Added Init.fallback option. Cleanup. (#3522)

bindings/bf/bgfx.bf

@@ -2237,6 +2237,7 @@ public static class bgfx
 		public uint64 capabilities;
 		public uint8 debug;
 		public uint8 profile;
+		public uint8 fallback;
 		public PlatformData platformData;
 		public Resolution resolution;
 		public Limits limits;

bindings/c3/bgfx.c3

@@ -1581,6 +1581,8 @@ struct Init
 	bool debug;
 	// Enable device for profiling.
 	bool profile;
+	// Enable fallback to next available renderer.
+	bool fallback;
 	// Platform data.
 	PlatformData platformData;
 	// Backbuffer resolution and reset parameters. See: `bgfx::Resolution`.

bindings/cs/bgfx.cs

@@ -2211,6 +2211,7 @@ public static partial class bgfx
 		public ulong capabilities;
 		public byte debug;
 		public byte profile;
+		public byte fallback;
 		public PlatformData platformData;
 		public Resolution resolution;
 		public Limits limits;

bindings/d/package.d

@@ -9,7 +9,7 @@ import bindbc.common.types: c_int64, c_uint64, va_list;
 import bindbc.bgfx.config;
 static import bgfx.impl;
 
-enum uint apiVersion = 135;
+enum uint apiVersion = 136;
 
 alias ViewID = ushort;
 
@@ -1199,6 +1199,7 @@ extern(C++, "bgfx") struct Init{
 	c_uint64 capabilities; ///Capabilities initialization mask (default: UINT64_MAX).
 	bool debug_; ///Enable device for debugging.
 	bool profile; ///Enable device for profiling.
+	bool fallback; ///Enable fallback to next available renderer.
 	PlatformData platformData; ///Platform data.
 	Resolution resolution; ///Backbuffer resolution and reset parameters. See: `bgfx::Resolution`.
 	Limits limits; ///Configurable runtime limits parameters.

bindings/zig/bgfx.zig

@@ -1419,6 +1419,7 @@ pub const Init = extern struct {
         capabilities: u64,
         debug: bool,
         profile: bool,
+        fallback: bool,
         platformData: PlatformData,
         resolution: Resolution,
         limits: Limits,

include/bgfx/bgfx.h

@@ -739,8 +739,9 @@ namespace bgfx
 
 		uint64_t capabilities; //!< Capabilities initialization mask (default: UINT64_MAX).
 
-		bool debug;   //!< Enable device for debugging.
-		bool profile; //!< Enable device for profiling.
+		bool debug;    //!< Enable device for debugging.
+		bool profile;  //!< Enable device for profiling.
+		bool fallback; //!< Enable fallback to next available renderer.
 
 		/// Platform data.
 		PlatformData platformData;

include/bgfx/c99/bgfx.h

@@ -742,6 +742,7 @@ typedef struct bgfx_init_s
     uint64_t             capabilities;       /** Capabilities initialization mask (default: UINT64_MAX). */
     bool                 debug;              /** Enable device for debugging.             */
     bool                 profile;            /** Enable device for profiling.             */
+    bool                 fallback;           /** Enable fallback to next available renderer. */
     bgfx_platform_data_t platformData;       /** Platform data.                           */
     bgfx_resolution_t    resolution;         /** Backbuffer resolution and reset parameters. See: `bgfx::Resolution`. */
     bgfx_init_limits_t   limits;             /** Configurable runtime limits parameters.  */

include/bgfx/defines.h

@@ -15,7 +15,7 @@
 #ifndef BGFX_DEFINES_H_HEADER_GUARD
 #define BGFX_DEFINES_H_HEADER_GUARD
 
-#define BGFX_API_VERSION UINT32_C(135)
+#define BGFX_API_VERSION UINT32_C(136)
 
 /**
  * Color RGB/alpha/depth write. When it's not specified write will be disabled.

