Skip to content

Commit 4b45264

Browse files
Esteban Padilla Cerdiofacebook-github-bot
Esteban Padilla Cerdio
authored and
facebook-github-bot
committed
Get device information from OpenCL (#4210)
Summary: Pull Request resolved: #4210 Some of the device's properties, like the number of available Streaming Processors, or the total size of the cache, are not supported by Vulkan as direct queries. These variables are essential for ArchProbe's algorithms, so we need to do some preprocessing with OpenCL to extract these values, before moving to the Vulkan implementation. This diff moves the GPUInfo implementation into a class to have better control of internal variables like the device's properties, and obtains important limits that will be used in the following diffs, like the device's SM Count, Cache size and number of logic threads. Reviewed By: jorgep31415 Differential Revision: D59636879 fbshipit-source-id: 1b9aa3a4ce48f360526e09576864514b2ac35429
1 parent 93a7725 commit 4b45264

File tree

2 files changed

+160
-99
lines changed

2 files changed

+160
-99
lines changed

‎backends/vulkan/tools/gpuinfo/include/utils.h

Lines changed: 30 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -10,6 +10,10 @@
1010

1111
#include <executorch/backends/vulkan/runtime/api/api.h>
1212

13+
#define CL_TARGET_OPENCL_VERSION 200
14+
#define CL_HPP_TARGET_OPENCL_VERSION CL_TARGET_OPENCL_VERSION
15+
#include <CL/opencl.hpp>
16+
1317
using namespace vkcompute;
1418
using namespace api;
1519

@@ -49,3 +53,29 @@ void ensure_min_niter(
4953
niter = uint32_t(niter * min_time_us / t);
5054
}
5155
}
56+
57+
cl_platform_id get_cl_platform_id() {
58+
cl_uint nplatform_id;
59+
clGetPlatformIDs(0, nullptr, &nplatform_id);
60+
std::vector<cl_platform_id> platform_ids;
61+
platform_ids.resize(nplatform_id);
62+
clGetPlatformIDs(nplatform_id, platform_ids.data(), nullptr);
63+
return platform_ids[0];
64+
}
65+
66+
cl_device_id get_cl_dev_id(cl_platform_id platform_id) {
67+
cl_uint ndev_id;
68+
clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_ALL, 0, nullptr, &ndev_id);
69+
std::vector<cl_device_id> dev_ids;
70+
dev_ids.resize(ndev_id);
71+
clGetDeviceIDs(
72+
platform_id, CL_DEVICE_TYPE_ALL, ndev_id, dev_ids.data(), nullptr);
73+
return dev_ids[0];
74+
}
75+
76+
cl::Device get_cl_device() {
77+
auto platform_id = get_cl_platform_id();
78+
auto dev_id = get_cl_dev_id(platform_id);
79+
cl::Device dev(dev_id);
80+
return dev;
81+
}

‎backends/vulkan/tools/gpuinfo/src/app.cpp

Lines changed: 130 additions & 99 deletions
Original file line numberDiff line numberDiff line change
@@ -13,117 +13,148 @@
1313
#include "stats.h"
1414
#include "utils.h"
1515

