Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 8896cf7cb7df34c699e7453a0f0c683d1202ed15 / . / src / core / CL / cl_kernels / common / gemm.cl
blob: 0c30c0e626011fe4f3464bf8ad57652a24d488c4 [file] [log] [blame]
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]1/*
Gian Marco Iodice10e88a72021-11-29 12:49:19 +0000[diff] [blame]2 * Copyright (c) 2017-2022 Arm Limited.
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]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 */
Usama Arif0681e3b2019-04-25 14:28:07 +0100[diff] [blame]24#include "gemm_helpers.h"
Vidhya Sudhan Loganathan17b0f8b2019-01-08 12:17:03 +0000[diff] [blame]25#include "repeat.h"
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]26
Ramy Elgammal451c3092022-02-01 23:01:27 +0000[diff] [blame]27#if defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE)
28
29#define CONCAT(a, b) a##b
30
31#define ARM_DOT1(a, b, c) \
32 ({ \
33 c = fma(a, b, c); \
34 })
35#define ARM_DOT2(a, b, c) \
36 ({ \
37 c = fma(a.s0, b.s0, c); \
38 c = fma(a.s1, b.s1, c); \
39 })
40#define ARM_DOT3(a, b, c) \
41 ({ \
42 ARM_DOT2(a, b, c); \
43 c = fma((a.s2), (b.s2), c); \
44 })
45#define ARM_DOT4(a, b, c) \
46 ({ \
47 ARM_DOT3(a, b, c); \
48 c = fma((a.s3), (b.s3), c); \
49 })
50#define ARM_DOT8(a, b, c) \
51 ({ \
52 ARM_DOT4((a.lo), (b.lo), c); \
53 ARM_DOT4((a.hi), (b.hi), c); \
54 })
55#define ARM_DOT16(a, b, c) \
56 ({ \
57 ARM_DOT8((a.lo), (b.lo), c); \
58 ARM_DOT8((a.hi), (b.hi), c); \
59 })
60
61#if N0 == 2
62#define ARM_DOT_K0XN0(k0, a, b, c) \
63 ({ \
64 CONCAT(ARM_DOT, k0) \
65 ((a), (b##0), (c.s0)); \
66 CONCAT(ARM_DOT, k0) \
67 ((a), (b##1), (c.s1)); \
68 })
69#elif N0 == 3 // N0 == 3
70#define ARM_DOT_K0XN0(k0, a, b, c) \
71 ({ \
72 CONCAT(ARM_DOT, k0) \
73 ((a), (b##0), (c.s0)); \
74 CONCAT(ARM_DOT, k0) \
75 ((a), (b##1), (c.s1)); \
76 CONCAT(ARM_DOT, k0) \
77 ((a), (b##2), (c.s2)); \
78 })
79#elif N0 == 4 // N0 == 4
80#define ARM_DOT_K0XN0(k0, a, b, c) \
81 ({ \
82 CONCAT(ARM_DOT, k0) \
83 ((a), (b##0), (c.s0)); \
84 CONCAT(ARM_DOT, k0) \
85 ((a), (b##1), (c.s1)); \
86 CONCAT(ARM_DOT, k0) \
87 ((a), (b##2), (c.s2)); \
88 CONCAT(ARM_DOT, k0) \
89 ((a), (b##3), (c.s3)); \
90 })
91#elif N0 == 8 // N0 == 8
92#define ARM_DOT_K0XN0(k0, a, b, c) \
93 ({ \
94 CONCAT(ARM_DOT, k0) \
95 ((a), (b##0), (c.s0)); \
96 CONCAT(ARM_DOT, k0) \
97 ((a), (b##1), (c.s1)); \
98 CONCAT(ARM_DOT, k0) \
99 ((a), (b##2), (c.s2)); \
100 CONCAT(ARM_DOT, k0) \
101 ((a), (b##3), (c.s3)); \
102 CONCAT(ARM_DOT, k0) \
103 ((a), (b##4), (c.s4)); \
104 CONCAT(ARM_DOT, k0) \
105 ((a), (b##5), (c.s5)); \
106 CONCAT(ARM_DOT, k0) \
107 ((a), (b##6), (c.s6)); \
108 CONCAT(ARM_DOT, k0) \
109 ((a), (b##7), (c.s7)); \
110 })
111#elif N0 == 16 // N0 == 16
112#define ARM_DOT_K0XN0(k0, a, b, c) \
113 ({ \
114 CONCAT(ARM_DOT, k0) \
115 ((a), (b##0), (c.s0)); \
116 CONCAT(ARM_DOT, k0) \
117 ((a), (b##1), (c.s1)); \
118 CONCAT(ARM_DOT, k0) \
119 ((a), (b##2), (c.s2)); \
120 CONCAT(ARM_DOT, k0) \
121 ((a), (b##3), (c.s3)); \
122 CONCAT(ARM_DOT, k0) \
123 ((a), (b##4), (c.s4)); \
124 CONCAT(ARM_DOT, k0) \
125 ((a), (b##5), (c.s5)); \
126 CONCAT(ARM_DOT, k0) \
127 ((a), (b##6), (c.s6)); \
128 CONCAT(ARM_DOT, k0) \
129 ((a), (b##7), (c.s7)); \
130 CONCAT(ARM_DOT, k0) \
131 ((a), (b##8), (c.s8)); \
132 CONCAT(ARM_DOT, k0) \
133 ((a), (b##9), (c.s9)); \
134 CONCAT(ARM_DOT, k0) \
135 ((a), (b##A), (c.sA)); \
136 CONCAT(ARM_DOT, k0) \
137 ((a), (b##B), (c.sB)); \
138 CONCAT(ARM_DOT, k0) \
139 ((a), (b##C), (c.sC)); \
140 CONCAT(ARM_DOT, k0) \
141 ((a), (b##D), (c.sD)); \
142 CONCAT(ARM_DOT, k0) \
143 ((a), (b##E), (c.sE)); \
144 CONCAT(ARM_DOT, k0) \
145 ((a), (b##F), (c.sF)); \
146 })
147#else // N0 not supported
148#error "N0 value not supported"
149#endif // N0 conditions
150
151#if defined(GEMM_MM_RESHAPED_ONLY_RHS_T)
152/** This OpenCL kernel computes the matrix multiplication between 2 matrices.
153 * The LHS matrix is NOT reshaped
154 * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed
Ramy Elgammal451c3092022-02-01 23:01:27 +0000[diff] [blame]155 *
156 * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
157 * @note The GEMM's dimensions (M,N and K) must be passed at runtime as kernel parameters.
158 * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
159 * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
160 * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
161 * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
162 * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
163 * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
164 * @note Only the following configurations of M0, N0 and K0 are currently supported:
165 * - M0 = 1, 2, 3, 4, 5, 6, 7, 8
166 * - N0 = 2, 3, 4, 8, 16
167 * - K0 = 2, 3, 4, 8, 16
168 * - H0 >= 1
169 *
170 * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
171 * The activation function is performed after the bias addition
172 * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
173 * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
174 * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
175 * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
176 * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
177 * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
178 *
179 * @param[in] lhs_ptr Pointer to the LHS matrix. Supported data type: F16/F32
180 * @param[in] lhs_stride_x Stride of the LHS matrix in X dimension (in bytes)
181 * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
182 * @param[in] lhs_stride_y Stride of the LHS matrix in Y dimension (in bytes)
183 * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
184 * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS matrix
185 * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
186 * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes)
187 * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
188 * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes)
189 * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
190 * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
191 * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
192 * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
193 * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
194 * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
195 * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
196 * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
197 * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
198 * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
199 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
200 * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
201 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
202 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
203 * @param[in] lhs_stride_z Stride of the LHS matrix in Z dimension (in bytes)
204 * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
205 * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
206 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
207 * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
208 * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
209 * @param[in] M Number of rows in LHS matrix not reshaped.
210 * @param[in] N Number of columns in RHS matrix not reshaped.
211 * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped.
212 */
213__kernel void gemm_mm_reshaped_only_rhs_t(IMAGE_DECLARATION(lhs),
214 IMAGE_DECLARATION(rhs),
215#if defined(BETA)
216 IMAGE_DECLARATION(bias),
217#endif // defined(BETA)
218 IMAGE_DECLARATION(dst),
219 uint lhs_stride_z,
220 uint rhs_stride_z,
221#if defined(BETA)
222 uint bias_stride_z,
223#endif //defined(BETA)
224 uint dst_stride_z
225#if defined(REINTERPRET_INPUT_AS_3D)
226 ,
227 uint lhs_cross_plane_pad
228#endif // REINTERPRET_INPUT_AS_3D
229#if defined(REINTERPRET_OUTPUT_AS_3D)
230 ,
231 uint dst_cross_plane_pad
232#endif // REINTERPRET_OUTPUT_AS_3D
233 ,
234 const int M,
235 const int N,
236 const int K)
237{
238 // Block size
239#define RHS_BLOCK_SIZE ((K0) * (N0))
240
241 // RHS offset and step X
242#if defined(RHS_INTERLEAVE)
243#define RHS_OFFSET_X (K0)
244#define RHS_STEP_X ((K0) * (H0))
245#define RHS_STEP_LOOP (1)
246#else // defined(RHS_INTERLEAVE)
247#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
248#define RHS_STEP_X (K0)
249#define RHS_STEP_LOOP (H0)
250#endif // defined(RHS_INTERLEAVE)
251
252 uint x = get_global_id(0);
253 uint y = get_global_id(1);
254 uint z = get_global_id(2);
255
256 const bool cond_y = y == 0;
257 const bool cond_x = ((x + 1) * N0 >= N);
258
259#if defined(DUMMY_WORK_ITEMS)
260 if((x * N0 >= N) || (y * M0 >= M))
261 {
262 return;
263 }
264#endif // defined(DUMMY_WORK_ITEMS)
265
266 // Compute LHS matrix address
267 uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
268
269 // Compute RHS reshaped matrix address
270 uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;
271
272#if defined(MATRIX_B_DEPTH)
273 // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
274 rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z;
275#else // defined(MATRIX_B_DEPTH)
276 rhs_offset += z * rhs_stride_z;
277#endif // defined(MATRIX_B_DEPTH)
278
279 REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0;
280 REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);
281
282#if defined(REINTERPRET_INPUT_AS_3D)
283 // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
284 CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
285
286 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
287 // multiply lhs_stride_z by DEPTH_GEMM3D
288 lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;
289
290#else // defined(REINTERPRET_INPUT_AS_3D)
291
292 // Add offset for batched GEMM
293 lhs_offset += z * lhs_stride_z;
294
295#endif // defined(REINTERPRET_INPUT_AS_3D)
296
297 // Initialize the accumulators
298 REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(M0-1)=0;
299
300 int i = 0;
301 for(; i <= (K - K0); i += K0)
302 {
303 // Supported cases (M0, K0):
304 // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
305 // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
306 // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
307 // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
308 // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
309 // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
310 // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
311 // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
312 // Load values from LHS matrix
313 LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);
314
315 // Load values from RHS reshaped matrix
316 LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero);
317
318 // Accumulate
319 ARM_DOT_K0XN0(K0, a0, b, c0);
320#if M0 > 1
321 ARM_DOT_K0XN0(K0, a1, b, c1);
322#endif // M0 > 1
323#if M0 > 2
324 ARM_DOT_K0XN0(K0, a2, b, c2);
325#endif // M0 > 2
326#if M0 > 3
327 ARM_DOT_K0XN0(K0, a3, b, c3);
328#endif // M0 > 3
329#if M0 > 4
330 ARM_DOT_K0XN0(K0, a4, b, c4);
331#endif // M0 > 4
332#if M0 > 5
333 ARM_DOT_K0XN0(K0, a5, b, c5);
334#endif // M0 > 5
335#if M0 > 6
336 ARM_DOT_K0XN0(K0, a6, b, c6);
337#endif // M0 > 6
338#if M0 > 7
339 ARM_DOT_K0XN0(K0, a7, b, c7);
340#endif // M0 > 7
341
342 lhs_offset += K0 * sizeof(DATA_TYPE);
343 rhs_offset += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE);
344 }
345
346 // Left-over accumulations
347 for(; i < K; ++i)
348 {
349 // Load values from LHS matrix
350 LOAD_BLOCK(M0, 1, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);
351
352 // Load values from RHS reshaped matrix
353 LOAD_BLOCK(N0, 1, DATA_TYPE, b, rhs_ptr, rhs_offset, RHS_STEP_X * sizeof(DATA_TYPE), zero);
354
355 // Accumulate
356 ARM_DOT_K0XN0(1, a0, b, c0);
357#if M0 > 1
358 ARM_DOT_K0XN0(1, a1, b, c1);
359#endif // M0 > 1
360#if M0 > 2
361 ARM_DOT_K0XN0(1, a2, b, c2);
362#endif // M0 > 2
363#if M0 > 3
364 ARM_DOT_K0XN0(1, a3, b, c3);
365#endif // M0 > 3
366#if M0 > 4
367 ARM_DOT_K0XN0(1, a4, b, c4);
368#endif // M0 > 4
369#if M0 > 5
370 ARM_DOT_K0XN0(1, a5, b, c5);
371#endif // M0 > 5
372#if M0 > 6
373 ARM_DOT_K0XN0(1, a6, b, c6);
374#endif // M0 > 6
375#if M0 > 7
376 ARM_DOT_K0XN0(1, a7, b, c7);
377#endif // M0 > 7
378
379 lhs_offset += sizeof(DATA_TYPE);
380 rhs_offset += sizeof(DATA_TYPE);
381 }
382
383 __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
384
385 REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
386
387#if defined(REINTERPRET_OUTPUT_AS_3D)
388
389 // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
390 CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
391
392 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
393 // multiply dst_stride_z by DEPTH_GEMM3D
394 dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
395
396#else // defined(REINTERPRET_OUTPUT_AS_3D)
397
398 // Add offset for batched GEMM
399 dst_addr += z * dst_stride_z;
400
401#endif // defined(REINTERPRET_OUTPUT_AS_3D)
402
403 // Multiply by the weight of matrix-matrix product and store the result
404#if defined(ALPHA)
405 SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
406#endif // defined(ALPHA)
407
408 // Add beta*bias
409#if defined(BETA)
410#if defined(BROADCAST_BIAS)
411 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
412
413 LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
414
415#ifndef UNIT_BETA
416 SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
417#endif // UNIT_BIAS
418
419 // c = c + bias[broadcasted]
420 ADD_BLOCK_BROADCAST(M0, c, bias0);
421
422#else // defined(BROADCAST_BIAS)
423 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
424
425 LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
426
427#ifndef UNIT_BETA
428 SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
429#endif // UNIT_BIAS
430
431 // c = c + bias
432 ADD_BLOCK(M0, c, bias);
433
434#endif // defined(BROADCAST_BIAS)
435#endif // defined(BETA)
436
437#if defined(ACTIVATION_TYPE)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]438 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, N0, c, A_VAL, B_VAL);
Ramy Elgammal451c3092022-02-01 23:01:27 +0000[diff] [blame]439#endif // defined(ACTIVATION_TYPE)
440
441 // Store output block
442 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
443
444#undef RHS_BLOCK_SIZE
445#undef RHS_OFFSET_X
446#undef RHS_STEP_X
447#undef RHS_STEP_LOOP
448}
449#endif // defined(GEMM_MM_RESHAPED_ONLY_RHS_T)
450
451#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_T_TEXTURE)
452/** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image
453 * The LHS matrix is NOT reshaped
454 * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is transposed
Ramy Elgammal451c3092022-02-01 23:01:27 +0000[diff] [blame]455 *
456 * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel
457 * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
458 * @note The GEMM's dimensions (M,N and K) must be passed at runtime as kernel parameters.
459 * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32)
460 * Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT
461 * could be different from the value returned by get_image_height(rhs_img).
462 * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
463 * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
464 * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
465 * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
466 * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
467 * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
468 * @note Only the following configurations of M0, N0 and K0 are currently supported:
469 * - M0 = 1, 2, 3, 4, 5, 6, 7, 8
470 * - N0 = 4, 8, 16
471 * - K0 = 4, 8, 16
472 * - H0 >= 1
473 *
474 * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
475 * The activation function is performed after the bias addition
476 * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
477 * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
478 * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
479 * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
480 * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
481 * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
482 *
483 * @param[in] lhs_ptr Pointer to the LHS matrix. Supported data type: F32
484 * @param[in] lhs_stride_x Stride of the LHS matrix in X dimension (in bytes)
485 * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
486 * @param[in] lhs_stride_y Stride of the LHS matrix in Y dimension (in bytes)
487 * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
488 * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS matrix
489 * @param[in] rhs_img The RHS reshaped matrix as OpenCL image object. Supported data type: same as @p lhs_ptr
490 * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
491 * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
492 * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
493 * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
494 * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
495 * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
496 * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
497 * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
498 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
499 * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
500 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
501 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
502 * @param[in] lhs_stride_z Stride of the LHS matrix in Z dimension (in bytes)
503 * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
504 * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
505 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
506 * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
507 * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
508 * @param[in] M Number of rows in LHS matrix not reshaped.
509 * @param[in] N Number of columns in RHS matrix not reshaped.
510 * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped.
