Fixes #566: Can now get and set kernel properties using an independen…

…t function * both for `kernel_t` and `void *` pointers, * underlying functionality also for `kernel::handle_t` * Typo fix
eyalroz · Jan 14, 2024 · 41a206e · 41a206e
1 parent d752042
commit 41a206e
Show file tree

Hide file tree

Showing 3 changed files with 69 additions and 14 deletions.
diff --git a/src/cuda/api/apriori_compiled_kernel.hpp b/src/cuda/api/apriori_compiled_kernel.hpp
@@ -33,7 +33,6 @@ inline handle_t get_handle(const void *, const char* = nullptr)
 #else
 inline handle_t get_handle(const void *kernel_function_ptr, const char* name = nullptr)
 {
-
 	handle_t handle;
 	auto status = cudaGetFuncBySymbol(&handle, kernel_function_ptr);
 	throw_if_error_lazy(status, "Failed obtaining a CUDA function handle for "
@@ -390,7 +389,7 @@ class apriori_compiled_kernel_t final : public kernel_t {
 			dynamic_shared_memory_per_block,
 			disable_caching_override);
 	}
-#endif // CAN_GET_APRIORI_KERNEL_HANDLE
+#endif // ! CAN_GET_APRIORI_KERNEL_HANDLE
 
 protected: // ctors & dtor
 	apriori_compiled_kernel_t(device::id_t device_id, context::handle_t primary_context_handle,
@@ -448,6 +447,39 @@ inline ::std::string identify(const apriori_compiled_kernel_t& kernel)
 
 } // namespace detail
 
+#if CAN_GET_APRIORI_KERNEL_HANDLE
+inline attribute_value_t get_attribute(const void* function_ptr, attribute_t attribute)
+{
+	auto handle = detail_::get_handle(function_ptr);
+	return kernel::detail_::get_attribute_in_current_context(handle, attribute);
+}
+
+inline void set_attribute(const void* function_ptr, attribute_t attribute, attribute_value_t value)
+{
+	auto handle = detail_::get_handle(function_ptr);
+	return detail_::set_attribute_in_current_context(handle, attribute, value);
+}
+
+inline attribute_value_t get_attribute(
+	const context_t&  context,
+	const void*       function_ptr,
+	attribute_t       attribute)
+{
+	CAW_SET_SCOPE_CONTEXT(context.handle());
+	return get_attribute(function_ptr, attribute);
+}
+
+inline void set_attribute(
+	const context_t&   context,
+	const void*        function_ptr,
+	attribute_t        attribute,
+	attribute_value_t  value)
+{
+	CAW_SET_SCOPE_CONTEXT(context.handle());
+	return set_attribute(function_ptr, attribute, value);
+}
+#endif // CAN_GET_APRIORI_KERNEL_HANDLE
+
 /**
  * @note The returned kernel proxy object will keep the device's primary
  * context active while the kernel exists.

diff --git a/src/cuda/api/kernel.hpp b/src/cuda/api/kernel.hpp
@@ -87,8 +87,23 @@ inline attribute_value_t get_attribute_in_current_context(handle_t handle, attri
 	return attribute_value;
 }
 
+inline void set_attribute_in_current_context(handle_t handle, attribute_t attribute, attribute_value_t value)
+{
+#if CUDA_VERSION >= 9000
+	auto result = cuFuncSetAttribute(handle, static_cast<CUfunction_attribute>(attribute), value);
+	throw_if_error_lazy(result,
+		"Setting CUDA device function attribute " +
+		::std::string(kernel::detail_::attribute_name(attribute)) + " of function at "
+		+ cuda::kernel::detail_::identify(handle) + " to value " + ::std::to_string(value));
+#else
+	throw(cuda::runtime_error {cuda::status::not_yet_implemented});
+#endif
+}
+
 } // namespace detail_
 
+inline attribute_value_t get_attribute(const kernel_t& kernel, attribute_t attribute);
+
 } // namespace kernel
 
 /**
@@ -149,8 +164,7 @@ class kernel_t {
 	VIRTUAL_UNLESS_CAN_GET_APRIORI_KERNEL_HANDLE
 	kernel::attribute_value_t get_attribute(kernel::attribute_t attribute) const
 	{
-		context::current::detail_::scoped_override_t set_context_for_this_context(context_handle_);
-		return kernel::detail_::get_attribute_in_current_context(handle(), attribute);
+		return kernel::get_attribute(*this, attribute);
 	}
 
 	VIRTUAL_UNLESS_CAN_GET_APRIORI_KERNEL_HANDLE
@@ -381,6 +395,24 @@ inline kernel_t wrap(
 	return kernel_t{ device_id, context_id, f, hold_primary_context_refcount_unit };
 }
 
+inline attribute_value_t get_attribute(const kernel_t& kernel, attribute_t attribute)
+{
+	CAW_SET_SCOPE_CONTEXT(kernel.context_handle());
+	return detail_::get_attribute_in_current_context(kernel.handle(), attribute);
+}
+
+inline void set_attribute(const kernel_t& kernel, attribute_t attribute, attribute_value_t value)
+{
+	CAW_SET_SCOPE_CONTEXT(kernel.context_handle());
+	return detail_::set_attribute_in_current_context(kernel.handle(), attribute, value);
+}
+
+inline attribute_value_t set_attribute(const kernel_t& kernel, attribute_t attribute)
+{
+	CAW_SET_SCOPE_CONTEXT(kernel.context_handle());
+	return kernel::detail_::get_attribute_in_current_context(kernel.handle(), attribute);
+}
+
 namespace occupancy {
 
 namespace detail_ {

diff --git a/src/cuda/api/multi_wrapper_impls/kernel.hpp b/src/cuda/api/multi_wrapper_impls/kernel.hpp
@@ -78,16 +78,7 @@ return device::get(device_id_);
 
 inline void kernel_t::set_attribute(kernel::attribute_t attribute, kernel::attribute_value_t value) const
 {
-#if CUDA_VERSION >= 9000
-	context::current::detail_::scoped_override_t set_context_for_this_context(context_handle_);
-	auto result = cuFuncSetAttribute(handle_, static_cast<CUfunction_attribute>(attribute), value);
-	throw_if_error_lazy(result,
-		"Setting CUDA device function attribute " +
-		::std::string(kernel::detail_::attribute_name(attribute)) +
-		" to value " + ::std::to_string(value));
-#else
-	throw(cuda::runtime_error {cuda::status::not_yet_implemented});
-#endif
+	kernel::detail_::set_attribute_in_current_context(handle_, attribute, value);
 }
 
 namespace detail_ {