Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 031d6a97de79fc3ca3eb6fca1611f03aa9b5893b / . / src / core / CL / cl_kernels / tile_helpers.h
blob: 496f2dd6648c27539871df00c2f5490d2f4238f0 [file] [log] [blame]
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1/*
2 * Copyright (c) 2021 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24
25/** Tile object
26 * A tile object is a 2D memory block and can be accessed using the following syntax:
27 * -# a[m0].v = access the the vector at row "m0" (OpenCL vector)
28 * -# a[m0].s[x] = access the scalar element at row "m0" and column "n0" (scalar access)
29 *
30 * @param[in] DATA_TYPE Data type of the tile
31 * @param[in] H Number of tile rows
32 * @param[in] W Number of tile colums
33 * @param[in] BASENAME Tile's name
34 */
35#define TILE(DATA_TYPE, H, W, BASENAME) TILE_STR(DATA_TYPE, H, W, BASENAME)
36#define TILE_STR(DATA_TYPE, H, W, BASENAME) \
37 union { \
38 DATA_TYPE s[W]; \
39 DATA_TYPE##W v; \
40 } BASENAME[H]
41
42#define TENSOR4D_IMAGE(name) \
43 __read_only image2d_t name##_img, \
44 __global uchar *name##_ptr, \
45 uint name##_stride_x, \
46 uint name##_step_x, \
47 uint name##_stride_y, \
48 uint name##_step_y, \
49 uint name##_stride_z, \
50 uint name##_step_z, \
51 uint name##_stride_w, \
52 uint name##_step_w, \
53 uint name##_offset_first_element_in_bytes
54
55#define TENSOR4D_BUFFER(name) \
56 __global uchar *name##_ptr, \
57 uint name##_stride_x, \
58 uint name##_step_x, \
59 uint name##_stride_y, \
60 uint name##_step_y, \
61 uint name##_stride_z, \
62 uint name##_step_z, \
63 uint name##_stride_w, \
64 uint name##_step_w, \
65 uint name##_offset_first_element_in_bytes
66
67#define TENSOR4D_STR(name, type) TENSOR4D_##type(name)
68#define TENSOR4D(name, type) TENSOR4D_STR(name, type)
69
70/** Loop unrolling */
71#define LOOP_UNROLLING(DATA_TYPE, VAR, START_IDX, NUM_ITERATIONS, STEP) \
72 _Pragma("unroll") for(DATA_TYPE VAR = START_IDX; VAR < NUM_ITERATIONS; VAR += STEP)
73
74/** Get the get_global_id with partial N0. This function is useful when the dimension is not multiple of N0 and we need to use a partial N0
75 * to avoid out-of-bound read/write
76 *
77 * @note PARTIAL_N0 is used for get_global_id(n) = 0.
78 *
79 * @param[in] IDX get_global_id index (0,1 and 2 only)
80 * @param[in] N0 Number of elements read/written on the IDX direction
81 * @param[in] PARTIAL_N0 Number of elements read/written on the IDX direction for get_global_id(IDX) = 0. If zero,
82 * the Number of elements read/written on the IDX direction for get_global_id(IDX) = 0 is N0
83 */
84#define GET_SPATIAL_IDX(IDX, N0, PARTIAL_N0) (max((int)(get_global_id(IDX) * N0 - (N0 - PARTIAL_N0) % N0), 0))
85
86/** Offset (in bytes) calculation for a 1D BUFFER (cl_buffer) tensor */
87#define OFFSET1D(base, data_type, x) (base##_offset_first_element_in_bytes + x * sizeof(data_type))
88
89/** Offset (in bytes) calculation for a 2D BUFFER (cl_buffer) tensor */
90#define OFFSET2D(base, data_type, x, y) (base##_offset_first_element_in_bytes + x * sizeof(data_type) + y * base##_stride_y)
91
92/** Offset (in bytes) calculation for a 3D BUFFER (cl_buffer) tensor */
93#define OFFSET3D(base, data_type, x, y, z) (base##_offset_first_element_in_bytes + x * sizeof(data_type) + y * base##_stride_y + z * base##_stride_z)
94
95/** Offset (in bytes) calculation for a 4D BUFFER (cl_buffer) tensor */
96#define OFFSET4D(base, data_type, x, y, z, w) (base##_offset_first_element_in_bytes + x * sizeof(data_type) + y * base##_stride_y + z * base##_stride_z + w * base##_stride_w)
97