511 */
512__kernel void gemm_mm_reshaped_only_rhs_t_texture(IMAGE_DECLARATION(lhs),
513 __read_only image2d_t rhs_img,
514#if defined(BETA)
515 IMAGE_DECLARATION(bias),
516#endif // defined(BETA)
517 IMAGE_DECLARATION(dst),
518 uint lhs_stride_z,
519 uint rhs_stride_z,
520#if defined(BETA)
521 uint bias_stride_z,
522#endif //defined(BETA)
523 uint dst_stride_z
524#if defined(REINTERPRET_INPUT_AS_3D)
525 ,
526 uint lhs_cross_plane_pad
527#endif // REINTERPRET_INPUT_AS_3D
528#if defined(REINTERPRET_OUTPUT_AS_3D)
529 ,
530 uint dst_cross_plane_pad
531#endif // REINTERPRET_OUTPUT_AS_3D
532 ,
533 const int M,
534 const int N,
535 const int K)
536{
537 // Pixel unit
538#define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0)
539
540 const uint LEFTOVER_K = K % K0;
541
542 // Block size
543#define RHS_BLOCK_SIZE (PIXEL_UNIT * (N0))
544
545 // RHS offset and step X
546#if defined(RHS_INTERLEAVE)
547#define RHS_OFFSET_X (PIXEL_UNIT)
548#define RHS_STEP_X (PIXEL_UNIT * (H0))
549#define RHS_STEP_LOOP (1)
550#else // defined(RHS_INTERLEAVE)
551#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
552#define RHS_STEP_X PIXEL_UNIT
553#define RHS_STEP_LOOP (H0)
554#endif // defined(RHS_INTERLEAVE)
555
556 uint x = get_global_id(0);
557 uint y = get_global_id(1);
558 uint z = get_global_id(2);
559
560 const bool cond_y = y == 0;
561 const bool cond_x = ((x + 1) * N0 >= N);
562
563#if defined(DUMMY_WORK_ITEMS)
564 if((x * N0 >= N) || (y * M0 >= M))
565 {
566 return;
567 }
568#endif // defined(DUMMY_WORK_ITEMS)
569
570 // Compute LHS matrix address
571 uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
572
573#if defined(MATRIX_B_DEPTH)
574 // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
575 const uint z_rhs = (get_global_id(2) % MATRIX_B_DEPTH);
576#else // defined(MATRIX_B_DEPTH)
577 const uint z_rhs = get_global_id(2);
578#endif // defined(MATRIX_B_DEPTH)
579
580 // Compute RHS matrix coordinates
581 uint x_rhs = (get_global_id(0) % H0) * (uint)RHS_OFFSET_X;
582 const uint y_rhs = (get_global_id(0) / (uint)H0) + z_rhs * RHS_HEIGHT;
583
584 REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0);
585 REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);
586
587#if defined(REINTERPRET_INPUT_AS_3D)
588 // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
589 CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
590
591 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
592 // multiply lhs_stride_z by DEPTH_GEMM3D
593 lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;
594
595#else // defined(REINTERPRET_INPUT_AS_3D)
596
597 // Add offset for batched GEMM
598 lhs_offset += z * lhs_stride_z;
599
600#endif // defined(REINTERPRET_INPUT_AS_3D)
601
602 // Initialize the accumulators
603 REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0);
604
605 int i = 0;
606 for(; i <= (K - K0); i += K0)
607 {
608 // Load values from LHS matrix
609 LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);
610
611 // Load values from RHS matrix stored in a cl_image
612 REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0);
613 LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0);
614
615 // Accumulate
616 ARM_DOT_K0XN0(K0, a0, b, c0);
617#if M0 > 1
618 ARM_DOT_K0XN0(K0, a1, b, c1);
619#endif // M0 > 1
620#if M0 > 2
621 ARM_DOT_K0XN0(K0, a2, b, c2);
622#endif // M0 > 2
623#if M0 > 3
624 ARM_DOT_K0XN0(K0, a3, b, c3);
625#endif // M0 > 3
626#if M0 > 4
627 ARM_DOT_K0XN0(K0, a4, b, c4);
628#endif // M0 > 4
629#if M0 > 5
630 ARM_DOT_K0XN0(K0, a5, b, c5);
631#endif // M0 > 5
632#if M0 > 6
633 ARM_DOT_K0XN0(K0, a6, b, c6);
634#endif // M0 > 6
635#if M0 > 7
636 ARM_DOT_K0XN0(K0, a7, b, c7);
637#endif // M0 > 7
638
639 lhs_offset += K0 * sizeof(DATA_TYPE);
640 x_rhs += N0 * RHS_STEP_X * RHS_STEP_LOOP;
641 }
642
643 if(LEFTOVER_K != 0)
644 {
645 // Note: We cannot read out-of-bound elements from the RHS matrix because
646 // the RHS width is always multiple of K0. This is not be true for the LHS matrix
647 // Left-over accumulations for LHS matrix
648
649 union UNION_VEC_TYPE
650 {
651 DATA_TYPE s[K0];
652 VEC_DATA_TYPE(DATA_TYPE, K0)
653 v;
654 };
655
656 union UNION_VEC_TYPE a0 = {.v = 0 };
657#if M0 > 1
658 union UNION_VEC_TYPE a1 = {.v = 0 };
659#endif // M0 > 1
660#if M0 > 2
661 union UNION_VEC_TYPE a2 = {.v = 0 };
662#endif // M0 > 2
663#if M0 > 3
664 union UNION_VEC_TYPE a3 = {.v = 0 };
665#endif // M0 > 3
666#if M0 > 4
667 union UNION_VEC_TYPE a4 = {.v = 0 };
668#endif // M0 > 4
669#if M0 > 5
670 union UNION_VEC_TYPE a5 = {.v = 0 };
671#endif // M0 > 5
672#if M0 > 6
673 union UNION_VEC_TYPE a6 = {.v = 0 };
674#endif // M0 > 6
675#if M0 > 7
676 union UNION_VEC_TYPE a7 = {.v = 0 };
677#endif // M0 > 7
678
679 REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0);
680
681 // Load from RHS matrix
682 LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0);
683
684 // Load from LHS matrix
685 for(int k = 0; k < LEFTOVER_K; ++k)
686 {
687 a0.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0);
688#if M0 > 1
689 a1.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1);
690#endif // M0 > 1
691#if M0 > 2
692 a2.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2);
693#endif // M0 > 2
694#if M0 > 3
695 a3.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3);
696#endif // M0 > 3
697#if M0 > 4
698 a4.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4);
699#endif // M0 > 4
700#if M0 > 5
701 a5.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5);
702#endif // M0 > 5
703#if M0 > 6
704 a6.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6);
705#endif // M0 > 6
706#if M0 > 7
707 a7.s[k] = *(__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7);
708#endif // M0 > 7
709
710 lhs_offset += sizeof(DATA_TYPE);
711 }
712
713 // Accumulate
714 ARM_DOT_K0XN0(K0, a0.v, b, c0);
715#if M0 > 1
716 ARM_DOT_K0XN0(K0, a1.v, b, c1);
717#endif // M0 > 1
718#if M0 > 2
719 ARM_DOT_K0XN0(K0, a2.v, b, c2);
720#endif // M0 > 2
721#if M0 > 3
722 ARM_DOT_K0XN0(K0, a3.v, b, c3);
723#endif // M0 > 3
724#if M0 > 4
725 ARM_DOT_K0XN0(K0, a4.v, b, c4);
726#endif // M0 > 4
727#if M0 > 5
728 ARM_DOT_K0XN0(K0, a5.v, b, c5);
729#endif // M0 > 5
730#if M0 > 6
731 ARM_DOT_K0XN0(K0, a6.v, b, c6);
732#endif // M0 > 6
733#if M0 > 7
734 ARM_DOT_K0XN0(K0, a7.v, b, c7);
735#endif // M0 > 7
736 }
737
738 __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
739
740 REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
741
742#if defined(REINTERPRET_OUTPUT_AS_3D)
743
744 // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
745 CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
746
747 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
748 // multiply dst_stride_z by DEPTH_GEMM3D
749 dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
750
751#else // defined(REINTERPRET_OUTPUT_AS_3D)
752
753 // Add offset for batched GEMM
754 dst_addr += z * dst_stride_z;
755
756#endif // defined(REINTERPRET_OUTPUT_AS_3D)
757
758 // Multiply by the weight of matrix-matrix product and store the result
759#if defined(ALPHA)
760 SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
761#endif // defined(ALPHA)
762
763 // Add beta*bias
764#if defined(BETA)
765#if defined(BROADCAST_BIAS)
766 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
767
768 LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
769
770#ifndef UNIT_BETA
771 SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
772#endif // UNIT_BIAS
773
774 // c = c + bias[broadcasted]
775 ADD_BLOCK_BROADCAST(M0, c, bias0);
776
777#else // defined(BROADCAST_BIAS)
778 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
779
780 LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
781
782#ifndef UNIT_BETA
783 SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
784#endif // UNIT_BIAS
785
786 // c = c + bias
787 ADD_BLOCK(M0, c, bias);
788
789#endif // defined(BROADCAST_BIAS)
790#endif // defined(BETA)
791
792#if defined(ACTIVATION_TYPE)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]793 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, N0, c, A_VAL, B_VAL);
Ramy Elgammal451c3092022-02-01 23:01:27 +0000[diff] [blame]794#endif // defined(ACTIVATION_TYPE)
795
796 // Store output block
797 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
798
799#undef RHS_BLOCK_SIZE
800#undef RHS_OFFSET_X
801#undef RHS_STEP_X
802#undef RHS_STEP_LOOP
803#undef PIXEL_UNIT
804}
805#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_T_TEXTURE)
806
807#define VFMA(a, b, c) \
808 ({ \
809 c = fma(a, b, c); \
810 })
811
812#if M0 == 1
813#define VFMA_M0xN0(i, a, b, c) \
814 ({ \
815 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
816 })
817#elif M0 == 2 // M0 == 2
818#define VFMA_M0xN0(i, a, b, c) \
819 ({ \
820 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
821 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
822 })
823#elif M0 == 3 // M0 == 3
824#define VFMA_M0xN0(i, a, b, c) \
825 ({ \
826 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
827 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
828 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
829 })
830#elif M0 == 4 // M0 == 4
831#define VFMA_M0xN0(i, a, b, c) \
832 ({ \
833 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
834 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
835 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
836 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
837 })
838#elif M0 == 5 // M0 == 5
839#define VFMA_M0xN0(i, a, b, c) \
840 ({ \
841 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
842 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
843 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
844 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
845 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
846 })
847#elif M0 == 6 // M0 == 6
848#define VFMA_M0xN0(i, a, b, c) \
849 ({ \
850 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
851 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
852 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
853 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
854 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
855 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
856 })
857#elif M0 == 7 // M0 == 7
858#define VFMA_M0xN0(i, a, b, c) \
859 ({ \
860 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
861 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
862 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
863 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
864 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
865 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
866 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \
867 })
868#elif M0 == 8 // M0 == 8
869#define VFMA_M0xN0(i, a, b, c) \
870 ({ \
871 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
872 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
873 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
874 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
875 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
876 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
877 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \
878 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##7).s##i), b, (c##7)); \
879 })
880#else // M0 not supported
881#error "M0 not supported"
882#endif // M0 not supported
883
884#if defined(GEMM_MM_RESHAPED_ONLY_RHS_NT)
885/** This OpenCL kernel computes the matrix multiplication between 2 matrices.
886 * The LHS matrix is NOT reshaped
887 * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is NOT transposed
Ramy Elgammal451c3092022-02-01 23:01:27 +0000[diff] [blame]888 *
889 * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
890 * @note The GEMM's dimensions (M,N and K) must be passed at runtime as kernel parameters.
891 * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
892 * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
893 * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
894 * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
895 * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
896 * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
897 * @note Only the following configurations of M0, N0 and K0 are currently supported:
898 * - M0 = 1, 2, 3, 4, 5, 6, 7, 8
899 * - N0 = 2, 3, 4, 8, 16
900 * - K0 = 2, 3, 4, 8, 16
901 * - H0 >= 1
902 *
903 * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
904 * The activation function is performed after the bias addition
905 * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
906 * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
907 * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
908 * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
909 * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
910 * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
911 *
912 * @param[in] lhs_ptr Pointer to the LHS matrix. Supported data type: F16/F32
913 * @param[in] lhs_stride_x Stride of the LHS matrix in X dimension (in bytes)
914 * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
915 * @param[in] lhs_stride_y Stride of the LHS matrix in Y dimension (in bytes)
916 * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
917 * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS matrix
918 * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
919 * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes)
920 * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
921 * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes)
922 * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
923 * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
924 * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
925 * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
926 * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
927 * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
928 * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
929 * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
930 * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
931 * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
932 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
933 * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
934 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
935 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
936 * @param[in] lhs_stride_z Stride of the LHS matrix in Z dimension (in bytes)
937 * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
938 * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
939 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
940 * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
941 * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
942 * @param[in] M Number of rows in LHS matrix not reshaped.
943 * @param[in] N Number of columns in RHS matrix not reshaped.
944 * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped.
945 */
946__kernel void gemm_mm_reshaped_only_rhs_nt(IMAGE_DECLARATION(lhs),
947 IMAGE_DECLARATION(rhs),
948#if defined(BETA)
949 IMAGE_DECLARATION(bias),
950#endif // defined(BETA)
951 IMAGE_DECLARATION(dst),
952 uint lhs_stride_z,
953 uint rhs_stride_z,
954#if defined(BETA)
955 uint bias_stride_z,
956#endif //defined(BETA)
957 uint dst_stride_z
958#if defined(REINTERPRET_INPUT_AS_3D)
959 ,
960 uint lhs_cross_plane_pad
961#endif // REINTERPRET_INPUT_AS_3D
962#if defined(REINTERPRET_OUTPUT_AS_3D)
963 ,
964 uint dst_cross_plane_pad
965#endif // REINTERPRET_OUTPUT_AS_3D
966 ,
967 const int M,
968 const int N,
969 const int K)
970{
971 // Block size
972#define RHS_BLOCK_SIZE ((K0) * (N0))
973
974 // RHS offset and step X
975#if defined(RHS_INTERLEAVE)
976#define RHS_OFFSET_X (N0)
977#define RHS_STEP_X ((N0) * (H0))
978#define RHS_STEP_LOOP (1)
979#else // defined(RHS_INTERLEAVE)
980#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
981#define RHS_STEP_X (N0)
982#define RHS_STEP_LOOP (H0)
983#endif // defined(RHS_INTERLEAVE)
984
985 uint x = get_global_id(0);
986 uint y = get_global_id(1);
987 uint z = get_global_id(2);
988
989 const bool cond_y = y == 0;
990 const bool cond_x = ((x + 1) * N0 >= N);
991
992#if defined(DUMMY_WORK_ITEMS)
993 if((x * N0 >= N) || (y * M0 >= M))
994 {
995 return;
996 }
997#endif // defined(DUMMY_WORK_ITEMS)
998
999 // Compute LHS matrix address
1000 uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
1001
1002 // Compute RHS reshaped matrix address
1003 uint rhs_offset = rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;
1004
1005#if defined(MATRIX_B_DEPTH)
1006 // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
1007 rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z;
1008#else // defined(MATRIX_B_DEPTH)
1009 rhs_offset += z * rhs_stride_z;
1010#endif // defined(MATRIX_B_DEPTH)
1011
1012 REPEAT_VAR_INIT_TO_CONST(8, uint, zin, 0); //uint zin0=0,zin1=0,zin2=0,... zin7=0;
1013 REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0); //uint zero0=0,zero1=0,zero2=0,... zero7=0;
1014
1015#if defined(REINTERPRET_INPUT_AS_3D)
1016
1017 // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
1018 CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
1019
1020 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
1021 // multiply lhs_stride_z by DEPTH_GEMM3D
1022 lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;
1023
1024#else // defined(REINTERPRET_INPUT_AS_3D)
1025
1026 // Add offset for batched GEMM
1027 lhs_offset += z * lhs_stride_z;
1028
1029#endif // defined(REINTERPRET_INPUT_AS_3D)
1030
1031 // Initialize the accumulators
1032 REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(N0-1)=0;
1033
1034 int i = 0;
1035 for(; i <= (K - K0); i += K0)
1036 {
1037 // Supported cases (M0, K0):
1038 // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
1039 // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
1040 // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
1041 // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
1042 // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
1043 // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
1044 // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
1045 // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
1046 // Load values from LHS matrix
1047 LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zin);
1048
1049 VEC_DATA_TYPE(DATA_TYPE, N0)
1050 b0;
1051
1052 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE)));
1053 VFMA_M0xN0(0, a, b0, c);
1054 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 1 * RHS_STEP_X * sizeof(DATA_TYPE)));
1055 VFMA_M0xN0(1, a, b0, c);
1056#if K0 > 2
1057 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 2 * RHS_STEP_X * sizeof(DATA_TYPE)));
1058 VFMA_M0xN0(2, a, b0, c);
1059#endif // K0 > 2
1060#if K0 > 3
1061 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 3 * RHS_STEP_X * sizeof(DATA_TYPE)));
1062 VFMA_M0xN0(3, a, b0, c);
1063#endif // K0 > 3
1064#if K0 > 4
1065 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 4 * RHS_STEP_X * sizeof(DATA_TYPE)));
1066 VFMA_M0xN0(4, a, b0, c);
1067 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 5 * RHS_STEP_X * sizeof(DATA_TYPE)));
1068 VFMA_M0xN0(5, a, b0, c);
1069 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 6 * RHS_STEP_X * sizeof(DATA_TYPE)));
1070 VFMA_M0xN0(6, a, b0, c);
1071 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 7 * RHS_STEP_X * sizeof(DATA_TYPE)));
1072 VFMA_M0xN0(7, a, b0, c);
1073#endif // K0 > 4
1074#if K0 > 8
1075 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 8 * RHS_STEP_X * sizeof(DATA_TYPE)));
1076 VFMA_M0xN0(8, a, b0, c);
1077 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 9 * RHS_STEP_X * sizeof(DATA_TYPE)));
1078 VFMA_M0xN0(9, a, b0, c);
1079 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 10 * RHS_STEP_X * sizeof(DATA_TYPE)));
1080 VFMA_M0xN0(A, a, b0, c);
1081 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 11 * RHS_STEP_X * sizeof(DATA_TYPE)));
1082 VFMA_M0xN0(B, a, b0, c);
1083 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 12 * RHS_STEP_X * sizeof(DATA_TYPE)));
1084 VFMA_M0xN0(C, a, b0, c);
1085 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 13 * RHS_STEP_X * sizeof(DATA_TYPE)));
1086 VFMA_M0xN0(D, a, b0, c);
1087 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 14 * RHS_STEP_X * sizeof(DATA_TYPE)));
1088 VFMA_M0xN0(E, a, b0, c);
1089 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 15 * RHS_STEP_X * sizeof(DATA_TYPE)));
1090 VFMA_M0xN0(F, a, b0, c);
1091#endif // K0 > 8
1092
1093 lhs_offset += K0 * sizeof(DATA_TYPE);
1094 rhs_offset += K0 * RHS_STEP_X * RHS_STEP_LOOP * sizeof(DATA_TYPE);
1095 }
1096
1097 // Left-over accumulations
1098 for(; i < K; ++i)
1099 {
1100 // Load values from LHS matrix
1101 VEC_DATA_TYPE(DATA_TYPE, 2)
1102 a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0));
1103#if M0 > 1
1104 VEC_DATA_TYPE(DATA_TYPE, 2)
1105 a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1));
1106#endif // M0 > 1
1107#if M0 > 2
1108 VEC_DATA_TYPE(DATA_TYPE, 2)
1109 a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2));
1110#endif // M0 > 2
1111#if M0 > 3
1112 VEC_DATA_TYPE(DATA_TYPE, 2)
1113 a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3));
1114#endif // M0 > 3
1115#if M0 > 4
1116 VEC_DATA_TYPE(DATA_TYPE, 2)
1117 a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4));
1118#endif // M0 > 4
1119#if M0 > 5
1120 VEC_DATA_TYPE(DATA_TYPE, 2)
1121 a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5));
1122#endif // M0 > 5
1123#if M0 > 6
1124 VEC_DATA_TYPE(DATA_TYPE, 2)
1125 a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6));
1126#endif // M0 > 6
1127#if M0 > 7
1128 VEC_DATA_TYPE(DATA_TYPE, 2)
1129 a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7));
1130#endif // M0 > 7
1131
1132 VEC_DATA_TYPE(DATA_TYPE, N0)
1133 b0;
1134
1135 b0 = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * RHS_STEP_X * sizeof(DATA_TYPE)));
1136 VFMA_M0xN0(0, a, b0, c);
1137
1138 lhs_offset += sizeof(DATA_TYPE);
1139 rhs_offset += RHS_STEP_X * sizeof(DATA_TYPE);
1140 }
1141
1142 __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
1143
1144 REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
1145
1146#if defined(REINTERPRET_OUTPUT_AS_3D)
1147 // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
1148 CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
1149
1150 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
1151 // multiply dst_stride_z by DEPTH_GEMM3D
1152 dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
1153
1154#else // defined(REINTERPRET_OUTPUT_AS_3D)
1155
1156 // Add offset for batched GEMM
1157 dst_addr += z * dst_stride_z;
1158
1159#endif // defined(REINTERPRET_OUTPUT_AS_3D)
1160
1161 // Multiply by the weight of matrix-matrix product and store the result
1162#if defined(ALPHA)
1163 SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
1164#endif // defined(ALPHA)
1165
1166 // Add beta*bias
1167#if defined(BETA)
1168#if defined(BROADCAST_BIAS)
1169 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
1170
1171 LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
1172
1173#ifndef UNIT_BETA
1174 SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
1175#endif // UNIT_BIAS
1176
1177 // c = c + bias[broadcasted]
1178 ADD_BLOCK_BROADCAST(M0, c, bias0);
1179
1180#else // defined(BROADCAST_BIAS)
1181 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
1182
1183 LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
1184
1185#ifndef UNIT_BETA
1186 SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
1187#endif // UNIT_BIAS
1188
1189 // c = c + bias
1190 ADD_BLOCK(M0, c, bias);
1191
1192#endif // defined(BROADCAST_BIAS)
1193#endif // defined(BETA)
1194
1195#if defined(ACTIVATION_TYPE)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]1196 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, N0, c, A_VAL, B_VAL);
Ramy Elgammal451c3092022-02-01 23:01:27 +0000[diff] [blame]1197#endif // defined(ACTIVATION_TYPE)
1198
1199 // Store output block
1200 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
1201
1202#undef RHS_BLOCK_SIZE
1203#undef RHS_OFFSET_X
1204#undef RHS_STEP_X
1205#undef RHS_STEP_LOOP
1206}
1207#endif // defined(GEMM_MM_RESHAPED_ONLY_RHS_NT)
1208
1209#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_TEXTURE)
1210/** This OpenCL kernel computes the matrix multiplication between 2 matrices.
