Merge pull request #55 from TNTwise/master

Fix padding and add 4.26

README.md

@@ -89,12 +89,13 @@ By default models are exported with ensemble=False and Fast=True
   - 66 = rife-v4.24 (ensemble=False)
   - 67 = rife-v4.24 (ensemble=True)
   - 68 = rife-v4.25 (ensemble=False)
+  - 69 = rife-v4.26 (ensemble=False)
 
   ## My experimental custom models (only works with 2x)
 
-  - 69 = sudo_rife4 (ensemble=False / fast=True)
-  - 70 = sudo_rife4 (ensemble=True / fast=False)
-  - 71 = sudo_rife4 (ensemble=True / fast=True)
+  - 70 = sudo_rife4 (ensemble=False / fast=True)
+  - 71 = sudo_rife4 (ensemble=True / fast=False)
+  - 72 = sudo_rife4 (ensemble=True / fast=True)
 
 - factor_num, factor_den: Factor of target frame rate. For example `factor_num=5, factor_den=2` will multiply input clip FPS by 2.5. Only rife-v4 model supports custom frame rate.
 

RIFE/plugin.cpp

@@ -203,8 +203,8 @@ static void VS_CC rifeCreate(const VSMap* in, VSMap* out, [[maybe_unused]] void*
         if (err)
             d->skipThreshold = 60.0;
 
-        if (model < 0 || model > 71)
-            throw "model must be between 0 and 71 (inclusive)";
+        if (model < 0 || model > 72)
+            throw "model must be between 0 and 72 (inclusive)";
 
         if (factorNum < 1)
             throw "factor_num must be at least 1";
@@ -278,7 +278,8 @@ static void VS_CC rifeCreate(const VSMap* in, VSMap* out, [[maybe_unused]] void*
                 ncnn::destroy_gpu_instance();
             return;
         }
-
+        bool extra_padding{};
+        extra_padding = false;
         if (modelPath.empty()) {
             std::string pluginPath{ vsapi->getPluginPath(vsapi->getPluginByID("com.holywu.rife", core)) };
             modelPath = pluginPath.substr(0, pluginPath.rfind('/')) + "/models";
@@ -493,12 +494,15 @@ static void VS_CC rifeCreate(const VSMap* in, VSMap* out, [[maybe_unused]] void*
                 modelPath += "/rife-v4.25_ensembleFalse";
                 break;
             case 69:
-                modelPath += "/sudo_rife4_ensembleFalse_fastTrue";
+                modelPath += "/rife-v4.26_ensembleFalse";
                 break;
             case 70:
-                modelPath += "/sudo_rife4_ensembleTrue_fastFalse";
+                modelPath += "/sudo_rife4_ensembleFalse_fastTrue";
                 break;
             case 71:
+                modelPath += "/sudo_rife4_ensembleTrue_fastFalse";
+                break;
+            case 72:
                 modelPath += "/sudo_rife4_ensembleTrue_fastTrue";
                 break;
 
@@ -512,7 +516,7 @@ static void VS_CC rifeCreate(const VSMap* in, VSMap* out, [[maybe_unused]] void*
 
         bool rife_v2{};
         bool rife_v4{};
-
+        
         if (modelPath.find("rife-v2") != std::string::npos)
             rife_v2 = true;
         else if (modelPath.find("rife-v3.9") != std::string::npos)
@@ -524,6 +528,11 @@ static void VS_CC rifeCreate(const VSMap* in, VSMap* out, [[maybe_unused]] void*
             rife_v4 = true;
         else if (modelPath.find("rife4") != std::string::npos)
             rife_v4 = true;
+            // rife 4.25 and 4.26 require more padding due to extra scales.
+            if (modelPath.find("rifev4.25") != std::string::npos)
+                extra_padding = true;
+            if (modelPath.find("rifev4.26") != std::string::npos)
+                extra_padding = true;
         else if (modelPath.find("rife") == std::string::npos)
             throw "unknown model dir type";
 