98/** Dot product integet 8bit function
99 *
100 * @note Performs: c += dot(a, b)
101 *
102 * @param[in] DST_DATA_TYPE Accumulator data type
103 * @param[in] K0 Number of accumulations
104 * @param[in] a OpenCL vector a
105 * @param[in] b OpenCL vector b
106 * @param[in] c Scalar variable c
107 */
108#define DOT_PRODUCT_INTEGER8(DST_DATA_TYPE, K0, a, b, c) DOT_PRODUCT_INTEGER8_STR(DST_DATA_TYPE, K0, a, b, c)
109#define DOT_PRODUCT_INTEGER8_STR(DST_DATA_TYPE, K0, a, b, c) DOT_PRODUCT##K0##_INTEGER8(DST_DATA_TYPE, a, b, c)
110#define DOT_PRODUCT1_INTEGER8(DST_DATA_TYPE, a, b, c) \
111 ({ \
112 c += (DST_DATA_TYPE)a * (DST_DATA_TYPE)b; \
113 })
114#define DOT_PRODUCT2_INTEGER8(DST_DATA_TYPE, a, b, c) \
115 ({ \
116 c += (DST_DATA_TYPE)a.s0 * (DST_DATA_TYPE)b.s0; \
117 c += (DST_DATA_TYPE)a.s1 * (DST_DATA_TYPE)b.s1; \
118 })
119#define DOT_PRODUCT3_INTEGER8(DST_DATA_TYPE, a, b, c) \
120 ({ \
121 DOT_PRODUCT2_INTEGER8(DST_DATA_TYPE, a, b, c); \
122 c += (DST_DATA_TYPE)a.s2 * (DST_DATA_TYPE)b.s2; \
123 })
124#if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
125#define DOT_PRODUCT4_INTEGER8(DST_DATA_TYPE, x, y, val) val = arm_dot_acc((x), (y), (val));
126#elif defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
127#define DOT_PRODUCT4_INTEGER8(DST_DATA_TYPE, x, y, val) val += arm_dot((x), (y));
128#else // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
129#define DOT_PRODUCT4_INTEGER8(DST_DATA_TYPE, x, y, val) \
130 ({ \
131 val += (DST_DATA_TYPE)x.s0 * (DST_DATA_TYPE)y.s0; \
132 val += (DST_DATA_TYPE)x.s1 * (DST_DATA_TYPE)y.s1; \
133 val += (DST_DATA_TYPE)x.s2 * (DST_DATA_TYPE)y.s2; \
134 val += (DST_DATA_TYPE)x.s3 * (DST_DATA_TYPE)y.s3; \
135 })
136#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
137#define DOT_PRODUCT8_INTEGER8(DST_DATA_TYPE, a, b, c) \
138 ({ \
139 DOT_PRODUCT4_INTEGER8((a.lo), (b.lo), c); \
140 DOT_PRODUCT4_INTEGER8((a.hi), (b.hi), c); \
141 })
142#define DOT_PRODUCT16_INTEGER8(DST_DATA_TYPE, a, b, c) \
143 ({ \
144 DOT_PRODUCT8_INTEGER8((a.lo), (b.lo), c); \
145 DOT_PRODUCT8_INTEGER8((a.hi), (b.hi), c); \
146 })
147
148/** Load a vector from global memory (tensor)
149 *
150 * @param[in] DATA_TYPE Data type
151 * @param[in] WIDTH Number of dst columns
152 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image).
153 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
154 * @param[in] TENSOR Tensor basename
155 * @param[in] X Starting X position
156 * @param[in] Y Starting Y position
157 * @param[in] STRIDE_Y Stride Y (in bytes)
158 */
159#define V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) V_LOAD_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y)
160#define V_LOAD_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) V_LOAD_##TENSOR_TYPE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y)
161#define V_LOAD_BUFFER(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) \
162 VLOAD(WIDTH) \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]163 (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (Y)*STRIDE_Y))
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]164#define V_LOAD_IMAGE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) READ_IMAGE2D(DATA_TYPE, CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(WIDTH), TENSOR##_img, (X) / 4, (Y))
165
166/** Load a tile from global memory (tensor)
167 *
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]168 * @param[in] DATA_TYPE Data type
169 * @param[in] HEIGHT Number of dst rows
170 * @param[in] WIDTH Number of dst columns
171 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image).