16-
void reg_count() {
17-
const uint32_t NREG_MIN = 1;
18-
const uint32_t NREG_MAX = 512;
19-
const uint32_t NREG_STEP = 1;
20-
21-
const double COMPENSATE = 0.01;
22-
const double THRESHOLD = 3;
23-
24-
const uint32_t NGRP_MIN = 1;
25-
const uint32_t NGRP_MAX = 64;
26-
const uint32_t NGRP_STEP = 1;
27-
28-
uint32_t NITER;
29-
30-
auto bench = [&](uint32_t ngrp, uint32_t nreg) {
31-
size_t len = sizeof(float);
32-
StorageBuffer buffer(context(), vkapi::kFloat, len);
33-
ParamsBuffer params(context(), int32_t(len));
34-
vkapi::PipelineBarrier pipeline_barrier{};
35-
36-
auto shader_name = "reg_count_" + std::to_string(nreg);
37-
38-
auto time = benchmark_on_gpu(shader_name, 100, [&]() {
39-
context()->submit_compute_job(
40-
VK_KERNEL_FROM_STR(shader_name),
41-
pipeline_barrier,
42-
{1, ngrp, 1},
43-
{1, 1, 1},
44-
{SV(NITER)},
45-
VK_NULL_HANDLE,
46-
0,
47-
buffer.buffer(),
48-
params.buffer());
49-
});
50-
return time;
51-
};
52-
53-
std::cout << "Calculating NITER..." << std::endl;
54-
ensure_min_niter(1000, NITER, [&]() { return bench(1, NREG_MIN); });
55-
std::cout << "NITER," << NITER << std::endl;
56-
57-
uint32_t nreg_max;
58-
59-
DtJumpFinder<5> dj(COMPENSATE, THRESHOLD);
60-
uint32_t nreg = NREG_MIN;
61-
for (; nreg <= NREG_MAX; nreg += NREG_STEP) {
62-
double time = bench(1, nreg);
63-
std::cout << "Testing nreg=\t" << nreg << "\tTime=\t" << time << std::endl;
64-
if (dj.push(time)) {
65-
nreg -= NREG_STEP;
66-
nreg_max = nreg;
67-
break;
68-
}
69-
}
70-
if (nreg >= NREG_MAX) {
71-
std::cout << "Unable to conclude a maximal register count" << std::endl;
72-
nreg_max = NREG_STEP;
73-
} else {
74-
std::cout << nreg_max << " registers are available at most" << std::endl;
16+
using namespace vkapi;
17+
18+
class App {
19+
private:
20+
size_t buf_cache_size_;
21+
uint32_t sm_count_;
22+
uint32_t nthread_logic_;
23+
24+
public:
25+
App() {
26+
context()->initialize_querypool();
27+
28+
std::cout << context()->adapter_ptr()->stringize() << std::endl
29+
<< std::endl;
30+
31+
auto cl_device = get_cl_device();
32+
33+
sm_count_ = cl_device.getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>();
34+
nthread_logic_ = cl_device.getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>();
35+
buf_cache_size_ = cl_device.getInfo<CL_DEVICE_GLOBAL_MEM_CACHE_SIZE>();
36+
37+
std::cout << std::endl;
38+
std::cout << "SM count," << sm_count_ << std::endl;
39+
std::cout << "Logic Thread Count," << nthread_logic_ << std::endl;
40+
std::cout << "Cache Size," << buf_cache_size_ << std::endl;
7541
}
7642

77-
auto find_ngrp_by_nreg = [&](const uint32_t nreg) {
78-
DtJumpFinder<5> dj(COMPENSATE, THRESHOLD);
79-
for (auto ngrp = NGRP_MIN; ngrp <= NGRP_MAX; ngrp += NGRP_STEP) {
80-
auto time = bench(ngrp, nreg);
81-
std::cout << "Testing occupation (nreg=" << nreg << "); ngrp=" << ngrp
82-
<< ", time=" << time << " us" << std::endl;
43+
void reg_count() {
44+
std::cout << std::endl;
45+
std::cout << "------ Register Count ------" << std::endl;
46+
const uint32_t NREG_MIN = 1;
47+
const uint32_t NREG_MAX = 512;
48+
const uint32_t NREG_STEP = 1;
49+
50+
const double COMPENSATE = 0.01;
51+
const double THRESHOLD = 3;
52+
53+
const uint32_t NGRP_MIN = 1;
54+
const uint32_t NGRP_MAX = 64;
55+
const uint32_t NGRP_STEP = 1;
56+
57+
uint32_t NITER;
58+
59+
auto bench = [&](uint32_t ngrp, uint32_t nreg) {
60+
size_t len = sizeof(float);
61+
StorageBuffer buffer(context(), vkapi::kFloat, len);
62+
ParamsBuffer params(context(), int32_t(len));
63+
vkapi::PipelineBarrier pipeline_barrier{};
64+
65+
auto shader_name = "reg_count_" + std::to_string(nreg);
66+
67+
auto time = benchmark_on_gpu(shader_name, 100, [&]() {
68+
context()->submit_compute_job(
69+
VK_KERNEL_FROM_STR(shader_name),
70+
pipeline_barrier,
71+
{1, ngrp, 1},
72+
{1, 1, 1},
73+
{SV(NITER)},
74+
VK_NULL_HANDLE,
75+
0,
76+
buffer.buffer(),
77+
params.buffer());
78+
});
79+
return time;
80+
};
81+
82+
std::cout << "Calculating NITER..." << std::endl;
83+
ensure_min_niter(1000, NITER, [&]() { return bench(1, NREG_MIN); });
84+
std::cout << "NITER," << NITER << std::endl;
85+
86+
uint32_t nreg_max;
8387

