vpp-nlmeansは32bitでは非対応にする。

原因不明のコンパイルエラーのため。
rigaya · May 1, 2024 · a9304e5 · a9304e5
1 parent 5ab07f4
commit a9304e5
Show file tree

Hide file tree

Showing 4 changed files with 24 additions and 2 deletions.
diff --git a/NVEncC_Options.en.md b/NVEncC_Options.en.md
@@ -2205,7 +2205,7 @@ Strong noise reduction filter.
   ```
 
 ### --vpp-nlmeans [&lt;param1&gt;=&lt;value1&gt;[,&lt;param2&gt;=&lt;value2&gt;]...]
-Non local means noise reduction filter.
+Non local means noise reduction filter. Only supported in 64bit binary.
 
 - **Parameters**
   - sigma=&lt;float&gt;  (default=0.005, 0.0 -)   

diff --git a/NVEncC_Options.ja.md b/NVEncC_Options.ja.md
@@ -2262,7 +2262,7 @@ yadifによるインタレ解除を行う。
   ```
 
 ### --vpp-nlmeans [&lt;param1&gt;=&lt;value1&gt;[,&lt;param2&gt;=&lt;value2&gt;]...]
-Non local meansを用いたノイズ除去フィルタ。
+Non local meansを用いたノイズ除去フィルタ。Windowsでは64bit版のみに対応。
 
 - **パラメータ**
   - sigma=&lt;float&gt;  (default=0.005, 0.0 -)   

diff --git a/NVEncCore/NVEncFilterDenoiseNLMeans.cu b/NVEncCore/NVEncFilterDenoiseNLMeans.cu
@@ -76,6 +76,8 @@ enum RGYFilterDenoiseNLMeansTmpBufIdx {
     TMP_TOTAL,
 };
 
+#if ENABLE_VPP_NLMEANS
+
 template<typename Type, int bit_depth>
 __device__ __inline__ float8 calc_sqdiff(Type val0, float8 val1) {
     float8 val0_1 = (float8)val0 - val1;
@@ -808,6 +810,8 @@ RGY_ERR NVEncFilterDenoiseNLMeans::denoiseFrame(RGYFrameInfo *pOutputFrame, cons
     return RGY_ERR_NONE;
 }
 
+#endif //ENABLE_VPP_NLMEANS
+
 NVEncFilterDenoiseNLMeans::NVEncFilterDenoiseNLMeans() : m_tmpBuf() {
     m_name = _T("nlmeans");
 }
@@ -817,6 +821,7 @@ NVEncFilterDenoiseNLMeans::~NVEncFilterDenoiseNLMeans() {
 }
 
 RGY_ERR NVEncFilterDenoiseNLMeans::init(shared_ptr<NVEncFilterParam> pParam, shared_ptr<RGYLog> pPrintMes) {
+#if ENABLE_VPP_NLMEANS
     RGY_ERR sts = RGY_ERR_NONE;
     m_pLog = pPrintMes;
     auto prm = std::dynamic_pointer_cast<NVEncFilterParamDenoiseNLMeans>(pParam);
@@ -921,13 +926,18 @@ RGY_ERR NVEncFilterDenoiseNLMeans::init(shared_ptr<NVEncFilterParam> pParam, sha
     setFilterInfo(pParam->print());
     m_param = pParam;
     return sts;
+#else
+    AddMessage(RGY_LOG_ERROR, _T("nlmeans not compiled in this build.\n"));
+    return RGY_ERR_UNSUPPORTED;
+#endif //ENABLE_VPP_NLMEANS
 }
 
 tstring NVEncFilterParamDenoiseNLMeans::print() const {
     return nlmeans.print();
 }
 
 RGY_ERR NVEncFilterDenoiseNLMeans::run_filter(const RGYFrameInfo *pInputFrame, RGYFrameInfo **ppOutputFrames, int *pOutputFrameNum, cudaStream_t stream) {
+#if ENABLE_VPP_NLMEANS
     RGY_ERR sts = RGY_ERR_NONE;
 
     if (pInputFrame->ptr[0] == nullptr) {
@@ -960,6 +970,10 @@ RGY_ERR NVEncFilterDenoiseNLMeans::run_filter(const RGYFrameInfo *pInputFrame, R
         return sts;
     }
     return sts;
+#else
+    AddMessage(RGY_LOG_ERROR, _T("nlmeans not compiled in this build.\n"));
+    return RGY_ERR_UNSUPPORTED;
+#endif
 }
 
 void NVEncFilterDenoiseNLMeans::close() {

diff --git a/NVEncCore/NVEncFilterDenoiseNLMeans.h b/NVEncCore/NVEncFilterDenoiseNLMeans.h
@@ -32,6 +32,12 @@
 #include "rgy_prm.h"
 #include <array>
 
+#if (defined(WIN32) || defined(WIN64)) && defined(_M_IX86)
+#define ENABLE_VPP_NLMEANS 0
+#else
+#define ENABLE_VPP_NLMEANS 1
+#endif
+
 // dxdyのペアを何並列で同時計算するか
 static const int RGY_NLMEANS_DXDY_STEP = 8;
 
@@ -53,13 +59,15 @@ class NVEncFilterDenoiseNLMeans : public NVEncFilter {
     virtual ~NVEncFilterDenoiseNLMeans();
     virtual RGY_ERR init(shared_ptr<NVEncFilterParam> pParam, shared_ptr<RGYLog> pPrintMes) override;
 protected:
+#if ENABLE_VPP_NLMEANS
     RGY_ERR denoisePlane(
         RGYFrameInfo *pOutputPlane,
         RGYFrameInfo *pTmpUPlane, RGYFrameInfo *pTmpVPlane,
         RGYFrameInfo *pTmpIWPlane,
         const RGYFrameInfo *pInputPlane,
         cudaStream_t stream);
     RGY_ERR denoiseFrame(RGYFrameInfo *pOutputFrame, const RGYFrameInfo *pInputFrame, cudaStream_t stream);
+#endif
     virtual RGY_ERR run_filter(const RGYFrameInfo *pInputFrame, RGYFrameInfo **ppOutputFrames, int *pOutputFrameNum, cudaStream_t stream) override;
     virtual void close() override;
     std::array<std::unique_ptr<CUFrameBuf>, 2 + 1 + RGY_NLMEANS_DXDY_STEP> m_tmpBuf;