Add workaround for Adreno copyImageToBuffer calls on Vulkan

src/gpu/vk/GrVkGpu.cpp

 /*
  * Copyright 2015 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "GrVkGpu.h"
 
 #include "GrContextOptions.h"
@@ skipping 1684 matching lines @@
         return false;
     }
 
     // Change layout of our target so it can be used as copy
     image->setImageLayout(this,
                           VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                           VK_ACCESS_TRANSFER_READ_BIT,
                           VK_PIPELINE_STAGE_TRANSFER_BIT,
                           false);
 
+    size_t bpp = GrBytesPerPixel(config);
+    size_t tightRowBytes = bpp * width;
+    bool flipY = kBottomLeft_GrSurfaceOrigin == surface->origin();
+
+    VkBufferImageCopy region;
+    memset(&region, 0, sizeof(VkBufferImageCopy));
+
+    bool copyFromOrigin = this->vkCaps().mustDoCopiesFromOrigin();
+    if (copyFromOrigin) {
+        region.imageOffset = { 0, 0, 0 };
+        region.imageExtent = { (uint32_t)(left + width),
+                               (uint32_t)(flipY ? surface->height() - top : top + height),
+                               1
+                             };
+    } else {
+        VkOffset3D offset = {
+            left,
+            flipY ? surface->height() - top - height : top,
+            0
+        };
+        region.imageOffset = offset;
+        region.imageExtent = { (uint32_t)width, (uint32_t)height, 1 };
+    }
+
+    size_t transBufferRowBytes = bpp * region.imageExtent.width;
     GrVkTransferBuffer* transferBuffer =
-        static_cast<GrVkTransferBuffer*>(this->createBuffer(rowBytes * height,
-                                                            kXferGpuToCpu_GrBufferType,
-                                                            kStream_GrAccessPattern));
-
-    bool flipY = kBottomLeft_GrSurfaceOrigin == surface->origin();
-    VkOffset3D offset = {
-        left,
-        flipY ? surface->height() - top - height : top,
-        0
-    };
+            static_cast<GrVkTransferBuffer*>(this->createBuffer(transBufferRowBytes * height,
+                                                                kXferGpuToCpu_GrBufferType,
+                                                                kStream_GrAccessPattern));
 
     // Copy the image to a buffer so we can map it to cpu memory
-    VkBufferImageCopy region;
-    memset(&region, 0, sizeof(VkBufferImageCopy));
     region.bufferOffset = transferBuffer->offset();
     region.bufferRowLength = 0; // Forces RowLength to be width. We handle the rowBytes below.
     region.bufferImageHeight = 0; // Forces height to be tightly packed. Only useful for 3d images.
     region.imageSubresource = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1 };
-    region.imageOffset = offset;
-    region.imageExtent = { (uint32_t)width, (uint32_t)height, 1 };
 
     fCurrentCmdBuffer->copyImageToBuffer(this,
                                          image,
                                          VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                                          transferBuffer,
                                          1,
                                          &region);
 
     // make sure the copy to buffer has finished
     transferBuffer->addMemoryBarrier(this,
                                      VK_ACCESS_TRANSFER_WRITE_BIT,
                                      VK_ACCESS_HOST_READ_BIT,
                                      VK_PIPELINE_STAGE_TRANSFER_BIT,
                                      VK_PIPELINE_STAGE_HOST_BIT,
                                      false);
 
     // We need to submit the current command buffer to the Queue and make sure it finishes before
     // we can copy the data out of the buffer.
     this->submitCommandBuffer(kForce_SyncQueue);
 
     void* mappedMemory = transferBuffer->map();
 
-    size_t tightRowBytes = GrBytesPerPixel(config) * width;
+    if (copyFromOrigin) {
+        uint32_t skipRows = region.imageExtent.height - height;
+        mappedMemory = (char*)mappedMemory + transBufferRowBytes * skipRows + bpp * left;
+    }
+
     if (flipY) {
         const char* srcRow = reinterpret_cast<const char*>(mappedMemory);
         char* dstRow = reinterpret_cast<char*>(buffer)+(height - 1) * rowBytes;
         for (int y = 0; y < height; y++) {
             memcpy(dstRow, srcRow, tightRowBytes);
-            srcRow += tightRowBytes;
+            srcRow += transBufferRowBytes;
             dstRow -= rowBytes;
         }
     } else {
-        if (tightRowBytes == rowBytes) {
+        if (transBufferRowBytes == rowBytes) {
             memcpy(buffer, mappedMemory, rowBytes*height);
         } else {
-            SkRectMemcpy(buffer, rowBytes, mappedMemory, tightRowBytes, tightRowBytes, height);
+            SkRectMemcpy(buffer, rowBytes, mappedMemory, transBufferRowBytes, tightRowBytes,
+                         height);
         }
     }
 
     transferBuffer->unmap();
     transferBuffer->unref();
-
     return true;
 }
 
 // The RenderArea bounds we pass into BeginRenderPass must have a start x value that is a multiple
 // of the granularity. The width must also be a multiple of the granularity or eaqual to the width
 // the the entire attachment. Similar requirements for the y and height components.
 void adjust_bounds_to_granularity(SkIRect* dstBounds, const SkIRect& srcBounds,
                                   const VkExtent2D& granularity, int maxWidth, int maxHeight) {
     // Adjust Width
     // Start with the right side of rect so we know if we end up going pass the maxWidth.
@@ skipping 52 matching lines @@
     // Currently it is fine for us to always pass in 1 for the clear count even if no attachment
     // uses it. In the current state, we also only use the LOAD_OP_CLEAR for the color attachment
     // which is always at the first attachment.
     fCurrentCmdBuffer->beginRenderPass(this, renderPass, 1, colorClear, *target, *pBounds, true);
     fCurrentCmdBuffer->executeCommands(this, buffer);
     fCurrentCmdBuffer->endRenderPass(this);
 
     this->didWriteToSurface(target, &bounds);
 }