intern/cycles/device/device.h

@@ -182,7 +182,7 @@ class Device {
   {
   }
 
-  /* Return true if device is ready for rendering, or report status if not. */
+  /* Report status and return true if device is ready for rendering. */
   virtual bool is_ready(string & /*status*/) const
   {
     return true;

intern/cycles/device/metal/device_impl.h

@@ -74,6 +74,7 @@ class MetalDevice : public Device {
   id<MTLBuffer> texture_bindings_3d = nil;
   std::vector<id<MTLTexture>> texture_slot_map;
 
+  bool have_mtlbuffer_textures = false;
   id<MTLArgumentEncoder> mtlTextureBufferArgEncoder = nil;
   id<MTLBuffer> texture_buffers = nil;
   std::vector<id<MTLBuffer>> texture_buffer_slot_map;

intern/cycles/device/metal/device_impl.mm

@@ -322,6 +322,11 @@ void MetalDevice::make_source(MetalPipelineType pso_type, const uint kernel_feat
     global_defines += "#define __KERNEL_LOCAL_ATOMIC_SORT__\n";
   }
 
+  if (pso_type == PSO_GENERIC || have_mtlbuffer_textures) {
+    /* Only enable MTLBuffer textures if needed as they add a small overhead. */ 
+    global_defines += "#define __KERNEL_METAL_BUFFER_TEXTURES__\n";
+  }
+
   if (use_metalrt) {
     global_defines += "#define __METALRT__\n";
     if (motion_blur) {
@@ -897,6 +902,17 @@ bool MetalDevice::is_ready(string &status) const
                            DEVICE_KERNEL_NUM);
     return false;
   }
+
+  if (int num_requests = MetalDeviceKernels::num_incomplete_specialization_requests()) {
+    status = string_printf("%d kernels to optimize", num_requests);
+  }
+  else if (kernel_specialization_level == PSO_SPECIALIZED_INTERSECT) {
+    status = "Using optimized intersection kernels";
+  }
+  else if (kernel_specialization_level == PSO_SPECIALIZED_SHADE) {
+    status = "Using optimized kernels";
+  }
+  
   metal_printf("MetalDevice::is_ready(...) --> true\n");
   return true;
 }
@@ -933,7 +949,7 @@ void MetalDevice::optimize_for_scene(Scene *scene)
   }
 
   if (specialize_in_background) {
-    if (!MetalDeviceKernels::any_specialization_happening_now()) {
+    if (MetalDeviceKernels::num_incomplete_specialization_requests() == 0) {
       dispatch_async(dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0),
                      specialize_kernels_fn);
     }
@@ -1103,7 +1119,7 @@ void MetalDevice::tex_alloc(device_texture &mem)
   if (mem.data_width > 16384 || mem.data_height > 16384) {
     use_tex = false;
   }
-  if (auto str = getenv("USE_TEX")) {
+  if (auto str = getenv("CYCLES_METAL_FORCE_MTLTEXTURE")) {
     use_tex = atoi(str);
   }
   if (use_tex) {
@@ -1189,12 +1205,12 @@ void MetalDevice::tex_alloc(device_texture &mem)
     mmem->mtlBuffer = mtlBuffer;
   }
   else {
+    have_mtlbuffer_textures = true;
     generic_alloc(mem);
     
     std::lock_guard<std::recursive_mutex> lock(metal_mem_map_mutex);
     mtlBuffer = metal_mem_map[&mem]->mtlBuffer;
     mem.device_pointer = (device_ptr)mtlBuffer;
-    //mem.host_pointer = 0;
   }
 
   /* Resize once */

intern/cycles/device/metal/kernel.h

@@ -104,7 +104,7 @@ struct MetalKernelPipeline {
 /* Cache of Metal kernels for each DeviceKernel. */
 namespace MetalDeviceKernels {
 
