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
|
/* SPDX-License-Identifier: Apache-2.0
* Copyright 2011-2022 Blender Foundation */
#include "integrator/shader_eval.h"
#include "device/device.h"
#include "device/queue.h"
#include "device/cpu/kernel.h"
#include "device/cpu/kernel_thread_globals.h"
#include "util/log.h"
#include "util/progress.h"
#include "util/tbb.h"
CCL_NAMESPACE_BEGIN
ShaderEval::ShaderEval(Device *device, Progress &progress) : device_(device), progress_(progress)
{
DCHECK_NE(device_, nullptr);
}
bool ShaderEval::eval(const ShaderEvalType type,
const int max_num_inputs,
const int num_channels,
const function<int(device_vector<KernelShaderEvalInput> &)> &fill_input,
const function<void(device_vector<float> &)> &read_output)
{
bool first_device = true;
bool success = true;
device_->foreach_device([&](Device *device) {
if (!first_device) {
VLOG_WORK << "Multi-devices are not yet fully implemented, will evaluate shader on a "
"single device.";
return;
}
first_device = false;
device_vector<KernelShaderEvalInput> input(device, "ShaderEval input", MEM_READ_ONLY);
device_vector<float> output(device, "ShaderEval output", MEM_READ_WRITE);
/* Allocate and copy device buffers. */
DCHECK_EQ(input.device, device);
DCHECK_EQ(output.device, device);
DCHECK_LE(output.size(), input.size());
input.alloc(max_num_inputs);
int num_points = fill_input(input);
if (num_points == 0) {
return;
}
input.copy_to_device();
output.alloc(num_points * num_channels);
output.zero_to_device();
/* Evaluate on CPU or GPU. */
success = (device->info.type == DEVICE_CPU) ?
eval_cpu(device, type, input, output, num_points) :
eval_gpu(device, type, input, output, num_points);
/* Copy data back from device if not canceled. */
if (success) {
output.copy_from_device(0, 1, output.size());
read_output(output);
}
input.free();
output.free();
});
return success;
}
bool ShaderEval::eval_cpu(Device *device,
const ShaderEvalType type,
device_vector<KernelShaderEvalInput> &input,
device_vector<float> &output,
const int64_t work_size)
{
vector<CPUKernelThreadGlobals> kernel_thread_globals;
device->get_cpu_kernel_thread_globals(kernel_thread_globals);
/* Find required kernel function. */
const CPUKernels &kernels = Device::get_cpu_kernels();
/* Simple parallel_for over all work items. */
KernelShaderEvalInput *input_data = input.data();
float *output_data = output.data();
bool success = true;
tbb::task_arena local_arena(device->info.cpu_threads);
local_arena.execute([&]() {
parallel_for(int64_t(0), work_size, [&](int64_t work_index) {
/* TODO: is this fast enough? */
if (progress_.get_cancel()) {
success = false;
return;
}
const int thread_index = tbb::this_task_arena::current_thread_index();
const KernelGlobalsCPU *kg = &kernel_thread_globals[thread_index];
switch (type) {
case SHADER_EVAL_DISPLACE:
kernels.shader_eval_displace(kg, input_data, output_data, work_index);
break;
case SHADER_EVAL_BACKGROUND:
kernels.shader_eval_background(kg, input_data, output_data, work_index);
break;
case SHADER_EVAL_CURVE_SHADOW_TRANSPARENCY:
kernels.shader_eval_curve_shadow_transparency(kg, input_data, output_data, work_index);
break;
}
});
});
return success;
}
bool ShaderEval::eval_gpu(Device *device,
const ShaderEvalType type,
device_vector<KernelShaderEvalInput> &input,
device_vector<float> &output,
const int64_t work_size)
{
/* Find required kernel function. */
DeviceKernel kernel;
switch (type) {
case SHADER_EVAL_DISPLACE:
kernel = DEVICE_KERNEL_SHADER_EVAL_DISPLACE;
break;
case SHADER_EVAL_BACKGROUND:
kernel = DEVICE_KERNEL_SHADER_EVAL_BACKGROUND;
break;
case SHADER_EVAL_CURVE_SHADOW_TRANSPARENCY:
kernel = DEVICE_KERNEL_SHADER_EVAL_CURVE_SHADOW_TRANSPARENCY;
break;
};
/* Create device queue. */
unique_ptr<DeviceQueue> queue = device->gpu_queue_create();
queue->init_execution();
/* Execute work on GPU in chunk, so we can cancel.
* TODO: query appropriate size from device. */
const int32_t chunk_size = 65536;
device_ptr d_input = input.device_pointer;
device_ptr d_output = output.device_pointer;
assert(work_size <= 0x7fffffff);
for (int32_t d_offset = 0; d_offset < int32_t(work_size); d_offset += chunk_size) {
int32_t d_work_size = std::min(chunk_size, int32_t(work_size) - d_offset);
DeviceKernelArguments args(&d_input, &d_output, &d_offset, &d_work_size);
queue->enqueue(kernel, d_work_size, args);
queue->synchronize();
if (progress_.get_cancel()) {
return false;
}
}
return true;
}
CCL_NAMESPACE_END
|