scripts/bgfx.idl

@@ -1,7 +1,7 @@
 -- vim: syntax=lua
 -- bgfx interface
 
-version(135)
+version(136)
 
 typedef "bool"
 typedef "char"
@@ -858,6 +858,7 @@ struct.Init { ctor }
 	.capabilities "uint64_t"           --- Capabilities initialization mask (default: UINT64_MAX).
 	.debug        "bool"               --- Enable device for debugging.
 	.profile      "bool"               --- Enable device for profiling.
+	.fallback     "bool"               --- Enable fallback to next available renderer.
 	.platformData "PlatformData"       --- Platform data.
 	.resolution   "Resolution"         --- Backbuffer resolution and reset parameters. See: `bgfx::Resolution`.
 	.limits       "Limits"             --- Configurable runtime limits parameters.

src/bgfx.cpp

@@ -737,10 +737,10 @@ namespace bgfx
 			.add(Attrib::TexCoord0, 2, AttribType::Float)
 			.end();
 
-		uint16_t width  = 2048;
-		uint16_t height = 24;
-		uint8_t  bpp    = 1;
-		uint32_t pitch  = width*bpp;
+		const uint16_t width  = 2048;
+		const uint16_t height = 24;
+		const uint8_t  bpp    = 1;
+		const uint32_t pitch  = width*bpp;
 
 		const Memory* mem;
 
@@ -767,6 +767,7 @@ namespace bgfx
 		m_vb = s_ctx->createTransientVertexBuffer(kNumBatchVertices*m_layout.m_stride, &m_layout);
 		m_ib = s_ctx->createTransientIndexBuffer(kNumBatchIndices*2);
 		m_scale = bx::max<uint8_t>(scale, 1);
+		m_usedData = 0;
 	}
 
 	void TextVideoMemBlitter::shutdown()
@@ -825,7 +826,7 @@ namespace bgfx
 	};
 	static_assert(BX_COUNTOF(s_paletteLinear) == 16);
 
-	void blit(RendererContextI* _renderCtx, TextVideoMemBlitter& _blitter, const TextVideoMem& _mem)
+	void dbgTextSubmit(RendererContextI* _renderCtx, TextVideoMemBlitter& _blitter, const TextVideoMem& _mem)
 	{
 		struct Vertex
 		{
@@ -848,7 +849,7 @@ namespace bgfx
 		const float fontHeight  = (_mem.m_small ? 8.0f : 16.0f)*_blitter.m_scale;
 		const float fontWidth   = 8.0f * _blitter.m_scale;
 
-		_renderCtx->blitSetup(_blitter);
+		_renderCtx->dbgTextRenderBegin(_blitter);
 
 		const uint32_t* palette = 0 != (s_ctx->m_init.resolution.reset & BGFX_RESET_SRGB_BACKBUFFER)
 			? s_paletteLinear
@@ -857,10 +858,10 @@ namespace bgfx
 
 		for (;yy < _mem.m_height;)
 		{
-			Vertex* vertex = (Vertex*)_blitter.m_vb->data;
+			Vertex*   vertex  = (Vertex  *)_blitter.m_vb->data;
 			uint16_t* indices = (uint16_t*)_blitter.m_ib->data;
 			uint32_t startVertex = 0;
-			uint32_t numIndices = 0;
+			uint32_t numIndices  = 0;
 
 			for (; yy < _mem.m_height && numIndices < kNumBatchIndices; ++yy)
 			{
@@ -883,7 +884,7 @@ namespace bgfx
 						const uint32_t fg = palette[attr&0xf];
 						const uint32_t bg = palette[(attr>>4)&0xf];
 
-						Vertex vert[4] =
+						const Vertex vert[4] =
 						{
 							{ (xx  )*fontWidth, (yy  )*fontHeight, 0.0f, fg, bg, (ch  )*8.0f*texelWidth, utop },
 							{ (xx+1)*fontWidth, (yy  )*fontHeight, 0.0f, fg, bg, (ch+1)*8.0f*texelWidth, utop },
@@ -916,8 +917,10 @@ namespace bgfx
 				}
 			}
 
