Fix non-coherent atom alignment issues

renderdoc/driver/vulkan/vk_initstate.cpp

@@ -1306,9 +1306,11 @@ bool WrappedVulkan::Serialise_InitialState(SerialiserType &ser, ResourceId id, V
     {
       if(initial && initial->mem.mem != VK_NULL_HANDLE)
       {
+        VkDeviceSize size = AlignUp(initial->mem.size, nonCoherentAtomSize);
+
         mappedMem = initial->mem;
-        vkr = ObjDisp(d)->MapMemory(Unwrap(d), Unwrap(mappedMem.mem), initial->mem.offs,
-                                    initial->mem.size, 0, (void **)&Contents);
+        vkr = ObjDisp(d)->MapMemory(Unwrap(d), Unwrap(mappedMem.mem), initial->mem.offs, size, 0,
+                                    (void **)&Contents);
         CheckVkResult(vkr);
 
         // invalidate the cpu cache for this memory range to avoid reading stale data
@@ -1317,7 +1319,7 @@ bool WrappedVulkan::Serialise_InitialState(SerialiserType &ser, ResourceId id, V
             NULL,
             Unwrap(mappedMem.mem),
             mappedMem.offs,
-            mappedMem.size,
+            size,
         };
 
         vkr = ObjDisp(d)->InvalidateMappedMemoryRanges(Unwrap(d), 1, &range);

renderdoc/driver/vulkan/vk_memory.cpp

@@ -207,6 +207,10 @@ MemoryAllocation WrappedVulkan::AllocateMemoryForResource(bool buffer, VkMemoryR
     // offs is where we can put our next sub-allocation
     VkDeviceSize offs = block.offs;
 
+    // for ease, ensure all allocations are allocated to the non-coherent atom size, so we can
+    // invalidate/flush safely. This is at most 256 bytes which is likely already satisfied.
+    offs = AlignUp(offs, nonCoherentAtomSize);
+
     // if we are on a buffer/image, account for any alignment we might have to do
     if(ret.buffer != block.buffer)
       offs = AlignUp(offs, m_PhysicalDeviceData.props.limits.bufferImageGranularity);