1211 * The LHS matrix is NOT reshaped
1212 * The RHS is reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the block K0xN0 is NOT transposed
Ramy Elgammal451c3092022-02-01 23:01:27 +0000[diff] [blame]1213 *
1214 * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel
1215 * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
1216 * @note The GEMM's dimensions (M,N and K) must be passed at runtime as kernel parameters.
1217 * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32)
1218 * Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT
1219 * could be different from the value returned by get_image_height(rhs_img).
1220 * @note The block's dimensions used for reshaping the RHS matrix (N0 and K0) must be passed at compile time using -DN0 and -DK0 (e.g. -DN0=8, -DK0=4).
1221 * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
1222 * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
1223 * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
1224 * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
1225 * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
1226 * @note Only the following configurations of M0, N0 and K0 are currently supported:
1227 * - M0 = 1, 2, 3, 4, 5, 6, 7, 8
1228 * - N0 = 4, 8, 16
1229 * - K0 = 4, 8, 16
1230 * - H0 >= 1
1231 *
1232 * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
1233 * The activation function is performed after the bias addition
1234 * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
1235 * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
1236 * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
1237 * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
1238 * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
1239 * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
1240 *
1241 * @param[in] lhs_ptr Pointer to the LHS matrix. Supported data type: F32
1242 * @param[in] lhs_stride_x Stride of the LHS matrix in X dimension (in bytes)
1243 * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
1244 * @param[in] lhs_stride_y Stride of the LHS matrix in Y dimension (in bytes)
1245 * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
1246 * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS matrix
1247 * @param[in] rhs_img The RHS reshaped matrix as OpenCL image object. Supported data type: same as @p lhs_ptr
1248 * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
1249 * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
1250 * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
1251 * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
1252 * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
1253 * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
1254 * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
1255 * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
1256 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
1257 * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
1258 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
1259 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
1260 * @param[in] lhs_stride_z Stride of the LHS matrix in Z dimension (in bytes)
1261 * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
1262 * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
1263 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
1264 * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
1265 * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
1266 * @param[in] M Number of rows in LHS matrix not reshaped.
1267 * @param[in] N Number of columns in RHS matrix not reshaped.
1268 * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped.
1269 */
1270__kernel void gemm_mm_reshaped_only_rhs_nt_texture(IMAGE_DECLARATION(lhs),
1271 __read_only image2d_t rhs_img,
1272#if defined(BETA)
1273 IMAGE_DECLARATION(bias),
1274#endif // defined(BETA)
1275 IMAGE_DECLARATION(dst),
1276 uint lhs_stride_z,
1277 uint rhs_stride_z,
1278#if defined(BETA)
1279 uint bias_stride_z,
1280#endif //defined(BETA)
1281 uint dst_stride_z
1282#if defined(REINTERPRET_INPUT_AS_3D)
1283 ,
1284 uint lhs_cross_plane_pad
1285#endif // REINTERPRET_INPUT_AS_3D
1286#if defined(REINTERPRET_OUTPUT_AS_3D)
1287 ,
1288 uint dst_cross_plane_pad
1289#endif // REINTERPRET_OUTPUT_AS_3D
1290 ,
1291 const int M,
1292 const int N,
1293 const int K)
1294{
1295 // Pixel unit
1296#define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0)
1297
1298 // Block size
1299#define RHS_BLOCK_SIZE ((K0) * (PIXEL_UNIT))
1300
1301 // RHS offset and step X
1302#if defined(RHS_INTERLEAVE)
1303#define RHS_OFFSET_X (PIXEL_UNIT)
1304#define RHS_STEP_X ((PIXEL_UNIT) * (H0))
1305#define RHS_STEP_LOOP 1
1306#else // defined(RHS_INTERLEAVE)
1307#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
1308#define RHS_STEP_X (PIXEL_UNIT)
1309#define RHS_STEP_LOOP (H0)
1310#endif // defined(RHS_INTERLEAVE)
1311
1312 uint x = get_global_id(0);
1313 uint y = get_global_id(1);
1314 uint z = get_global_id(2);
1315
1316 const bool cond_y = y == 0;
1317 const bool cond_x = ((x + 1) * N0 >= N);
1318
1319#if defined(DUMMY_WORK_ITEMS)
1320 if((x * N0 >= N) || (y * M0 >= M))
1321 {
1322 return;
1323 }
1324#endif // defined(DUMMY_WORK_ITEMS)
1325
1326 // Compute LHS matrix address
1327 uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
1328
1329#if defined(MATRIX_B_DEPTH)
1330 // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
1331 const uint z_rhs = (z % MATRIX_B_DEPTH);
1332#else // defined(MATRIX_B_DEPTH)
1333 const uint z_rhs = z;
1334#endif // defined(MATRIX_B_DEPTH)
1335
1336 // Compute RHS matrix coordinates
1337 uint x_rhs = (x % H0) * (uint)RHS_OFFSET_X;
1338 const uint y_rhs = (x / (uint)H0) + z_rhs * RHS_HEIGHT;
1339
1340 REPEAT_VAR_INIT_TO_CONST(8, uint, zin, 0);
1341 REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);
1342
1343#if defined(REINTERPRET_INPUT_AS_3D)
1344
1345 // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
1346 CALCULATE_Z_OFFSET(M0, uint, zin, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
1347
1348 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
1349 // multiply lhs_stride_z by DEPTH_GEMM3D
1350 lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;
1351
1352#else // defined(REINTERPRET_INPUT_AS_3D)
1353
1354 // Add offset for batched GEMM
1355 lhs_offset += z * lhs_stride_z;
1356
1357#endif // defined(REINTERPRET_INPUT_AS_3D)
1358
1359 // Initialize the accumulators
1360 REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0);
1361
1362 int i = 0;
1363 for(; i <= (K - K0); i += K0)
1364 {
1365 // Load values from LHS matrix
1366 LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zin);
1367
1368 VEC_DATA_TYPE(DATA_TYPE, N0)
1369 b0;
1370
1371 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs));
1372 VFMA_M0xN0(0, a, b0, c);
1373 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 1 * RHS_STEP_X), (y_rhs));
1374 VFMA_M0xN0(1, a, b0, c);
1375#if K0 > 2
1376 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 2 * RHS_STEP_X), (y_rhs));
1377 VFMA_M0xN0(2, a, b0, c);
1378#endif // K0 > 2
1379#if K0 > 3
1380 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 3 * RHS_STEP_X), (y_rhs));
1381 VFMA_M0xN0(3, a, b0, c);
1382#endif // K0 > 3
1383#if K0 > 4
1384 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 4 * RHS_STEP_X), (y_rhs));
1385 VFMA_M0xN0(4, a, b0, c);
1386 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 5 * RHS_STEP_X), (y_rhs));
1387 VFMA_M0xN0(5, a, b0, c);
1388 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 6 * RHS_STEP_X), (y_rhs));
1389 VFMA_M0xN0(6, a, b0, c);
1390 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 7 * RHS_STEP_X), (y_rhs));
1391 VFMA_M0xN0(7, a, b0, c);
1392#endif // K0 > 4
1393#if K0 > 8
1394 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 8 * RHS_STEP_X), (y_rhs));
1395 VFMA_M0xN0(8, a, b0, c);
1396 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 9 * RHS_STEP_X), (y_rhs));
1397 VFMA_M0xN0(9, a, b0, c);
1398 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 10 * RHS_STEP_X), (y_rhs));
1399 VFMA_M0xN0(A, a, b0, c);
1400 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 11 * RHS_STEP_X), (y_rhs));
1401 VFMA_M0xN0(B, a, b0, c);
1402 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 12 * RHS_STEP_X), (y_rhs));
1403 VFMA_M0xN0(C, a, b0, c);
1404 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 13 * RHS_STEP_X), (y_rhs));
1405 VFMA_M0xN0(D, a, b0, c);
1406 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 14 * RHS_STEP_X), (y_rhs));
1407 VFMA_M0xN0(E, a, b0, c);
1408 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 15 * RHS_STEP_X), (y_rhs));
1409 VFMA_M0xN0(F, a, b0, c);
1410#endif // K0 > 8
1411
1412 lhs_offset += K0 * sizeof(DATA_TYPE);
1413 x_rhs += K0 * RHS_STEP_X * RHS_STEP_LOOP;
1414 }
1415
1416 // Left-over accumulations
1417 for(; i < K; ++i)
1418 {
1419 // Load values from LHS matrix
1420 VEC_DATA_TYPE(DATA_TYPE, 2)
1421 a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zin0));
1422#if M0 > 1
1423 VEC_DATA_TYPE(DATA_TYPE, 2)
1424 a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zin1));
1425#endif // M0 > 1
1426#if M0 > 2
1427 VEC_DATA_TYPE(DATA_TYPE, 2)
1428 a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zin2));
1429#endif // M0 > 2
1430#if M0 > 3
1431 VEC_DATA_TYPE(DATA_TYPE, 2)
1432 a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zin3));
1433#endif // M0 > 3
1434#if M0 > 4
1435 VEC_DATA_TYPE(DATA_TYPE, 2)
1436 a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zin4));
1437#endif // M0 > 4
1438#if M0 > 5
1439 VEC_DATA_TYPE(DATA_TYPE, 2)
1440 a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zin5));
1441#endif // M0 > 5
1442#if M0 > 6
1443 VEC_DATA_TYPE(DATA_TYPE, 2)
1444 a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zin6));
1445#endif // M0 > 6
1446#if M0 > 7
1447 VEC_DATA_TYPE(DATA_TYPE, 2)
1448 a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zin7));
1449#endif // M0 > 7
1450
1451 VEC_DATA_TYPE(DATA_TYPE, N0)
1452 b0;
1453 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs));
1454
1455 VFMA_M0xN0(0, a, b0, c);
1456
1457 lhs_offset += sizeof(DATA_TYPE);
1458 x_rhs += RHS_STEP_X;
1459 }
1460
1461 __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
1462
1463 REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
1464
1465#if defined(REINTERPRET_OUTPUT_AS_3D)
1466 // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
1467 CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
1468
1469 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
1470 // multiply dst_stride_z by DEPTH_GEMM3D
1471 dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
1472
1473#else // defined(REINTERPRET_OUTPUT_AS_3D)
1474
1475 // Add offset for batched GEMM
1476 dst_addr += z * dst_stride_z;
1477
1478#endif // defined(REINTERPRET_OUTPUT_AS_3D)
1479
1480 // Multiply by the weight of matrix-matrix product and store the result
1481#if defined(ALPHA)
1482 SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
1483#endif // defined(ALPHA)
1484
1485 // Add beta*bias
1486#if defined(BETA)
1487#if defined(BROADCAST_BIAS)
1488 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
1489
1490 LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
1491
1492#ifndef UNIT_BETA
1493 SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
1494#endif // UNIT_BIAS
1495
1496 // c = c + bias[broadcasted]
1497 ADD_BLOCK_BROADCAST(M0, c, bias0);
1498
1499#else // defined(BROADCAST_BIAS)
1500 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
1501
1502 LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
1503
1504#ifndef UNIT_BETA
1505 SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
1506#endif // UNIT_BIAS
1507
1508 // c = c + bias
1509 ADD_BLOCK(M0, c, bias);
1510
1511#endif // defined(BROADCAST_BIAS)
1512#endif // defined(BETA)
1513
1514#if defined(ACTIVATION_TYPE)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]1515 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, N0, c, A_VAL, B_VAL);
Ramy Elgammal451c3092022-02-01 23:01:27 +0000[diff] [blame]1516#endif // defined(ACTIVATION_TYPE)
1517
1518 // Store output block
1519 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
1520
1521#undef RHS_BLOCK_SIZE
1522#undef RHS_OFFSET_X
1523#undef RHS_STEP_X
1524#undef RHS_STEP_LOOP
1525}
1526#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_ONLY_RHS_NT_TEXTURE)
1527#endif // defined(M0) && defined(N0) && defined(K0) && defined(H0) && defined(DATA_TYPE)
1528
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]1529#if defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR)
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1530
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]1531#if defined(MIXED_PRECISION)
1532#if K0 == 2
1533#define ARM_DOT_K0(a, b, c) \
1534 ({ \
1535 c += a.s0 * b.s0; \
1536 c += a.s1 * b.s1; \
1537 })
1538#elif K0 == 3 // K0 == 3
1539#define ARM_DOT_K0(a, b, c) \
1540 ({ \
1541 c += a.s0 * b.s0; \
1542 c += a.s1 * b.s1; \
1543 c += a.s2 * b.s2; \
1544 })
1545#elif K0 == 4 // K0 == 4
1546#define ARM_DOT_K0(a, b, c) \
1547 ({ \
1548 c += a.s0 * b.s0; \
1549 c += a.s1 * b.s1; \
1550 c += a.s2 * b.s2; \
1551 c += a.s3 * b.s3; \
1552 })
1553#elif K0 == 8 // K0 == 8
1554#define ARM_DOT_K0(a, b, c) \
1555 ({ \
1556 c += a.s0 * b.s0; \
1557 c += a.s1 * b.s1; \
1558 c += a.s2 * b.s2; \
1559 c += a.s3 * b.s3; \
1560 c += a.s4 * b.s4; \
1561 c += a.s5 * b.s5; \
1562 c += a.s6 * b.s6; \
1563 c += a.s7 * b.s7; \
1564 })
1565#elif K0 == 16 // K0 == 16
1566#define ARM_DOT_K0(a, b, c) \
1567 ({ \
1568 c += a.s0 * b.s0; \
1569 c += a.s1 * b.s1; \
1570 c += a.s2 * b.s2; \
1571 c += a.s3 * b.s3; \
1572 c += a.s4 * b.s4; \
1573 c += a.s5 * b.s5; \
1574 c += a.s6 * b.s6; \
1575 c += a.s7 * b.s7; \
1576 c += a.s8 * b.s8; \
1577 c += a.s9 * b.s9; \
1578 c += a.sA * b.sA; \
1579 c += a.sB * b.sB; \
1580 c += a.sC * b.sC; \
1581 c += a.sD * b.sD; \
1582 c += a.sE * b.sE; \
1583 c += a.sF * b.sF; \
1584 })
1585#else // K0 not supported
1586#error "K0 value not supported"
1587#endif // K0 conditions
1588#else // defined(MIXED_PRECISION)
Gian Marco Iodicebacfec52019-01-11 11:30:55 +0000[diff] [blame]1589#if K0 == 2
1590#define ARM_DOT_K0(a, b, c) \
1591 ({ \
1592 c = fma(a.s0, b.s0, c); \
1593 c = fma(a.s1, b.s1, c); \
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1594 })
Gian Marco Iodicebacfec52019-01-11 11:30:55 +0000[diff] [blame]1595#elif K0 == 3 // K0 == 3
1596#define ARM_DOT_K0(a, b, c) \
1597 ({ \
1598 c = fma(a.s0, b.s0, c); \
1599 c = fma(a.s1, b.s1, c); \
1600 c = fma(a.s2, b.s2, c); \
1601 })
1602#elif K0 == 4 // K0 == 4
1603#define ARM_DOT_K0(a, b, c) \
1604 ({ \
1605 c = fma(a.s0, b.s0, c); \
1606 c = fma(a.s1, b.s1, c); \
1607 c = fma(a.s2, b.s2, c); \
1608 c = fma(a.s3, b.s3, c); \
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1609 })
1610#elif K0 == 8 // K0 == 8
Gian Marco Iodicebacfec52019-01-11 11:30:55 +0000[diff] [blame]1611#define ARM_DOT_K0(a, b, c) \
1612 ({ \
1613 c = fma(a.s0, b.s0, c); \
1614 c = fma(a.s1, b.s1, c); \
1615 c = fma(a.s2, b.s2, c); \
1616 c = fma(a.s3, b.s3, c); \
1617 c = fma(a.s4, b.s4, c); \
1618 c = fma(a.s5, b.s5, c); \
1619 c = fma(a.s6, b.s6, c); \
1620 c = fma(a.s7, b.s7, c); \
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1621 })
1622#elif K0 == 16 // K0 == 16
Gian Marco Iodicebacfec52019-01-11 11:30:55 +0000[diff] [blame]1623#define ARM_DOT_K0(a, b, c) \
1624 ({ \
1625 c = fma(a.s0, b.s0, c); \
1626 c = fma(a.s1, b.s1, c); \
1627 c = fma(a.s2, b.s2, c); \
1628 c = fma(a.s3, b.s3, c); \
1629 c = fma(a.s4, b.s4, c); \
1630 c = fma(a.s5, b.s5, c); \
1631 c = fma(a.s6, b.s6, c); \
1632 c = fma(a.s7, b.s7, c); \
1633 c = fma(a.s8, b.s8, c); \
1634 c = fma(a.s9, b.s9, c); \
1635 c = fma(a.sA, b.sA, c); \
1636 c = fma(a.sB, b.sB, c); \
1637 c = fma(a.sC, b.sC, c); \
1638 c = fma(a.sD, b.sD, c); \
1639 c = fma(a.sE, b.sE, c); \
1640 c = fma(a.sF, b.sF, c); \
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1641 })
1642#else // K0 not supported
1643#error "K0 value not supported"
1644#endif // K0 conditions
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]1645#endif // defined(MIXED_PRECISION)
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1646
Giorgio Arena7d2f69f2021-05-11 16:39:33 +0100[diff] [blame]1647#if defined(ARM_DOT_K0XN0)
1648#undef ARM_DOT_K0XN0
1649#endif // defined(ARM_DOT_K0XN0)
1650
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1651#if N0 == 2
1652#define ARM_DOT_K0XN0(a, b, c) \
1653 ({ \
1654 ARM_DOT_K0((a), (b##0), (c.s0)); \
1655 ARM_DOT_K0((a), (b##1), (c.s1)); \
1656 })
Gian Marco Iodicebacfec52019-01-11 11:30:55 +0000[diff] [blame]1657#elif N0 == 3 // N0 == 3
1658#define ARM_DOT_K0XN0(a, b, c) \
1659 ({ \
1660 ARM_DOT_K0((a), (b##0), (c.s0)); \
1661 ARM_DOT_K0((a), (b##1), (c.s1)); \
1662 ARM_DOT_K0((a), (b##2), (c.s2)); \
1663 })
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1664#elif N0 == 4 // N0 == 4