-			_renderCtx->blitRender(_blitter, numIndices);
+			_renderCtx->dbgTextRender(_blitter, numIndices);
 		}
+
+		_renderCtx->dbgTextRenderEnd(_blitter);
 	}
 
 	void ClearQuad::init()
@@ -926,7 +929,8 @@ namespace bgfx
 
 		if (RendererType::Noop != g_caps.rendererType)
 		{
-			m_layout
+			VertexLayout layout;
+			layout
 				.begin()
 				.add(Attrib::Position, 2, AttribType::Float)
 				.end();
@@ -954,7 +958,7 @@ namespace bgfx
 				float m_y;
 			};
 
-			const uint16_t stride = m_layout.m_stride;
+			const uint16_t stride = layout.m_stride;
 			const bgfx::Memory* mem = bgfx::alloc(4 * stride);
 			Vertex* vertex = (Vertex*)mem->data;
 			BX_ASSERT(stride == sizeof(Vertex), "Stride/Vertex mismatch (stride %d, sizeof(Vertex) %d)", stride, sizeof(Vertex));
@@ -971,7 +975,8 @@ namespace bgfx
 			vertex->m_x = 1.0f;
 			vertex->m_y = 1.0f;
 
-			m_vb = s_ctx->createVertexBuffer(mem, m_layout, 0);
+			m_layout = s_ctx->createVertexLayout(layout);
+			m_vb     = s_ctx->createVertexBuffer(mem, layout, 0);
 		}
 	}
 
@@ -991,6 +996,7 @@ namespace bgfx
 			}
 
 			s_ctx->destroyVertexBuffer(m_vb);
+			s_ctx->destroyVertexLayout(m_layout);
 		}
 	}
 
@@ -1351,15 +1357,9 @@ namespace bgfx
 		if (UINT8_MAX != m_draw.m_streamMask)
 		{
 			uint32_t numVertices = UINT32_MAX;
-			for (uint32_t idx = 0, streamMask = m_draw.m_streamMask
-				; 0 != streamMask
-				; streamMask >>= 1, idx += 1
-				)
+			for (BitMaskToIndexIteratorT it(m_draw.m_streamMask); !it.isDone(); it.next() )
 			{
-				const uint32_t ntz = bx::uint32_cnttz(streamMask);
-				streamMask >>= ntz;
-				idx         += ntz;
-				numVertices = bx::min(numVertices, m_numVertices[idx]);
+				numVertices = bx::min(numVertices, m_numVertices[it.idx]);
 			}
 
 			m_draw.m_numVertices = numVertices;
@@ -1444,30 +1444,31 @@ namespace bgfx
 			, "Exceed number of available blit items per frame. BGFX_CONFIG_MAX_BLIT_ITEMS is %d. Skipping blit."
 			, BGFX_CONFIG_MAX_BLIT_ITEMS
 			);
-		if (m_frame->m_numBlitItems < BGFX_CONFIG_MAX_BLIT_ITEMS)
+		const uint32_t blitItemIdx = bx::atomicFetchAndAddsat<uint32_t>(&m_frame->m_numBlitItems, 1, BGFX_CONFIG_MAX_BLIT_ITEMS);
+		if (BGFX_CONFIG_MAX_BLIT_ITEMS-1 <= blitItemIdx)
 		{
-			uint16_t item = m_frame->m_numBlitItems++;
-
-			BlitItem& bi = m_frame->m_blitItem[item];
-			bi.m_srcX    = _srcX;
-			bi.m_srcY    = _srcY;
-			bi.m_srcZ    = _srcZ;
-			bi.m_dstX    = _dstX;
-			bi.m_dstY    = _dstY;
-			bi.m_dstZ    = _dstZ;
-			bi.m_width   = _width;
-			bi.m_height  = _height;
-			bi.m_depth   = _depth;
-			bi.m_srcMip  = _srcMip;
-			bi.m_dstMip  = _dstMip;
-			bi.m_src     = _src;
-			bi.m_dst     = _dst;
-
-			BlitKey key;
-			key.m_view = _id;
-			key.m_item = item;
-			m_frame->m_blitKeys[item] = key.encode();
+			return;
 		}
+
+		BlitItem& bi = m_frame->m_blitItem[blitItemIdx];
+		bi.m_srcX   = _srcX;
+		bi.m_srcY   = _srcY;
+		bi.m_srcZ   = _srcZ;
+		bi.m_dstX   = _dstX;
+		bi.m_dstY   = _dstY;
+		bi.m_dstZ   = _dstZ;
+		bi.m_width  = _width;
+		bi.m_height = _height;
+		bi.m_depth  = _depth;
+		bi.m_srcMip = _srcMip;
+		bi.m_dstMip = _dstMip;
+		bi.m_src    = _src;
+		bi.m_dst    = _dst;
+
+		BlitKey key;
+		key.m_view = _id;
+		key.m_item = bx::narrowCast<uint16_t>(blitItemIdx);
+		m_frame->m_blitKeys[blitItemIdx] = key.encode();
 	}
 
