1 /*
2 * Copyright (c) 2022-2023 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.h"
25
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/CL/OpenCL.h"
30 #include "arm_compute/core/Helpers.h"
31 #include "arm_compute/core/TensorInfo.h"
32 #include "arm_compute/core/Utils.h"
33 #include "arm_compute/core/Validate.h"
34 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
35 #include "src/core/CL/CLUtils.h"
36 #include "src/core/helpers/AutoConfiguration.h"
37 #include "src/core/helpers/WindowHelpers.h"
38 #include "src/core/utils/helpers/float_ops.h"
39 #include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h"
40 #include "support/Cast.h"
41 #include "support/StringSupport.h"
42
43 namespace arm_compute
44 {
45 namespace opencl
46 {
47 namespace kernels
48 {
49 namespace
50 {
51 using ElementsProcessed = Steps;
52
53 // Block size dimensions for the MMUL extension
54 constexpr int mmul_m0 = 4;
55 constexpr int mmul_n0 = 4;
56 constexpr int mmul_k0 = 4;
57
validate_arguments(const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * src2,const ITensorInfo * dst,float alpha,float beta,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)58 Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
59 const GEMMRHSMatrixInfo &rhs_info,
60 const GEMMKernelInfo &gemm_info)
61 {
62 ARM_COMPUTE_UNUSED(alpha);
63 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst);
64 ARM_COMPUTE_RETURN_ERROR_ON_MSG(!arm_matrix_multiply_supported(CLKernelLibrary::get().get_device()), "The extension cl_arm_matrix_multiply is not supported on the target platform");
65 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::F16, DataType::F32);
66 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1);
67 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4");
68 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3");
69 ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs_info.m0 < 1, "Only values greater than 0 are supported for m0");
70 ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.n0 != 1 && rhs_info.n0 != 2 && rhs_info.n0 != 3 && rhs_info.n0 != 4 && rhs_info.n0 != 8 && rhs_info.n0 != 16, "Only 1,2,3,4,8, and 16 are supported for n0");
71 ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.k0 != 1 || lhs_info.k0 != 1), "Only 1 is supported for k0");
72 ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.h0 != 4), "Only 4 is supported for h0");
73 ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.interleave != true, "Only true is supported for interleave with mmul extension enabled");
74 ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.transpose != false, "Only false is supported for transpose with mmul extension enabled");
75 ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported");
76 ARM_COMPUTE_RETURN_ON_ERROR(gemm::validate_image2d_support_on_rhs(*src1, rhs_info));
77
78 const unsigned int m = gemm_info.m;
79 const unsigned int n = gemm_info.n;
80 const unsigned int k = gemm_info.k;
81
82 ARM_COMPUTE_UNUSED(m);
83 ARM_COMPUTE_UNUSED(n);
84 ARM_COMPUTE_UNUSED(k);
85
86 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(0) != k);
87
88 // Validate the reinterpreted-as-3D-case
89 if(gemm_info.depth_output_gemm3d != 0)
90 {
91 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) * src0->dimension(2) != m);
92 }
93 else
94 {
95 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) != m);
96 }
97
98 // Validate the gemm-batched case
99 if(src1->num_dimensions() > 2)
100 {
101 if(gemm_info.depth_output_gemm3d != 0)
102 {
103 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(3) != src1->dimension(2));
104 }
105 else
106 {
107 ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(2) != src1->dimension(2));
108 }
109 }
110
111 if(src2 != nullptr && !(helpers::float_ops::is_zero(beta)))
112 {
113 const unsigned int src2_dim0 = src2->dimension(0);
114 const unsigned int src2_dim1 = src2->dimension(1);
115
116 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src2, src1);
117 if(gemm_info.broadcast_bias)
118 {
119 ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim1 != 1 || src2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
120 }
121 else
122 {
123 ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim0 != n || src2_dim1 != m), "Incorrect dimension of bias matrix");
124 }
125 }
126
127 TensorShape tensor_shape1{ src1->tensor_shape() };
128 tensor_shape1.set(0, n);
129 tensor_shape1.set(1, k);
130
131 const TensorInfo tensor_info1 = src1->clone()->set_tensor_shape(tensor_shape1);
132 const TensorInfo tensor_info_reshaped1 = src1->clone()->set_tensor_shape(misc::shape_calculator::compute_rhs_reshaped_shape(tensor_info1, rhs_info));
133
134 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src1, &tensor_info_reshaped1);
135
136 if(dst->total_size() != 0)
137 {
138 const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info));
139 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst);
140 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst);
141 }
142
143 return Status{};
144 }
145
validate_and_configure_window(ITensorInfo * src0,ITensorInfo * src1,ITensorInfo * src2,ITensorInfo * dst,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)146 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info,
147 const GEMMRHSMatrixInfo &rhs_info,
148 const GEMMKernelInfo &gemm_info)
149 {
150 ARM_COMPUTE_UNUSED(src0, src1, src2);
151 bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
152
153 // dst tensor auto initialization if not yet initialized
154 auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)));
155
156 TensorInfo tmp_info(*dst);
157
158 if(reinterpret_output_as_3d)
159 {
160 // Since the dst tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM,
161 // the window needs to be constructed on the 2D collapsed version of the tensor
162 TensorShape tmp_shape(dst->tensor_shape());
163 tmp_shape.collapse(2U, 1U);
164 tmp_info.set_tensor_shape(tmp_shape);
165 }
166
167 Window win = calculate_max_window(tmp_info, Steps(1, 1));
168
169 // Collapse along the Z direction
170 // This collapse needs to be here in order to tune the Z dimension of LWS
171 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
172 Window collapsed = win.collapse(win, dimension_to_collapse);
173
174 // Reconfigure window size, one arm_matrix_multiply kernel needs 16 threads to finish.
