mirror of
https://github.com/fumiama/base16384-sycl.git
synced 2026-06-08 12:00:36 +08:00
optimize: add xeinfo class & more compl. kernel
This commit is contained in:
源文雨
@@ -5,36 +5,56 @@
|
||||
#include <iostream>
|
||||
#include <sycl/sycl.hpp>
|
||||
|
||||
typedef enum {
|
||||
base16384_errors_code_ok,
|
||||
base16384_errors_code_sync_sycl_exception,
|
||||
base16384_errors_code_std_exception,
|
||||
base16384_errors_code_unknown_exception,
|
||||
} base16384_errors_code_enum_t;
|
||||
|
||||
template <typename T>
|
||||
concept base16384_has_what_concept_t = requires(T t) { t.what(); };
|
||||
|
||||
template <base16384_has_what_concept_t T>
|
||||
static void base16384_print_what(T e, std::string msg) {
|
||||
std::cerr << msg << e.what() << std::endl;
|
||||
}
|
||||
template <typename F>
|
||||
concept base16384_callable_concept_t = requires(F f) { f(); };
|
||||
|
||||
// failed try to exec fn, catch and print .what() when exception is thrown.
|
||||
static base16384_errors_code_enum_t base16384_try_failed(std::function<void(void)> fn) {
|
||||
try {
|
||||
fn();
|
||||
} catch (sycl::exception &e) {
|
||||
base16384_print_what(e, "Caught sync SYCL exception: ");
|
||||
return base16384_errors_code_sync_sycl_exception;
|
||||
} catch (std::exception &e) {
|
||||
base16384_print_what(e, "Caught std exception: ");
|
||||
return base16384_errors_code_std_exception;
|
||||
} catch (...) {
|
||||
std::cerr << "Caught unknown exception." << std::endl;
|
||||
return base16384_errors_code_unknown_exception;
|
||||
}
|
||||
return base16384_errors_code_ok;
|
||||
}
|
||||
namespace base16384 {
|
||||
class errors {
|
||||
private:
|
||||
errors() = default;
|
||||
|
||||
template <base16384_has_what_concept_t T>
|
||||
static void print_what(T e, std::string msg) {
|
||||
std::cerr << msg << e.what() << std::endl;
|
||||
};
|
||||
|
||||
public:
|
||||
errors(const errors &) = delete;
|
||||
errors(errors &&) = delete;
|
||||
errors &operator=(const errors &) = delete;
|
||||
errors &operator=(errors &&) = delete;
|
||||
auto operator<=>(const errors &) const = delete;
|
||||
~errors() noexcept = default;
|
||||
|
||||
typedef enum {
|
||||
code_ok,
|
||||
code_sync_sycl_exception,
|
||||
code_std_exception,
|
||||
code_unknown_exception,
|
||||
} code_enum_t;
|
||||
|
||||
// failed try to exec fn, catch and print .what() when exception is thrown.
|
||||
template <base16384_callable_concept_t F>
|
||||
static code_enum_t try_failed(F &&fn) {
|
||||
try {
|
||||
fn();
|
||||
} catch (sycl::exception &e) {
|
||||
print_what(e, "Caught sync SYCL exception: ");
|
||||
return code_sync_sycl_exception;
|
||||
} catch (std::exception &e) {
|
||||
print_what(e, "Caught std exception: ");
|
||||
return code_std_exception;
|
||||
} catch (...) {
|
||||
std::cerr << "Caught unknown exception." << std::endl;
|
||||
return code_unknown_exception;
|
||||
}
|
||||
return code_ok;
|
||||
};
|
||||
};
|
||||
|
||||
} // namespace base16384
|
||||
|
||||
#endif
|
||||
|
||||
101
include/xeinfo.hpp
Normal file
101
include/xeinfo.hpp
Normal file
@@ -0,0 +1,101 @@
|
||||
#ifndef _XEINFO_HPP_
|
||||
#define _XEINFO_HPP_
|
||||
|
||||
#include <iostream>
|
||||
#include <sstream>
|
||||
#include <string>
|
||||
#include <sycl/sycl.hpp>
|
||||
#include <vector>
|
||||
|
||||
namespace base16384 {
|
||||
|
||||
class xeinfo {
|
||||
private:
|
||||
std::pair<size_t, int> calculate_optimal_sizes() const {
|
||||
size_t best_sub_group_size = sub_group_sizes[0];
|
||||
int best_work_group_size = 0;
|
||||
|
||||
for (auto sg_size : sub_group_sizes) {
|
||||
int wg_size = num_thread_per_xecore * sg_size;
|
||||