172 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
173 * @param[in] TENSOR Tensor basename
174 * @param[in] X Starting X position
175 * @param[in] Y Starting Y position
176 * @param[in] YI_MULTIPLIER Parameter used to multiply the internal row increment (_i).
177 * In common cases should be 1 but it becomes useful when we want to load rows which are multiple of STRIDE_Y. (e.g. loading the weights of convolution layer).
178 * In this case the address calculation is performed as: (Y + _i * Y_MULTIPLIER) * STRIDE_Y
179 * @param[in] STRIDE_Y Stride Y (in bytes) used to load each row.
180 * @param[out] dst Output tile
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]181 */
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]182#define T_LOAD(DATA_TYPE, HEIGHT, WIDTH, TENSOR_TYPE, TENSOR, X, Y, YI_MULTIPLIER, STRIDE_Y, dst) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]183 ({ \
184 LOOP_UNROLLING(int, _i, 0, HEIGHT, 1) \
185 { \
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]186 dst[_i].v = V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, ((Y) + _i * (int)(YI_MULTIPLIER)), STRIDE_Y); \
187 } \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]188 })
189
190/** Load a tile from global memory (tensor) using an indirect Y index tile
191 *
192 * @param[in] DATA_TYPE Data type
193 * @param[in] HEIGHT Number of dst rows
194 * @param[in] WIDTH Number of dst columns
195 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported
196 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
197 * @param[in] TENSOR Tensor basename
198 * @param[in] X Starting X position
199 * @param[in] STRIDE_Y Stride Y (in bytes)
200 * @param[in] indirect_y Indirect Y index tile
201 * @param[out] dst Output tile
202 */
203#define T_LOAD_INDIRECT(DATA_TYPE, HEIGHT, WIDTH, TENSOR_TYPE, TENSOR, X, STRIDE_Y, indirect_y, dst) \
204 ({ \
205 LOOP_UNROLLING(int, _i, 0, HEIGHT, 1) \
206 { \
207 dst[_i].v = V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, (indirect_y[_i].v), STRIDE_Y); \
208 } \
209 })
210
Gian Marco Iodice534b8892021-04-01 16:17:16 +0100[diff] [blame]211/** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout
212 *
213 * @param[in] DATA_TYPE Data type
214 * @param[in] TILE_HEIGHT Number of elements to load from Y (height) dimension
215 * @param[in] TILE_WIDTH Number of elements to load from X (width) dimension
216 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension
217 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported
218 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16)
219 * @param[in] TENSOR Tensor basename
220 * @param[in] B Starting batch index
221 * @param[in] Y Starting Y index
222 * @param[in] X Starting X index
223 * @param[in] C Starting C index
224 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension
225 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension
226 * @param[in] STRIDE_Y Stride Y (in bytes)
227 * @param[out] dst Output tile
228 */
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]229#define T_LOAD_NHWC(DATA_TYPE, TILE_HEIGHT, TILE_WIDTH, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, STRIDE_Y, dst) \
Gian Marco Iodice534b8892021-04-01 16:17:16 +0100[diff] [blame]230 ({ \
231 LOOP_UNROLLING(int, _yk, 0, (TILE_HEIGHT), 1) \
232 { \
233 LOOP_UNROLLING(int, _xk, 0, (TILE_WIDTH), 1) \
234 { \
235 int _src_y = (X) + _xk + ((Y) + _yk) * (TENSOR_WIDTH); \
236 _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT); \