@@ -614,7 +623,7 @@ static void VS_CC rifeCreate(const VSMap* in, VSMap* out, [[maybe_unused]] void*
             vsapi->freeMap(ret);
         }
 
-        d->rife = std::make_unique<RIFE>(gpuId, tta, uhd, 1, rife_v2, rife_v4);
+        d->rife = std::make_unique<RIFE>(gpuId, tta, uhd, 1, rife_v2, rife_v4, extra_padding);
 
 #ifdef _WIN32
         auto bufferSize{ MultiByteToWideChar(CP_UTF8, 0, modelPath.c_str(), -1, nullptr, 0) };

RIFE/rife.cpp

@@ -21,7 +21,7 @@
 
 DEFINE_LAYER_CREATOR(Warp)
 
-RIFE::RIFE(int gpuid, bool _tta_mode, bool _uhd_mode, int _num_threads, bool _rife_v2, bool _rife_v4)
+RIFE::RIFE(int gpuid, bool _tta_mode, bool _uhd_mode, int _num_threads, bool _rife_v2, bool _rife_v4, bool _extra_padding)
 {
     vkdev = gpuid == -1 ? 0 : ncnn::get_gpu_device(gpuid);
 
@@ -41,6 +41,7 @@ RIFE::RIFE(int gpuid, bool _tta_mode, bool _uhd_mode, int _num_threads, bool _ri
     num_threads = _num_threads;
     rife_v2 = _rife_v2;
     rife_v4 = _rife_v4;
+    extra_padding = _extra_padding;
 }
 
 RIFE::~RIFE()
@@ -1177,10 +1178,14 @@ int RIFE::process_v4(const float* src0R, const float* src0G, const float* src0B,
     opt.blob_vkallocator = blob_vkallocator;
     opt.workspace_vkallocator = blob_vkallocator;
     opt.staging_vkallocator = staging_vkallocator;
-
-    // pad to 32n
-    int w_padded = (w + 31) / 32 * 32;
-    int h_padded = (h + 31) / 32 * 32;
+    int w_padded, h_padded;
+    if (extra_padding) {
+        w_padded = (w + 63) / 64 * 64;
+        h_padded = (h + 63) / 64 * 64;
+    } else {
+        w_padded = (w + 31) / 32 * 32;
+        h_padded = (h + 31) / 32 * 32;
+    }
 
     const size_t in_out_tile_elemsize = opt.use_fp16_storage ? 2u : 4u;
 

RIFE/rife.h

@@ -11,7 +11,7 @@
 class RIFE
 {
 public:
-    RIFE(int gpuid, bool tta_mode = false, bool uhd_mode = false, int num_threads = 1, bool rife_v2 = false, bool rife_v4 = false);
+    RIFE(int gpuid, bool tta_mode = false, bool uhd_mode = false, int num_threads = 1, bool rife_v2 = false, bool rife_v4 = false, bool extra_padding = false);
     ~RIFE();
 
 #if _WIN32
@@ -51,6 +51,7 @@ class RIFE
     int num_threads;
     bool rife_v2;
     bool rife_v4;
+    bool extra_padding;
 };
 
 #endif // RIFE_H

meson.build

@@ -60,6 +60,8 @@ else
   opt_var.add_cmake_defines({'NCNN_AVX512': false})
   opt_var.add_cmake_defines({'NCNN_AVX512VNNI': false})
   opt_var.add_cmake_defines({'NCNN_XOP': false})
+  opt_var.add_cmake_defines({'NCNN_DISABLE_RTTI': false})
+  opt_var.add_cmake_defines({'NCNN_DISABLE_EXCEPTION': false})
 
   opt_var.add_cmake_defines({'WITH_LAYER_absval': false})
   opt_var.add_cmake_defines({'WITH_LAYER_argmax': false})