5 files changed, 62 insertions, 20 deletions
diff --git a/intern/cycles/device/metal/device_impl.mm b/intern/cycles/device/metal/device_impl.mm
index 6a16d4bb3b4..6f1042b1e55 100644
--- a/intern/cycles/device/metal/device_impl.mm
+++ b/intern/cycles/device/metal/device_impl.mm
@@ -296,9 +296,11 @@ void MetalDevice::make_source(MetalPipelineType pso_type, const uint kernel_feat
   }
 
   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. */
@@ -339,6 +341,14 @@ bool MetalDevice::compile_and_load(MetalPipelineType pso_type)
 
   MTLCompileOptions *options = [[MTLCompileOptions alloc] init];
 
+#  if defined(MAC_OS_VERSION_13_0)
+  if (@available(macos 13.0, *)) {
+    if (device_vendor == METAL_GPU_INTEL) {
+      [options setOptimizationLevel:MTLLibraryOptimizationLevelSize];
+    }
+  }
+#  endif
+
   options.fastMathEnabled = YES;
   if (@available(macOS 12.0, *)) {
     options.languageVersion = MTLLanguageVersion2_4;
diff --git a/intern/cycles/device/metal/kernel.mm b/intern/cycles/device/metal/kernel.mm
index 5e0cb6d18f4..55938d1a03a 100644
--- a/intern/cycles/device/metal/kernel.mm
+++ b/intern/cycles/device/metal/kernel.mm
@@ -162,6 +162,13 @@ bool ShaderCache::should_load_kernel(DeviceKernel device_kernel,
     }
   }
 
+  if (device_kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_MNEE) {
+    if ((device->kernel_features & KERNEL_FEATURE_MNEE) == 0) {
+      /* Skip shade_surface_mnee kernel if the scene doesn't require it. */
+      return false;
+    }
+  }
+
   if (pso_type != PSO_GENERIC) {
     /* Only specialize kernels where it can make an impact. */
     if (device_kernel < DEVICE_KERNEL_INTEGRATOR_INTERSECT_CLOSEST ||
@@ -317,6 +324,12 @@ bool MetalKernelPipeline::should_use_binary_archive() const
       }
     }
 
+    /* Workaround for Intel GPU having issue using Binary Archives */
+    MetalGPUVendor gpu_vendor = MetalInfo::get_device_vendor(mtlDevice);
+    if (gpu_vendor == METAL_GPU_INTEL) {
+      return false;
+    }
+
     if (pso_type == PSO_GENERIC) {
       /* Archive the generic kernels. */
       return true;
diff --git a/intern/cycles/device/metal/queue.h b/intern/cycles/device/metal/queue.h
index fc32740f3e1..2a6c12e2a60 100644
--- a/intern/cycles/device/metal/queue.h
+++ b/intern/cycles/device/metal/queue.h
@@ -23,7 +23,7 @@ class MetalDeviceQueue : public DeviceQueue {
   ~MetalDeviceQueue();
 
   virtual int num_concurrent_states(const size_t) const override;
-  virtual int num_concurrent_busy_states() const override;
+  virtual int num_concurrent_busy_states(const size_t) const override;
   virtual int num_sort_partition_elements() const override;
 
   virtual void init_execution() override;
diff --git a/intern/cycles/device/metal/queue.mm b/intern/cycles/device/metal/queue.mm
index 5ac63a16c61..c0df2c8553f 100644
--- a/intern/cycles/device/metal/queue.mm
+++ b/intern/cycles/device/metal/queue.mm
@@ -264,33 +264,46 @@ MetalDeviceQueue::~MetalDeviceQueue()
   }
 }
 
-int MetalDeviceQueue::num_concurrent_states(const size_t /*state_size*/) const
+int MetalDeviceQueue::num_concurrent_states(const size_t state_size) const
 {
-  /* METAL_WIP */
-  /* TODO: compute automatically. */
-  /* TODO: must have at least num_threads_per_block. */
-  int result = 1048576;
-  if (metal_device_->device_vendor == METAL_GPU_AMD) {
-    result *= 2;
+  static int result = 0;
+  if (result) {
+    return result;
   }
-  else if (metal_device_->device_vendor == METAL_GPU_APPLE) {
+
+  result = 1048576;
+  if (metal_device_->device_vendor == METAL_GPU_APPLE) {
     result *= 4;
+
+    if (MetalInfo::get_apple_gpu_architecture(metal_device_->mtlDevice) == APPLE_M2) {
+      size_t system_ram = system_physical_ram();
+      size_t allocated_so_far = [metal_device_->mtlDevice currentAllocatedSize];
+      size_t max_recommended_working_set = [metal_device_->mtlDevice recommendedMaxWorkingSetSize];
+
+      /* Determine whether we can double the state count, and leave enough GPU-available memory
+       * (1/8 the system RAM or 1GB - whichever is largest). Enlarging the state size allows us to
+       * keep dispatch sizes high and minimize work submission overheads. */
+      size_t min_headroom = std::max(system_ram / 8, size_t(1024 * 1024 * 1024));
+      size_t total_state_size = result * state_size;
+      if (max_recommended_working_set - allocated_so_far - total_state_size * 2 >= min_headroom) {
+        result *= 2;
+        metal_printf("Doubling state count to exploit available RAM (new size = %d)\n", result);
+      }
+    }
+  }
+  else if (metal_device_->device_vendor == METAL_GPU_AMD) {
+    /* METAL_WIP */
+    /* TODO: compute automatically. */
+    /* TODO: must have at least num_threads_per_block. */
+    result *= 2;
   }
   return result;
 }
 
-int MetalDeviceQueue::num_concurrent_busy_states() const
+int MetalDeviceQueue::num_concurrent_busy_states(const size_t state_size) const
 {
-  /* METAL_WIP */
-  /* TODO: compute automatically. */
-  int result = 65536;
-  if (metal_device_->device_vendor == METAL_GPU_AMD) {
-    result *= 2;
-  }
-  else if (metal_device_->device_vendor == METAL_GPU_APPLE) {
-    result *= 4;
-  }
-  return result;
+  /* A 1:4 busy:total ratio gives best rendering performance, independent of total state count. */
+  return num_concurrent_states(state_size) / 4;
 }
 
 int MetalDeviceQueue::num_sort_partition_elements() const
diff --git a/intern/cycles/device/metal/util.mm b/intern/cycles/device/metal/util.mm
index 65c67c400fe..f47638fac15 100644
--- a/intern/cycles/device/metal/util.mm
+++ b/intern/cycles/device/metal/util.mm
@@ -110,6 +110,12 @@ vector<id<MTLDevice>> const &MetalInfo::get_usable_devices()
       usable |= (vendor == METAL_GPU_AMD);
     }
 
+#  if defined(MAC_OS_VERSION_13_0)
+    if (@available(macos 13.0, *)) {
+      usable |= (vendor == METAL_GPU_INTEL);
+    }
+#  endif
+
     if (usable) {
       metal_printf("- %s\n", device_name.c_str());
       [device retain];