237 int _src_valid_y = (((X) + _xk) >= 0 && ((X) + _xk) < (int)(TENSOR_WIDTH) && ((Y) + _yk) >= 0 && ((Y) + _yk) < (int)(TENSOR_HEIGHT)); \
238 if(_src_valid_y != 0) \
239 { \
240 dst[_xk + _yk * (TILE_WIDTH)].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]241 } \
242 } \
243 } \
244 })
245
246/** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout using indirect X and Y coordinates
247 *
248 * @param[in] DATA_TYPE Data type
249 * @param[in] TILE_HEIGHT Number of elements to load from Y (height) dimension
250 * @param[in] TILE_WIDTH Number of elements to load from X (width) dimension
251 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension
252 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported
253 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16)
254 * @param[in] TENSOR Tensor basename
255 * @param[in] B Starting batch index
256 * @param[in] Y Starting Y index
257 * @param[in] X Starting X index
258 * @param[in] C Starting C index
259 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension
260 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension
261 * @param[in] STRIDE_Y Stride Y (in bytes)
262 * @param[out] xi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect X coordinate
263 * @param[out] yi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect Y coordinate
264 * @param[out] dst Output tile
265 */
266#define T_LOAD_NHWC_INDIRECT(DATA_TYPE, TILE_HEIGHT, TILE_WIDTH, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, STRIDE_Y, xi, yi, dst) \
267 ({ \
268 LOOP_UNROLLING(int, _i, 0, (TILE_WIDTH * TILE_HEIGHT), 1) \
269 { \
270 int _src_y = (X) + xi[_i].v + ((Y) + yi[_i].v) * (TENSOR_WIDTH); \
271 _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT); \
272 int _src_valid_y = (((X) + xi[_i].v) >= 0 && ((X) + xi[_i].v) < (int)(TENSOR_WIDTH) && ((Y) + yi[_i].v) >= 0 && ((Y) + yi[_i].v) < (int)(TENSOR_HEIGHT)); \
273 if(_src_valid_y != 0) \
274 { \
275 dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \
276 } \
277 } \
Gian Marco Iodice534b8892021-04-01 16:17:16 +0100[diff] [blame]278 })
279
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]280/** Store a tile to global memory (tensor) using an indirect Y index tile and conditionally use a different length for the store
281 *
282 * @note If WIDTH1_CONDITION is true, the store will use the WIDTH1 length for the store
283 * @note The vectors are stored in reverse order so the invalid rows are overwritten by the valid ones
284 *
285 * @param[in] DATA_TYPE Data type
286 * @param[in] HEIGHT Number of src rows
287 * @param[in] WIDTH0 Store width to use if WIDTH1_CONDITION = false
288 * @param[in] WIDTH1 Store width to use if WIDTH1_CONDITION = true
289 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported
290 * cl_image is not supported.
291 * @param[in] TENSOR Tensor basename
292 * @param[in] X Starting X position
293 * @param[in] STRIDE_Y Stride Y (in bytes)
294 * @param[in] WIDTH1_CONDITION Condition to select the WIDTH1 store
295 * @param[in] src Input tile
296 * @param[in] indirect_y Indirect Y index tile
297 */
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]298#define T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, STRIDE_Y, WIDTH1_CONDITION, src, indirect_y) \
299 ({ \
300 if(WIDTH1_CONDITION) \
301 { \
302 LOOP_UNROLLING(int, _i, 0, HEIGHT, 1) \
303 { \
304 VSTORE_PARTIAL(WIDTH0, WIDTH1) \
305 (src[HEIGHT - 1 - _i].v, 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \
306 } \
307 } \
308 else \
309 { \
310 LOOP_UNROLLING(int, _i, 0, HEIGHT, 1) \
311 { \
312 VSTORE(WIDTH0) \
