intern/cycles/device/metal/device_impl.h

@@ -29,7 +29,8 @@ class MetalDevice : public Device {
   id<MTLArgumentEncoder> mtlAncillaryArgEncoder =
       nil; /* encoder used for fetching device pointers from MTLBuffers */
   string source[PSO_NUM];
-  string source_md5[PSO_NUM];
+  string kernels_md5[PSO_NUM];
+  string global_defines_md5[PSO_NUM];
 
   bool capture_enabled = false;
 
@@ -115,6 +116,8 @@ class MetalDevice : public Device {
 
   bool use_local_atomic_sort() const;
 
+  string preprocess_source(MetalPipelineType pso_type, const uint kernel_features, string* source = nullptr);
+
   bool make_source_and_check_if_compile_needed(MetalPipelineType pso_type);
 
   void make_source(MetalPipelineType pso_type, const uint kernel_features);

intern/cycles/device/metal/device_impl.mm

@@ -310,7 +310,7 @@ bool MetalDevice::use_local_atomic_sort() const
   return DebugFlags().metal.use_local_atomic_sort;
 }
 
-void MetalDevice::make_source(MetalPipelineType pso_type, const uint kernel_features)
+string MetalDevice::preprocess_source(MetalPipelineType pso_type, const uint kernel_features, string* source)
 {
   string global_defines;
   if (use_adaptive_compilation()) {
@@ -353,6 +353,61 @@ void MetalDevice::make_source(MetalPipelineType pso_type, const uint kernel_feat
   NSOperatingSystemVersion macos_ver = [processInfo operatingSystemVersion];
   global_defines += "#define __KERNEL_METAL_MACOS__ " + to_string(macos_ver.majorVersion) + "\n";
 
