Macro to implement windowing mean operation
*
* Computes the average of values in each window. Padding areas are treated as 0.
* Only available for floating-point types.
*
* @param IN_TYPE C type of input elements (float type)
* @param OUT_TYPE C type of output elements (float type)
* @param TYPE_SUFFIX Suffix for function name
*/
#define REDUCE_WINDOW_MEAN_OP(IN_TYPE, OUT_TYPE, TYPE_SUFFIX) \
void reduce_window_mean_##TYPE_SUFFIX(const void *input_ptr, void *output_ptr, \
const size_t *metadata) { \
const IN_TYPE *input = (const IN_TYPE *)input_ptr; \
OUT_TYPE *output = (OUT_TYPE *)output_ptr; \
\
const size_t num_els = metadata[0]; \
const size_t num_dims = metadata[1]; \
\
const size_t *input_shape = metadata + 2; \
const size_t *input_strides = metadata + 2 + num_dims; \
const size_t input_offset = metadata[2 + 2 * num_dims]; \
const size_t *window_shape = metadata + 2 + 2 * num_dims + 1; \
const size_t *strides = metadata + 2 + 3 * num_dims + 1; \
const size_t *padding = metadata + 2 + 4 * num_dims + 1; \
const size_t *output_shape = metadata + 2 + 6 * num_dims + 1; \
\
for (size_t out_idx = 0; out_idx < num_els; out_idx++) { \
size_t out_coords[16]; \
size_t tmp = out_idx; \
for (int i = (int)num_dims - 1; i >= 0; i--) { \
out_coords[i] = tmp % output_shape[i]; \
tmp /= output_shape[i]; \
} \
\
OUT_TYPE sum = 0; \
size_t window_size = 1; \
for (size_t i = 0; i < num_dims; i++) { \
window_size *= window_shape[i]; \
} \
\
for (size_t win_idx = 0; win_idx < window_size; win_idx++) { \
size_t window_coords[16]; \
size_t tmp_win = win_idx; \
for (int i = (int)num_dims - 1; i >= 0; i--) { \
window_coords[i] = tmp_win % window_shape[i]; \
tmp_win /= window_shape[i]; \
} \
\
bool in_bounds = true; \
size_t input_coords[16]; \
for (size_t i = 0; i < num_dims; i++) { \
size_t padded_pos = out_coords[i] * strides[i] + window_coords[i]; \
size_t pad_before = padding[i * 2]; \
if (padded_pos < pad_before) { \
in_bounds = false; \
break; \
} \
size_t input_pos = padded_pos - pad_before; \
if (input_pos >= input_shape[i]) { \
in_bounds = false; \
break; \
} \
input_coords[i] = input_pos; \
} \
\
if (in_bounds) { \
size_t idx = input_offset; \
for (size_t i = 0; i < num_dims; i++) { \
idx += input_coords[i] * input_strides[i]; \
} \
sum += (OUT_TYPE)input[idx]; \
} \
} \
\
output[out_idx] = sum / (OUT_TYPE)window_size; \
} \
}
REDUCE_WINDOW_OP(f8e4m3_t, f8e4m3_t, max_f8e4m3, F8E4M3_NEG_INF, acc = MAX(acc, val))
REDUCE_WINDOW_OP(f8e5m2_t, f8e5m2_t, max_f8e5m2, F8E5M2_NEG_INF, acc = MAX(acc, val))
REDUCE_WINDOW_OP(bf16_t, bf16_t, max_bf16, BF16_NEG_INF, acc = MAX(acc, val))
REDUCE_WINDOW_OP(f16_t, f16_t, max_f16, F16_NEG_INF, acc = MAX(acc, val))
REDUCE_WINDOW_OP(float, float, max_f32, -INFINITY, acc = MAX(acc, val))
REDUCE_WINDOW_OP(double, double, max_f64, -INFINITY, acc = MAX(acc, val))