if (wg_size <= max_work_group_size && wg_size > best_work_group_size) {
|
||||
best_sub_group_size = sg_size;
|
||||
best_work_group_size = 1 << (31 - __builtin_clz(static_cast<unsigned>(wg_size)));
|
||||
}
|
||||
}
|
||||
|
||||
return {std::move(best_sub_group_size), std::move(best_work_group_size)};
|
||||
}
|
||||
|
||||
public:
|
||||
xeinfo(sycl::device device) noexcept
|
||||
: num_slices(device.get_info<sycl::ext::intel::info::device::gpu_slices>()),
|
||||
num_subslices_per_slice(
|
||||
device.get_info<sycl::ext::intel::info::device::gpu_subslices_per_slice>()),
|
||||
num_eus_per_subslice(
|
||||
device.get_info<sycl::ext::intel::info::device::gpu_eu_count_per_subslice>()),
|
||||
num_threads_per_eu(
|
||||
device.get_info<sycl::ext::intel::info::device::gpu_hw_threads_per_eu>()),
|
||||
global_mem_size(device.get_info<sycl::info::device::global_mem_size>()),
|
||||
local_mem_size(device.get_info<sycl::info::device::local_mem_size>()),
|
||||
max_work_group_size(device.get_info<sycl::info::device::max_work_group_size>()),
|
||||
sub_group_sizes(device.get_info<sycl::info::device::sub_group_sizes>()),
|
||||
num_thread_per_xecore(num_eus_per_subslice * num_threads_per_eu),
|
||||
total_xecores(num_slices * num_subslices_per_slice),
|
||||
total_vector_engines(num_slices * num_subslices_per_slice * num_eus_per_subslice),
|
||||
total_hardware_threads(num_slices * num_subslices_per_slice * num_eus_per_subslice *
|
||||
num_threads_per_eu),
|
||||
optimal_sizes(calculate_optimal_sizes()),
|
||||
sub_group_size(optimal_sizes.first),
|
||||
work_group_size(optimal_sizes.second) {}
|
||||
|
||||
xeinfo(const xeinfo&) = delete;
|
||||
xeinfo(xeinfo&&) = delete;
|
||||
xeinfo& operator=(const xeinfo&) = delete;
|
||||
xeinfo& operator=(xeinfo&&) = delete;
|
||||
auto operator<=>(const xeinfo&) const = delete;
|
||||
~xeinfo() noexcept = default;
|
||||
|
||||
const int num_slices;
|
||||
const int num_subslices_per_slice;
|
||||
const int num_eus_per_subslice;
|
||||
const int num_threads_per_eu;
|
||||
const int global_mem_size;
|
||||
const int local_mem_size;
|
||||
const int max_work_group_size;
|
||||
const std::vector<unsigned long long> sub_group_sizes;
|
||||
const int num_thread_per_xecore;
|
||||
const int total_xecores;
|
||||
const int total_vector_engines;
|
||||
const int total_hardware_threads;
|
||||
|
||||
private:
|
||||
const std::pair<size_t, int> optimal_sizes;
|
||||
|
||||
public:
|
||||
const size_t sub_group_size;
|
||||
const int work_group_size;
|
||||
|
||||
std::string string() const {
|
||||
std::ostringstream builder;
|
||||
builder << "Intel GPU 特性:\n";
|
||||
builder << " XeCore 数量: " << total_xecores << "\n";
|
||||
builder << " 每个 XeCore 的向量引擎数: " << num_eus_per_subslice << "\n";
|
||||
builder << " 向量引擎总数: " << total_vector_engines << "\n";
|
||||
builder << " 每个 XeCore 的硬件线程数: " << num_thread_per_xecore << "\n";
|
||||
builder << " 每个向量引擎的硬件线程数: " << num_threads_per_eu << "\n";
|
||||
builder << " 硬件线程总数: " << total_hardware_threads << "\n";
|
||||
builder << " GPU 内存大小: " << global_mem_size << " 字节\n";
|
||||
builder << " 每个工作组的共享本地内存: " << local_mem_size << " 字节\n";
|
||||
builder << " 最大工作组大小: " << max_work_group_size << "\n";
|
||||
builder << " 支持的子组大小:";
|
||||
for (size_t i = 0; i < sub_group_sizes.size(); i++) builder << " " << sub_group_sizes[i];
|
||||
builder << "\n";
|
||||
builder << " 推荐选择子组大小: " << sub_group_size << "\n";
|
||||
builder << " 100% 占用率工作组大小: " << work_group_size;
|
||||
return builder.str();
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace base16384
|
||||
|
||||
#endif // _XEINFO_HPP_
|
||||
Reference in New Issue
Block a user