+  /* Replace specific KernelData "dot" dereferences with a Metal function_constant identifier of
+   * the same character length. Build a string of all active constant values which is then hashed
+   * in order to identify the PSO.
+   */
+  if (pso_type != PSO_GENERIC) {
+    if (source) {
+      const double starttime = time_dt();
+
+#  define KERNEL_STRUCT_BEGIN(name, parent) \
+      string_replace_same_length(*source, "kernel_data." #parent ".", "kernel_data_" #parent "_");
+
+      bool next_member_is_specialized = true;
+
+#  define KERNEL_STRUCT_MEMBER_DONT_SPECIALIZE next_member_is_specialized = false;
+
+#  define KERNEL_STRUCT_MEMBER(parent, _type, name) \
+      if (!next_member_is_specialized) { \
+        string_replace( \
+            *source, "kernel_data_" #parent "_" #name, "kernel_data." #parent ".__unused_" #name); \
+        next_member_is_specialized = true; \
+      }
+
+#  include "kernel/data_template.h"
+
+#  undef KERNEL_STRUCT_MEMBER
+#  undef KERNEL_STRUCT_MEMBER_DONT_SPECIALIZE
+#  undef KERNEL_STRUCT_BEGIN
+
+      metal_printf("KernelData patching took %.1f ms\n", (time_dt() - starttime) * 1000.0);
+    }
+
+    /* Opt in to all of available specializations. This can be made more granular for the
+     * PSO_SPECIALIZED_INTERSECT case in order to minimize the number of specialization requests,
+     * but the overhead should be negligible as these are very quick to (re)build and aren't
+     * serialized to disk via MTLBinaryArchives.
+     */
+    global_defines += "#define __KERNEL_USE_DATA_CONSTANTS__\n";
+  }
+
+#  if 0
+  metal_printf("================\n%s================\n",
+               global_defines.c_str());
+#  endif
+
+  if (source) {
+    *source = global_defines + *source;
+  }
+
+  MD5Hash md5;
+  md5.append(global_defines);
+  return md5.get_hex();
+}
+
+void MetalDevice::make_source(MetalPipelineType pso_type, const uint kernel_features)
+{
   string &source = this->source[pso_type];
   source = "\n#include \"kernel/device/metal/kernel.metal\"\n";
   source = path_source_replace_includes(source, path_get("source"));
@@ -361,62 +416,7 @@ void MetalDevice::make_source(MetalPipelineType pso_type, const uint kernel_feat
    * With Metal function constants we can generate a single variant of the kernel source which can
    * be repeatedly respecialized.
    */
-  string baked_constants;
-
-  /* Replace specific KernelData "dot" dereferences with a Metal function_constant identifier of
-   * the same character length. Build a string of all active constant values which is then hashed
-   * in order to identify the PSO.
-   */
-  if (pso_type != PSO_GENERIC) {
-    const double starttime = time_dt();
-
-#  define KERNEL_STRUCT_BEGIN(name, parent) \
-    string_replace_same_length(source, "kernel_data." #parent ".", "kernel_data_" #parent "_");
-
-    bool next_member_is_specialized = true;
-
-#  define KERNEL_STRUCT_MEMBER_DONT_SPECIALIZE next_member_is_specialized = false;
-
-    /* Add constants to md5 so that 'get_best_pipeline' is able to return a suitable match. */
-#  define KERNEL_STRUCT_MEMBER(parent, _type, name) \
-    if (next_member_is_specialized) { \
-      baked_constants += string(#parent "." #name "=") + \
-                         to_string(_type(launch_params.data.parent.name)) + "\n"; \
-    } \
-    else { \
-      string_replace( \
-          source, "kernel_data_" #parent "_" #name, "kernel_data." #parent ".__unused_" #name); \
-      next_member_is_specialized = true; \
-    }
-
-#  include "kernel/data_template.h"
-
-    /* Opt in to all of available specializations. This can be made more granular for the
-     * PSO_SPECIALIZED_INTERSECT case in order to minimize the number of specialization requests,
-     * but the overhead should be negligible as these are very quick to (re)build and aren't
-     * serialized to disk via MTLBinaryArchives.
-     */
-    global_defines += "#define __KERNEL_USE_DATA_CONSTANTS__\n";
-
-    metal_printf("KernelData patching took %.1f ms\n", (time_dt() - starttime) * 1000.0);
-  }
-
-  source = global_defines + source;
-#  if 0
-  metal_printf("================\n%s================\n\%s================\n",
-               global_defines.c_str(),
-               baked_constants.c_str());
-#  endif
-
-  /* Generate an MD5 from the source and include any baked constants. This is used when caching
-   * PSOs. */
-  MD5Hash md5;
-  md5.append(baked_constants);
-  md5.append(source);
-  if (use_metalrt) {
-    md5.append(std::to_string(kernel_features & METALRT_FEATURE_MASK));
-  }
-  source_md5[pso_type] = md5.get_hex();
+  global_defines_md5[pso_type] = preprocess_source(pso_type, kernel_features, &source);
 }
 
 bool MetalDevice::load_kernels(const uint _kernel_features)
@@ -450,9 +450,45 @@ bool MetalDevice::load_kernels(const uint _kernel_features)
 
 bool MetalDevice::make_source_and_check_if_compile_needed(MetalPipelineType pso_type)
 {
-  if (this->source[pso_type].empty()) {
+  string defines_md5 = preprocess_source(pso_type, kernel_features);
+
+  /* Rebuild the source string if the injected block of #defines has changed. */
+  if (global_defines_md5[pso_type] != defines_md5) {
     make_source(pso_type, kernel_features);
   }
+
+  string constant_values;
+  if (pso_type != PSO_GENERIC) {
+    bool next_member_is_specialized = true;
+
+#  define KERNEL_STRUCT_MEMBER_DONT_SPECIALIZE next_member_is_specialized = false;
+
+    /* Add specialization constants to md5 so that 'get_best_pipeline' is able to return a suitable match. */
+#  define KERNEL_STRUCT_MEMBER(parent, _type, name) \
+    if (next_member_is_specialized) { \
+      constant_values += string(#parent "." #name "=") + \
+                         to_string(_type(launch_params.data.parent.name)) + "\n"; \
+    } \
+    else { \
+      next_member_is_specialized = true; \
+    }
+
+#  include "kernel/data_template.h"
+
+#  undef KERNEL_STRUCT_MEMBER
+#  undef KERNEL_STRUCT_MEMBER_DONT_SPECIALIZE
+
+#  if 0
+    metal_printf("================\n%s================\n",
+                constant_values.c_str());
+#  endif
+  }
+
+  MD5Hash md5;
+  md5.append(constant_values);
+  md5.append(source[pso_type]);
+  kernels_md5[pso_type] = md5.get_hex();
+
   return MetalDeviceKernels::should_load_kernels(this, pso_type);
 }
 

