[mtl,vk] prepare api for host-visible device local memory

include/nicegraf.h

@@ -1938,11 +1938,15 @@ typedef enum ngf_buffer_storage_type {
   /**
    * \ingroup ngf
    *
-   * Private memory that cannot be accessed by the host directly. The contents of a
+   * Memory that is local to the device (GPU). Depending on device capabilities,
+   * this memory may or may not be accessed by the host directly. See 
+   * \ref ngf_device_capabilities::device_local_memory_is_host_visible.
+   * 
+   * When the device-local memory isn't host-visible, the contents of a
    * buffer backed by this type of memory can only be modified by executing a
    * \ref ngf_cmd_copy_buffer.
    */
-  NGF_BUFFER_STORAGE_PRIVATE
+  NGF_BUFFER_STORAGE_DEVICE_LOCAL
 } ngf_buffer_storage_type;
 
 /**
@@ -2444,6 +2448,16 @@ typedef struct ngf_device_capabilities {
    * This value is derived from \ref texture_depth_sample_counts.
    */
   ngf_sample_count max_supported_texture_depth_sample_count;
+
+  /**
+   * Indicates whether the device-local storage is also host visible.
+   * Examples of cases where this may be supported are iGPU systems with unified memory,
+   * or discrete GPUs with ReBAR enabled.
+   * On systems with this capability, device-local storage can be mapped directly into
+   * the host address space, removing the need for host-visible staging buffers in certain
+   * cases.
+   */
+  bool device_local_memory_is_host_visible;
 } ngf_device_capabilities;
 
 /**

misc/common/mesh-loader.cpp

@@ -108,7 +108,7 @@ mesh load_mesh_from_file(const char* mesh_file_name, ngf_xfer_encoder xfenc) {
    */
   const ngf_buffer_info vertex_data_buffer_info = {
       .size         = vertex_data_staging_buffer_info.size,
-      .storage_type = NGF_BUFFER_STORAGE_PRIVATE,
+      .storage_type = NGF_BUFFER_STORAGE_DEVICE_LOCAL,
       .buffer_usage = NGF_BUFFER_USAGE_VERTEX_BUFFER | NGF_BUFFER_USAGE_XFER_DST,
   };
   NGF_MISC_CHECK_NGF_ERROR(result.vertex_data.initialize(vertex_data_buffer_info));
@@ -122,7 +122,7 @@ mesh load_mesh_from_file(const char* mesh_file_name, ngf_xfer_encoder xfenc) {
   if (result.num_indices > 0) {
     const ngf_buffer_info index_data_buffer_info = {
         .size         = sizeof(uint32_t) * result.num_indices,
-        .storage_type = NGF_BUFFER_STORAGE_PRIVATE,
+        .storage_type = NGF_BUFFER_STORAGE_DEVICE_LOCAL,
         .buffer_usage = NGF_BUFFER_USAGE_INDEX_BUFFER | NGF_BUFFER_USAGE_XFER_DST,
     };
     NGF_MISC_CHECK_NGF_ERROR(result.index_data.initialize(index_data_buffer_info));

samples/06-vertex-attribs/vertex-attribs.cpp

@@ -215,12 +215,12 @@ void* sample_initialize(
    */
   const ngf_buffer_info vertex_buffer_info = {
       .size         = sizeof(vertex_attribs::vertex_data),
-      .storage_type = NGF_BUFFER_STORAGE_PRIVATE,
+      .storage_type = NGF_BUFFER_STORAGE_DEVICE_LOCAL,
       .buffer_usage = NGF_BUFFER_USAGE_VERTEX_BUFFER | NGF_BUFFER_USAGE_XFER_DST,
   };
   const ngf_buffer_info index_buffer_info = {
       .size         = sizeof(vertex_attribs::index_data),
-      .storage_type = NGF_BUFFER_STORAGE_PRIVATE,
+      .storage_type = NGF_BUFFER_STORAGE_DEVICE_LOCAL,
       .buffer_usage = NGF_BUFFER_USAGE_INDEX_BUFFER | NGF_BUFFER_USAGE_XFER_DST,
   };
   const ngf_buffer_info vertex_staging_buffer_info = {
@@ -269,7 +269,7 @@ void* sample_initialize(
    */
   const ngf_buffer_info instance_data_buffer_info = {
       .size         = vertex_attribs::INSTANCE_DATA_SIZE,
-      .storage_type = NGF_BUFFER_STORAGE_PRIVATE,
+      .storage_type = NGF_BUFFER_STORAGE_DEVICE_LOCAL,
       .buffer_usage = NGF_BUFFER_USAGE_TEXEL_BUFFER | NGF_BUFFER_USAGE_XFER_DST,
   };
   const ngf_buffer_info instance_data_staging_buffer_info = {

