Add FunctionNode.write_timestamps to allow writing timestamps before/after dispatch

Author: skallweitNV

slangpy/core/function.py

@@ -10,7 +10,15 @@
 )
 
 from slangpy.reflection import SlangFunction, SlangType
-from slangpy import CommandEncoder, TypeConformance, uint3, Logger, NativeHandle, NativeHandleType
+from slangpy import (
+    CommandEncoder,
+    QueryPool,
+    TypeConformance,
+    uint3,
+    Logger,
+    NativeHandle,
+    NativeHandleType,
+)
 from slangpy.slangpy import Shape
 from slangpy.bindings.typeregistry import PYTHON_SIGNATURES
 
@@ -139,6 +147,12 @@ def cuda_stream(self, stream: NativeHandle) -> "FunctionNode":
         """
         return FunctionNodeCUDAStream(self, stream)
 
+    def write_timestamps(self, write_timestamps: tuple[QueryPool, int, int]) -> "FunctionNode":
+        """
+        Specify a query pool and and a before/after query index to write timestamps before/after the dispatch.
+        """
+        return FunctionNodeWriteTimestamps(self, write_timestamps)
+
     def constants(self, constants: dict[str, Any]):
         """
         Specify link time constants that should be set when the function is compiled. These are
@@ -427,6 +441,21 @@ def _populate_build_info(self, info: FunctionBuildInfo):
         info.options["cuda_stream"] = self.stream
 
 
+class FunctionNodeWriteTimestamps(FunctionNode):
+    def __init__(
+        self, parent: NativeFunctionNode, write_timestamps: tuple[QueryPool, int, int]
+    ) -> None:
+        super().__init__(parent, FunctionNodeType.write_timestamps, write_timestamps)
+        self.slangpy_signature = str(write_timestamps)
+
+    @property
+    def write_timestamps(self):
+        return cast(tuple[QueryPool, int, int], self._native_data)
+
+    def _populate_build_info(self, info: FunctionBuildInfo):
+        info.options["write_timestamps"] = self.write_timestamps
+
+
 class FunctionNodeConstants(FunctionNode):
     def __init__(self, parent: NativeFunctionNode, constants: dict[str, Any]) -> None:
         super().__init__(parent, FunctionNodeType.kernelgen, constants)

src/slangpy_ext/utils/slangpy.cpp

@@ -10,6 +10,7 @@
 #include "sgl/utils/slangpy.h"
 #include "sgl/device/device.h"
 #include "sgl/device/kernel.h"
+#include "sgl/device/query.h"
 #include "sgl/device/command.h"
 #include "sgl/stl/bit.h" // Replace with <bit> when available on all platforms.
 
@@ -643,10 +644,25 @@ nb::object NativeCallData::exec(
 
     if (command_encoder == nullptr) {
         // If we are not appending to a command encoder, we can dispatch directly.
-        m_kernel->dispatch(uint3(total_threads, 1, 1), bind_vars, CommandQueueType::graphics, cuda_stream);
+        m_kernel->dispatch(
+            uint3(total_threads, 1, 1),
+            bind_vars,
+            CommandQueueType::graphics,
+            cuda_stream,
+            opts->get_query_pool(),
+            opts->get_query_before_index(),
+            opts->get_query_after_index()
+        );
     } else {
         // If we are appending to a command encoder, we need to use the command encoder.
-        m_kernel->dispatch(uint3(total_threads, 1, 1), bind_vars, command_encoder);
+        m_kernel->dispatch(
+            uint3(total_threads, 1, 1),
+            bind_vars,
+            command_encoder,
+            opts->get_query_pool(),
+            opts->get_query_before_index(),
+            opts->get_query_after_index()
+        );
     }
 
     // If command_buffer is not null, return early.

src/slangpy_ext/utils/slangpy.h

@@ -598,10 +598,25 @@ class NativeCallRuntimeOptions : Object {
     /// Set the CUDA stream.
     void set_cuda_stream(NativeHandle cuda_stream) { m_cuda_stream = cuda_stream; }
 
+    QueryPool* get_query_pool() const { return m_query_pool.get(); }
+
+    void set_query_pool(QueryPool* query_pool) { m_query_pool = ref(query_pool); }
+
+    uint32_t get_query_before_index() const { return m_query_before_index; }
+
+    void set_query_before_index(uint32_t query_before_index) { m_query_before_index = query_before_index; }
+
+    uint32_t get_query_after_index() const { return m_query_after_index; }
+
+    void set_query_after_index(uint32_t query_after_index) { m_query_after_index = query_after_index; }
+
 private:
     nb::list m_uniforms;
     nb::object m_this{nb::none()};
     NativeHandle m_cuda_stream;
+    ref<QueryPool> m_query_pool;
+    uint32_t m_query_before_index{0};
+    uint32_t m_query_after_index{0};
 };
 
 /// Defines the common logging functions for a given log level.

src/slangpy_ext/utils/slangpyfunction.h

@@ -14,19 +14,21 @@
 
 #include "sgl/device/fwd.h"
 #include "sgl/device/resource.h"
+#include "sgl/device/query.h"
 
 #include "utils/slangpy.h"
 
 namespace sgl::slangpy {
 
-enum class FunctionNodeType { unknown, uniforms, kernelgen, this_, cuda_stream };
+enum class FunctionNodeType { unknown, uniforms, kernelgen, this_, cuda_stream, write_timestamps };
 SGL_ENUM_INFO(
     FunctionNodeType,
     {{FunctionNodeType::unknown, "unknown"},
      {FunctionNodeType::uniforms, "uniforms"},
      {FunctionNodeType::kernelgen, "kernelgen"},
      {FunctionNodeType::this_, "this"},
-     {FunctionNodeType::cuda_stream, "cuda_stream"}}
+     {FunctionNodeType::cuda_stream, "cuda_stream"},
+     {FunctionNodeType::write_timestamps, "write_timestamps"}}
 );
 SGL_ENUM_REGISTER(FunctionNodeType);
 
@@ -72,6 +74,13 @@ class NativeFunctionNode : NativeObject {
         case sgl::slangpy::FunctionNodeType::cuda_stream:
             options->set_cuda_stream(nb::cast<NativeHandle>(m_data));
             break;
+        case sgl::slangpy::FunctionNodeType::write_timestamps: {
+            nb::tuple t = nb::cast<nb::tuple>(m_data);
+            options->set_query_pool(nb::cast<QueryPool*>(t[0]));
+            options->set_query_before_index(nb::cast<uint32_t>(t[1]));
+            options->set_query_after_index(nb::cast<uint32_t>(t[2]));
+            break;
+        }
         default:
             break;
         }