intern/cycles/device/metal/kernel.h

@@ -76,7 +76,7 @@ struct MetalKernelPipeline {
 
   id<MTLLibrary> mtlLibrary = nil;
   MetalPipelineType pso_type;
-  string source_md5;
+  string kernels_md5;
   size_t usage_count = 0;
 
   KernelData kernel_data_;

intern/cycles/device/metal/kernel.mm

@@ -292,7 +292,7 @@ bool ShaderCache::should_load_kernel(DeviceKernel device_kernel,
     /* check whether the kernel has already been requested / cached */
     thread_scoped_lock lock(cache_mutex);
     for (auto &pipeline : pipelines[device_kernel]) {
-      if (pipeline->source_md5 == device->source_md5[pso_type]) {
+      if (pipeline->kernels_md5 == device->kernels_md5[pso_type]) {
         return false;
       }
     }
@@ -332,7 +332,7 @@ void ShaderCache::load_kernel(DeviceKernel device_kernel,
   memcpy(&pipeline->kernel_data_, &device->launch_params.data, sizeof(pipeline->kernel_data_));
   pipeline->pso_type = pso_type;
   pipeline->mtlDevice = mtlDevice;
-  pipeline->source_md5 = device->source_md5[pso_type];
+  pipeline->kernels_md5 = device->kernels_md5[pso_type];
   pipeline->mtlLibrary = device->mtlLibrary[pso_type];
   pipeline->device_kernel = device_kernel;
   pipeline->threads_per_threadgroup = device->max_threads_per_threadgroup;
@@ -392,8 +392,8 @@ MetalKernelPipeline *ShaderCache::get_best_pipeline(DeviceKernel kernel, const M
         }
 
         if (pipeline->pso_type != PSO_GENERIC) {
-          if (pipeline->source_md5 == device->source_md5[PSO_SPECIALIZED_INTERSECT] ||
-              pipeline->source_md5 == device->source_md5[PSO_SPECIALIZED_SHADE]) {
+          if (pipeline->kernels_md5 == device->kernels_md5[PSO_SPECIALIZED_INTERSECT] ||
+              pipeline->kernels_md5 == device->kernels_md5[PSO_SPECIALIZED_SHADE]) {
             best_pipeline = pipeline.get();
           }
         }
@@ -674,7 +674,7 @@ void MetalKernelPipeline::compile()
     NSProcessInfo *processInfo = [NSProcessInfo processInfo];
     string osVersion = [[processInfo operatingSystemVersionString] UTF8String];
     MD5Hash local_md5;
-    local_md5.append(source_md5);
+    local_md5.append(kernels_md5);
     local_md5.append(osVersion);
     local_md5.append((uint8_t *)&this->threads_per_threadgroup,
                      sizeof(this->threads_per_threadgroup));

source/blender/editors/space_graph/graph_buttons.c

@@ -942,6 +942,7 @@ static void graph_panel_drivers_header(const bContext *C, Panel *panel)
   }
 
   graph_draw_driven_property_enabled_btn(panel->layout, ale->id, fcu, IFACE_("Driver"));
+  MEM_freeN(ale);
 }
 
 static void graph_draw_driven_property_panel(uiLayout *layout, ID *id, FCurve *fcu)