175 Window::Dimension x_dimension = collapsed.x();
176 Window::Dimension y_dimension = collapsed.y();
177
178 // Make M and N multiple of M0 and N0 respectively
179 const unsigned int ceil_to_multiple_n_n0 = ceil_to_multiple(x_dimension.end(), rhs_info.n0);
180 const unsigned int ceil_to_multiple_m_m0 = ceil_to_multiple(y_dimension.end(), lhs_info.m0);
181
182 // Divide M and N by M0 and N0 respectively
183 const unsigned int n_div_n0 = ceil_to_multiple_n_n0 / rhs_info.n0;
184 const unsigned int m_div_m0 = ceil_to_multiple_m_m0 / lhs_info.m0;
185
186 // Make n_div_n0 and m_div_m0 multiple of mmul_n0 and mmul_k0 respectively
187 const unsigned int ceil_to_multiple_n_div_n0_mmul_n0 = ceil_to_multiple(n_div_n0, mmul_n0);
188 const unsigned int ceil_to_multiple_m_div_m0_mmul_k0 = ceil_to_multiple(m_div_m0, mmul_k0);
189
190 // Ensure x_dimension is multiple of MMUL block size (mmul_n0 * mmul_k0)
191 x_dimension.set_end(ceil_to_multiple_n_div_n0_mmul_n0 * mmul_k0);
192 y_dimension.set_end(ceil_to_multiple_m_div_m0_mmul_k0 / mmul_k0);
193
194 collapsed.set(Window::DimX, x_dimension);
195 collapsed.set(Window::DimY, y_dimension);
196
197 return std::make_pair(Status{}, collapsed);
198 }
199 } // namespace
200
ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel()201 ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel::ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel()
202 {
203 _type = CLKernelType::GEMM;
204 }
205
configure(const CLCompileContext & compile_context,ITensorInfo * src0,ITensorInfo * src1,ITensorInfo * src2,ITensorInfo * dst,float alpha,float beta,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)206 void ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha,
207 float beta,
208 const GEMMLHSMatrixInfo &lhs_info,
209 const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
210 {
211 ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
212
213 // dst tensor auto initialization if not yet initialized
214 auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)));
215
216 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
217
218 auto padding_info = get_padding_info({ src0, src1, src2, dst });
219 _add_bias = src2 != nullptr;
220 _export_to_cl_image = rhs_info.export_to_cl_image;
221
222 // Configure kernel window
223 auto win_config = validate_and_configure_window(src0, src1, src2, dst, lhs_info, rhs_info, gemm_info);
224 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
225
226 IClKernel::configure_internal(win_config.second);
227
228 _m = gemm_info.m;
229 _n = gemm_info.n;
230 _k = gemm_info.k;
231
232 const unsigned int m0_leftover = _m % lhs_info.m0;
233 const unsigned int n0_leftover = _n % rhs_info.n0;
234
235 // Create build options
236 CLBuildOptions build_opts;
237 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src0->data_type()));
238 build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
239 build_opts.add_option_if(src2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
240 build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
241 build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
242 build_opts.add_option_if(src0->data_type() == DataType::F16, "-DHALF_PRECISION");
243 build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
244 build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
245 build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0));
246 build_opts.add_option("-DM0_LEFTOVER=" + support::cpp11::to_string(m0_leftover));
247 build_opts.add_option("-DN0_LEFTOVER=" + support::cpp11::to_string(n0_leftover));
248 build_opts.add_option("-DMMUL_M0=" + support::cpp11::to_string(mmul_m0));
249 build_opts.add_option("-DMMUL_N0=" + support::cpp11::to_string(mmul_n0));
250 build_opts.add_option("-DMMUL_K0=" + support::cpp11::to_string(mmul_k0));
251 build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
252 build_opts.add_option("-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
253 build_opts.add_option("-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
254
255 std::string kernel_name("gemm_mm_reshaped_only_rhs_nt_mmul");
256 kernel_name += rhs_info.export_to_cl_image ? "_texture" : "";