313 (src[HEIGHT - 1 - _i].v, 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \
314 } \
315 } \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]316 })
317
318/** Offset correction for the QASYMM8 computation
319 *
320 * @param[in] ACC_DATA_TYPE Accumulator data type
321 * @param[in] M0 Number of src/dst rows
322 * @param[in] N0 Number of src/dst columns
323 * @param[in] K0 Number of src columns
324 * @param[in] SRC_OFFSET Source quantization offset
325 * @param[in] WEI_OFFSET Weights quantization shift
326 * @param[in] lhs LHS tile
327 * @param[in] rhs RHS tile
328 * @param[out] dst DST tile
329 */
330#define T_OFFSET_CORRECTION(ACC_DATA_TYPE, M0, N0, K0, SRC_OFFSET, WEI_OFFSET, lhs, rhs, dst) \
331 ({ \
332 LOOP_UNROLLING(int, _m0, 0, M0, 1) \
333 { \
334 ACC_DATA_TYPE _tm = 0; \
335 LOOP_UNROLLING(int, _k0, 0, K0, 1) \
336 { \
337 _tm += ((ACC_DATA_TYPE)lhs[_m0].s[_k0] * (ACC_DATA_TYPE)WEI_OFFSET); \
338 } \
339 LOOP_UNROLLING(int, _n0, 0, N0, 1) \
340 { \
341 dst[_m0].s[_n0] += _tm; \
342 LOOP_UNROLLING(int, _k0, 0, K0, 1) \
343 { \
344 dst[_m0].s[_n0] += ((ACC_DATA_TYPE)rhs[_n0].s[_k0] * (ACC_DATA_TYPE)SRC_OFFSET); \
345 } \
346 } \
347 } \
348 })
349
350/** Quantized the tile (ASYMMETRIC) with fixed-point scale
351 *
352 * @param[in] SRC_DATA_TYPE SRC data type
353 * @param[in] DST_DATA_TYPE DST data type
354 * @param[in] M0 Number of src/dst rows
355 * @param[in] N0 Number of src/dst columns
356 * @param[in] DST_OFFSET Quantization offset
357 * @param[in] DST_SHIFT Quantization shift
358 * @param[in] DST_MULTIPLIER Quantization multiplier
359 * @param[in] src Input tile
360 * @param[out] dst Output tile
361 */
362#define T_QUANTIZE8_ASYMMETRIC(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst) \
363 ({ \
364 LOOP_UNROLLING(int, _m0, 0, M0, 1) \
365 { \
366 LOOP_UNROLLING(int, _n0, 0, N0, 1) \
367 { \
368 SRC_DATA_TYPE _tmp = 0; \
369 if(DST_SHIFT < 0) \
370 { \
371 _tmp = ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(src[_m0].s[_n0], DST_MULTIPLIER, DST_SHIFT, 1); \
372 } \
373 else \
374 { \
375 _tmp = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(src[_m0].s[_n0], DST_MULTIPLIER, DST_SHIFT, 1); \
376 } \
377 _tmp += DST_OFFSET; \
378 dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \
379 } \
380 } \
381 })
382
383/** Conditional rowset (memset by row)
384 *
385 * @note Set the row to VALUE_TO_SET if the corresponding mask == 0
386 *
387 * @param[in] DATA_TYPE Data type
388 * @param[in] M0 Number of LHS rows
389 * @param[in] N0 Number of LHS columns
390 * @param[in] VALUE_TO_SET Value to set the row
391 * @param[in, out] a Input/output tile
392 * @param[out] mask Mask to check for setting the row to VALUE_TO_SET
393 */
394#define T_ROWSET_MASK(DATA_TYPE, M0, N0, VALUE_TO_SET, a, mask) \
395 ({ \
396 LOOP_UNROLLING(int, _m0, 0, M0, 1) \
397 { \
398 LOOP_UNROLLING(int, _n0, 0, N0, 1) \
399 { \
400 a[_m0].s[_n0] = select((DATA_TYPE)(a[_m0].s[_n0]), (DATA_TYPE)(VALUE_TO_SET), (SELECT_DATA_TYPE(DATA_TYPE))(mask[_m0].v == (DATA_TYPE)0)); \
401 } \
402 } \
403 })
404
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]405/** Element-wise activation
406 *
407 * @note Performs: activation(LHS) = DST
408 *
409 * @param[in] DATA_TYPE SRC/DST data type
410 * @param[in] M0 Number of SRC/DST rows
411 * @param[in] N0 Number of SRC/DST columns
412 * @param[in] ACTIVATION_TYPE Activation type
413 * @param[in] A_VAL A value used for the activation (e.g. tanh_op, brelu,..)
414 * @param[in] B_VAL B value used for the activation (e.g. tanh_op, brelu,..)
