Vulkan: Support non-pow2 buffer alignments

Bug: angleproject:2898
Change-Id: I7c8e548a1a449f48225516c37d830ecc201b064d
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/1762496
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>

src/libANGLE/renderer/vulkan/FramebufferVk.cpp

@@ -38,8 +38,9 @@ constexpr unsigned int kEmulatedAlphaValue = 1;
 constexpr size_t kMinReadPixelsBufferSize = 128000;
 
 // Alignment value to accommodate the largest known, for now, uncompressed Vulkan format
-// VK_FORMAT_R64G64B64A64_SFLOAT
-constexpr size_t kReadPixelsBufferAlignment = 32;
+// VK_FORMAT_R64G64B64A64_SFLOAT, while supporting 3-component types such as
+// VK_FORMAT_R16G16B16_SFLOAT.
+constexpr size_t kReadPixelsBufferAlignment = 32 * 3;
 
 // Clear values are only used when loadOp=Clear is set in clearWithRenderPassOp.  When starting a
 // new render pass, the clear value is set to an unlikely value (bright pink) to stand out better

src/libANGLE/renderer/vulkan/vk_helpers.cpp

@@ -330,7 +330,7 @@ angle::Result DynamicBuffer::allocate(ContextVk *contextVk,
                                       VkDeviceSize *offsetOut,
                                       bool *newBufferAllocatedOut)
 {
-    size_t sizeToAllocate = roundUpPow2(sizeInBytes, mAlignment);
+    size_t sizeToAllocate = roundUp(sizeInBytes, mAlignment);
 
     angle::base::CheckedNumeric<size_t> checkedNextWriteOffset = mNextAllocationOffset;
     checkedNextWriteOffset += sizeToAllocate;
@@ -546,17 +546,30 @@ void DynamicBuffer::updateAlignment(RendererVk *renderer, size_t alignment)
     size_t atomSize =
         static_cast<size_t>(renderer->getPhysicalDeviceProperties().limits.nonCoherentAtomSize);
 
-    // We need lcm(alignment, atomSize), we are assuming one divides the other so std::max() could
-    // be used instead.
-    ASSERT(alignment % atomSize == 0 || atomSize % alignment == 0);
-    alignment = std::max(alignment, atomSize);
-    ASSERT(gl::isPow2(alignment));
+    // We need lcm(alignment, atomSize).  Usually, one divides the other so std::max() could be used
+    // instead.  Only known case where this assumption breaks is for 3-component types with 16- or
+    // 32-bit channels, so that's special-cased to avoid a full-fledged lcm implementation.
+
+    if (gl::isPow2(alignment))
+    {
+        ASSERT(alignment % atomSize == 0 || atomSize % alignment == 0);
+        ASSERT(gl::isPow2(atomSize));
+
+        alignment = std::max(alignment, atomSize);
+    }
+    else
+    {
+        ASSERT(gl::isPow2(atomSize));
+        ASSERT(alignment % 3 == 0);
+        ASSERT(gl::isPow2(alignment / 3));
+
+        alignment = std::max(alignment / 3, atomSize) * 3;
+    }
 
     // If alignment has changed, make sure the next allocation is done at an aligned offset.
     if (alignment != mAlignment)
     {
-        mNextAllocationOffset =
-            roundUpPow2(mNextAllocationOffset, static_cast<uint32_t>(alignment));
+        mNextAllocationOffset = roundUp(mNextAllocationOffset, static_cast<uint32_t>(alignment));
     }
 
     mAlignment = alignment;

src/libANGLE/renderer/vulkan/vk_helpers.h

@@ -93,7 +93,6 @@ class DynamicBuffer : angle::NonCopyable
 
     BufferHelper *getCurrentBuffer() { return mBuffer; }
 
-    size_t getAlignment() { return mAlignment; }
     void updateAlignment(RendererVk *renderer, size_t alignment);
 
     // For testing only!