1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
|
/*
* Copyright (c) 2023-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "GpuCkwCast.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/utils/helpers/AdjustVecSize.h"
#include "arm_compute/core/Validate.h"
#include "src/core/helpers/WindowHelpers.h"
#include "src/dynamic_fusion/sketch/gpu/ckw_driver/components/utils/CkwHelper.h"
#include "src/dynamic_fusion/sketch/gpu/ckw_driver/components/utils/type_converter/Common.h"
#include "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwScopedKernelWriter.h"
#include "src/dynamic_fusion/sketch/gpu/ckw_driver/GpuCkwVariableTable.h"
#include "src/dynamic_fusion/sketch/gpu/GpuKernelArgument.h"
#include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h"
#include "compute_kernel_writer/include/ckw/KernelWriter.h"
#include <cstdint>
#include <string>
namespace arm_compute
{
namespace experimental
{
namespace dynamic_fusion
{
GpuCkwCast::GpuCkwCast(ComponentId id, const ArgumentPack<ITensorInfo> &tensors, const Attributes &attributes)
: IGpuCkwComponentDriver{id, tensors}, _src{}, _dst{}, _attributes{attributes}
{
_src = this->tensors().get_const_tensor(TensorType::ACL_SRC_0);
_dst = this->tensors().get_const_tensor(TensorType::ACL_DST_0);
ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _dst);
ARM_COMPUTE_ERROR_ON_MSG(is_data_type_float(_src->data_type()) == false,
"The source data type must be a floating-point data type");
}
void GpuCkwCast::write_component_code(const ComponentGroup &comp_group,
GpuCkwVariableTable &vtable,
GpuCkwScopedKernelWriter writer) const
{
/********************************************************************************
* 1 - Define tensors
********************************************************************************/
GpuCkwComponentArgument *src = vtable.declare_variable(comp_group, writer, _src, "src");
GpuCkwComponentArgument *dst = vtable.declare_variable(comp_group, writer, _dst, "dst");
/********************************************************************************
* 2 - Define CKW constants
********************************************************************************/
const auto dst_h = static_cast<int32_t>(_dst->dimension(1));
// CKW constants
auto const_dst_h_i32 = writer->declare_constant_tile(ckw::ConstantData({{dst_h}}, ckw::DataType::Int32));
auto const_pos_1_i32 = writer->declare_constant_tile(ckw::ConstantData({{1}}, ckw::DataType::Int32));
auto const_0_i32 = writer->declare_constant_tile(ckw::ConstantData({{0}}, ckw::DataType::Int32));
/********************************************************************************
* 3 - Define the compute block parameters and destination tile (if not root component)
* Bind the tile to the tensor to share it among different components and
* initialize the compute block parameters
********************************************************************************/
// The compute block parameters depend on the employed tensor format
// Destination compute block size
int32_t dst_n0 = -1;
int32_t dst_m0 = -1;
// Destination compute block size left-over
int32_t dst_n0_partial = -1;
int32_t dst_m0_partial = -1;
// Shift-back for the overlapping-min strategy
int32_t dst_shift_back = -1;
if (!dst->has_tile())
{
// If ROOT component, we use ckw::TensorSamplerFormat::Dim0_Dim1xDim2_1
// as tensor format
const auto root_window = comp_group.get_root_component()->ckw_component_driver()->get_window();
dst_n0 = root_window.x().step();
dst_m0 = root_window.y().step();
dst_n0_partial = _dst->dimension(0) % dst_n0;
dst_m0_partial = (_dst->dimension(1) * _dst->dimension(2)) % dst_m0;
dst_shift_back = (dst_n0 - dst_n0_partial) % dst_n0;
ckw::TensorSampler sampler_dst;
sampler_dst.format(ckw::TensorSamplerFormat::Dim0_Dim1xDim2_1);
if (dst_n0_partial == 0)
{
sampler_dst.address_mode_x(ckw::TensorSamplerAddressModeX::None);
}
else
{
sampler_dst.address_mode_x(ckw::TensorSamplerAddressModeX::OverlappingMin);
}
if (dst_m0_partial == 0)
{
sampler_dst.address_mode_y(ckw::TensorSamplerAddressModeY::None);
}
else
{
sampler_dst.address_mode_y(ckw::TensorSamplerAddressModeY::ClampToBorderMaxOnly);
}
sampler_dst.address_mode_z(ckw::TensorSamplerAddressModeZ::None);
sampler_dst.storage(ckw::TensorStorageType::BufferUint8Ptr);
// Declare destination tile
ckw::DataType dst_dt = to_ckw(_dst->data_type());
auto tile_dst = writer->declare_tile("dst", ckw::TileInfo(dst_dt, dst_m0, dst_n0));
// Bind tile to the tensor
dst->init_virtual_tensor(tile_dst, sampler_dst);
}
else
{
// Change dst_n0 and dst_m0 if NOT root component!