REDUCE_WINDOW_OP(int8_t, int8_t, max_i8, INT8_MIN, acc = MAX(acc, val))
REDUCE_WINDOW_OP(int16_t, int16_t, max_i16, INT16_MIN, acc = MAX(acc, val))
REDUCE_WINDOW_OP(int32_t, int32_t, max_i32, INT32_MIN, acc = MAX(acc, val))
REDUCE_WINDOW_OP(int64_t, int64_t, max_i64, INT64_MIN, acc = MAX(acc, val))
REDUCE_WINDOW_OP(uint8_t, uint8_t, max_u8, 0u, acc = MAX(acc, val))
REDUCE_WINDOW_OP(uint16_t, uint16_t, max_u16, 0u, acc = MAX(acc, val))
REDUCE_WINDOW_OP(uint32_t, uint32_t, max_u32, 0u, acc = MAX(acc, val))
REDUCE_WINDOW_OP(uint64_t, uint64_t, max_u64, 0u, acc = MAX(acc, val))
REDUCE_WINDOW_MEAN_OP(f8e4m3_t, f8e4m3_t, f8e4m3)
REDUCE_WINDOW_MEAN_OP(f8e5m2_t, f8e5m2_t, f8e5m2)
REDUCE_WINDOW_MEAN_OP(bf16_t, bf16_t, bf16)
REDUCE_WINDOW_MEAN_OP(f16_t, f16_t, f16)
REDUCE_WINDOW_MEAN_OP(float, float, f32)
REDUCE_WINDOW_MEAN_OP(double, double, f64)
REDUCE_WINDOW_OP(f8e4m3_t, f8e4m3_t, sum_f8e4m3, 0, acc += val)
REDUCE_WINDOW_OP(f8e5m2_t, f8e5m2_t, sum_f8e5m2, 0, acc += val)
REDUCE_WINDOW_OP(bf16_t, bf16_t, sum_bf16, 0, acc += val)
REDUCE_WINDOW_OP(f16_t, f16_t, sum_f16, 0, acc += val)
REDUCE_WINDOW_OP(float, float, sum_f32, 0.0f, acc += val)
REDUCE_WINDOW_OP(double, double, sum_f64, 0.0, acc += val)
REDUCE_WINDOW_OP(int8_t, int8_t, sum_i8, 0, acc += val)
REDUCE_WINDOW_OP(int16_t, int16_t, sum_i16, 0, acc += val)
REDUCE_WINDOW_OP(int32_t, int32_t, sum_i32, 0, acc += val)
REDUCE_WINDOW_OP(int64_t, int64_t, sum_i64, 0, acc += val)
REDUCE_WINDOW_OP(uint8_t, uint8_t, sum_u8, 0u, acc += val)
REDUCE_WINDOW_OP(uint16_t, uint16_t, sum_u16, 0u, acc += val)
REDUCE_WINDOW_OP(uint32_t, uint32_t, sum_u32, 0u, acc += val)
REDUCE_WINDOW_OP(uint64_t, uint64_t, sum_u64, 0u, acc += val)
REDUCE_WINDOW_OP(f8e4m3_t, f8e4m3_t, min_f8e4m3, F8E4M3_POS_INF, acc = MIN(acc, val))
REDUCE_WINDOW_OP(f8e5m2_t, f8e5m2_t, min_f8e5m2, F8E5M2_POS_INF, acc = MIN(acc, val))
REDUCE_WINDOW_OP(bf16_t, bf16_t, min_bf16, BF16_POS_INF, acc = MIN(acc, val))
REDUCE_WINDOW_OP(f16_t, f16_t, min_f16, F16_POS_INF, acc = MIN(acc, val))
REDUCE_WINDOW_OP(float, float, min_f32, INFINITY, acc = MIN(acc, val))
REDUCE_WINDOW_OP(double, double, min_f64, INFINITY, acc = MIN(acc, val))
REDUCE_WINDOW_OP(int8_t, int8_t, min_i8, INT8_MAX, acc = MIN(acc, val))
REDUCE_WINDOW_OP(int16_t, int16_t, min_i16, INT16_MAX, acc = MIN(acc, val))
REDUCE_WINDOW_OP(int32_t, int32_t, min_i32, INT32_MAX, acc = MIN(acc, val))
REDUCE_WINDOW_OP(int64_t, int64_t, min_i64, INT64_MAX, acc = MIN(acc, val))
REDUCE_WINDOW_OP(uint8_t, uint8_t, min_u8, UINT8_MAX, acc = MIN(acc, val))
REDUCE_WINDOW_OP(uint16_t, uint16_t, min_u16, UINT16_MAX, acc = MIN(acc, val))
REDUCE_WINDOW_OP(uint32_t, uint32_t, min_u32, UINT32_MAX, acc = MIN(acc, val))
REDUCE_WINDOW_OP(uint64_t, uint64_t, min_u64, UINT64_MAX, acc = MIN(acc, val))