 	void Frame::sort()
@@ -2479,7 +2480,7 @@ namespace bgfx
 			, m_init.resolution.height
 			);
 
-		int64_t now = bx::getHPCounter();
+		const int64_t now = bx::getHPCounter();
 		m_submit->m_perfStats.cpuTimeFrame = now - m_frameTimeLast;
 		m_frameTimeLast = now;
 	}
@@ -2640,8 +2641,6 @@ namespace bgfx
 		{
 			const uint32_t pos = _cmdbuf.m_pos;
 
-			uint32_t currentKey = UINT32_MAX;
-
 			for (uint32_t ii = 0, num = m_textureUpdateBatch.m_num; ii < num; ++ii)
 			{
 				_cmdbuf.m_pos = m_textureUpdateBatch.m_values[ii];
@@ -2670,27 +2669,11 @@ namespace bgfx
 				const Memory* mem;
 				_cmdbuf.read(mem);
 
-				uint32_t key = m_textureUpdateBatch.m_keys[ii];
-				if (key != currentKey)
-				{
-					if (currentKey != UINT32_MAX)
-					{
-						m_renderCtx->updateTextureEnd();
-					}
-					currentKey = key;
-					m_renderCtx->updateTextureBegin(handle, side, mip);
-				}
-
 				m_renderCtx->updateTexture(handle, side, mip, rect, zz, depth, pitch, mem);
 
 				release(mem);
 			}
 
-			if (currentKey != UINT32_MAX)
-			{
-				m_renderCtx->updateTextureEnd();
-			}
-
 			m_textureUpdateBatch.reset();
 
 			_cmdbuf.m_pos = pos;
@@ -2716,6 +2699,7 @@ namespace bgfx
 	BGFX_RENDERER_CONTEXT(nvn);
 	BGFX_RENDERER_CONTEXT(gl);
 	BGFX_RENDERER_CONTEXT(vk);
+	BGFX_RENDERER_CONTEXT(wgpu);
 
 #undef BGFX_RENDERER_CONTEXT
 
@@ -2948,7 +2932,8 @@ namespace bgfx
 		{
 			RendererType::Enum renderer = RendererType::Enum(scores[ii] & 0xff);
 			renderCtx = s_rendererCreator[renderer].createFn(_init);
-			if (NULL != renderCtx)
+			if (NULL != renderCtx
+			||  !_init.fallback)
 			{
 				break;
 			}
@@ -3643,6 +3628,7 @@ namespace bgfx
 		, capabilities(UINT64_MAX)
 		, debug(BX_ENABLED(BGFX_CONFIG_DEBUG) )
 		, profile(BX_ENABLED(BGFX_CONFIG_DEBUG_ANNOTATION) )
+		, fallback(true)
 		, callback(NULL)
 		, allocator(NULL)
 	{

src/bgfx_p.h

@@ -496,7 +496,7 @@ namespace bgfx
 		uint16_t m_flags;
 	};
 
