VRTWarpedDataset::IRasterIO(): avoid a memory allocation and pixel copy when possible

alg/gdalwarper.h

@@ -557,6 +557,9 @@ class CPL_DLL GDALWarpOperation
     CPLErr Initialize(const GDALWarpOptions *psNewOptions);
     void *CreateDestinationBuffer(int nDstXSize, int nDstYSize,
                                   int *pbWasInitialized = nullptr);
+    void InitializeDestinationBuffer(void *pDstBuffer, int nDstXSize,
+                                     int nDstYSize,
+                                     int *pbWasInitialized = nullptr);
     static void DestroyDestinationBuffer(void *pDstBuffer);
 
     const GDALWarpOptions *GetOptions();

alg/gdalwarpoperation.cpp

@@ -694,10 +694,40 @@ void *GDALWarpOperation::CreateDestinationBuffer(int nDstXSize, int nDstYSize,
     void *pDstBuffer = VSI_MALLOC3_VERBOSE(
         cpl::fits_on<int>(nWordSize * psOptions->nBandCount), nDstXSize,
         nDstYSize);
-    if (pDstBuffer == nullptr)
+    if (pDstBuffer)
     {
-        return nullptr;
+        InitializeDestinationBuffer(pDstBuffer, nDstXSize, nDstYSize,
+                                    pbInitialized);
     }
+    return pDstBuffer;
+}
+
+/**
+ * This method initializes a destination buffer for use with WarpRegionToBuffer.
+ *
+ * It is initialized based on the INIT_DEST settings.
+ *
+ * This method is called by CreateDestinationBuffer().
+ * It is meant at being used by callers that have already allocated the
+ * destination buffer without using CreateDestinationBuffer().
+ *
+ * @param pDstBuffer Buffer of size
+ *                   GDALGetDataTypeSizeBytes(psOptions->eWorkingDataType) *
+ *                   nDstXSize * nDstYSize * psOptions->nBandCount bytes.
+ * @param nDstXSize Width of output window on destination buffer to be produced.
+ * @param nDstYSize Height of output window on destination buffer to be
+ *                  produced.
+ * @param pbInitialized Filled with boolean indicating if the buffer was
+ *                      initialized.
+ * @since 3.10
+ */
+void GDALWarpOperation::InitializeDestinationBuffer(void *pDstBuffer,
+                                                    int nDstXSize,
+                                                    int nDstYSize,
+                                                    int *pbInitialized)
+{
+    const int nWordSize = GDALGetDataTypeSizeBytes(psOptions->eWorkingDataType);
+
     const GPtrDiff_t nBandSize =
         static_cast<GPtrDiff_t>(nWordSize) * nDstXSize * nDstYSize;
 
@@ -713,8 +743,7 @@ void *GDALWarpOperation::CreateDestinationBuffer(int nDstXSize, int nDstYSize,
         {
             *pbInitialized = FALSE;
         }
-
-        return pDstBuffer;
+        return;
     }
 
     if (pbInitialized != nullptr)
@@ -776,8 +805,6 @@ void *GDALWarpOperation::CreateDestinationBuffer(int nDstXSize, int nDstYSize,
     }
 
     CSLDestroy(papszInitValues);
-
-    return pDstBuffer;
 }
 
 /**

frmts/vrt/vrtwarped.cpp

@@ -1968,8 +1968,16 @@ CPLErr VRTWarpedDataset::IRasterIO(
 
     // Build a map from warped output bands to their index
     std::map<int, int> oMapBandToWarpingBandIndex;
+    bool bAllBandsIncreasingOrder =
+        (psWO->nBandCount == nBands && nBands == nBandCount);
     for (int i = 0; i < psWO->nBandCount; ++i)
+    {
         oMapBandToWarpingBandIndex[psWO->panDstBands[i]] = i;
+        if (psWO->panDstBands[i] != i + 1 || panBandMap[i] != i + 1)
+        {
+            bAllBandsIncreasingOrder = false;
+        }
+    }
 
     // Check that all requested bands are actually warped output bands.
     for (int i = 0; i < nBandCount; ++i)
@@ -2054,51 +2062,72 @@ CPLErr VRTWarpedDataset::IRasterIO(
     CPLDebugOnly("VRT",
                  "Using optimized VRTWarpedDataset::IRasterIO() code path");
 
