option owns vulkan device index (#5973)

Author: nihui

src/gpu.cpp

@@ -266,6 +266,8 @@ class GpuInfoPrivate
     void query_extension_properties();
 
 public:
+    int device_index;
+
     // physical device
     VkPhysicalDevice physicalDevice;
 
@@ -1141,11 +1143,21 @@ GpuInfo& GpuInfo::operator=(const GpuInfo&)
     return *this;
 }
 
+int GpuInfo::device_index() const
+{
+    return d->device_index;
+}
+
 VkPhysicalDevice GpuInfo::physicalDevice() const
 {
     return d->physicalDevice;
 }
 
+VkPhysicalDevice GpuInfo::physical_device() const
+{
+    return d->physicalDevice;
+}
+
 const VkPhysicalDeviceFeatures& GpuInfo::physicalDevicefeatures() const
 {
     return d->physicalDevicefeatures;
@@ -1161,6 +1173,11 @@ const VkPhysicalDeviceMemoryProperties& GpuInfo::physicalDeviceMemoryProperties(
     return d->physicalDeviceMemoryProperties;
 }
 
+const VkPhysicalDeviceMemoryProperties& GpuInfo::physical_device_memory_properties() const
+{
+    return d->physicalDeviceMemoryProperties;
+}
+
 const std::vector<VkExtensionProperties>& GpuInfo::deviceExtensionProperties() const
 {
     return d->deviceExtensionProperties;
@@ -2209,10 +2226,11 @@ int create_gpu_instance(const char* driver_path)
         const VkPhysicalDevice& physicalDevice = physicalDevices[i];
         delete g_gpu_infos[gpu_info_index];
         g_gpu_infos[gpu_info_index] = new GpuInfo;
-        // GpuInfoPrivate& gpu_info = *(g_gpu_infos[gpu_info_index]->d);
 
         GpuInfo& gpu_info = *g_gpu_infos[gpu_info_index];
 
+        gpu_info.d->device_index = gpu_info_index;
+
         gpu_info.d->physicalDevice = physicalDevice;
 
         gpu_info.d->query_features();
@@ -2591,12 +2609,11 @@ const ncnn::Packing_vulkan* VulkanDevicePrivate::get_utility_operator(int storag
 
     opt.use_vulkan_compute = true;
 
-    opt.blob_vkallocator = opt.workspace_vkallocator = vkdev->acquire_blob_allocator();
-    opt.staging_vkallocator = vkdev->acquire_staging_allocator();
-
     // cache uop pipeline as device member explicitly
     opt.pipeline_cache = 0;
 
+    opt.vulkan_device_index = vkdev->info.device_index();
+
     ncnn::Packing_vulkan* uop = new ncnn::Packing_vulkan;
     uop->vkdev = vkdev;
 
@@ -2613,9 +2630,6 @@ const ncnn::Packing_vulkan* VulkanDevicePrivate::get_utility_operator(int storag
 
     uop_packing[storage_type_from][storage_type_to][cast_type_from_index][cast_type_to_index][packing_type_to_index] = uop;
 
-    vkdev->reclaim_blob_allocator(opt.blob_vkallocator);
-    vkdev->reclaim_staging_allocator(opt.staging_vkallocator);
-
     return uop;
 }
 
@@ -2627,6 +2641,7 @@ void VulkanDevicePrivate::destroy_utility_operator()
     opt.use_int8_arithmetic = false;
     opt.use_cooperative_matrix = false;
     opt.pipeline_cache = 0;
+    opt.vulkan_device_index = vkdev->info.device_index();
 
     // from buffer | image
     // to buffer | image
@@ -4576,11 +4591,13 @@ int compile_spirv_module(const char* comp_data, int comp_data_size, const Option
 
     bool support_shader_int64 = false;
 
-    if (opt.blob_vkallocator)
+    // fill device macros
     {
-        const VulkanDevice* vkdev = opt.blob_vkallocator->vkdev;
+        int device_index = opt.vulkan_device_index;
+        if (device_index < 0 || device_index >= get_gpu_count())
+            device_index = get_default_gpu_index();
 
-        const GpuInfo& info = vkdev->info;
+        const GpuInfo& info = get_gpu_info(device_index);
 
         support_shader_int64 = info.physicalDevicefeatures().shaderInt64;
 
@@ -4935,10 +4952,6 @@ int compile_spirv_module(const char* comp_data, int comp_data_size, const Option
 
 #undef DD_APPEND_PROPERTY
     }
-    else
-    {
-        NCNN_LOGE("opt.blob_vkallocator is null");
-    }
 
     std::string define_macro_data;
 

src/gpu.h

@@ -193,8 +193,11 @@ class NCNN_EXPORT GpuInfo
     explicit GpuInfo();
     virtual ~GpuInfo();
 
