Vulkan: Optimize memory allocation offset rounding

src/common/PoolAlloc.cpp

@@ -65,7 +65,7 @@ PoolAllocator::PoolAllocator(int growthIncrement, int allocationAlignment)
         mHeaderSkip = minAlign;
         if (mHeaderSkip < sizeof(Header))
         {
-            mHeaderSkip = rx::roundUp(sizeof(Header), mAlignment);
+            mHeaderSkip = rx::roundUpPow2(sizeof(Header), mAlignment);
         }
     }
     //
@@ -331,4 +331,4 @@ void Allocation::checkAllocList() const
         alloc->check();
 }
 
-}  // namespace angle
\ No newline at end of file
+}  // namespace angle

src/common/mathutil.h

@@ -34,7 +34,7 @@ const unsigned int Float32One   = 0x3F800000;
 const unsigned short Float16One = 0x3C00;
 
 template <typename T>
-inline bool isPow2(T x)
+inline constexpr bool isPow2(T x)
 {
     static_assert(std::is_integral<T>::value, "isPow2 must be called on an integer type.");
     return (x & (x - 1)) == 0 && (x != 0);
@@ -1249,6 +1249,13 @@ T roundUp(const T value, const T alignment)
 }
 
 template <typename T>
+constexpr T roundUpPow2(const T value, const T alignment)
+{
+    ASSERT(gl::isPow2(alignment));
+    return (value + alignment - 1) & ~(alignment - 1);
+}
+
+template <typename T>
 angle::CheckedNumeric<T> CheckedRoundUp(const T value, const T alignment)
 {
     angle::CheckedNumeric<T> checkedValue(value);

src/libANGLE/renderer/d3d/VertexBuffer.cpp

@@ -103,7 +103,7 @@ angle::Result VertexBufferInterface::getSpaceRequired(const gl::Context *context
                                                &spaceRequired));
 
     // Align to 16-byte boundary
-    unsigned int alignedSpaceRequired = roundUp(spaceRequired, 16u);
+    unsigned int alignedSpaceRequired = roundUpPow2(spaceRequired, 16u);
     ANGLE_CHECK_GL_ALLOC(GetImplAs<ContextD3D>(context), alignedSpaceRequired >= spaceRequired);
 
     *spaceInBytesOut = alignedSpaceRequired;

src/libANGLE/renderer/vulkan/BufferVk.cpp

@@ -26,6 +26,7 @@ namespace
 // On some hardware, reading 4 bytes from address 4k returns 0, making it impossible to read the
 // last n bytes.  By rounding up the buffer sizes to a multiple of 4, the problem is alleviated.
 constexpr size_t kBufferSizeGranularity = 4;
+static_assert(gl::isPow2(kBufferSizeGranularity), "use as alignment, must be power of two");
 
 // Start with a fairly small buffer size. We can increase this dynamically as we convert more data.
 constexpr size_t kConvertedArrayBufferInitialSize = 1024 * 8;
@@ -109,7 +110,7 @@ angle::Result BufferVk::setData(const gl::Context *context,
         VkBufferCreateInfo createInfo    = {};
         createInfo.sType                 = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
         createInfo.flags                 = 0;
-        createInfo.size                  = roundUp(size, kBufferSizeGranularity);
+        createInfo.size                  = roundUpPow2(size, kBufferSizeGranularity);
         createInfo.usage                 = usageFlags;
         createInfo.sharingMode           = VK_SHARING_MODE_EXCLUSIVE;
         createInfo.queueFamilyIndexCount = 0;

src/libANGLE/renderer/vulkan/vk_helpers.cpp

@@ -343,7 +343,7 @@ angle::Result DynamicBuffer::allocate(ContextVk *contextVk,
                                       VkDeviceSize *offsetOut,
                                       bool *newBufferAllocatedOut)
 {
-    size_t sizeToAllocate = roundUp(sizeInBytes, mAlignment);
+    size_t sizeToAllocate = roundUpPow2(sizeInBytes, mAlignment);
 
     angle::base::CheckedNumeric<size_t> checkedNextWriteOffset = mNextAllocationOffset;
     checkedNextWriteOffset += sizeToAllocate;
@@ -563,11 +563,13 @@ void DynamicBuffer::updateAlignment(RendererVk *renderer, size_t alignment)
     // be used instead.
     ASSERT(alignment % atomSize == 0 || atomSize % alignment == 0);
     alignment = std::max(alignment, atomSize);
+    ASSERT(gl::isPow2(alignment));
 
     // If alignment has changed, make sure the next allocation is done at an aligned offset.
     if (alignment != mAlignment)
     {
-        mNextAllocationOffset = roundUp(mNextAllocationOffset, static_cast<uint32_t>(alignment));
+        mNextAllocationOffset =
+            roundUpPow2(mNextAllocationOffset, static_cast<uint32_t>(alignment));
     }
 
     mAlignment = alignment;