1665#define ARM_DOT_K0XN0(a, b, c) \
1666 ({ \
1667 ARM_DOT_K0((a), (b##0), (c.s0)); \
1668 ARM_DOT_K0((a), (b##1), (c.s1)); \
1669 ARM_DOT_K0((a), (b##2), (c.s2)); \
1670 ARM_DOT_K0((a), (b##3), (c.s3)); \
1671 })
1672#elif N0 == 8 // N0 == 8
1673#define ARM_DOT_K0XN0(a, b, c) \
1674 ({ \
1675 ARM_DOT_K0((a), (b##0), (c.s0)); \
1676 ARM_DOT_K0((a), (b##1), (c.s1)); \
1677 ARM_DOT_K0((a), (b##2), (c.s2)); \
1678 ARM_DOT_K0((a), (b##3), (c.s3)); \
1679 ARM_DOT_K0((a), (b##4), (c.s4)); \
1680 ARM_DOT_K0((a), (b##5), (c.s5)); \
1681 ARM_DOT_K0((a), (b##6), (c.s6)); \
1682 ARM_DOT_K0((a), (b##7), (c.s7)); \
1683 })
1684#elif N0 == 16 // N0 == 16
1685#define ARM_DOT_K0XN0(a, b, c) \
1686 ({ \
1687 ARM_DOT_K0((a), (b##0), (c.s0)); \
1688 ARM_DOT_K0((a), (b##1), (c.s1)); \
1689 ARM_DOT_K0((a), (b##2), (c.s2)); \
1690 ARM_DOT_K0((a), (b##3), (c.s3)); \
1691 ARM_DOT_K0((a), (b##4), (c.s4)); \
1692 ARM_DOT_K0((a), (b##5), (c.s5)); \
1693 ARM_DOT_K0((a), (b##6), (c.s6)); \
1694 ARM_DOT_K0((a), (b##7), (c.s7)); \
1695 ARM_DOT_K0((a), (b##8), (c.s8)); \
1696 ARM_DOT_K0((a), (b##9), (c.s9)); \
1697 ARM_DOT_K0((a), (b##A), (c.sA)); \
1698 ARM_DOT_K0((a), (b##B), (c.sB)); \
1699 ARM_DOT_K0((a), (b##C), (c.sC)); \
1700 ARM_DOT_K0((a), (b##D), (c.sD)); \
1701 ARM_DOT_K0((a), (b##E), (c.sE)); \
1702 ARM_DOT_K0((a), (b##F), (c.sF)); \
1703 })
1704#else // N0 not supported
1705#error "N0 value not supported"
1706#endif // N0 conditions
1707
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]1708#if defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T)
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1709/** This OpenCL kernel computes the matrix multiplication between 2 matrices.
1710 * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed
1711 * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed
1712 *
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]1713 * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
1714 * @note The data type used for the accumulators must be passed at compile time using -DDATA_TYPE_ACCUMULATOR (e.g. -DDATA_TYPE_ACCUMULATOR=float)
1715 * @note The F16 computation also supports mixed precision through the option -DMIXED_PRECISION passed at compile time. If enabled, DATA_TYPE_ACCUMULATOR should be set to float
Gian Marco Iodiceb0c50372019-03-15 10:13:05 +0000[diff] [blame]1716 * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]1717 * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24).
Gian Marco Iodiced1f54762019-07-19 09:54:47 +0100[diff] [blame]1718 * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4).
1719 * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2)
1720 * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1721 * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time.
1722 * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
Gian Marco Iodice088d63a2020-08-11 14:14:06 +0100[diff] [blame]1723 * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
1724 * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1725 * @note Only the following configurations of M0, N0 and K0 are currently supported:
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]1726 * - M0 = 2, 3, 4, 5, 6, 7, 8
Gian Marco Iodicebacfec52019-01-11 11:30:55 +0000[diff] [blame]1727 * - N0 = 2, 3, 4, 8, 16
1728 * - K0 = 2, 3, 4, 8, 16
Gian Marco Iodice62251f72019-03-11 16:07:12 +0000[diff] [blame]1729 * - V0 >= 1
1730 * - H0 >= 1
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1731 *
Gian Marco Iodiced1f54762019-07-19 09:54:47 +0100[diff] [blame]1732 * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
Gian Marco Iodiceca1f4602019-07-16 15:46:48 +0100[diff] [blame]1733 * The activation function is performed after the bias addition
Gian Marco Iodiced1f54762019-07-19 09:54:47 +0100[diff] [blame]1734 * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time:
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1735 * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
1736 * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
1737 * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
1738 * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped
1739 *
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1740 * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F16/F32
1741 * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes)
1742 * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
1743 * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes)
1744 * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
1745 * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix
1746 * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
1747 * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes)
1748 * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
1749 * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes)
1750 * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
1751 * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
1752 * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
1753 * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
1754 * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
1755 * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
1756 * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
1757 * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
1758 * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
1759 * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
1760 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
1761 * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
1762 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
1763 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1764 * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes)
1765 * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
1766 * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
1767 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
1768 * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]1769 * @param[in] M Number of rows in LHS matrix not reshaped.
1770 * @param[in] N Number of columns in RHS matrix not reshaped.
1771 * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped.
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1772 */
1773__kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs),
1774 IMAGE_DECLARATION(rhs),
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1775#if defined(BETA)
1776 IMAGE_DECLARATION(bias),
1777#endif // defined(BETA)
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1778 IMAGE_DECLARATION(dst),
1779 uint lhs_stride_z,
1780 uint rhs_stride_z,
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1781#if defined(BETA)
1782 uint bias_stride_z,
1783#endif //defined(BETA)
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1784 uint dst_stride_z
1785#if defined(REINTERPRET_OUTPUT_AS_3D)
1786 ,
1787 uint dst_cross_plane_pad
1788#endif // REINTERPRET_OUTPUT_AS_3D
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]1789 ,
1790 const int M,
1791 const int N,
1792 const int K)
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1793{
1794 // Block size
1795#define LHS_BLOCK_SIZE ((K0) * (M0))
1796
1797#if defined(LHS_INTERLEAVE)
1798#define LHS_OFFSET_X (K0)
1799#define LHS_STEP_X ((K0) * (V0))
1800#define LHS_STEP_LOOP (1)
1801#else // defined(INTERLEAVE)
1802#define LHS_OFFSET_X (LHS_BLOCK_SIZE)
1803#define LHS_STEP_X (K0)
1804#define LHS_STEP_LOOP (V0)
1805#endif // defined(INTERLEAVE)
1806
1807 // Block size
1808#define RHS_BLOCK_SIZE ((K0) * (N0))
1809
1810 // RHS offset and step X
1811#if defined(RHS_INTERLEAVE)
1812#define RHS_OFFSET_X (K0)
1813#define RHS_STEP_X ((K0) * (H0))
1814#define RHS_STEP_LOOP (1)
1815#else // defined(RHS_INTERLEAVE)
1816#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
1817#define RHS_STEP_X (K0)
1818#define RHS_STEP_LOOP (H0)
1819#endif // defined(RHS_INTERLEAVE)
1820
Gian Marco Iodiceb0c50372019-03-15 10:13:05 +0000[diff] [blame]1821#if defined(DUMMY_WORK_ITEMS)
1822 if((get_global_id(0) * N0 >= N) || (get_global_id(1) * M0 >= M))
1823 {
1824 return;
1825 }
1826#endif // defined(DUMMY_WORK_ITEMS)
1827
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1828 // Compute LHS matrix address
1829 __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (get_global_id(1) % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (get_global_id(1) / V0) * (uint)lhs_stride_y +
1830 (get_global_id(2) * lhs_stride_z);
1831
1832 // Compute RHS matrix address
1833 __global uchar *rhs_addr = rhs_ptr + rhs_offset_first_element_in_bytes + (get_global_id(0) % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (get_global_id(0) / (uint)H0) * rhs_stride_y;
1834
1835#if defined(MATRIX_B_DEPTH)
1836 // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
1837 rhs_addr += (get_global_id(2) % MATRIX_B_DEPTH) * rhs_stride_z;
1838#else // defined(MATRIX_B_DEPTH)
1839 rhs_addr += get_global_id(2) * rhs_stride_z;
1840#endif // defined(MATRIX_B_DEPTH)
1841
1842 // Initialize the accumulators
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]1843 REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0);
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1844
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1845 REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0;
1846 REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);
Usama Arif0681e3b2019-04-25 14:28:07 +0100[diff] [blame]1847
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]1848 for(int i = 0; i < K; i += K0)
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1849 {
1850 // Supported cases (M0, K0):
Gian Marco Iodiceadc53952019-02-15 11:10:31 +0000[diff] [blame]1851 // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
1852 // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
1853 // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
1854 // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
1855 // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
1856 // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
1857 // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
1858 // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1859 // Load values from LHS matrix
Usama Arif0681e3b2019-04-25 14:28:07 +0100[diff] [blame]1860 LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_addr, 0, LHS_STEP_X * sizeof(DATA_TYPE), zlhs);
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1861
1862 // Load values from RHS matrix
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1863 LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_addr, 0, RHS_STEP_X * sizeof(DATA_TYPE), zero);
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1864
1865 // Accumulate
1866 ARM_DOT_K0XN0(a0, b, c0);
1867#if M0 > 1
1868 ARM_DOT_K0XN0(a1, b, c1);
1869#endif // M0 > 1
1870#if M0 > 2
1871 ARM_DOT_K0XN0(a2, b, c2);
1872#endif // M0 > 2
1873#if M0 > 3
1874 ARM_DOT_K0XN0(a3, b, c3);
1875#endif // M0 > 3
1876#if M0 > 4
1877 ARM_DOT_K0XN0(a4, b, c4);
1878#endif // M0 > 4
1879#if M0 > 5
1880 ARM_DOT_K0XN0(a5, b, c5);
1881#endif // M0 > 5
1882#if M0 > 6
1883 ARM_DOT_K0XN0(a6, b, c6);
1884#endif // M0 > 6
1885#if M0 > 7
1886 ARM_DOT_K0XN0(a7, b, c7);
1887#endif // M0 > 7
1888
1889 lhs_addr += (M0 * LHS_STEP_X * LHS_STEP_LOOP) * sizeof(DATA_TYPE);
1890 rhs_addr += (N0 * RHS_STEP_X * RHS_STEP_LOOP) * sizeof(DATA_TYPE);
1891 }
1892
1893 __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * dst_stride_y);
1894
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1895 REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1896
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]1897 const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
1898 const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);
1899
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1900#if defined(REINTERPRET_OUTPUT_AS_3D)
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1901
1902 // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
Michele Di Giorgio5fa963f2020-11-23 15:05:12 +0000[diff] [blame]1903 CALCULATE_Z_OFFSET(M0, uint, zout, get_global_id(1) * (uint)M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1904 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
1905 // multiply dst_stride_z by DEPTH_GEMM3D
1906 dst_addr += get_global_id(2) * dst_stride_z * DEPTH_GEMM3D;
1907
1908#else // defined(REINTERPRET_OUTPUT_AS_3D)
1909
1910 // Add offset for batched GEMM
1911 dst_addr += get_global_id(2) * dst_stride_z;
1912
1913#endif // defined(REINTERPRET_OUTPUT_AS_3D)
1914
1915 // Multiply by the weight of matrix-matrix product and store the result
1916#if defined(ALPHA)
Usama Arif0681e3b2019-04-25 14:28:07 +0100[diff] [blame]1917 SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1918#endif // defined(ALPHA)
1919
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1920 // Add beta*bias
1921#if defined(BETA)
1922#if defined(BROADCAST_BIAS)
1923 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
1924
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]1925 LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1926
1927#ifndef UNIT_BETA
1928 SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
1929#endif // UNIT_BIAS
1930
1931 // c = c + bias[broadcasted]
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]1932#if defined(MIXED_PRECISION)
1933 CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
1934 ADD_BLOCK_BROADCAST(M0, c, bias_hp0);
1935#else // defined(MIXED_PRECISION)
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1936 ADD_BLOCK_BROADCAST(M0, c, bias0);
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]1937#endif // defined(MIXED_PRECISION)
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1938
1939#else // defined(BROADCAST_BIAS)
1940 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id(
1941 2) * bias_stride_z;
1942
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]1943 LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1944
1945#ifndef UNIT_BETA
1946 SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
1947#endif // UNIT_BIAS
1948
1949 // c = c + bias
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]1950#if defined(MIXED_PRECISION)
1951 CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
1952 ADD_BLOCK(M0, c, bias_hp);
1953#else // defined(MIXED_PRECISION)
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1954 ADD_BLOCK(M0, c, bias);
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]1955#endif // defined(MIXED_PRECISION)
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1956
1957#endif // defined(BROADCAST_BIAS)
1958#endif // defined(BETA)
1959
Gian Marco Iodiceca1f4602019-07-16 15:46:48 +0100[diff] [blame]1960#if defined(ACTIVATION_TYPE)
Georgios Pinitasa07ce152019-10-11 17:38:50 +0100[diff] [blame]1961#if defined(MIXED_PRECISION)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]1962 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, N0, c, A_VAL, B_VAL);
Georgios Pinitasa07ce152019-10-11 17:38:50 +0100[diff] [blame]1963#else // defined(MIXED_PRECISION)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]1964 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, N0, c, A_VAL, B_VAL);
Georgios Pinitasa07ce152019-10-11 17:38:50 +0100[diff] [blame]1965#endif // defined(MIXED_PRECISION)
Gian Marco Iodiceca1f4602019-07-16 15:46:48 +0100[diff] [blame]1966#endif // defined(ACTIVATION_TYPE)
1967
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1968 // Store output block
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]1969#if defined(MIXED_PRECISION)
Gian Marco Iodice088d63a2020-08-11 14:14:06 +0100[diff] [blame]1970 CONVERT_BLOCK(M0, N0, DATA_TYPE, c, c_lp);
Giorgio Arena1e2af2a2020-10-15 17:39:41 +0100[diff] [blame]1971 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c_lp, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]1972#else // defined(MIXED_PRECISION)
Giorgio Arena1e2af2a2020-10-15 17:39:41 +0100[diff] [blame]1973 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]1974#endif // defined(MIXED_PRECISION)
Gian Marco Iodicee16c8902019-06-14 16:11:10 +0100[diff] [blame]1975
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1976#undef LHS_BLOCK_SIZE
1977#undef LHS_OFFSET_X
1978#undef LHS_STEP_X
1979#undef RHS_BLOCK_SIZE
1980#undef RHS_OFFSET_X
1981#undef RHS_STEP_X
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]1982#undef LHS_STEP_LOOP
1983#undef RHS_STEP_LOOP
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]1984}
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]1985#endif // defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T)
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]1986
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]1987#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T_TEXTURE)
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]1988/** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image object.
