SWDEV-301823 - Optimize hipMemset2D/3D

Change-Id: Ibe560149a263c2ac6b08e4eb1a1d331d2aeac78c

rocclr/device/rocm/rocvirtual.cpp

@@ -2211,10 +2211,23 @@ void VirtualGPU::submitFillMemory(amd::FillMemoryCommand& cmd) {
   amd::ScopedLock lock(execution());
 
   profilingBegin(cmd);
-
-  if (!fillMemory(cmd.type(), &cmd.memory(), cmd.pattern(), cmd.patternSize(), cmd.origin(),
-                  cmd.size())) {
-    cmd.setStatus(CL_INVALID_OPERATION);
+  size_t width  = cmd.size().c[0];
+  size_t height = cmd.size().c[1];
+  size_t depth  = cmd.size().c[2];
+  size_t pitch  = cmd.surface().c[0];
+  amd::Coord3D origin = cmd.origin();
+  amd::Coord3D region{cmd.surface().c[1], cmd.surface().c[2], depth};
+  amd::BufferRect rect;
+  rect.create(static_cast<size_t*>(origin), static_cast<size_t*>(region),
+      pitch, 0);
+  for (size_t slice = 0; slice < depth; slice++) {
+    for (size_t row = 0; row < height; row++) {
+      const size_t rowOffset = rect.offset(0, row, slice);
+      if (!fillMemory(cmd.type(), &cmd.memory(), cmd.pattern(), cmd.patternSize(),
+          amd::Coord3D{rowOffset, 0, 0}, amd::Coord3D{width, 1, 1})) {
+        cmd.setStatus(CL_INVALID_OPERATION);
+      }
+    }
   }
   profilingEnd(cmd);
 }

rocclr/platform/command.hpp

@@ -601,16 +601,18 @@ class FillMemoryCommand : public OneMemoryArgCommand {
  private:
   Coord3D origin_;                   //!< Origin of the region to write to.
   Coord3D size_;                     //!< Size of the region to write to.
+  Coord3D surface_;                  //!< Total surface
   char pattern_[MaxFillPatterSize];  //!< The fill pattern
   size_t patternSize_;               //!< Pattern size
 
  public:
   FillMemoryCommand(HostQueue& queue, cl_command_type cmdType, const EventWaitList& eventWaitList,
-                    Memory& memory, const void* pattern, size_t patternSize, Coord3D origin,
-                    Coord3D size)
+                    Memory& memory, const void* pattern, size_t patternSize, const Coord3D& origin,
+                    const Coord3D& size, const Coord3D& surface)
       : OneMemoryArgCommand(queue, cmdType, eventWaitList, memory),
         origin_(origin),
         size_(size),
+        surface_(surface),
         patternSize_(patternSize) {
     // Sanity checks
     assert(pattern != NULL && "pattern cannot be null");
@@ -632,6 +634,9 @@ class FillMemoryCommand : public OneMemoryArgCommand {
   //! Return the region size
   const Coord3D& size() const { return size_; }
 
+  //! Return the surface
+  const Coord3D& surface() const { return surface_; }
+
   //! Return true if the entire memory object is written.
   bool isEntireMemory() const;
 };