intern/cycles/kernel/device/gpu/kernel.h

@@ -279,6 +279,27 @@ ccl_gpu_kernel(GPU_KERNEL_BLOCK_NUM_THREADS, GPU_KERNEL_MAX_REGISTERS)
 }
 ccl_gpu_kernel_postfix
 
+#ifdef __KERNEL_ONEAPI__
+/* The NODE_VOLUME feature adds a lot of code, including the entire NanoVDB library,
+ * causing GPU compilers to spend additional registers while the feature is not used often,
+ * leading to suboptimal execution.
+ * The use of a specialized version below gives a noticeable speed-up for oneAPI execution. */
+ccl_gpu_kernel(GPU_KERNEL_BLOCK_NUM_THREADS, GPU_KERNEL_MAX_REGISTERS)
+    ccl_gpu_kernel_signature(integrator_shade_surface_no_volume,
+                             ccl_global const int *path_index_array,
+                             ccl_global float *render_buffer,
+                             const int work_size)
+{
+  const int global_index = ccl_gpu_global_id_x();
+
+  if (ccl_gpu_kernel_within_bounds(global_index, work_size)) {
+    const int state = (path_index_array) ? path_index_array[global_index] : global_index;
+    ccl_gpu_kernel_call(integrator_shade_surface_no_volume(NULL, state, render_buffer));
+  }
+}
+ccl_gpu_kernel_postfix
+#endif
+
 #if defined(__KERNEL_METAL_APPLE__) && defined(__METALRT__)
 constant int __dummy_constant [[function_constant(Kernel_DummyConstant)]];
 #endif

intern/cycles/kernel/device/oneapi/kernel.cpp

@@ -463,8 +463,18 @@ bool oneapi_enqueue_kernel(KernelContext *kernel_context,
           break;
         }
         case DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE: {
-          oneapi_call(
-              kg, cgh, global_size, local_size, args, oneapi_kernel_integrator_shade_surface);
+          if (kernel_features & KERNEL_FEATURE_NODE_VOLUME) {
+            oneapi_call(
+                kg, cgh, global_size, local_size, args, oneapi_kernel_integrator_shade_surface);
+          }
+          else {
+            oneapi_call(kg,
+                        cgh,
+                        global_size,
+                        local_size,
+                        args,
+                        oneapi_kernel_integrator_shade_surface_no_volume);
+          }
           break;
         }
         case DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE: {

intern/cycles/kernel/integrator/shade_surface.h

@@ -799,6 +799,15 @@ ccl_device_forceinline void integrator_shade_surface(KernelGlobals kg,
   integrator_shade_surface_next_kernel<current_kernel>(kg, state);
 }
 
+#ifdef __KERNEL_ONEAPI__
+ccl_device_forceinline void integrator_shade_surface_no_volume(
+    KernelGlobals kg, IntegratorState state, ccl_global float *ccl_restrict render_buffer)
+{
+  integrator_shade_surface<KERNEL_FEATURE_NODE_MASK_SURFACE & ~KERNEL_FEATURE_NODE_RAYTRACE &
+                           ~KERNEL_FEATURE_NODE_VOLUME>(kg, state, render_buffer);
+}
+#endif
+
 ccl_device_forceinline void integrator_shade_surface_raytrace(
     KernelGlobals kg, IntegratorState state, ccl_global float *ccl_restrict render_buffer)
 {

intern/cycles/scene/shader_nodes.cpp

@@ -1944,6 +1944,19 @@ void PointDensityTextureNode::compile(OSLCompiler &compiler)
   }
 }
 
+int PointDensityTextureNode::get_feature()
+{
+  ShaderOutput *density_out = output("Density");
+  ShaderOutput *color_out = output("Color");
+
+  const bool use_density = !density_out->links.empty();
+  const bool use_color = !color_out->links.empty();
+
+  /* NOTE: A need for NODE_VOLUME feature is conditional and based on
+   * the coresponding logic in ::compile implementation. */
+  return ShaderNode::get_feature() | ((use_density || use_color) ? KERNEL_FEATURE_NODE_VOLUME : 0);
+}
+
 /* Normal */
 
 NODE_DEFINE(NormalNode)

intern/cycles/scene/shader_nodes.h

@@ -346,6 +346,8 @@ class PointDensityTextureNode : public ShaderNode {
  public:
   SHADER_NODE_NO_CLONE_CLASS(PointDensityTextureNode)
 
+  virtual int get_feature();
+
   ~PointDensityTextureNode();
   ShaderNode *clone(ShaderGraph *graph) const;
   void attributes(Shader *shader, AttributeRequestSet *attributes);