+    int device_index() const;
+
     // vulkan physical device
     VkPhysicalDevice physicalDevice() const;
+    VkPhysicalDevice physical_device() const; // api compatibility
 
     // features
     const VkPhysicalDeviceFeatures& physicalDevicefeatures() const;
@@ -204,6 +207,7 @@ class NCNN_EXPORT GpuInfo
 
     // memory properties
     const VkPhysicalDeviceMemoryProperties& physicalDeviceMemoryProperties() const;
+    const VkPhysicalDeviceMemoryProperties& physical_device_memory_properties() const; // api compatibility
 
     // extension properties
     const std::vector<VkExtensionProperties>& deviceExtensionProperties() const;

src/net.cpp

@@ -1380,7 +1380,14 @@ int Net::load_param(const DataReader& dr)
 
     if (opt.use_vulkan_compute)
     {
-        if (!d->vkdev) d->vkdev = get_gpu_device();
+        if (!d->vkdev)
+        {
+            int device_index = opt.vulkan_device_index;
+            if (device_index < 0 || device_index >= get_gpu_count())
+                device_index = get_default_gpu_index();
+
+            d->vkdev = get_gpu_device(device_index);
+        }
         if (!d->vkdev || !d->vkdev->is_valid()) opt.use_vulkan_compute = false; // no valid vulkan device, fallback to cpu
     }
     if (opt.use_vulkan_compute)
@@ -1683,7 +1690,14 @@ int Net::load_param_bin(const DataReader& dr)
 
     if (opt.use_vulkan_compute)
     {
-        if (!d->vkdev) d->vkdev = get_gpu_device();
+        if (!d->vkdev)
+        {
+            int device_index = opt.vulkan_device_index;
+            if (device_index < 0 || device_index >= get_gpu_count())
+                device_index = get_default_gpu_index();
+
+            d->vkdev = get_gpu_device(device_index);
+        }
         if (!d->vkdev || !d->vkdev->is_valid()) opt.use_vulkan_compute = false; // no valid vulkan device, fallback to cpu
     }
     if (opt.use_vulkan_compute)
@@ -2305,11 +2319,13 @@ std::vector<Layer*>& Net::mutable_layers()
 #if NCNN_VULKAN
 void Net::set_vulkan_device(int device_index)
 {
+    opt.vulkan_device_index = device_index;
     d->vkdev = get_gpu_device(device_index);
 }
 
 void Net::set_vulkan_device(const VulkanDevice* _vkdev)
 {
+    opt.vulkan_device_index = _vkdev->info.device_index();
     d->vkdev = _vkdev;
 }
 

src/option.cpp

@@ -21,6 +21,10 @@ namespace ncnn {
 Option::Option()
 {
     lightmode = true;
+    use_shader_pack8 = false;
+    use_subgroup_ops = false;
+    use_reserved_0 = false;
+
     num_threads = get_physical_big_cpu_count();
     blob_allocator = 0;
     workspace_allocator = 0;
@@ -50,18 +54,13 @@ Option::Option()
 
     use_packing_layout = true;
 
-    use_shader_pack8 = false;
-
-    use_subgroup_ops = false;
-
-    use_subgroup_reserved_1 = false;
-    use_subgroup_reserved_2 = false;
-    use_subgroup_reserved_3 = false;
+    vulkan_device_index = -1;
+    use_reserved_1 = false;
 
     use_image_storage = false;
     use_tensor_storage = false;
 
-    use_reserved_0 = false;
+    use_reserved_2 = false;
 
     flush_denormals = 3;
 
@@ -78,6 +77,10 @@ Option::Option()
 
     use_fp16_uniform = true;
     use_int8_uniform = true;
+
+    use_reserved_9 = false;
+    use_reserved_10 = false;
+    use_reserved_11 = false;
 }
 
 } // namespace ncnn

src/option.h

@@ -37,6 +37,14 @@ class NCNN_EXPORT Option
     // enabled by default
     bool lightmode;
 
+    // use pack8 shader
+    bool use_shader_pack8;
+
+    // enable subgroup in shader
+    bool use_subgroup_ops;
+
+    bool use_reserved_0;
+
     // thread count
     // default value is the one returned by get_cpu_count()
     int num_threads;
@@ -105,20 +113,16 @@ class NCNN_EXPORT Option
     // enabled by default
     bool use_packing_layout;
 
-    bool use_shader_pack8;
+    // the vulkan device
+    int vulkan_device_index;
 
-    // subgroup option
-    bool use_subgroup_ops;
-
-    bool use_subgroup_reserved_1;
-    bool use_subgroup_reserved_2;
-    bool use_subgroup_reserved_3;
+    bool use_reserved_1;
 
     // turn on for adreno
     bool use_image_storage;
     bool use_tensor_storage;
 
-    bool use_reserved_0;
+    bool use_reserved_2;
 
     // enable DAZ(Denormals-Are-Zero) and FTZ(Flush-To-Zero)
     // default value is 3