Add GrVkCopyPipeline to handle vulkan copies as draws

src/gpu/vk/GrVkGpuCommandBuffer.cpp

 /*
 * Copyright 2016 Google Inc.
 *
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */
 
 #include "GrVkGpuCommandBuffer.h"
 
 #include "GrFixedClip.h"
@@ skipping 100 matching lines @@
         // We have sumbitted no actual draw commands to the command buffer and we are not using
         // the render pass to do a clear so there is no need to submit anything.
         return;
     }
 #endif
 
     // Change layout of our render target so it can be used as the color attachment. Currently
     // we don't attach the resolve to the framebuffer so no need to change its layout.
     GrVkImage* targetImage = fRenderTarget->msaaImage() ? fRenderTarget->msaaImage()
                                                         : fRenderTarget;
+
+    // Change layout of our render target so it can be used as the color attachment
     targetImage->setImageLayout(fGpu,
                                 VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
                                 VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
-                                VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                                VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,

[jvanverth1, 2016/09/26 17:26:11]

Why this change? (just curious)

[egdaniel, 2016/09/26 21:12:39]

Just from threads about these barriers with other vulkan folk I got the feeling
that TOP/Bottom of pipe weren't really "real" stages and that just using
ALL_GRAPHICS or ALL_STAGES is a safer way to get the results I want (even though
they really should be equivalent).

                                 false);
 
     // If we are using a stencil attachment we also need to update its layout
     if (GrStencilAttachment* stencil = fRenderTarget->renderTargetPriv().getStencilAttachment()) {
         GrVkStencilAttachment* vkStencil = (GrVkStencilAttachment*)stencil;
         vkStencil->setImageLayout(fGpu,
                                   VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL,
                                   VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
                                   VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT,
-                                  VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                                  VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,
                                   false);
     }
 
     fGpu->submitSecondaryCommandBuffer(fCommandBuffer, fRenderPass, &fColorClearValue,
                                        fRenderTarget, bounds);
 }
 
 void GrVkGpuCommandBuffer::discard(GrRenderTarget* target) {
     if (fIsEmpty) {
         // We will change the render pass to do a clear load instead
@@ skipping 306 matching lines @@
             fGpu->stats()->incNumDraws();
         } while ((nonIdxMesh = iter.next()));
     }
 
     // Technically we don't have to call this here (since there is a safety check in
     // pipelineState:setData but this will allow for quicker freeing of resources if the
     // pipelineState sits in a cache for a while.
     pipelineState->freeTempResources(fGpu);
 }