-bool any_specialization_happening_now();
+int num_incomplete_specialization_requests();
 int get_loaded_kernel_count(MetalDevice const *device, MetalPipelineType pso_type);
 bool should_load_kernels(MetalDevice const *device, MetalPipelineType pso_type);
 bool load(MetalDevice *device, MetalPipelineType pso_type);

intern/cycles/device/metal/kernel.mm

@@ -857,16 +857,15 @@ void MetalDeviceKernels::wait_for_all()
   }
 }
 
-bool MetalDeviceKernels::any_specialization_happening_now()
+int MetalDeviceKernels::num_incomplete_specialization_requests()
 {
   /* Return true if any ShaderCaches have ongoing specialization requests (typically there will be
    * only 1). */
+  int total = 0;
   for (int i = 0; i < g_shaderCacheCount; i++) {
-    if (g_shaderCache[i].second->incomplete_specialization_requests > 0) {
-      return true;
-    }
+    total += g_shaderCache[i].second->incomplete_specialization_requests;
   }
-  return false;
+  return total;
 }
 
 int MetalDeviceKernels::get_loaded_kernel_count(MetalDevice const *device,

intern/cycles/kernel/device/metal/context_begin.h

@@ -18,10 +18,7 @@ class MetalKernelContext {
     : launch_params_metal(_launch_params_metal)
     {}
 
-    /* texture fetch adapter functions */
-    typedef uint64_t ccl_gpu_tex_object_2D;
-    typedef uint64_t ccl_gpu_tex_object_3D;
-
+#ifdef __KERNEL_METAL_BUFFER_TEXTURES__
     template<typename T> ccl_device_forceinline T tex_fetch(device void* data, int64_t index)
     {
       return reinterpret_cast<ccl_global T *>(data)[index];
@@ -162,13 +159,18 @@ class MetalKernelContext {
         u[3] = (1.0f / 6.0f) * t * t * t; \
       } \
       (void)0
-  
+
+#endif /* __KERNEL_METAL_BUFFER_TEXTURES__ */
+
     ccl_device float4 kernel_tex_image_interp(KernelGlobals kg, int tex_id, float x, float y)
     {
       device const TextureInfo &info = kernel_data_fetch(texture_info, tex_id);
       
       const uint tid(info.data);
       const uint sid(info.data >> 32);
+#ifndef __KERNEL_METAL_BUFFER_TEXTURES__
+      return metal_ancillaries->textures_2d[tid].tex.sample(metal_samplers[sid], make_float2(x,y));
+#else
       if (sid < 256) {
         return metal_ancillaries->textures_2d[tid].tex.sample(metal_samplers[sid], make_float2(x,y));
       }
@@ -216,6 +218,7 @@ class MetalKernelContext {
         }
         return r;
       }
+ #endif
  #endif
     }
     
@@ -225,6 +228,10 @@ class MetalKernelContext {
       
       const uint tid(info.data);
       const uint sid(info.data >> 32);
+
+#ifndef __KERNEL_METAL_BUFFER_TEXTURES__
+      return metal_ancillaries->textures_3d[tid].tex.sample(metal_samplers[sid], P);
+#else
       if (sid < 256) {
         return metal_ancillaries->textures_3d[tid].tex.sample(metal_samplers[sid], P);
       }
@@ -323,6 +330,7 @@ class MetalKernelContext {
         return r;
       }
   #endif
+  #endif
   }
 
   // clang-format on

intern/cycles/session/session.cpp

@@ -706,6 +706,12 @@ void Session::update_status_time(bool show_pause, bool show_done)
                               string_printf("Sample %d/%d", current_sample, num_samples));
   }
 
+  /* Append any device-specific status (such as background kernel optimization) */
+  string device_status;
+  if (device->is_ready(device_status) && !device_status.empty()) {
+    substatus += string_printf(" (%s)", device_status.c_str());
+  }
+
   /* TODO(sergey): Denoising status from the path trace. */
 
   if (show_pause) {