1989 * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be NOT transposed
1990 * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be transposed
1991 *
1992 * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel
1993 * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
1994 * @note The data type used for the accumulators must be passed at compile time using -DDATA_TYPE_ACCUMULATOR (e.g. -DDATA_TYPE_ACCUMULATOR=float)
1995 * @note The F16 computation also supports mixed precision through the option -DMIXED_PRECISION passed at compile time. If enabled, DATA_TYPE_ACCUMULATOR should be set to float
1996 * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
1997 * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24).
Gian Marco Iodice781cba72020-06-19 16:56:57 +0100[diff] [blame]1998 * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32)
1999 * Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT
2000 * could be different from the value returned by get_image_height(rhs_img).
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2001 * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4).
2002 * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2)
2003 * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
2004 * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time.
2005 * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
Gian Marco Iodice088d63a2020-08-11 14:14:06 +0100[diff] [blame]2006 * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
2007 * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2008 * @note Only the following configurations of M0, N0 and K0 are currently supported:
2009 * - M0 = 2, 3, 4, 5, 6, 7, 8
2010 * - N0 = 4, 8, 16
2011 * - K0 = 4, 8, 16
2012 * - V0 >= 1
2013 * - H0 >= 1
2014 *
2015 * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
2016 * The activation function is performed after the bias addition
2017 * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time:
2018 * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
2019 * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
2020 * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
2021 * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped
2022 *
2023 * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F32
2024 * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes)
2025 * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
2026 * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes)
2027 * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
2028 * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix
2029 * @param[in] rhs_img The RHS reshaped matrix as OpenCL image object. Supported data type: same as @p lhs_ptr
2030 * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
2031 * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
2032 * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
2033 * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
2034 * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
2035 * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
2036 * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
2037 * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
2038 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
2039 * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
2040 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
2041 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
2042 * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes)
2043 * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
2044 * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
2045 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
2046 * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2047 * @param[in] M Number of rows in LHS matrix not reshaped.
2048 * @param[in] N Number of columns in RHS matrix not reshaped.
2049 * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped.
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2050 */
2051__kernel void gemm_mm_reshaped_lhs_nt_rhs_t_texture(IMAGE_DECLARATION(lhs),
2052 __read_only image2d_t rhs_img,
2053#if defined(BETA)
2054 IMAGE_DECLARATION(bias),
2055#endif // defined(BETA)
2056 IMAGE_DECLARATION(dst),
2057 uint lhs_stride_z,
2058 uint rhs_stride_z,
2059#if defined(BETA)
2060 uint bias_stride_z,
2061#endif //defined(BETA)
2062 uint dst_stride_z
2063#if defined(REINTERPRET_OUTPUT_AS_3D)
2064 ,
2065 uint dst_cross_plane_pad
2066#endif // REINTERPRET_OUTPUT_AS_3D
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2067 ,
2068 const int M,
2069 const int N,
2070 const int K)
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2071{
2072 // Pixel unit
2073#define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(K0)
2074
2075 // Block size
2076#define LHS_BLOCK_SIZE ((K0) * (M0))
2077
2078#if defined(LHS_INTERLEAVE)
2079#define LHS_OFFSET_X (K0)
2080#define LHS_STEP_X ((K0) * (V0))
2081#define LHS_STEP_LOOP (1)
2082#else // defined(INTERLEAVE)
2083#define LHS_OFFSET_X (LHS_BLOCK_SIZE)
2084#define LHS_STEP_X (K0)
2085#define LHS_STEP_LOOP (V0)
2086#endif // defined(INTERLEAVE)
2087
2088 // Block size
2089#define RHS_BLOCK_SIZE (PIXEL_UNIT * (N0))
2090
2091 // RHS offset and step X
2092#if defined(RHS_INTERLEAVE)
2093#define RHS_OFFSET_X (PIXEL_UNIT)
2094#define RHS_STEP_X (PIXEL_UNIT * (H0))
2095#define RHS_STEP_LOOP (1)
2096#else // defined(RHS_INTERLEAVE)
2097#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
2098#define RHS_STEP_X PIXEL_UNIT
2099#define RHS_STEP_LOOP (H0)
2100#endif // defined(RHS_INTERLEAVE)
2101
2102#if defined(DUMMY_WORK_ITEMS)
2103 if((get_global_id(0) * N0 >= N) || (get_global_id(1) * M0 >= M))
2104 {
2105 return;
2106 }
2107#endif // defined(DUMMY_WORK_ITEMS)
2108
2109 // Compute LHS matrix address
2110 __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (get_global_id(1) % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (get_global_id(1) / V0) * (uint)lhs_stride_y +
2111 (get_global_id(2) * lhs_stride_z);
2112
2113#if defined(MATRIX_B_DEPTH)
2114 // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
2115 const uint z_rhs = (get_global_id(2) % MATRIX_B_DEPTH);
2116#else // defined(MATRIX_B_DEPTH)
2117 const uint z_rhs = get_global_id(2);
2118#endif // defined(MATRIX_B_DEPTH)
2119
2120 // Compute RHS matrix coordinates
2121 uint x_rhs = (get_global_id(0) % H0) * (uint)RHS_OFFSET_X;
2122 const uint y_rhs = (get_global_id(0) / (uint)H0) + z_rhs * RHS_HEIGHT;
2123
2124 // Initialize the accumulators
2125 REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0);
2126
2127 REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0;
2128 REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);
2129
2130 for(int i = 0; i < K; i += K0)
2131 {
2132 // Load values from LHS matrix
2133 LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_addr, 0, LHS_STEP_X * sizeof(DATA_TYPE), zlhs);
2134
2135 // Load values from RHS matrix stored in a cl_image
2136 REPEAT_VAR_INIT_TO_CONST(N0, VEC_DATA_TYPE(DATA_TYPE, K0), b, 0);
2137 LOAD_TEXTURE2D(N0, PIXEL_UNIT, DATA_TYPE, b, rhs_img, x_rhs, y_rhs, RHS_STEP_X, 0);
2138
2139 // Accumulate
2140 ARM_DOT_K0XN0(a0, b, c0);
2141#if M0 > 1
2142 ARM_DOT_K0XN0(a1, b, c1);
2143#endif // M0 > 1
2144#if M0 > 2
2145 ARM_DOT_K0XN0(a2, b, c2);
2146#endif // M0 > 2
2147#if M0 > 3
2148 ARM_DOT_K0XN0(a3, b, c3);
2149#endif // M0 > 3
2150#if M0 > 4
2151 ARM_DOT_K0XN0(a4, b, c4);
2152#endif // M0 > 4
2153#if M0 > 5
2154 ARM_DOT_K0XN0(a5, b, c5);
2155#endif // M0 > 5
2156#if M0 > 6
2157 ARM_DOT_K0XN0(a6, b, c6);
2158#endif // M0 > 6
2159#if M0 > 7
2160 ARM_DOT_K0XN0(a7, b, c7);
2161#endif // M0 > 7
2162
2163 lhs_addr += (M0 * LHS_STEP_X * LHS_STEP_LOOP) * sizeof(DATA_TYPE);
2164
2165 x_rhs += N0 * RHS_STEP_X * RHS_STEP_LOOP;
2166 }
2167
2168 __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * dst_stride_y);
2169
2170 REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
2171
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]2172 const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
2173 const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);
2174
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2175#if defined(REINTERPRET_OUTPUT_AS_3D)
2176
2177 // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
Michele Di Giorgio5fa963f2020-11-23 15:05:12 +0000[diff] [blame]2178 CALCULATE_Z_OFFSET(M0, uint, zout, get_global_id(1) * (uint)M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2179 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
2180 // multiply dst_stride_z by DEPTH_GEMM3D
2181 dst_addr += get_global_id(2) * dst_stride_z * DEPTH_GEMM3D;
2182
2183#else // defined(REINTERPRET_OUTPUT_AS_3D)
2184
2185 // Add offset for batched GEMM
2186 dst_addr += get_global_id(2) * dst_stride_z;
2187
2188#endif // defined(REINTERPRET_OUTPUT_AS_3D)
2189
2190 // Multiply by the weight of matrix-matrix product and store the result
2191#if defined(ALPHA)
2192 SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
2193#endif // defined(ALPHA)
2194
2195 // Add beta*bias
2196#if defined(BETA)
2197#if defined(BROADCAST_BIAS)
2198 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
2199
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]2200 LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2201
2202#ifndef UNIT_BETA
2203 SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
2204#endif // UNIT_BIAS
2205
2206 // c = c + bias[broadcasted]
2207#if defined(MIXED_PRECISION)
2208 CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
2209 ADD_BLOCK_BROADCAST(M0, c, bias_hp0);
2210#else // defined(MIXED_PRECISION)
2211 ADD_BLOCK_BROADCAST(M0, c, bias0);
2212#endif // defined(MIXED_PRECISION)
2213
2214#else // defined(BROADCAST_BIAS)
2215 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id(
2216 2) * bias_stride_z;
2217
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]2218 LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2219
2220#ifndef UNIT_BETA
2221 SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
2222#endif // UNIT_BIAS
2223
2224 // c = c + bias
2225#if defined(MIXED_PRECISION)
2226 CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
2227 ADD_BLOCK(M0, c, bias_hp);
2228#else // defined(MIXED_PRECISION)
2229 ADD_BLOCK(M0, c, bias);
2230#endif // defined(MIXED_PRECISION)
2231
2232#endif // defined(BROADCAST_BIAS)
2233#endif // defined(BETA)
2234
2235#if defined(ACTIVATION_TYPE)
2236#if defined(MIXED_PRECISION)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]2237 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, N0, c, A_VAL, B_VAL);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2238#else // defined(MIXED_PRECISION)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]2239 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, N0, c, A_VAL, B_VAL);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2240#endif // defined(MIXED_PRECISION)
2241#endif // defined(ACTIVATION_TYPE)
2242
2243 // Store output block
2244#if defined(MIXED_PRECISION)
Gian Marco Iodice088d63a2020-08-11 14:14:06 +0100[diff] [blame]2245 CONVERT_BLOCK(M0, N0, DATA_TYPE, c, c_lp);
Giorgio Arena1e2af2a2020-10-15 17:39:41 +0100[diff] [blame]2246 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c_lp, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2247#else // defined(MIXED_PRECISION)
Giorgio Arena1e2af2a2020-10-15 17:39:41 +0100[diff] [blame]2248 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2249#endif // defined(MIXED_PRECISION)
2250
2251#undef LHS_BLOCK_SIZE
2252#undef LHS_OFFSET_X
2253#undef LHS_STEP_X
2254#undef RHS_BLOCK_SIZE
2255#undef RHS_OFFSET_X
2256#undef RHS_STEP_X
2257#undef PIXEL_UNIT
2258#undef LHS_STEP_LOOP
2259#undef RHS_STEP_LOOP
2260}
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2261#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_NT_RHS_T_TEXTURE)
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2262
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2263#if defined(LHS_TRANSPOSE)
2264
2265#define VTYPE(TYPE, SIZE) VEC_DATA_TYPE(TYPE, SIZE)
2266
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2267#if defined(MIXED_PRECISION)
2268
2269#if(GPU_ARCH == GPU_ARCH_MIDGARD)
2270#define ARM_VFMA(N0, a, b, c) c += (CONVERT(a, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0))) * (CONVERT(b, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0)));
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2271#else // GPU_ARCH == GPU_ARCH_MIDGARD
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2272#define ARM_VFMA(N0, a, b, c) c = fma((CONVERT(a, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0))), (CONVERT(b, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0))), (c));
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2273#endif // GPU_ARCH == GPU_ARCH_MIDGARD
2274
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2275#else // defined(MIXED_PRECISION
2276
2277#if(GPU_ARCH == GPU_ARCH_MIDGARD)
2278#define ARM_VFMA(N0, a, b, c) c += (a) * (b);
2279#else // GPU_ARCH == GPU_ARCH_MIDGARD
2280#define ARM_VFMA(N0, a, b, c) c = fma((a), (b), (c));
2281#endif // GPU_ARCH == GPU_ARCH_MIDGARD
2282
2283#endif // defined(MIXED_PRECISION)
2284
2285#define ARM_VVM_T_NT_1xN0x1(N0, TYPE, a, b, C) \
2286 ({ \
2287 ARM_VFMA(N0, (VTYPE(TYPE, N0))(a), b, (C##0)); \
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2288 })
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2289#define ARM_VVM_T_NT_2xN0x1(N0, TYPE, a, b, C) \
2290 ({ \
2291 ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s0), b, (C##0)); \
2292 ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s1), b, (C##1)); \
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2293 })
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2294#define ARM_VVM_T_NT_3xN0x1(N0, TYPE, a, b, C) \
2295 ({ \
2296 ARM_VVM_T_NT_2xN0x1(N0, TYPE, a, b, C); \
2297 ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s2), b, (C##2)); \
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2298 })
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2299#define ARM_VVM_T_NT_4xN0x1(N0, TYPE, a, b, C) \
2300 ({ \
2301 ARM_VVM_T_NT_3xN0x1(N0, TYPE, a, b, C); \
2302 ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s3), b, (C##3)); \
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2303 })
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2304#define ARM_VVM_T_NT_8xN0x1(N0, TYPE, a, b, C) \
2305 ({ \
2306 ARM_VVM_T_NT_4xN0x1(N0, TYPE, a, b, C); \
2307 ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s4), b, (C##4)); \
2308 ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s5), b, (C##5)); \
2309 ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s6), b, (C##6)); \
2310 ARM_VFMA(N0, (VTYPE(TYPE, N0))(a.s7), b, (C##7)); \
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2311 })
2312
2313// Factory macro for the column-vector (transposed) by row-vector (not transposed) multiplication. K0 = 1
2314// a is the column-vector (transposed)
2315// b is the row-vector (not transposed)
2316// C is the output matrix
2317// Lower case is a vector (a, b)
2318// Upper case is a matrix (C)
2319#define ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, a, b, C) ARM_VVM_T_NT_##M0##xN0x1(N0, TYPE, a, b, C)
2320
2321#define ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, A, B, C) \
2322 ({ \
2323 ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##0), (B##0), C); \
2324 })
2325#define ARM_MM_T_NT_M0xN0x2(M0, N0, TYPE, A, B, C) \
2326 ({ \
2327 ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, A, B, C); \
2328 ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##1), (B##1), C); \
2329 })
2330#define ARM_MM_T_NT_M0xN0x3(M0, N0, TYPE, A, B, C) \
2331 ({ \
2332 ARM_MM_T_NT_M0xN0x2(M0, N0, TYPE, A, B, C); \
2333 ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##2), (B##2), C); \
2334 })
2335#define ARM_MM_T_NT_M0xN0x4(M0, N0, TYPE, A, B, C) \
2336 ({ \
2337 ARM_MM_T_NT_M0xN0x3(M0, N0, TYPE, A, B, C); \
2338 ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##3), (B##3), C); \
2339 })
2340#define ARM_MM_T_NT_M0xN0x8(M0, N0, TYPE, A, B, C) \
2341 ({ \
2342 ARM_MM_T_NT_M0xN0x4(M0, N0, TYPE, A, B, C); \
2343 ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##4), (B##4), C); \
2344 ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##5), (B##5), C); \
2345 ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##6), (B##6), C); \
2346 ARM_VVM_T_NT_M0xN0x1(M0, N0, TYPE, (A##7), (B##7), C); \
2347 })
2348#define ARM_MM_T_NT_M0xN0x16(M0, N0, TYPE, A, B, C) \
2349 ({ \
2350 ARM_MM_T_NT_M0xN0x8(M0, N0, TYPE, A, B, C); \
2351 ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##8), (B##8), C); \
2352 ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##9), (B##9), C); \
2353 ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##A), (B##A), C); \
2354 ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##B), (B##B), C); \
2355 ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##C), (B##C), C); \
2356 ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##D), (B##D), C); \
2357 ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##E), (B##E), C); \
2358 ARM_MM_T_NT_M0xN0x1(M0, N0, TYPE, (A##F), (B##F), C); \
2359 })
2360
2361// Factory macro for the matrix (transposed) by matrix (not transposed) multiplication.
2362// The dimensions for this matrix multiplications are defined through M0, N0 and K0
2363// The dimensions supported are:
2364// M0: 1, 2, 3, 4, 8
2365// N0: 1, 2, 3, 4, 8, 16
2366// K0: 1, 2, 3, 4, 8, 16
2367// This macro calls the vector-by-matrix macro K0 times
2368// A, B and C are matrices
Gian Marco Iodice05639f62019-09-24 12:05:06 +0100[diff] [blame]2369#define ARM_MM_T_NT(M0, N0, K0, TYPE, A, B, C) \
2370 CONCAT(ARM_MM_T_NT_M0xN0x, K0) \
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2371 (M0, N0, TYPE, A, B, C)
2372
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2373#if defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2374/** This OpenCL kernel computes the matrix multiplication between 2 matrices.