88+
DtJumpFinder<5> dj(COMPENSATE, THRESHOLD);
89+
uint32_t nreg = NREG_MIN;
90+
for (; nreg <= NREG_MAX; nreg += NREG_STEP) {
91+
double time = bench(1, nreg);
92+
std::cout << "Testing nreg=\t" << nreg << "\tTime=\t" << time
93+
<< std::endl;
8494
if (dj.push(time)) {
85-
ngrp -= NGRP_STEP;
86-
std::cout << "Using " << nreg << " registers can have " << ngrp
87-
<< " concurrent single-thread workgroups" << std::endl;
88-
return ngrp;
95+
nreg -= NREG_STEP;
96+
nreg_max = nreg;
97+
break;
8998
}
9099
}
91-
std::cout
92-
<< "Unable to conclude a maximum number of concurrent single-thread workgroups when "
93-
<< nreg << " registers are occupied" << std::endl;
94-
return (uint32_t)1;
95-
};
96-
97-
uint32_t ngrp_full, ngrp_half;
98-
ngrp_full = find_ngrp_by_nreg(nreg_max);
99-
ngrp_half = find_ngrp_by_nreg(nreg_max / 2);
100-
101-
std::string reg_ty;
100+
if (nreg >= NREG_MAX) {
101+
std::cout << "Unable to conclude a maximal register count" << std::endl;
102+
nreg_max = NREG_STEP;
103+
} else {
104+
std::cout << nreg_max << " registers are available at most" << std::endl;
105+
}
102106

103-
if (ngrp_full * 1.5 < ngrp_half) {
104-
std::cout << "All physical threads in an sm share " << nreg_max
105-
<< " registers" << std::endl;
106-
reg_ty = "Pooled";
107+
auto find_ngrp_by_nreg = [&](const uint32_t nreg) {
108+
DtJumpFinder<5> dj(COMPENSATE, THRESHOLD);
109+
for (auto ngrp = NGRP_MIN; ngrp <= NGRP_MAX; ngrp += NGRP_STEP) {
110+
auto time = bench(ngrp, nreg);
111+
std::cout << "Testing occupation (nreg=" << nreg << "); ngrp=" << ngrp
112+
<< ", time=" << time << " us" << std::endl;
113+
114+
if (dj.push(time)) {
115+
ngrp -= NGRP_STEP;
116+
std::cout << "Using " << nreg << " registers can have " << ngrp
117+
<< " concurrent single-thread workgroups" << std::endl;
118+
return ngrp;
119+
}
120+
}
121+
std::cout
122+
<< "Unable to conclude a maximum number of concurrent single-thread workgroups when "
123+
<< nreg << " registers are occupied" << std::endl;
124+
return (uint32_t)1;
125+
};
126+
127+
uint32_t ngrp_full, ngrp_half;
128+
ngrp_full = find_ngrp_by_nreg(nreg_max);
129+
ngrp_half = find_ngrp_by_nreg(nreg_max / 2);
130+
131+
std::string reg_ty;
132+
133+
if (ngrp_full * 1.5 < ngrp_half) {
134+
std::cout << "All physical threads in an sm share " << nreg_max
135+
<< " registers" << std::endl;
136+
reg_ty = "Pooled";
137+
138+
} else {
139+
std::cout << "Each physical thread has " << nreg_max << " registers"
140+
<< std::endl;
141+
reg_ty = "Dedicated";
142+
}
107143

108-
} else {
109-
std::cout << "Each physical thread has " << nreg_max << " registers"
144+
std::cout << std::endl << std::endl;
145+
std::cout << "NITER," << NITER << std::endl;
146+
std::cout << "Max registers," << nreg_max << std::endl;
147+
std::cout << "Concurrent full single thread workgroups," << ngrp_full
148+
<< std::endl;
149+
std::cout << "Concurrent half single thread workgroups," << ngrp_half
110150
<< std::endl;
111-
reg_ty = "Dedicated";
151+
std::cout << "Register type," << reg_ty << std::endl;
112152
}
113-
114-
std::cout << "\n\nNITER," << NITER << std::endl;
115-
std::cout << "Max registers," << nreg_max << std::endl;
116-
std::cout << "Concurrent full single thread workgroups," << ngrp_full
117-
<< std::endl;
118-
std::cout << "Concurrent half single thread workgroups," << ngrp_half
119-
<< std::endl;
120-
std::cout << "Register type," << reg_ty << std::endl;
121-
}
153+
};
122154

123155
int main(int argc, const char** argv) {
124-
context()->initialize_querypool();
125-
126-
reg_count();
156+
App app;
127157

158+
app.reg_count();
128159
return 0;
129160
}

0 commit comments

Comments
 (0)