samples/0b-compute-vertices/compute-vertices.cpp

@@ -143,7 +143,7 @@ void* sample_initialize(
       .buffer_usage = NGF_BUFFER_USAGE_XFER_SRC};
   const ngf_buffer_info index_buffer_info {
       .size         = compute_verts::ntotal_indices * sizeof(uint32_t),
-      .storage_type = NGF_BUFFER_STORAGE_PRIVATE,
+      .storage_type = NGF_BUFFER_STORAGE_DEVICE_LOCAL,
       .buffer_usage = NGF_BUFFER_USAGE_XFER_DST | NGF_BUFFER_USAGE_INDEX_BUFFER};
   ngf::buffer staging_index_buffer;
   NGF_MISC_CHECK_NGF_ERROR(staging_index_buffer.initialize(staging_index_buffer_info));
@@ -176,7 +176,7 @@ void* sample_initialize(
    */
   const ngf_buffer_info vertex_buffer_info {
       .size         = compute_verts::ntotal_verts * (4u * sizeof(float)) * 2,
-      .storage_type = NGF_BUFFER_STORAGE_PRIVATE,
+      .storage_type = NGF_BUFFER_STORAGE_DEVICE_LOCAL,
       .buffer_usage = NGF_BUFFER_USAGE_VERTEX_BUFFER | NGF_BUFFER_USAGE_STORAGE_BUFFER};
   NGF_MISC_CHECK_NGF_ERROR(state->vertex_buffer.initialize(vertex_buffer_info));
   state->compute_buffer_slice.buffer = state->vertex_buffer.get();

source/ngf-mtl/impl.cpp

@@ -939,6 +939,7 @@ static void ngfmtl_populate_ngf_device(uint32_t handle, ngf_device& ngfdev, MTL:
   caps.max_3d_image_dimension                   = 2048;
   caps.max_image_layers                         = 2048;
   caps.max_uniform_buffer_range                 = NGF_DEVICE_LIMIT_UNKNOWN;
+  caps.device_local_memory_is_host_visible      = false; // TODO
 
   if (gpu_family_idx >= ngfmtl_gpufam_idx(MTL::GPUFamilyApple6)) {
     caps.max_sampled_images_per_stage = 128;
@@ -1655,7 +1656,7 @@ ngf_id<MTL::Buffer> ngfmtl_create_buffer(const ngf_buffer_info& info) {
   case NGF_BUFFER_STORAGE_HOST_WRITEABLE:
     options = MTL::ResourceCPUCacheModeWriteCombined | MTL::ResourceStorageModeShared;
     break;
-  case NGF_BUFFER_STORAGE_PRIVATE:
+  case NGF_BUFFER_STORAGE_DEVICE_LOCAL:
     options = MTL::ResourceStorageModePrivate;
     break;
   default:

source/ngf-vk/impl.c

@@ -810,7 +810,7 @@ static VkMemoryPropertyFlags get_vk_memory_flags(ngf_buffer_storage_type s) {
   case NGF_BUFFER_STORAGE_HOST_WRITEABLE:
   case NGF_BUFFER_STORAGE_HOST_READABLE_WRITEABLE:
     return VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT;
-  case NGF_BUFFER_STORAGE_PRIVATE:
+  case NGF_BUFFER_STORAGE_DEVICE_LOCAL:
     return VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
   }
   return 0;
@@ -3525,6 +3525,8 @@ ngf_error ngf_get_device_list(const ngf_device** devices, uint32_t* ndevices) {
           ngfi_get_highest_sample_count(devcaps->texture_color_sample_counts);
       devcaps->max_supported_texture_depth_sample_count =
           ngfi_get_highest_sample_count(devcaps->texture_depth_sample_counts);
+
+      devcaps->device_local_memory_is_host_visible = false; // TODO
     }
 ngf_enumerate_devices_cleanup:
     if (tmp_instance != VK_NULL_HANDLE) { destroy_vk_instance(tmp_instance, NULL); }
@@ -5702,7 +5704,7 @@ ngf_error ngf_create_buffer(const ngf_buffer_info* info, ngf_buffer* result) {
   const VkBufferUsageFlags    vk_usage_flags = get_vk_buffer_usage(info->buffer_usage);
   const VkMemoryPropertyFlags vk_mem_flags   = get_vk_memory_flags(info->storage_type);
   const bool     vk_mem_is_host_visible      = vk_mem_flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT;
-  const uint32_t vma_usage_flags             = info->storage_type == NGF_BUFFER_STORAGE_PRIVATE
+  const uint32_t vma_usage_flags             = info->storage_type == NGF_BUFFER_STORAGE_DEVICE_LOCAL
                                                    ? VMA_MEMORY_USAGE_GPU_ONLY
                                                    : VMA_MEMORY_USAGE_CPU_ONLY;
   ngf_error      err                         = NGF_ERROR_OK;