2375 * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be transposed
2376 * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be NOT transposed
2377 *
2378 * @note LHS_TRANSPOSE should be passed at compile time in order to compile this OpenCL kernel (e.g. -DLHS_TRANSPOSE).
2379 * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2380 * @note The GEMM's dimensions M, N and K must be passed at compile time using -DM, -DN and -DK (e.g. -DM=52, -DN=90 and -DK=24).
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2381 * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4).
2382 * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2)
2383 * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
2384 * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time.
2385 * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
Gian Marco Iodice088d63a2020-08-11 14:14:06 +0100[diff] [blame]2386 * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
2387 * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2388 * @note Only the following configurations of M0, N0 and K0 are currently supported:
2389 * - M0 = 2, 3, 4, 8
2390 * - N0 = 2, 3, 4, 8, 16
2391 * - K0 = 2, 3, 4, 8, 16
2392 * - V0 >= 1
2393 * - H0 >= 1
2394 *
2395 * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
2396 * The activation function is performed after the bias addition
2397 * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time:
2398 * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
2399 * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
2400 * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
2401 * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped
2402 *
2403 * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F16/F32
2404 * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes)
2405 * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
2406 * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes)
2407 * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
2408 * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix
2409 * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
2410 * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes)
2411 * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
2412 * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes)
2413 * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
2414 * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
2415 * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
2416 * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
2417 * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
2418 * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
2419 * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
2420 * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
2421 * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
2422 * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
2423 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
2424 * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
2425 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
2426 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2427 * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes)
2428 * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
2429 * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
2430 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
2431 * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2432 * @param[in] M Number of rows in LHS matrix not reshaped.
2433 * @param[in] N Number of columns in RHS matrix not reshaped.
2434 * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped.
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2435 */
2436__kernel void gemm_mm_reshaped_lhs_t_rhs_nt(IMAGE_DECLARATION(lhs),
2437 IMAGE_DECLARATION(rhs),
2438#if defined(BETA)
2439 IMAGE_DECLARATION(bias),
2440#endif // defined(BETA)
2441 IMAGE_DECLARATION(dst),
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2442 uint lhs_stride_z,
2443 uint rhs_stride_z,
2444#if defined(BETA)
2445 uint bias_stride_z,
2446#endif //defined(BETA)
2447 uint dst_stride_z
2448#if defined(REINTERPRET_OUTPUT_AS_3D)
2449 ,
2450 uint dst_cross_plane_pad
2451#endif // REINTERPRET_OUTPUT_AS_3D
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2452 ,
2453 const int M,
2454 const int N,
2455 const int K)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2456{
2457 // Block size
2458#define LHS_BLOCK_SIZE ((K0) * (M0))
2459
2460#if defined(LHS_INTERLEAVE)
2461#define LHS_OFFSET_X (M0)
2462#define LHS_STEP_X ((M0) * (V0))
2463#define LHS_STEP_LOOP (1)
2464#else // defined(INTERLEAVE)
2465#define LHS_OFFSET_X (LHS_BLOCK_SIZE)
2466#define LHS_STEP_X (M0)
2467#define LHS_STEP_LOOP (V0)
2468#endif // defined(INTERLEAVE)
2469
2470 // Block size
2471#define RHS_BLOCK_SIZE ((K0) * (N0))
2472
2473 // RHS offset and step X
2474#if defined(RHS_INTERLEAVE)
2475#define RHS_OFFSET_X (N0)
2476#define RHS_STEP_X ((N0) * (H0))
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2477#else // defined(RHS_INTERLEAVE)
2478#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
2479#define RHS_STEP_X (N0)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2480#endif // defined(RHS_INTERLEAVE)
2481
2482 const uint x = get_global_id(0);
2483 const uint y = get_global_id(1);
2484 const uint z = get_global_id(2);
2485
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]2486 const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
2487 const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);
2488
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2489#if defined(DUMMY_WORK_ITEMS)
2490 if((x * N0 >= N) || (y * M0 >= M))
2491 {
2492 return;
2493 }
2494#endif // defined(DUMMY_WORK_ITEMS)
2495
2496 // Compute LHS matrix address
2497 __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (y % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (y / V0) * (uint)lhs_stride_y + (z * lhs_stride_z);
2498
2499 // Compute RHS matrix address
2500 __global uchar *rhs_addr = rhs_ptr + rhs_offset_first_element_in_bytes + (x % H0) * (uint)RHS_OFFSET_X * sizeof(DATA_TYPE) + (x / (uint)H0) * rhs_stride_y;
2501
2502#if defined(MATRIX_B_DEPTH)
2503 // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
2504 rhs_addr += (z % MATRIX_B_DEPTH) * rhs_stride_z;
2505#else // defined(MATRIX_B_DEPTH)
2506 rhs_addr += z * rhs_stride_z;
2507#endif // defined(MATRIX_B_DEPTH)
2508
2509 // Initialize the accumulators
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2510 REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0);
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2511
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2512 REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0);
2513
Gian Marco Iodice05639f62019-09-24 12:05:06 +0100[diff] [blame]2514 __global DATA_TYPE *lhs = (__global DATA_TYPE *)(lhs_addr);
2515 __global DATA_TYPE *rhs = (__global DATA_TYPE *)(rhs_addr);
2516
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2517 for(int i = 0; i < K; i += K0)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2518 {
Gian Marco Iodice05639f62019-09-24 12:05:06 +0100[diff] [blame]2519 VEC_DATA_TYPE(DATA_TYPE, M0)
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2520 a0;
Gian Marco Iodice05639f62019-09-24 12:05:06 +0100[diff] [blame]2521 VEC_DATA_TYPE(DATA_TYPE, N0)
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2522 b0;
2523
2524 a0 = VLOAD(M0)(0, lhs);
Gian Marco Iodice05639f62019-09-24 12:05:06 +0100[diff] [blame]2525 b0 = VLOAD(N0)(0, rhs);
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2526
Gian Marco Iodice05639f62019-09-24 12:05:06 +0100[diff] [blame]2527 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2528
Gian Marco Iodice05639f62019-09-24 12:05:06 +0100[diff] [blame]2529 lhs += LHS_STEP_X;
2530 rhs += RHS_STEP_X;
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2531
Gian Marco Iodice05639f62019-09-24 12:05:06 +0100[diff] [blame]2532#if K0 > 1
2533 a0 = VLOAD(M0)(0, lhs);
2534 b0 = VLOAD(N0)(0, rhs);
2535
2536 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2537
2538 lhs += LHS_STEP_X;
2539 rhs += RHS_STEP_X;
2540#endif // K0 > 1
2541
2542#if K0 > 2
2543 a0 = VLOAD(M0)(0, lhs);
2544 b0 = VLOAD(N0)(0, rhs);
2545
2546 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2547
2548 lhs += LHS_STEP_X;
2549 rhs += RHS_STEP_X;
2550#endif // K0 > 2
2551
2552#if K0 > 3
2553 a0 = VLOAD(M0)(0, lhs);
2554 b0 = VLOAD(N0)(0, rhs);
2555
2556 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2557
2558 lhs += LHS_STEP_X;
2559 rhs += RHS_STEP_X;
2560#endif // K0 > 3
2561
2562#if K0 > 4
2563 a0 = VLOAD(M0)(0, lhs);
2564 b0 = VLOAD(N0)(0, rhs);
2565
2566 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2567
2568 lhs += LHS_STEP_X;
2569 rhs += RHS_STEP_X;
2570
2571 a0 = VLOAD(M0)(0, lhs);
2572 b0 = VLOAD(N0)(0, rhs);
2573
2574 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2575
2576 lhs += LHS_STEP_X;
2577 rhs += RHS_STEP_X;
2578
2579 a0 = VLOAD(M0)(0, lhs);
2580 b0 = VLOAD(N0)(0, rhs);
2581
2582 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2583
2584 lhs += LHS_STEP_X;
2585 rhs += RHS_STEP_X;
2586
2587 a0 = VLOAD(M0)(0, lhs);
2588 b0 = VLOAD(N0)(0, rhs);
2589
2590 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2591
2592 lhs += LHS_STEP_X;
2593 rhs += RHS_STEP_X;
2594#endif // K0 > 4
2595
2596#if K0 > 8
2597 a0 = VLOAD(M0)(0, lhs);
2598 b0 = VLOAD(N0)(0, rhs);
2599
2600 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2601
2602 lhs += LHS_STEP_X;
2603 rhs += RHS_STEP_X;
2604
2605 a0 = VLOAD(M0)(0, lhs);
2606 b0 = VLOAD(N0)(0, rhs);
2607
2608 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2609
2610 lhs += LHS_STEP_X;
2611 rhs += RHS_STEP_X;
2612
2613 a0 = VLOAD(M0)(0, lhs);
2614 b0 = VLOAD(N0)(0, rhs);
2615
2616 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2617
2618 lhs += LHS_STEP_X;
2619 rhs += RHS_STEP_X;
2620
2621 a0 = VLOAD(M0)(0, lhs);
2622 b0 = VLOAD(N0)(0, rhs);
2623
2624 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2625
2626 lhs += LHS_STEP_X;
2627 rhs += RHS_STEP_X;
2628
2629 a0 = VLOAD(M0)(0, lhs);
2630 b0 = VLOAD(N0)(0, rhs);
2631
2632 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2633
2634 lhs += LHS_STEP_X;
2635 rhs += RHS_STEP_X;
2636
2637 a0 = VLOAD(M0)(0, lhs);
2638 b0 = VLOAD(N0)(0, rhs);
2639
2640 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2641
2642 lhs += LHS_STEP_X;
2643 rhs += RHS_STEP_X;
2644
2645 a0 = VLOAD(M0)(0, lhs);
2646 b0 = VLOAD(N0)(0, rhs);
2647
2648 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2649
2650 lhs += LHS_STEP_X;
2651 rhs += RHS_STEP_X;
2652
2653 a0 = VLOAD(M0)(0, lhs);
2654 b0 = VLOAD(N0)(0, rhs);
2655
2656 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2657
2658 lhs += LHS_STEP_X;
2659 rhs += RHS_STEP_X;
2660#endif // K0 > 8
2661
2662#ifndef LHS_INTERLEAVE
2663 lhs += (M0 * K0 * (V0 - 1));
2664#endif // LHS_INTERLEAVE
2665
2666#ifndef RHS_INTERLEAVE
2667 rhs += (N0 * K0 * (H0 - 1));
2668#endif // RHS_INTERLEAVE
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2669 }
2670
2671 __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * dst_stride_y);
2672
2673 REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
2674
2675#if defined(REINTERPRET_OUTPUT_AS_3D)
2676
2677 // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
Michele Di Giorgio5fa963f2020-11-23 15:05:12 +0000[diff] [blame]2678 CALCULATE_Z_OFFSET(M0, uint, zout, y * (uint)M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2679 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
2680 // multiply dst_stride_z by DEPTH_GEMM3D
2681 dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
2682
2683#else // defined(REINTERPRET_OUTPUT_AS_3D)
2684
2685 // Add offset for batched GEMM
2686 dst_addr += z * dst_stride_z;
2687
2688#endif // defined(REINTERPRET_OUTPUT_AS_3D)
2689
2690 // Multiply by the weight of matrix-matrix product and store the result
2691#if defined(ALPHA)
2692 SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
2693#endif // defined(ALPHA)
2694
2695 // Add beta*bias
2696#if defined(BETA)
2697#if defined(BROADCAST_BIAS)
2698 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE));
2699
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]2700 LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2701
2702#ifndef UNIT_BETA
2703 SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
2704#endif // UNIT_BIAS
2705
2706 // c = c + bias[broadcasted]
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2707#if defined(MIXED_PRECISION)
2708 CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
2709 ADD_BLOCK_BROADCAST(M0, c, bias_hp0);
2710#else // defined(MIXED_PRECISION)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2711 ADD_BLOCK_BROADCAST(M0, c, bias0);
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2712#endif // defined(MIXED_PRECISION)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2713
2714#else // defined(BROADCAST_BIAS)
Gian Marco Iodice088d63a2020-08-11 14:14:06 +0100[diff] [blame]2715 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id(
2716 2) * bias_stride_z;
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2717
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]2718 LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2719
2720#ifndef UNIT_BETA
2721 SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
2722#endif // UNIT_BIAS
2723
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2724#if defined(MIXED_PRECISION)
2725 CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
2726 ADD_BLOCK(M0, c, bias_hp);
2727#else // defined(MIXED_PRECISION)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2728 ADD_BLOCK(M0, c, bias);
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2729#endif // defined(MIXED_PRECISION)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2730
2731#endif // defined(BROADCAST_BIAS)
2732#endif // defined(BETA)
2733
2734#if defined(ACTIVATION_TYPE)
Georgios Pinitasa07ce152019-10-11 17:38:50 +0100[diff] [blame]2735#if defined(MIXED_PRECISION)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]2736 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, N0, c, A_VAL, B_VAL);
Georgios Pinitasa07ce152019-10-11 17:38:50 +0100[diff] [blame]2737#else // defined(MIXED_PRECISION)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]2738 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, N0, c, A_VAL, B_VAL);
Georgios Pinitasa07ce152019-10-11 17:38:50 +0100[diff] [blame]2739#endif // defined(MIXED_PRECISION)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2740#endif // defined(ACTIVATION_TYPE)
2741
2742 // Store output block
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2743#if defined(MIXED_PRECISION)
Gian Marco Iodice088d63a2020-08-11 14:14:06 +0100[diff] [blame]2744 CONVERT_BLOCK(M0, N0, DATA_TYPE, c, c_lp);
Giorgio Arena1e2af2a2020-10-15 17:39:41 +0100[diff] [blame]2745 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c_lp, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2746#else // defined(MIXED_PRECISION)
Giorgio Arena1e2af2a2020-10-15 17:39:41 +0100[diff] [blame]2747 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Gian Marco Iodice0c17aa22019-09-27 09:23:15 +0100[diff] [blame]2748#endif // defined(MIXED_PRECISION)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2749
2750#undef LHS_BLOCK_SIZE
2751#undef LHS_OFFSET_X
2752#undef LHS_STEP_X
2753#undef RHS_BLOCK_SIZE
2754#undef RHS_OFFSET_X
2755#undef RHS_STEP_X
2756}
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2757#endif // defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT)
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]2758
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2759#if defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT_TEXTURE)
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2760/** This OpenCL kernel computes the matrix multiplication between 2 matrices. The RHS matrix is stored in OpenCL image object.
2761 * The LHS matrix must be reshaped with @ref CLGEMMReshapeLHSMatrixKernel and the M0xK0 must be transposed
2762 * The RHS matrix must be reshaped with @ref CLGEMMReshapeRHSMatrixKernel and the K0xN0 must be NOT transposed
2763 *
2764 * @note -DOPENCL_IMAGE_SUPPORT must be passed at compile time in order to compile this OpenCL kernel
2765 * @note LHS_TRANSPOSE should be passed at compile time in order to compile this OpenCL kernel (e.g. -DLHS_TRANSPOSE).
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2766 * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2767 * @note The GEMM's dimensions M, N and K must be passed at runtime.
Gian Marco Iodice781cba72020-06-19 16:56:57 +0100[diff] [blame]2768 * @note The height of the RHS matrix, defined before creating the OpenCL image object from the OpenCL buffer, should be passed at compile time using -DRHS_HEIGHT=<value> (e.g. -DRHS_HEIGHT=32)
2769 * Since we cannot create a 3d image from a buffer, the third dimension could be collapsed with the second dimension so RHS_HEIGHT
2770 * could be different from the value returned by get_image_height(rhs_img).
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2771 * @note The block's dimensions used for reshaping the LHS matrix and the RHS matrix (M0, N0 and K0) must be passed at compile time using -DM0, -DN0 and -DK0 (e.g. -DM0=4, -DN0=8, -DK0=4).
2772 * @note The number of M0xK0 vertical blocks stored on the same output row of the reshaped LHS matrix must be passed at compile time using -DV0 (e.g. -DV0=2)
2773 * @note The number of K0xN0 horizontal blocks stored on the same output row of the reshaped RHS matrix must be passed at compile time using -DH0 (e.g. -DH0=2)
2774 * @note If the M0xK0 blocks in the reshaped LHS matrix have been interleaved, the option -DLHS_INTERLEAVE must passed at compile time.
2775 * @note If the K0xN0 blocks in the reshaped RHS matrix have been interleaved, the option -DRHS_INTERLEAVE must passed at compile time.
Gian Marco Iodice088d63a2020-08-11 14:14:06 +0100[diff] [blame]2776 * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
2777 * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2778 * @note Only the following configurations of M0, N0 and K0 are currently supported:
2779 * - M0 = 2, 3, 4, 8
2780 * - N0 = 4, 8, 16
2781 * - K0 = 4, 8, 16
2782 * - V0 >= 1
2783 * - H0 >= 1
2784 *
2785 * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
2786 * The activation function is performed after the bias addition
2787 * @note In case the output has to be reinterpreted as a 3D tensor (e.g. output of convolution layer), the following information must be passed at compile time:
2788 * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
2789 * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
2790 * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
2791 * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped
2792 *
2793 * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F32
2794 * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes)
2795 * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
2796 * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes)
2797 * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
2798 * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix
2799 * @param[in] rhs_img The RHS reshaped matrix as cl_image 2d. Supported data type: same as @p lhs_ptr
2800 * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
2801 * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
2802 * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
2803 * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
2804 * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
2805 * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
2806 * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
2807 * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
2808 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
2809 * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
2810 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
2811 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
2812 * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes)
2813 * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
2814 * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
2815 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
2816 * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2817 * @param[in] M Number of rows in LHS matrix not reshaped.
2818 * @param[in] N Number of columns in RHS matrix not reshaped.
2819 * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped.
