Vulkan: improve staging data performance by using scratch buffers per…

… frame. (#3295)

* Vulkan: improve staging data performance by using scratch buffers per frame.

* vulkan: Add alignment parameter to request scratch space.

* Align staging buffers to texel block size.

* Fix scratch buffer allocation bug.

* Fix some non-deterministic behavior found by Valgrind. Paranoid printing.

* Remove debugging printing

* Fix alignment of converted formats.

* Remove forgotten debug print.

src/bgfx.cpp

@@ -2537,6 +2537,7 @@ namespace bgfx
 
 	void Context::flushTextureUpdateBatch(CommandBuffer& _cmdbuf)
 	{
+		BGFX_PROFILER_SCOPE("flushTextureUpdateBatch", 0xff2040ff);
 		if (m_textureUpdateBatch.sort() )
 		{
 			const uint32_t pos = _cmdbuf.m_pos;

src/bgfx_p.h

@@ -1700,6 +1700,9 @@ namespace bgfx
 					bind.m_idx = kInvalidHandle;
 					bind.m_type = 0;
 					bind.m_samplerFlags = 0;
+					bind.m_format = 0;
+					bind.m_access = 0;
+					bind.m_mip = 0;
 				}
 			}
 		};
@@ -2168,6 +2171,8 @@ namespace bgfx
 			bx::memSet(m_occlusion, 0xff, sizeof(m_occlusion) );
 
 			m_perfStats.viewStats = m_viewStats;
+
+			bx::memSet(&m_renderItemBind[0], 0, sizeof(m_renderItemBind));
 		}
 
 		~Frame()
@@ -2445,6 +2450,13 @@ namespace bgfx
 	{
 		EncoderImpl()
 		{
+			// Although it will be cleared by the discard(), the fact that the
+			// struct is padded to have a size equal to the cache line size,
+			// will leaves bytes uninitialized. This will influence the hashing
+			// as it reads those bytes too. To make this deterministic, we will
+			// clear all bytes (inclusively the padding) before we start.
+			bx::memSet(&m_bind, 0, sizeof(m_bind));
+
 			discard(BGFX_DISCARD_ALL);
 		}
 
@@ -2725,6 +2737,9 @@ namespace bgfx
 				? BGFX_SAMPLER_INTERNAL_DEFAULT
 				: _flags
 				;
+			bind.m_format = 0;
+			bind.m_access = 0;
+			bind.m_mip    = 0;
 
 			if (isValid(_sampler) )
 			{

src/config.h

@@ -324,6 +324,22 @@ BX_STATIC_ASSERT(bx::isPowerOf2(BGFX_CONFIG_MAX_VIEWS), "BGFX_CONFIG_MAX_VIEWS m
 #	define BGFX_CONFIG_TRANSIENT_INDEX_BUFFER_SIZE (2<<20)
 #endif // BGFX_CONFIG_TRANSIENT_INDEX_BUFFER_SIZE
 
+#ifndef BGFX_CONFIG_PER_FRAME_SCRATCH_STAGING_BUFFER_SIZE
+/// Amount of scratch buffer size (per in-flight frame) that will be reserved
+/// for staging data for copying to the device (such as vertex buffer data,
+/// texture data, etc). This buffer will be used instead of allocating memory
+/// on device separately for every data copy.
+/// Note: Currently only used by the Vulkan backend.
+#   define BGFX_CONFIG_PER_FRAME_SCRATCH_STAGING_BUFFER_SIZE (32<<20)
+#endif
+
+#ifndef BGFX_CONFIG_MAX_STAGING_SIZE_FOR_SCRACH_BUFFER
+/// The threshold of data size above which the staging scratch buffer will
+/// not be used, but instead a separate device memory allocation will take
+/// place to stage the data for copying to device.
+#   define BGFX_CONFIG_MAX_STAGING_SIZE_FOR_SCRACH_BUFFER (16 << 20)
+#endif
+
 #ifndef BGFX_CONFIG_MAX_INSTANCE_DATA_COUNT
 #	define BGFX_CONFIG_MAX_INSTANCE_DATA_COUNT 5
 #endif // BGFX_CONFIG_MAX_INSTANCE_DATA_COUNT