// ATTENTION:
// dst_m0_partial depends on the TensorSamplerFormat
dst_n0 = dst->tile().tile_info().width();
dst_m0 = dst->tile().tile_info().height();
dst_n0_partial = _dst->dimension(0) % dst_n0;
ckw::TensorSampler sampler_dst = dst->tensor_sampler();
if (sampler_dst.format() == ckw::TensorSamplerFormat::Dim0_Dim1xDim2_1)
{
dst_m0_partial = (_dst->dimension(1) * _dst->dimension(2)) % dst_m0;
}
else if (sampler_dst.format() == ckw::TensorSamplerFormat::Dim0_Dim1_Dim2)
{
dst_m0_partial = _dst->dimension(1) % dst_m0;
}
// Shift-back for the overlapping-min strategy
dst_shift_back = (dst_n0 - dst_n0_partial) % dst_n0;
}
const auto &tile_dst = dst->tile();
/********************************************************************************
* 4 - Define the compute block parameters CKW constants
********************************************************************************/
// Only now we can declare the N0 and M0 as constant
auto const_dst_n0_i32 = writer->declare_constant_tile(ckw::ConstantData({{dst_n0}}, ckw::DataType::Int32));
auto const_dst_m0_i32 = writer->declare_constant_tile(ckw::ConstantData({{dst_m0}}, ckw::DataType::Int32));
auto const_dst_shift_back_n0_i32 =
writer->declare_constant_tile(ckw::ConstantData({{dst_shift_back}}, ckw::DataType::Int32));
/********************************************************************************
* 5 - Define the sampler for the input tensor
********************************************************************************/
if (!src->has_tile())
{
// Sampler
ckw::TensorSampler sampler_src = dst->tensor_sampler();
auto tile_gid_0 = writer->declare_tile("gid_0", ckw::TileInfo(ckw::DataType::Int32));
auto tile_gid_1 = writer->declare_tile("gid_1", ckw::TileInfo(ckw::DataType::Int32));
auto tile_gid_2 = writer->declare_tile("gid_2", ckw::TileInfo(ckw::DataType::Int32));
writer->op_get_global_id(tile_gid_0, 0);
writer->op_get_global_id(tile_gid_1, 1);
writer->op_get_global_id(tile_gid_2, 2);
auto tile_cout0 = writer->declare_tile("cout0", ckw::TileInfo(ckw::DataType::Int32)); // OFM
auto tile_mout0 = writer->declare_tile("mout0", ckw::TileInfo(ckw::DataType::Int32)); // WIDTH or WIDTH x HEIGHT
auto tile_mout1 = writer->declare_tile("mout1", ckw::TileInfo(ckw::DataType::Int32)); // HEIGHT or 0
auto tile_bout0 = writer->declare_tile("bout0", ckw::TileInfo(ckw::DataType::Int32)); // BATCH SIZE IDX
// Calculate coordinates
get_coordinate_from_gws_overlapping_min(writer, tile_cout0, tile_gid_0, const_dst_n0_i32,
const_dst_shift_back_n0_i32, const_0_i32);
get_coordinate_from_gws(writer, tile_mout0, tile_gid_1, const_dst_m0_i32);
// Get the boundary aware coordinates at each global dimension index
if (sampler_src.format() == ckw::TensorSamplerFormat::Dim0_Dim1xDim2_1)
{
writer->op_assign(tile_mout1, const_0_i32);
get_coordinate_from_gws(writer, tile_bout0, tile_gid_2, const_pos_1_i32);
}
else if (sampler_src.format() == ckw::TensorSamplerFormat::Dim0_Dim1_Dim2)
{
writer->op_binary(tile_mout1, ckw::BinaryOp::Mod, tile_gid_2, const_dst_h_i32);
writer->op_binary(tile_bout0, ckw::BinaryOp::Div, tile_gid_2, const_dst_h_i32);
}
ckw::DataType src_dt = to_ckw(_src->data_type());
auto tile_src = writer->declare_tile("src", ckw::TileInfo(src_dt, dst_m0, dst_n0));
writer->op_load(tile_src, src->tensor(), sampler_src, tile_cout0, tile_mout0, tile_mout1, tile_bout0);
// Here, init_virtual_tensor() it is used to bring the tile_src outside the compound statement
src->init_virtual_tensor(tile_src, sampler_src);
}
auto tile_src = src->tile();
/********************************************************************************
* 6 - Extra operations required before writing the main code (optional)
********************************************************************************/
// Not required
/********************************************************************************
* 7 - Write the rest of the code
********************************************************************************/
// Only None ConvertPolicy is supported for floating-point data types
ckw::ConvertPolicy convert_policy = ckw::ConvertPolicy::None;
writer->op_cast(tile_dst, tile_src, convert_policy);
ARM_COMPUTE_ERROR_ON_MSG(dst->has_tile() == false, "You must bind a tile before appending another component");
}
Window GpuCkwCast::get_window() const
{
ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized");
TensorShape output_shape = _dst->tensor_shape();
// Collapse Dim 1 (W) and Dim 2 (H) together, leave Dim 0 (C) unchanged
// This is in line with the collapsing convention used by operators like Conv2d
output_shape.collapse(2U, 1U);
constexpr uint32_t vector_size_byte_opencl = 16;
const uint32_t num_elems_processed_per_iteration =
adjust_vec_size(vector_size_byte_opencl / _dst->element_size(), _dst->dimension(0));
Window win = calculate_max_window(output_shape, Steps(num_elems_processed_per_iteration));
return win;
}
} // namespace dynamic_fusion
} // namespace experimental
} // namespace arm_compute
|