-    // Allocate a warping destination buffer
-    // Note: we could potentially use the pData target buffer argument of this
-    // function in some circumstances... but probably not worth the complication
-    GByte *pabyWarpBuffer = static_cast<GByte *>(
-        m_poWarper->CreateDestinationBuffer(nXSize, nYSize));
+    // Allocate a warping destination buffer if needed.
+    // We can use directly the output buffer pData if:
+    // - we request exactly all warped bands, and that there are as many
+    //   warped bands as dataset bands (no alpha)
+    // - the output buffer data atype is the warping working data type
+    // - the output buffer has a band-sequential layout.
+    GByte *pabyWarpBuffer;
 
-    if (pabyWarpBuffer == nullptr)
+    if (bAllBandsIncreasingOrder && psWO->eWorkingDataType == eBufType &&
+        nPixelSpace == GDALGetDataTypeSizeBytes(eBufType) &&
+        nLineSpace == nPixelSpace * nXSize &&
+        (nBands == 1 || nBandSpace == nLineSpace * nYSize))
     {
-        return CE_Failure;
+        pabyWarpBuffer = static_cast<GByte *>(pData);
+        m_poWarper->InitializeDestinationBuffer(pabyWarpBuffer, nXSize, nYSize);
+    }
+    else
+    {
+        pabyWarpBuffer = static_cast<GByte *>(
+            m_poWarper->CreateDestinationBuffer(nXSize, nYSize));
+
+        if (pabyWarpBuffer == nullptr)
+        {
+            return CE_Failure;
+        }
     }
 
     const CPLErr eErr = m_poWarper->WarpRegionToBuffer(
         nXOff, nYOff, nXSize, nYSize, pabyWarpBuffer, psWO->eWorkingDataType,
         nSrcXOff, nSrcYOff, nSrcXSize, nSrcYSize, dfSrcXExtraSize,
         dfSrcYExtraSize);
-    if (eErr == CE_None)
+
+    if (pabyWarpBuffer != pData)
     {
-        // Copy warping buffer into user destination buffer
-        for (int i = 0; i < nBandCount; i++)
+        if (eErr == CE_None)
         {
-            const int nRasterIOBand = panBandMap[i];
-            const auto oIterToWarpingBandIndex =
-                oMapBandToWarpingBandIndex.find(nRasterIOBand);
-            // cannot happen due to earlier check
-            CPLAssert(oIterToWarpingBandIndex !=
-                      oMapBandToWarpingBandIndex.end());
-
-            const GByte *const pabyWarpBandBuffer =
-                pabyWarpBuffer +
-                static_cast<GPtrDiff_t>(oIterToWarpingBandIndex->second) *
-                    nXSize * nYSize * nWarpDTSize;
-            GByte *const pabyDstBand = pabyDst + i * nBandSpace;
-
-            for (int iY = 0; iY < nYSize; iY++)
+            // Copy warping buffer into user destination buffer
+            for (int i = 0; i < nBandCount; i++)
             {
-                GDALCopyWords(pabyWarpBandBuffer + static_cast<GPtrDiff_t>(iY) *
-                                                       nXSize * nWarpDTSize,
-                              psWO->eWorkingDataType, nWarpDTSize,
-                              pabyDstBand + iY * nLineSpace, eBufType,
-                              static_cast<int>(nPixelSpace), nXSize);
+                const int nRasterIOBand = panBandMap[i];
+                const auto oIterToWarpingBandIndex =
+                    oMapBandToWarpingBandIndex.find(nRasterIOBand);
+                // cannot happen due to earlier check
+                CPLAssert(oIterToWarpingBandIndex !=
+                          oMapBandToWarpingBandIndex.end());
+
+                const GByte *const pabyWarpBandBuffer =
+                    pabyWarpBuffer +
+                    static_cast<GPtrDiff_t>(oIterToWarpingBandIndex->second) *
+                        nXSize * nYSize * nWarpDTSize;
+                GByte *const pabyDstBand = pabyDst + i * nBandSpace;
+
+                for (int iY = 0; iY < nYSize; iY++)
+                {
+                    GDALCopyWords(pabyWarpBandBuffer +
+                                      static_cast<GPtrDiff_t>(iY) * nXSize *
+                                          nWarpDTSize,
+                                  psWO->eWorkingDataType, nWarpDTSize,
+                                  pabyDstBand + iY * nLineSpace, eBufType,
+                                  static_cast<int>(nPixelSpace), nXSize);
+                }
             }
         }
-    }
 
-    m_poWarper->DestroyDestinationBuffer(pabyWarpBuffer);
+        m_poWarper->DestroyDestinationBuffer(pabyWarpBuffer);
+    }
 
     return eErr;
 }