composite: use 16 bit where appropriate

Author: Grok Compression

src/lib/core/codestream/decompress/CodeStreamDecompress.cpp

@@ -1757,6 +1757,41 @@ bool CodeStreamDecompress::activateScratch(bool singleTile, GrkImage* scratch)
 {
   multiTileComposite_->copyHeaderTo(scratch);
 
+  // Set int16 data_type on scratch components when all components are eligible
+  // for 16-bit DWT. This ensures the composite buffer is allocated as int16,
+  // avoiding int16→int32 widening during tile compositing.
+  // Mirror the eligibility logic from TileProcessor::createDecompressTileComponentWindows.
+  // Only do this when there's no region decode, because partial tiles have
+  // wholeTileDecompress_=false and fall back to int32 DWT.
+  if(scratch->has_multiple_tiles && region_.empty())
+  {
+    bool allEligible = true;
+    bool hasMct =
+        defaultTcp_->mct_ == 1 && scratch->numcomps >= 3 && scratch->componentsEqual(3, false);
+    for(uint16_t i = 0; i < scratch->numcomps; i++)
+    {
+      auto tccp = defaultTcp_->tccps_ + i;
+      auto comp = scratch->comps + i;
+      if(tccp->qmfbid_ != 1)
+      {
+        allEligible = false;
+        break;
+      }
+      bool isMctComp = hasMct && (i <= 2);
+      uint32_t headroom = isMctComp ? 5 : 4;
+      if(comp->prec + headroom > 16)
+      {
+        allEligible = false;
+        break;
+      }
+    }
+    if(allEligible)
+    {
+      for(uint16_t i = 0; i < scratch->numcomps; i++)
+        scratch->comps[i].data_type = GRK_INT_16;
+    }
+  }
+
   // no need to allocate composite data if there is only one tile
   // i.e. no compositing — UNLESS band callback is active, which needs
   // a destination buffer for compositing before writing.

src/lib/core/util/GrkImage.cpp

@@ -396,9 +396,10 @@ bool GrkImage::allocData(grk_image_comp* comp, bool clear)
   if(!comp || comp->w == 0 || comp->h == 0)
     return false;
   single_component_data_free(comp);
-  uint32_t stride = grk_make_aligned_width<int32_t>(comp->w);
+  uint32_t stride = (comp->data_type == GRK_INT_16) ? grk_make_aligned_width<int16_t>(comp->w)
+                                                    : grk_make_aligned_width<int32_t>(comp->w);
   size_t dataSize = (uint64_t)stride * comp->h * sizeOfDataType(comp->data_type);
-  auto data = (int32_t*)grk_aligned_malloc(dataSize);
+  auto data = grk_aligned_malloc(dataSize);
   if(!data)
   {
     grk::grklog.error("Failed to allocate aligned memory buffer of dimensions %u x %u",

src/lib/core/util/GrkImage.h

@@ -155,8 +155,8 @@ class GrkImage : public grk_image
     {
       case GRK_INT_32:
         return compositeInterleaved_T<int32_t>(src, yBegin, yEnd);
-      // case GRK_INT_16:
-      //   return compositeInterleaved_T<int16_t>(src,yBegin,yEnd);
+      case GRK_INT_16:
+        return compositeInterleaved_T<int16_t>(src, yBegin, yEnd);
       // case GRK_INT_8:
       //   return compositeInterleaved_T<int8_t>(src,yBegin,yEnd);
       // case GRK_FLOAT:

src/lib/core/util/GrkImageSIMD.cpp

@@ -1127,8 +1127,7 @@ void hwy_copy_tile_to_swath(const grk_image* tile_img, const grk_swath_buffer* b
             }
             else if(buf->prec <= 16 && buf->sgnd)
             {
-              int16_t w =
-                  v <= -32768 ? (int16_t)-32768 : v >= 32767 ? (int16_t)32767 : (int16_t)v;
+              int16_t w = v <= -32768 ? (int16_t)-32768 : v >= 32767 ? (int16_t)32767 : (int16_t)v;
               memcpy(dstPx, &w, sizeof(int16_t));
             }
             else if(buf->prec <= 16)
@@ -1186,8 +1185,7 @@ void hwy_copy_tile_to_swath(const grk_image* tile_img, const grk_swath_buffer* b
             }
             else if(buf->prec <= 16 && buf->sgnd)
             {
-              int16_t w =
-                  v <= -32768 ? (int16_t)-32768 : v >= 32767 ? (int16_t)32767 : (int16_t)v;
+              int16_t w = v <= -32768 ? (int16_t)-32768 : v >= 32767 ? (int16_t)32767 : (int16_t)v;
               memcpy(dstPx, &w, sizeof(int16_t));
             }
             else if(buf->prec <= 16)