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2820 */
2821__kernel void gemm_mm_reshaped_lhs_t_rhs_nt_texture(IMAGE_DECLARATION(lhs),
2822 __read_only image2d_t rhs_img,
2823#if defined(BETA)
2824 IMAGE_DECLARATION(bias),
2825#endif // defined(BETA)
2826 IMAGE_DECLARATION(dst),
2827 uint lhs_stride_z,
2828 uint rhs_stride_z,
2829#if defined(BETA)
2830 uint bias_stride_z,
2831#endif //defined(BETA)
2832 uint dst_stride_z
2833#if defined(REINTERPRET_OUTPUT_AS_3D)
2834 ,
2835 uint dst_cross_plane_pad
2836#endif // REINTERPRET_OUTPUT_AS_3D
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]2837 ,
2838 const int M,
2839 const int N,
2840 const int K)
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]2841{
2842 // Pixel unit
2843#define PIXEL_UNIT CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(N0)
2844
2845 // Block size
2846#define LHS_BLOCK_SIZE ((K0) * (M0))
2847
2848#if defined(LHS_INTERLEAVE)
2849#define LHS_OFFSET_X (M0)
2850#define LHS_STEP_X ((M0) * (V0))
2851#define LHS_STEP_LOOP (1)
2852#else // defined(INTERLEAVE)
2853#define LHS_OFFSET_X (LHS_BLOCK_SIZE)
2854#define LHS_STEP_X (M0)
2855#define LHS_STEP_LOOP (V0)
2856#endif // defined(INTERLEAVE)
2857
2858 // Block size
2859#define RHS_BLOCK_SIZE ((K0) * (PIXEL_UNIT))
2860
2861 // RHS offset and step X
2862#if defined(RHS_INTERLEAVE)
2863#define RHS_OFFSET_X (PIXEL_UNIT)
2864#define RHS_STEP_X ((PIXEL_UNIT) * (H0))
2865#else // defined(RHS_INTERLEAVE)
2866#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
2867#define RHS_STEP_X (PIXEL_UNIT)
2868#endif // defined(RHS_INTERLEAVE)
2869
2870 const uint x = get_global_id(0);
2871 const uint y = get_global_id(1);
2872 const uint z = get_global_id(2);
2873
2874#if defined(DUMMY_WORK_ITEMS)
2875 if((x * N0 >= N) || (y * M0 >= M))
2876 {
2877 return;
2878 }
2879#endif // defined(DUMMY_WORK_ITEMS)
2880
2881 // Compute LHS matrix address
2882 __global uchar *lhs_addr = lhs_ptr + lhs_offset_first_element_in_bytes + (y % V0) * (uint)LHS_OFFSET_X * sizeof(DATA_TYPE) + (y / V0) * (uint)lhs_stride_y + (z * lhs_stride_z);
2883
2884#if defined(MATRIX_B_DEPTH)
2885 // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
2886 const uint z_rhs = (z % MATRIX_B_DEPTH);
2887#else // defined(MATRIX_B_DEPTH)
2888 const uint z_rhs = z;
2889#endif // defined(MATRIX_B_DEPTH)
2890
2891 // Compute RHS matrix coordinates
2892 uint x_rhs = (x % H0) * (uint)RHS_OFFSET_X;
2893 const uint y_rhs = (x / (uint)H0) + z_rhs * RHS_HEIGHT;
2894
2895 // Initialize the accumulators
2896 REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE_ACCUMULATOR, N0), c, 0);
2897
2898 REPEAT_VAR_INIT_TO_CONST(M0, uint, zero, 0);
2899
2900 __global DATA_TYPE *lhs = (__global DATA_TYPE *)(lhs_addr);
2901
2902 for(int i = 0; i < K; i += K0)
2903 {
2904 VEC_DATA_TYPE(DATA_TYPE, M0)
2905 a0;
2906 VEC_DATA_TYPE(DATA_TYPE, N0)
2907 b0;
2908
2909 a0 = VLOAD(M0)(0, lhs);
2910 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 0 * RHS_STEP_X), (y_rhs));
2911
2912 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2913
2914 lhs += LHS_STEP_X;
2915
2916#if K0 > 1
2917 a0 = VLOAD(M0)(0, lhs);
2918 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 1 * RHS_STEP_X), (y_rhs));
2919
2920 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2921
2922 lhs += LHS_STEP_X;
2923#endif // K0 > 1
2924
2925#if K0 > 2
2926 a0 = VLOAD(M0)(0, lhs);
2927 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 2 * RHS_STEP_X), (y_rhs));
2928
2929 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2930
2931 lhs += LHS_STEP_X;
2932#endif // K0 > 2
2933
2934#if K0 > 3
2935 a0 = VLOAD(M0)(0, lhs);
2936 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 3 * RHS_STEP_X), (y_rhs));
2937
2938 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2939
2940 lhs += LHS_STEP_X;
2941#endif // K0 > 3
2942
2943#if K0 > 4
2944 a0 = VLOAD(M0)(0, lhs);
2945 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 4 * RHS_STEP_X), (y_rhs));
2946
2947 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2948
2949 lhs += LHS_STEP_X;
2950
2951 a0 = VLOAD(M0)(0, lhs);
2952 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 5 * RHS_STEP_X), (y_rhs));
2953
2954 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2955
2956 lhs += LHS_STEP_X;
2957
2958 a0 = VLOAD(M0)(0, lhs);
2959 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 6 * RHS_STEP_X), (y_rhs));
2960
2961 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2962
2963 lhs += LHS_STEP_X;
2964
2965 a0 = VLOAD(M0)(0, lhs);
2966 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 7 * RHS_STEP_X), (y_rhs));
2967
2968 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2969
2970 lhs += LHS_STEP_X;
2971#endif // K0 > 4
2972
2973#if K0 > 8
2974 a0 = VLOAD(M0)(0, lhs);
2975 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 8 * RHS_STEP_X), (y_rhs));
2976
2977 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2978
2979 lhs += LHS_STEP_X;
2980
2981 a0 = VLOAD(M0)(0, lhs);
2982 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 9 * RHS_STEP_X), (y_rhs));
2983
2984 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2985
2986 lhs += LHS_STEP_X;
2987
2988 a0 = VLOAD(M0)(0, lhs);
2989 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 10 * RHS_STEP_X), (y_rhs));
2990
2991 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2992
2993 lhs += LHS_STEP_X;
2994
2995 a0 = VLOAD(M0)(0, lhs);
2996 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 11 * RHS_STEP_X), (y_rhs));
2997
2998 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
2999
3000 lhs += LHS_STEP_X;
3001
3002 a0 = VLOAD(M0)(0, lhs);
3003 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 12 * RHS_STEP_X), (y_rhs));
3004
3005 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
3006
3007 lhs += LHS_STEP_X;
3008
3009 a0 = VLOAD(M0)(0, lhs);
3010 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 13 * RHS_STEP_X), (y_rhs));
3011
3012 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
3013
3014 lhs += LHS_STEP_X;
3015
3016 a0 = VLOAD(M0)(0, lhs);
3017 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 14 * RHS_STEP_X), (y_rhs));
3018
3019 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
3020
3021 lhs += LHS_STEP_X;
3022
3023 a0 = VLOAD(M0)(0, lhs);
3024 b0 = READ_IMAGE2D(DATA_TYPE, PIXEL_UNIT, rhs_img, (x_rhs + 15 * RHS_STEP_X), (y_rhs));
3025
3026 ARM_MM_T_NT(M0, N0, 1, DATA_TYPE, a, b, c);
3027
3028 lhs += LHS_STEP_X;
3029#endif // K0 > 8
3030
3031#ifndef LHS_INTERLEAVE
3032 lhs += (M0 * K0 * (V0 - 1));
3033#endif // LHS_INTERLEAVE
3034
3035 x_rhs += K0 * RHS_STEP_X;
3036#ifndef RHS_INTERLEAVE
3037 x_rhs += (PIXEL_UNIT * K0 * (H0 - 1));
3038#endif // RHS_INTERLEAVE
3039 }
3040
3041 __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * dst_stride_y);
3042
3043 REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
3044
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]3045 const bool cond_y = ((get_global_id(1) + 1) * M0 >= M);
3046 const bool cond_x = ((get_global_id(0) + 1) * N0 >= N);
3047
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]3048#if defined(REINTERPRET_OUTPUT_AS_3D)
3049
3050 // The plane (zin) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
Michele Di Giorgio5fa963f2020-11-23 15:05:12 +0000[diff] [blame]3051 CALCULATE_Z_OFFSET(M0, uint, zout, y * (uint)M0, HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]3052 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
3053 // multiply dst_stride_z by DEPTH_GEMM3D
3054 dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
3055
3056#else // defined(REINTERPRET_OUTPUT_AS_3D)
3057
3058 // Add offset for batched GEMM
3059 dst_addr += z * dst_stride_z;
3060
3061#endif // defined(REINTERPRET_OUTPUT_AS_3D)
3062
3063 // Multiply by the weight of matrix-matrix product and store the result
3064#if defined(ALPHA)
3065 SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
3066#endif // defined(ALPHA)
3067
3068 // Add beta*bias
3069#if defined(BETA)
3070#if defined(BROADCAST_BIAS)
3071 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE));
3072
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]3073 LOAD_BLOCK_BOUNDARY_AWARE(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, 1, PARTIAL_STORE_N0, false, cond_x);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]3074
3075#ifndef UNIT_BETA
3076 SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
3077#endif // UNIT_BIAS
3078
3079 // c = c + bias[broadcasted]
3080#if defined(MIXED_PRECISION)
3081 CONVERT_BLOCK(1, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
3082 ADD_BLOCK_BROADCAST(M0, c, bias_hp0);
3083#else // defined(MIXED_PRECISION)
3084 ADD_BLOCK_BROADCAST(M0, c, bias0);
3085#endif // defined(MIXED_PRECISION)
3086
3087#else // defined(BROADCAST_BIAS)
3088 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (y * (uint)M0 * bias_stride_y) + z * bias_stride_z;
3089
Giorgio Arenabde2f352021-09-07 14:15:28 +0100[diff] [blame]3090 LOAD_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]3091
3092#ifndef UNIT_BETA
3093 SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
3094#endif // UNIT_BIAS
3095
3096#if defined(MIXED_PRECISION)
3097 CONVERT_BLOCK(M0, N0, DATA_TYPE_ACCUMULATOR, bias, bias_hp);
3098 ADD_BLOCK(M0, c, bias_hp);
3099#else // defined(MIXED_PRECISION)
3100 ADD_BLOCK(M0, c, bias);
3101#endif // defined(MIXED_PRECISION)
3102
3103#endif // defined(BROADCAST_BIAS)
3104#endif // defined(BETA)
3105
3106#if defined(ACTIVATION_TYPE)
3107#if defined(MIXED_PRECISION)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]3108 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE_ACCUMULATOR, N0, c, A_VAL, B_VAL);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]3109#else // defined(MIXED_PRECISION)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]3110 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, N0, c, A_VAL, B_VAL);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]3111#endif // defined(MIXED_PRECISION)
3112#endif // defined(ACTIVATION_TYPE)
3113
3114 // Store output block
3115#if defined(MIXED_PRECISION)
Gian Marco Iodice088d63a2020-08-11 14:14:06 +0100[diff] [blame]3116 CONVERT_BLOCK(M0, N0, DATA_TYPE, c, c_lp);
Giorgio Arena1e2af2a2020-10-15 17:39:41 +0100[diff] [blame]3117 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c_lp, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]3118#else // defined(MIXED_PRECISION)
Giorgio Arena1e2af2a2020-10-15 17:39:41 +0100[diff] [blame]3119 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]3120#endif // defined(MIXED_PRECISION)
3121
3122#undef LHS_BLOCK_SIZE
3123#undef LHS_OFFSET_X
3124#undef LHS_STEP_X
3125#undef RHS_BLOCK_SIZE
3126#undef RHS_OFFSET_X
3127#undef RHS_STEP_X
3128#undef PIXEL_UNIT
3129#undef LHS_STEP_LOOP
3130#undef RHS_STEP_LOOP
3131}
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]3132#endif // defined(OPENCL_IMAGE_SUPPORT) && defined(GEMM_MM_RESHAPED_LHS_T_RHS_NT_TEXTURE)
Gian Marco Iodicee3a849a2020-06-10 17:59:30 +0100[diff] [blame]3133
Giorgio Arenaae99b6e2019-08-01 14:22:12 +0100[diff] [blame]3134#endif // defined(LHS_TRANSPOSE)
3135
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]3136#endif // defined(M0) && defined(N0) && defined(K0) && defined(V0) && defined(H0) && defined(DATA_TYPE) && defined(DATA_TYPE_ACCUMULATOR)
Gian Marco Iodicebf9731e2018-12-12 10:18:04 +0000[diff] [blame]3137
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]3138#if defined(M0) && defined(N0) && defined(K0) && defined(DATA_TYPE)
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3139
3140#define VFMA(a, b, c) \
3141 ({ \
3142 c = fma(a, b, c); \
3143 })
3144
3145#if M0 == 1
3146#define RHS_VFMA_M0xN0(i, a, b, c) \
3147 ({ \
3148 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
3149 })
3150#elif M0 == 2 // M0 == 2
3151#define RHS_VFMA_M0xN0(i, a, b, c) \
3152 ({ \
3153 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
3154 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
3155 })
3156#elif M0 == 3 // M0 == 3
3157#define RHS_VFMA_M0xN0(i, a, b, c) \
3158 ({ \
3159 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
3160 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
3161 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
3162 })
3163#elif M0 == 4 // M0 == 4
3164#define RHS_VFMA_M0xN0(i, a, b, c) \
3165 ({ \
3166 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
3167 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
3168 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
3169 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
3170 })
3171#elif M0 == 5 // M0 == 5
3172#define RHS_VFMA_M0xN0(i, a, b, c) \
3173 ({ \
3174 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
3175 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
3176 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
3177 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
3178 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
3179 })
3180#elif M0 == 6 // M0 == 6
3181#define RHS_VFMA_M0xN0(i, a, b, c) \
3182 ({ \
3183 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
3184 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
3185 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
3186 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
3187 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
3188 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
3189 })
3190#elif M0 == 7 // M0 == 7
3191#define RHS_VFMA_M0xN0(i, a, b, c) \
3192 ({ \
3193 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
3194 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
3195 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
3196 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
3197 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
3198 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
3199 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \
3200 })
3201#elif M0 == 8 // M0 == 8
3202#define RHS_VFMA_M0xN0(i, a, b, c) \
3203 ({ \
3204 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##0).s##i), b, (c##0)); \
3205 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##1).s##i), b, (c##1)); \
3206 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##2).s##i), b, (c##2)); \
3207 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##3).s##i), b, (c##3)); \
3208 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##4).s##i), b, (c##4)); \
3209 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##5).s##i), b, (c##5)); \
3210 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##6).s##i), b, (c##6)); \
3211 VFMA((VEC_DATA_TYPE(DATA_TYPE, N0))((a##7).s##i), b, (c##7)); \
3212 })
3213#else // M0 not supported
3214#error "M0 not supported"
3215#endif // M0 not supported
3216
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]3217#if defined(GEMM_MM_NATIVE)
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3218/** This OpenCL kernel computes the matrix multiplication between 2 matrices.
3219 * The LHS matrix is NOT reshaped
3220 * The RHS matrix is NOT reshaped
3221 *
3222 * @note If the first two dimensions of NDRange have been dispatched with "dummy_work_items" support, the option -DDUMMY_WORK_ITEMS must be passed at compile time.
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]3223 * @note The GEMM's dimensions (M,N and K) must be passed at runtime as kernel parameters.
Gian Marco Iodiced1f54762019-07-19 09:54:47 +0100[diff] [blame]3224 * @note The number of M0 rows to process must be passed at compile time using -DM0 (e.g. -DM0=2)
3225 * @note The number of K0 partial accumulations must be passed at compile time using -DK0 (e.g., -DK0=2)
3226 * @note The number of N0 columns to process must be passed at compile time using -DN0 (e.g. -DN0=2)
SiCong Li3a501662020-06-26 10:02:06 +0100[diff] [blame]3227 * @note The size of the partial store block in y must be passed at compile time using -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_M0=1)
3228 * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_STORE_N0 (e.g. -DPARTIAL_STORE_N0=1)
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3229 * @note Only the following configurations of M0, N0 and K0 are currently supported:
3230 * - M0 = 1, 2, 3, 4, 5, 6, 7, 8
3231 * - N0 = 2, 3, 4, 8, 16
3232 * - K0 = 2, 3, 4, 8, 16
3233 *
Gian Marco Iodiced1f54762019-07-19 09:54:47 +0100[diff] [blame]3234 * @note If the activation type were passed at compile time through -DACTIVATION_TYPE (e.g. -DACTIVATION_TYPE=RELU), A, B variables, required by some activation functions, should be passed at compile time as well using -DA_VAL= and -DB_VAL= respectively.
Gian Marco Iodiceca1f4602019-07-16 15:46:48 +0100[diff] [blame]3235 * The activation function is performed after the bias addition
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3236 * @note In case the input or output have to be reinterpreted as a 3D tensor, the following information must be passed at compile time:
3237 * -# REINTERPRET_INPUT_AS_3D: To reinterpret the input as 3D
3238 * -# REINTERPRET_OUTPUT_AS_3D: To reinterpret the output as 3D
3239 * -# HEIGHT_GEMM3D: The height of the output in case it has to be reinterpreted as a 3D tensor.
3240 * -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
3241 * (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix
3242 *
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3243 * @param[in] lhs_ptr Pointer to the LHS matrix. Supported data type: F16/F32
3244 * @param[in] lhs_stride_x Stride of the LHS matrix in X dimension (in bytes)
3245 * @param[in] lhs_step_x lhs_stride_x * number of elements along X processed per workitem(in bytes)
3246 * @param[in] lhs_stride_y Stride of the LHS matrix in Y dimension (in bytes)
3247 * @param[in] lhs_step_y lhs_stride_y * number of elements along Y processed per workitem(in bytes)
3248 * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS matrix
3249 * @param[in] rhs_ptr Pointer to the RHS matrix. Supported data type: same as @p lhs_ptr
3250 * @param[in] rhs_stride_x Stride of the RHS matrix in X dimension (in bytes)
3251 * @param[in] rhs_step_x rhs_stride_x * number of elements along X processed per workitem(in bytes)
3252 * @param[in] rhs_stride_y Stride of the RHS matrix in Y dimension (in bytes)
3253 * @param[in] rhs_step_y rhs_stride_y * number of elements along Y processed per workitem(in bytes)
3254 * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS matrix
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3255 * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
3256 * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
3257 * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
3258 * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
3259 * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
3260 * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
3261 * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
3262 * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
3263 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
3264 * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
3265 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
3266 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
3267 * @param[in] lhs_stride_z Stride of the LHS matrix in Z dimension (in bytes)
3268 * @param[in] rhs_stride_z Stride of the RHS matrix in Z dimension (in bytes)
3269 * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
3270 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]3271 * @param[in] M Number of rows in LHS matrix not reshaped.