-	struct Rect
+	BX_ALIGN_DECL(8, struct) Rect
 	{
 		Rect()
 		{
@@ -520,21 +520,34 @@ namespace bgfx
 
 		bool isZero() const
 		{
+			static_assert(8 == sizeof(Rect), "");
+
 			uint64_t ui64 = *( (uint64_t*)this);
 			return UINT64_C(0) == ui64;
 		}
 
 		bool isZeroArea() const
 		{
-			return 0 == m_width
+			return false
+				|| 0 == m_width
 				|| 0 == m_height
 				;
 		}
 
+		bool isEqual(const Rect& _other) const
+		{
+			return true
+				&& m_x      == _other.m_x
+				&& m_y      == _other.m_y
+				&& m_width  == _other.m_width
+				&& m_height == _other.m_height
+				;
+		}
+
 		void set(uint16_t _x, uint16_t _y, uint16_t _width, uint16_t _height)
 		{
-			m_x = _x;
-			m_y = _y;
+			m_x      = _x;
+			m_y      = _y;
 			m_width  = _width;
 			m_height = _height;
 		}
@@ -545,8 +558,8 @@ namespace bgfx
 			const uint16_t sy = bx::max<uint16_t>(_a.m_y, _b.m_y);
 			const uint16_t ex = bx::min<uint16_t>(_a.m_x + _a.m_width,  _b.m_x + _b.m_width );
 			const uint16_t ey = bx::min<uint16_t>(_a.m_y + _a.m_height, _b.m_y + _b.m_height);
-			m_x = sx;
-			m_y = sy;
+			m_x      = sx;
+			m_y      = sy;
 			m_width  = (uint16_t)bx::uint32_satsub(ex, sx);
 			m_height = (uint16_t)bx::uint32_satsub(ey, sy);
 		}
@@ -627,6 +640,12 @@ namespace bgfx
 		return uint32_t( (_stencil >> (32*_0or1) ) );
 	}
 
+	static constexpr uint64_t kStencilNoRefMask = packStencil(~BGFX_STENCIL_FUNC_REF_MASK, ~BGFX_STENCIL_FUNC_REF_MASK);
+	static constexpr uint64_t kStencilDisabled  = packStencil(
+		  BGFX_STENCIL_TEST_ALWAYS | BGFX_STENCIL_OP_FAIL_S_KEEP | BGFX_STENCIL_OP_FAIL_Z_KEEP | BGFX_STENCIL_OP_PASS_Z_KEEP
+		, BGFX_STENCIL_TEST_ALWAYS | BGFX_STENCIL_OP_FAIL_S_KEEP | BGFX_STENCIL_OP_FAIL_Z_KEEP | BGFX_STENCIL_OP_PASS_Z_KEEP
+		);
+
 	inline constexpr bool needBorderColor(uint64_t _flags)
 	{
 		return BGFX_SAMPLER_U_BORDER == (_flags & BGFX_SAMPLER_U_BORDER)
@@ -765,19 +784,21 @@ namespace bgfx
 
 		TextureHandle m_texture;
 		TransientVertexBuffer* m_vb;
-		TransientIndexBuffer* m_ib;
-		VertexLayout m_layout;
+		TransientIndexBuffer*  m_ib;
+		VertexLayout  m_layout;
 		ProgramHandle m_program;
 		uint8_t m_scale;
+
+		uintptr_t m_usedData;
 	};
 
 	struct RendererContextI;
 
-	extern void blit(RendererContextI* _renderCtx, TextVideoMemBlitter& _blitter, const TextVideoMem& _mem);
+	extern void dbgTextSubmit(RendererContextI* _renderCtx, TextVideoMemBlitter& _blitter, const TextVideoMem& _mem);
 