415 * @param[out] src SRC tile
416 * @param[out] dst DST tile
417 */
418#define T_ACTIVATION(DATA_TYPE, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, src, dst) \
419 ({ \
420 LOOP_UNROLLING(int, _m0, 0, M0, 1) \
421 { \
422 dst[_m0].v = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, N0, src[_m0].v, A_VAL, B_VAL); \
423 } \
424 })
425
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]426/** Element-wise addition with a constant value
427 *
428 * @note Performs: LHS + constant = DST
429 *
430 * @param[in] DATA_TYPE LHS/RHS/DST data type
431 * @param[in] M0 Number of LHS rows
432 * @param[in] N0 Number of LHS columns
433 * @param[in] lhs LHS tile
434 * @param[in] rhs_constant Constant value
435 * @param[out] dst DST tile
436 */
437#define T_ADD_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \
438 ({ \
439 LOOP_UNROLLING(int, _m0, 0, M0, 1) \
440 { \
441 LOOP_UNROLLING(int, _n0, 0, N0, 1) \
442 { \
443 dst[_m0].s[_n0] = lhs[_m0].s[_n0] + rhs_constant; \
444 } \
445 } \
446 })
447
448/** Element-wise addition with RHS broadcasted (RHS has the X dimension only)
449 *
450 * @note Performs: LHS + RHS[broadcasted] = DST
451 * @note Both tiles must have same data type
452 *
453 * @param[in] DATA_TYPE LHS/RHS/DST data type
454 * @param[in] M0 Number of LHS rows
455 * @param[in] N0 Number of LHS columns
456 * @param[in] lhs LHS tile
457 * @param[in] rhs RHS tile
458 * @param[out] dst DST tile
459 */
460#define T_ADD_BROADCAST_X(DATA_TYPE, M0, N0, lhs, rhs, dst) \
461 ({ \
462 LOOP_UNROLLING(int, _m0, 0, M0, 1) \
463 { \
464 dst[_m0].v = lhs[_m0].v + rhs[0].v; \
465 } \
466 })
467
468/** Matrix multiplication
469 *
470 * @note Performs: LHS X RHS + DST = DST
471 *
472 * @param[in] LHS_DATA_TYPE LHS tile data type
473 * @param[in] RHS_DATA_TYPE RHS tile data type
474 * @param[in] DST_DATA_TYPE RHS tile data type
475 * @param[in] M0 Number of LHS rows
476 * @param[in] N0 Number of RHS columns
477 * @param[in] K0 Number of LHS columns
478 * @param[in] LHS_LAYOUT LHS layout (T= transposed, NT= not transposed)
479 * @param[in] RHS_LAYOUT RHS layout (T= transposed, NT= not transposed)
480 * @param[in] lhs LHS tile
481 * @param[in] rhs RHS tile
482 * @param[in, out] dst DST tile
483 */
484#define T_MMUL(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, LHS_LAYOUT, RHS_LAYOUT, lhs, rhs, dst) T_MMUL_##LHS_LAYOUT##_##RHS_LAYOUT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
485#define T_MMUL_NT_T(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_##LHS_DATA_TYPE##_##RHS_DATA_TYPE##_##DST_DATA_TYPE(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
486#define T_MMUL_NT_T_float_float_float(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
487#define T_MMUL_NT_T_half_half_half(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
488#define T_MMUL_NT_T_char_char_int(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
489#define T_MMUL_NT_T_uchar_uchar_uint(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
490#define T_MMUL_NT_T_uchar_uchar_int(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
491#define T_MMUL_NT_T_FLOAT(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \
492 { \
493 LOOP_UNROLLING(int, _m, 0, M0, 1) \
494 { \
495 LOOP_UNROLLING(int, _n, 0, N0, 1) \
496 { \
497 LOOP_UNROLLING(int, _k, 0, K0, 1) \
498 { \
499 dst[_m].s[_n] = fma((lhs[_m].s[_k]), (rhs[_n].s[_k]), dst[_m].s[_n]); \
500 } \
501 } \
502 } \
503 }
504#define T_MMUL_NT_T_INTEGER8(DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \
505 ({ \
506 LOOP_UNROLLING(int, _m, 0, M0, 1) \
507 { \
508 LOOP_UNROLLING(int, _n, 0, N0, 1) \
509 { \
510 DOT_PRODUCT_INTEGER8(DST_DATA_TYPE, K0, (lhs[_m].v), (rhs[_n].v), dst[_m].s[_n]); \
511 } \
512 } \
513 })