257
258 // A macro guard to compile ONLY the kernel of interest
259 build_opts.add_option("-D" + upper_string(kernel_name));
260
261 // Create kernel
262 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
263
264 // Set config_id for enabling LWS tuning
265 _config_id = kernel_name;
266 _config_id += "_";
267 _config_id += (_add_bias ? "add_bias_" : "");
268 _config_id += (gemm_info.broadcast_bias ? "broadcast_bias_" : "");
269 _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
270 _config_id += lower_string(string_from_data_type(src0->data_type()));
271 _config_id += "_";
272 _config_id += support::cpp11::to_string(_m);
273 _config_id += "_";
274 _config_id += support::cpp11::to_string(_n);
275 _config_id += "_";
276 _config_id += support::cpp11::to_string(_k);
277 _config_id += "_";
278 _config_id += support::cpp11::to_string(lhs_info.m0);
279 _config_id += "_";
280 _config_id += support::cpp11::to_string(rhs_info.n0);
281
282 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
283 }
284
validate(const ITensorInfo * src0,const ITensorInfo * src1,const ITensorInfo * src2,const ITensorInfo * dst,float alpha,float beta,const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const GEMMKernelInfo & gemm_info)285 Status ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta,
286 const GEMMLHSMatrixInfo &lhs_info,
287 const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
288 {
289 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
290 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src0->clone().get(),
291 src1->clone().get(),
292 src2 != nullptr ? src2->clone().get() : nullptr,
293 dst->clone().get(),
294 lhs_info,
295 rhs_info,
296 gemm_info)
297 .first);
298
299 return Status{};
300 }
301
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)302 void ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
303 {
304 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
305 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
306
307 const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
308 const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
309 const auto src2 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
310 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
311
312 ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);
313 ARM_COMPUTE_ERROR_ON(_add_bias && src2 == nullptr);
314
315 if(src1->info()->num_dimensions() < 3)
316 {
317 // The stride_z for matrix B must be zero if we do not slice
318 ARM_COMPUTE_ERROR_ON(src1->info()->strides_in_bytes()[3] != 0);
319 }
320
321 cl::Image2D src1_image2d;
322
323 if(_export_to_cl_image)
324 {
325 const TensorShape shape2d(src1->info()->dimension(0) / 4, src1->info()->dimension(1) * src1->info()->dimension(2));
326 const size_t image_row_pitch = src1->info()->strides_in_bytes()[1];
327
328 src1_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), src1->cl_buffer(), shape2d, src1->info()->data_type(), image_row_pitch, CLImage2DType::ReadOnly);
329 }
330
331 Window slice = window.first_slice_window_3D();
332
333 do
334 {
335 unsigned int idx = 0;
336
337 add_3d_tensor_nhw_argument(idx, src0);
338 if(_export_to_cl_image)
339 {
340 _kernel.setArg(idx++, src1_image2d);
341 }
342 add_3d_tensor_nhw_argument(idx, src1);
343
344 // Bias buffer (_add_bias == true)
345 if(_add_bias)
346 {
347 add_3d_tensor_nhw_argument(idx, src2);
348 }
349 // dst buffer
350 add_3d_tensor_nhw_argument(idx, dst);
351
352 // Pass m, n and k at runtime as signed ints, to ensure results of any subtractions they could be operand in, would still be signed.
353 _kernel.setArg<cl_int>(idx++, _m);
354 _kernel.setArg<cl_int>(idx++, _n);
355 _kernel.setArg<cl_int>(idx++, _k);
356
357 // LWS_x should be multiple of 16 at least. (32, 2) has been chosen to have more work-items on a single core
358 // LWS also enforces the order of execution of the workitems which improves cache utilization
359 enqueue(queue, *this, slice, cl::NDRange(32, 2), false);
360 }
361 while(window.slide_window_slice_3D(slice));
362 }
363 } // namespace kernels
364 } // namespace opencl
365 } // namespace arm_compute
366