3272 * @param[in] N Number of columns in RHS matrix not reshaped.
3273 * @param[in] K Number of columns in LHS matrix and rows in RHS matrix not reshaped.
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3274 * @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
3275 * @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3276 */
3277__kernel void gemm_mm_native(IMAGE_DECLARATION(lhs),
3278 IMAGE_DECLARATION(rhs),
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3279#if defined(BETA)
3280 IMAGE_DECLARATION(bias),
3281#endif // defined(BETA)
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3282 IMAGE_DECLARATION(dst),
3283 uint lhs_stride_z,
3284 uint rhs_stride_z,
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3285#if defined(BETA)
3286 uint bias_stride_z,
3287#endif //defined(BETA)
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]3288 uint dst_stride_z,
3289 const int M,
3290 const int N,
3291 const int K
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3292#if defined(REINTERPRET_INPUT_AS_3D)
3293 ,
3294 uint lhs_cross_plane_pad
3295#endif // REINTERPRET_INPUT_AS_3D
3296#if defined(REINTERPRET_OUTPUT_AS_3D)
3297 ,
3298 uint dst_cross_plane_pad
3299#endif // REINTERPRET_OUTPUT_AS_3D
3300 )
3301{
3302 // Block size
3303#define RHS_BLOCK_SIZE ((K0) * (N0))
3304
3305 // RHS offset and step X
3306#define RHS_OFFSET_X (RHS_BLOCK_SIZE)
3307
3308 uint x = get_global_id(0);
3309 uint y = get_global_id(1);
3310 uint z = get_global_id(2);
3311
3312#if defined(DUMMY_WORK_ITEMS)
3313 if((x * N0 >= N) || (y * M0 >= M))
3314 {
3315 return;
3316 }
3317#endif // defined(DUMMY_WORK_ITEMS)
3318
3319 // Compute LHS matrix address
SiCongLi71cbd282021-11-03 12:17:06 +0000[diff] [blame]3320 uint lhs_offset = lhs_offset_first_element_in_bytes + COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * (uint)lhs_stride_y;
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3321
3322 // Compute RHS matrix address
3323 uint rhs_offset = rhs_offset_first_element_in_bytes + x * N0 * sizeof(DATA_TYPE);
3324
3325#if defined(MATRIX_B_DEPTH)
3326 // Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3
3327 rhs_offset += (z % MATRIX_B_DEPTH) * rhs_stride_z;
3328#else // defined(MATRIX_B_DEPTH)
3329 rhs_offset += z * rhs_stride_z;
3330#endif // defined(MATRIX_B_DEPTH)
3331
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3332 REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0);
3333 REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3334
3335#if defined(REINTERPRET_INPUT_AS_3D)
3336 // The plane (zlhs) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
SiCongLi71cbd282021-11-03 12:17:06 +0000[diff] [blame]3337 CALCULATE_Z_OFFSET(M0, uint, zlhs, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, lhs_cross_plane_pad, lhs_stride_y);
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3338
3339 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
3340 // multiply lhs_stride_z by DEPTH_GEMM3D
3341 lhs_offset += z * lhs_stride_z * DEPTH_GEMM3D;
3342
3343#else // defined(REINTERPRET_INPUT_AS_3D)
3344
3345 // Add offset for batched GEMM
3346 lhs_offset += z * lhs_stride_z;
3347
3348#endif // defined(REINTERPRET_INPUT_AS_3D)
3349
3350 // Initialize the accumulators
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3351 REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(M0-1)=0;
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3352
3353 int i = 0;
Gian Marco Iodicec9cecc02021-10-15 10:23:24 +0100[diff] [blame]3354#if K0 > 1
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3355 for(; i <= (K - K0); i += K0)
3356 {
3357 // Supported cases (M0, K0):
3358 // 1,2 - 1,3 - 1,4 - 1,8 - 1,16
3359 // 2,2 - 2,3 - 2,4 - 2,8 - 2,16
3360 // 3,2 - 3,3 - 3,4 - 3,8 - 3,16
3361 // 4,2 - 4,3 - 4,4 - 4,8 - 4,16
3362 // 5,2 - 5,3 - 5,4 - 5,8 - 5,16
3363 // 6,2 - 6,3 - 6,4 - 6,8 - 6,16
3364 // 7,2 - 7,3 - 7,4 - 7,8 - 7,16
3365 // 8,2 - 8,3 - 8,4 - 8,8 - 8,16
3366 // Load values from LHS matrix
3367 LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_ptr, lhs_offset, lhs_stride_y, zlhs);
3368
3369 // Load values from RHS matrix
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3370 LOAD_BLOCK(K0, N0, DATA_TYPE, b, rhs_ptr, rhs_offset, rhs_stride_y, zero);
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3371
3372 RHS_VFMA_M0xN0(0, a, b0, c);
3373 RHS_VFMA_M0xN0(1, a, b1, c);
3374#if K0 > 2
3375 RHS_VFMA_M0xN0(2, a, b2, c);
3376#endif // K0 > 2
3377#if K0 > 3
3378 RHS_VFMA_M0xN0(3, a, b3, c);
3379#endif // K0 > 3
3380#if K0 > 4
3381 RHS_VFMA_M0xN0(4, a, b4, c);
3382 RHS_VFMA_M0xN0(5, a, b5, c);
3383 RHS_VFMA_M0xN0(6, a, b6, c);
3384 RHS_VFMA_M0xN0(7, a, b7, c);
3385#endif // K0 > 4
3386#if K0 > 8
3387 RHS_VFMA_M0xN0(8, a, b8, c);
3388 RHS_VFMA_M0xN0(9, a, b9, c);
Gian Marco Iodice7b9d7ca2019-09-19 16:37:39 +0100[diff] [blame]3389 RHS_VFMA_M0xN0(A, a, bA, c);
3390 RHS_VFMA_M0xN0(B, a, bB, c);
3391 RHS_VFMA_M0xN0(C, a, bC, c);
3392 RHS_VFMA_M0xN0(D, a, bD, c);
3393 RHS_VFMA_M0xN0(E, a, bE, c);
3394 RHS_VFMA_M0xN0(F, a, bF, c);
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3395#endif // K0 > 8
3396
3397 lhs_offset += K0 * sizeof(DATA_TYPE);
3398 rhs_offset += K0 * rhs_stride_y;
3399 }
Gian Marco Iodicec9cecc02021-10-15 10:23:24 +0100[diff] [blame]3400#endif // K0 > 1
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3401 // Left-over accumulations
3402 for(; i < K; ++i)
3403 {
3404 // Load values from LHS matrix
3405 VEC_DATA_TYPE(DATA_TYPE, 2)
3406 a0 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 0 * lhs_stride_y + zlhs0));
3407#if M0 > 1
3408 VEC_DATA_TYPE(DATA_TYPE, 2)
3409 a1 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 1 * lhs_stride_y + zlhs1));
3410#endif // M0 > 1
3411#if M0 > 2
3412 VEC_DATA_TYPE(DATA_TYPE, 2)
3413 a2 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 2 * lhs_stride_y + zlhs2));
3414#endif // M0 > 2
3415#if M0 > 3
3416 VEC_DATA_TYPE(DATA_TYPE, 2)
3417 a3 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 3 * lhs_stride_y + zlhs3));
3418#endif // M0 > 3
3419#if M0 > 4
3420 VEC_DATA_TYPE(DATA_TYPE, 2)
3421 a4 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 4 * lhs_stride_y + zlhs4));
3422#endif // M0 > 4
3423#if M0 > 5
3424 VEC_DATA_TYPE(DATA_TYPE, 2)
3425 a5 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 5 * lhs_stride_y + zlhs5));
3426#endif // M0 > 5
3427#if M0 > 6
3428 VEC_DATA_TYPE(DATA_TYPE, 2)
3429 a6 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 6 * lhs_stride_y + zlhs6));
3430#endif // M0 > 6
3431#if M0 > 7
3432 VEC_DATA_TYPE(DATA_TYPE, 2)
3433 a7 = *((__global DATA_TYPE *)(lhs_ptr + lhs_offset + 7 * lhs_stride_y + zlhs7));
3434#endif // M0 > 7
3435
3436 VEC_DATA_TYPE(DATA_TYPE, N0)
3437 b = VLOAD(N0)(0, (__global DATA_TYPE *)(rhs_ptr + rhs_offset + 0 * rhs_stride_y));
3438 RHS_VFMA_M0xN0(0, a, b, c);
3439
3440 lhs_offset += sizeof(DATA_TYPE);
3441 rhs_offset += rhs_stride_y;
3442 }
3443
SiCongLi71cbd282021-11-03 12:17:06 +0000[diff] [blame]3444 __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * dst_stride_y);
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3445
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3446 REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3447
3448#if defined(REINTERPRET_OUTPUT_AS_3D)
3449 // The plane (zout) is calculated dividing M (y * M0) by HEIGHT_GEMM3D
SiCongLi71cbd282021-11-03 12:17:06 +0000[diff] [blame]3450 CALCULATE_Z_OFFSET(M0, uint, zout, COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0), HEIGHT_GEMM3D, DEPTH_GEMM3D, dst_cross_plane_pad, dst_stride_y);
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3451
3452 // Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we
3453 // multiply dst_stride_z by DEPTH_GEMM3D
3454 dst_addr += z * dst_stride_z * DEPTH_GEMM3D;
3455
3456#else // defined(REINTERPRET_OUTPUT_AS_3D)
3457
3458 // Add offset for batched GEMM
3459 dst_addr += z * dst_stride_z;
3460
3461#endif // defined(REINTERPRET_OUTPUT_AS_3D)
3462
3463 // Multiply by the weight of matrix-matrix product and store the result
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3464#if defined(ALPHA)
3465 SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
3466#endif // defined(ALPHA)
3467
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3468 // Add beta*bias
3469#if defined(BETA)
3470#if defined(BROADCAST_BIAS)
3471 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
3472
SiCongLi71cbd282021-11-03 12:17:06 +0000[diff] [blame]3473 LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3474
3475#ifndef UNIT_BETA
3476 SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
3477#endif // UNIT_BIAS
3478
3479 // c = c + bias[broadcasted]
3480 ADD_BLOCK_BROADCAST(M0, c, bias0);
3481
3482#else // defined(BROADCAST_BIAS)
SiCongLi71cbd282021-11-03 12:17:06 +0000[diff] [blame]3483 __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (x * (uint)N0 * sizeof(DATA_TYPE)) + (COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) * bias_stride_y) + z * bias_stride_z;
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3484
SiCongLi71cbd282021-11-03 12:17:06 +0000[diff] [blame]3485 LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
Gian Marco Iodice944170e2019-06-24 14:40:30 +0100[diff] [blame]3486
3487#ifndef UNIT_BETA
3488 SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
3489#endif // UNIT_BIAS
3490
3491 // c = c + bias
3492 ADD_BLOCK(M0, c, bias);
3493
3494#endif // defined(BROADCAST_BIAS)
3495#endif // defined(BETA)
3496
Gian Marco Iodiceca1f4602019-07-16 15:46:48 +0100[diff] [blame]3497#if defined(ACTIVATION_TYPE)
Gian Marco Iodice635013a2022-11-03 09:30:56 +0000[diff] [blame]3498 ACTIVATION_BLOCK(M0, ACTIVATION_TYPE, DATA_TYPE, N0, c, A_VAL, B_VAL);
Gian Marco Iodiceca1f4602019-07-16 15:46:48 +0100[diff] [blame]3499#endif // defined(ACTIVATION_TYPE)
3500
SiCongLi71cbd282021-11-03 12:17:06 +0000[diff] [blame]3501 const bool cond_y = y == 0;
3502 const bool cond_x = ((x + 1) * N0 >= N);
3503
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3504 // Store output block
Giorgio Arena1e2af2a2020-10-15 17:39:41 +0100[diff] [blame]3505 STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout, PARTIAL_STORE_M0, PARTIAL_STORE_N0, cond_y, cond_x);
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3506}
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]3507#endif // defined(GEMM_MM_NATIVE)
3508#endif // defined(M0) && defined(N0) && defined(K0) && defined(DATA_TYPE)
giuros01b3204e72019-04-01 13:50:22 +0100[diff] [blame]3509
Anton Lokhmotov3e80c7f2017-11-20 11:02:10 +0000[diff] [blame]3510#if defined(BETA)
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3511/** This OpenCL kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta:
3512 *
Gian Marco19835e52018-01-30 13:35:54 +0000[diff] [blame]3513 * @note The beta's value need to be passed at compile time using -DBETA
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3514 *
3515 * @param[in] src_ptr Pointer to the source matrix. Supported data types: F32
3516 * @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes)
3517 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
3518 * @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes)
3519 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
Isabella Gottardi8e74f442018-03-01 16:42:00 +0000[diff] [blame]3520 * @param[in] src_stride_z Stride of the destination tensor in Z dimension (in bytes)
3521 * @param[in] src_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3522 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]3523 * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3524 * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
3525 * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
3526 * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
3527 * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
Isabella Gottardi8e74f442018-03-01 16:42:00 +0000[diff] [blame]3528 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
3529 * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3530 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
3531 */
Isabella Gottardi8e74f442018-03-01 16:42:00 +0000[diff] [blame]3532__kernel void gemm_ma_f32(TENSOR3D_DECLARATION(src),
3533 TENSOR3D_DECLARATION(dst))
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3534{
Anton Lokhmotov3e80c7f2017-11-20 11:02:10 +0000[diff] [blame]3535 // Compute source and destination addresses
Isabella Gottardi8e74f442018-03-01 16:42:00 +0000[diff] [blame]3536 Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
3537 Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3538
Anton Lokhmotov3e80c7f2017-11-20 11:02:10 +0000[diff] [blame]3539 // Load values from A x B
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3540 float4 alpha_ab = vload4(0, (__global float *)dst.ptr);
3541
Anton Lokhmotov3e80c7f2017-11-20 11:02:10 +0000[diff] [blame]3542 // Load values from Matrix C
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3543 float4 c = vload4(0, (__global float *)src.ptr);
3544
Anton Lokhmotov3e80c7f2017-11-20 11:02:10 +0000[diff] [blame]3545 // Computes alpha * axb + beta * c
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3546 float4 out = alpha_ab + (float4)BETA * c;
3547
Anton Lokhmotov3e80c7f2017-11-20 11:02:10 +0000[diff] [blame]3548 // Store final result in axb matrix
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3549 vstore4(out, 0, (__global float *)dst.ptr);
3550}
3551
Vidhya Sudhan Loganathan76c85642018-05-25 13:53:02 +0100[diff] [blame]3552#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3553/** This OpenCL kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta:
3554 *
Gian Marco19835e52018-01-30 13:35:54 +0000[diff] [blame]3555 * @note The beta's value need to be passed at compile time using -DBETA
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]3556 *
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3557 * @param[in] src_ptr Pointer to the source matrix. Supported data types: F16
3558 * @param[in] src_stride_x Stride of the source matrix in X dimension (in bytes)
3559 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
3560 * @param[in] src_stride_y Stride of the source matrix in Y dimension (in bytes)
3561 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
Isabella Gottardi8e74f442018-03-01 16:42:00 +0000[diff] [blame]3562 * @param[in] src_stride_z Stride of the destination tensor in Z dimension (in bytes)
3563 * @param[in] src_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3564 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source matrix
Gian Marco Iodice3a3066b2017-06-23 13:38:14 +0100[diff] [blame]3565 * @param[out] dst_ptr Pointer to the destination matrix Supported data types: same as @p src_ptr
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3566 * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
3567 * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
3568 * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
3569 * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
Isabella Gottardi8e74f442018-03-01 16:42:00 +0000[diff] [blame]3570 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
3571 * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3572 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
3573 */
Isabella Gottardi8e74f442018-03-01 16:42:00 +0000[diff] [blame]3574__kernel void gemm_ma_f16(TENSOR3D_DECLARATION(src),
3575 TENSOR3D_DECLARATION(dst))
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3576{
Anton Lokhmotov3e80c7f2017-11-20 11:02:10 +0000[diff] [blame]3577 // Compute source and destination addresses
Isabella Gottardi8e74f442018-03-01 16:42:00 +0000[diff] [blame]3578 Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
3579 Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3580
Anton Lokhmotov3e80c7f2017-11-20 11:02:10 +0000[diff] [blame]3581 // Load values from A x B
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3582 half8 alpha_ab = vload8(0, (__global half *)dst.ptr);
3583
Anton Lokhmotov3e80c7f2017-11-20 11:02:10 +0000[diff] [blame]3584 // Load values from Matrix C
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3585 half8 c = vload8(0, (__global half *)src.ptr);
3586
Anton Lokhmotov3e80c7f2017-11-20 11:02:10 +0000[diff] [blame]3587 // Computes alpha * axb + beta * c
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3588 half8 out = alpha_ab + (half8)BETA * c;
3589
Anton Lokhmotov3e80c7f2017-11-20 11:02:10 +0000[diff] [blame]3590 // Store final result in axb matrix
Anthony Barbier6ff3b192017-09-04 18:44:23 +0100[diff] [blame]3591 vstore8(out, 0, (__global half *)dst.ptr);
3592}
Vidhya Sudhan Loganathan76c85642018-05-25 13:53:02 +0100[diff] [blame]3593#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)
ramelg019cca5922021-11-11 10:05:00 +0000[diff] [blame]3594#endif // defined(BETA)