Skip to content

[GPU] improve writing about enable_kernels_reuse #36912

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
isanghao wants to merge 2 commits into
base: master
Choose a base branch
from
+24 −8
Open

[GPU] improve writing about enable_kernels_reuse #36912

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
dae0ba4
[GPU] improve writing about enable_kernels_reuse
isanghao Jul 16, 2026
1df5652
add architecture information
isanghao Jul 16, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions src/inference/include/openvino/runtime/intel_gpu/properties.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -146,10 +146,10 @@ static constexpr Property<bool> enable_sdpa_optimization{"GPU_ENABLE_SDPA_OPTIMI
static constexpr Property<bool> enable_lora_operation{"GPU_ENABLE_LORA_OPERATION"};

/**
* @brief Turning on this property enables kernels reuse between implementations, resulting in a lower memory footprint.
* However, as a drawback, OpenCL set_arguments() call will be made more often, resulting in higher host pressure
* and slower execution in some host-bottleneck cases.
* This property is available only for single-stream scenarios and will be ignored in other cases.
* @brief Enables kernel reuse across implementations to reduce memory footprint.
* As a trade-off, OpenCL set_arguments() is invoked more frequently, which increases host CPU usage
* and may slow down execution.
* This property applies only to single-stream scenarios and is ignored otherwise.
* @ingroup ov_runtime_ocl_gpu_prop_cpp_api
*/
static constexpr Property<bool> enable_kernels_reuse{"GPU_ENABLE_KERNELS_REUSE"};
Expand Down
2 changes: 1 addition & 1 deletion src/plugins/intel_gpu/include/intel_gpu/runtime/options.inl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,7 @@ OV_CONFIG_RELEASE_OPTION(ov, cache_encryption_callbacks, ov::EncryptionCallbacks
OV_CONFIG_RELEASE_OPTION(ov::hint, dynamic_quantization_group_size, 0, "Dynamic quantization group size")
OV_CONFIG_RELEASE_OPTION(ov::intel_gpu::hint, dynamic_quantization_group_size_max, UINT64_MAX, "Maximum dynamic quantization group size. When group_size is set as a higher value than this number, dynamic quantization will be turned off")
OV_CONFIG_RELEASE_OPTION(ov::hint, kv_cache_precision, ov::element::dynamic, "")
OV_CONFIG_RELEASE_OPTION(ov::intel_gpu::hint, enable_kernels_reuse, false, "")
OV_CONFIG_RELEASE_OPTION(ov::intel_gpu::hint, enable_kernels_reuse, false, "Enables kernel reuse across implementations to reduce memory footprint.")
OV_CONFIG_RELEASE_OPTION(ov, weights_path, "", "Path to the model weights file used for weightless caching")
OV_CONFIG_RELEASE_OPTION(ov::hint, activations_scale_factor, -1.0f, "Scalar floating point value that is used for runtime activation tensor scaling with fp16 inference precision")
OV_CONFIG_RELEASE_OPTION(ov::internal, enable_lp_transformations, false, "Enable/Disable Low precision transformations set")
Expand Down
22 changes: 19 additions & 3 deletions src/plugins/intel_gpu/src/runtime/ocl/ocl_device.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -310,9 +310,6 @@ device_info init_device_info(const cl::Device& device, const cl::Context& contex
info.gfx_ver.major > 12 || (info.gfx_ver.major == 12 && info.gfx_ver.minor >= 70)) {
info.has_separate_cache = true;
}
GPU_DEBUG_INFO << "GPU version: "
<< static_cast<int>(info.gfx_ver.major) << "." << static_cast<int>(info.gfx_ver.minor) << "." << static_cast<int>(info.gfx_ver.revision)
<< (info.has_separate_cache ? " with separate cache" : "") << std::endl;
} else if (nv_device_attr_supported) {
info.gfx_ver = {static_cast<uint16_t>(device.getInfo<CL_DEVICE_COMPUTE_CAPABILITY_MAJOR_NV>()),
static_cast<uint8_t>(device.getInfo<CL_DEVICE_COMPUTE_CAPABILITY_MINOR_NV>()),
Expand Down Expand Up @@ -377,6 +374,25 @@ device_info init_device_info(const cl::Device& device, const cl::Context& contex
info.arch = gpu_arch::unknown;
#endif // ENABLE_ONEDNN_FOR_GPU

auto arch_to_string = [](gpu_arch arch) -> const char* {
switch (arch) {
case gpu_arch::unknown: return "unknown";
case gpu_arch::gen9: return "gen9";
case gpu_arch::gen11: return "gen11";
case gpu_arch::xe_lp: return "xe_lp";
case gpu_arch::xe_hp: return "xe_hp";
case gpu_arch::xe_hpg: return "xe_hpg";
case gpu_arch::xe_hpc: return "xe_hpc";
case gpu_arch::xe2: return "xe2";
case gpu_arch::xe3: return "xe3";
case gpu_arch::xe3p: return "xe3p";
}
return "unknown";
};
GPU_DEBUG_INFO << "GPU architecture: " << arch_to_string(info.arch) << ", version: "
<< static_cast<int>(info.gfx_ver.major) << "." << static_cast<int>(info.gfx_ver.minor) << "." << static_cast<int>(info.gfx_ver.revision)
<< (info.has_separate_cache ? " with separate cache" : "") << std::endl;

return info;
}

Expand Down
Loading