-	inline void blit(RendererContextI* _renderCtx, TextVideoMemBlitter& _blitter, const TextVideoMem* _mem)
+	inline void dbgTextSubmit(RendererContextI* _renderCtx, TextVideoMemBlitter& _blitter, const TextVideoMem* _mem)
 	{
-		blit(_renderCtx, _blitter, *_mem);
+		dbgTextSubmit(_renderCtx, _blitter, *_mem);
 	}
 
 	template <uint32_t maxKeys>
@@ -790,8 +811,8 @@ namespace bgfx
 
 		void add(uint32_t _key, uint32_t _value)
 		{
-			uint32_t num = m_num++;
-			m_keys[num] = _key;
+			const uint32_t num = m_num++;
+			m_keys  [num] = _key;
 			m_values[num] = _value;
 		}
 
@@ -799,7 +820,7 @@ namespace bgfx
 		{
 			if (0 < m_num)
 			{
-				uint32_t* tempKeys = (uint32_t*)BX_STACK_ALLOC(sizeof(m_keys) );
+				uint32_t* tempKeys   = (uint32_t*)BX_STACK_ALLOC(sizeof(m_keys) );
 				uint32_t* tempValues = (uint32_t*)BX_STACK_ALLOC(sizeof(m_values) );
 				bx::radixSort(m_keys, tempKeys, m_values, tempValues, m_num);
 				return true;
@@ -823,6 +844,32 @@ namespace bgfx
 		uint32_t m_values[maxKeys];
 	};
 
+	template<typename MaskT>
+	struct BitMaskToIndexIteratorT
+	{
+		BitMaskToIndexIteratorT(MaskT _mask)
+		{
+			const uint8_t ntz = bx::countTrailingZeros(_mask);
+			mask = _mask >> ntz;
+			idx  = ntz;
+		}
+
+		void next()
+		{
+			const uint8_t ntzPlus1 = bx::countTrailingZeros(mask>>1) + 1;
+			mask >>= ntzPlus1;
+			idx   += ntzPlus1;
+		}
+
+		bool isDone() const
+		{
+			return 0 == mask;
+		}
+
+		MaskT   mask;
+		uint8_t idx;
+	};
+
 	struct ClearQuad
 	{
 		ClearQuad()
@@ -837,7 +884,7 @@ namespace bgfx
 		void shutdown();
 
 		VertexBufferHandle m_vb;
-		VertexLayout m_layout;
+		VertexLayoutHandle m_layout;
 		ProgramHandle m_program[BGFX_CONFIG_MAX_FRAME_BUFFER_ATTACHMENTS];
 	};
 
@@ -2537,7 +2584,7 @@ namespace bgfx
 		UniformBuffer** m_uniformBuffer;
 
 		uint32_t m_numRenderItems;
-		uint16_t m_numBlitItems;
+		uint32_t m_numBlitItems;
 
 		uint32_t m_iboffset;
 		uint32_t m_vboffset;
@@ -3564,9 +3611,7 @@ namespace bgfx
 		virtual void createProgram(ProgramHandle _handle, ShaderHandle _vsh, ShaderHandle _fsh) = 0;
 		virtual void destroyProgram(ProgramHandle _handle) = 0;
 		virtual void* createTexture(TextureHandle _handle, const Memory* _mem, uint64_t _flags, uint8_t _skip) = 0;
-		virtual void updateTextureBegin(TextureHandle _handle, uint8_t _side, uint8_t _mip) = 0;
 		virtual void updateTexture(TextureHandle _handle, uint8_t _side, uint8_t _mip, const Rect& _rect, uint16_t _z, uint16_t _depth, uint16_t _pitch, const Memory* _mem) = 0;
-		virtual void updateTextureEnd() = 0;
 		virtual void readTexture(TextureHandle _handle, void* _data, uint8_t _mip) = 0;
 		virtual void resizeTexture(TextureHandle _handle, uint16_t _width, uint16_t _height, uint8_t _numMips, uint16_t _numLayers) = 0;
 		virtual void overrideInternal(TextureHandle _handle, uintptr_t _ptr, uint16_t _layerIndex) = 0;
@@ -3584,8 +3629,9 @@ namespace bgfx
 		virtual void setMarker(const char* _name, uint16_t _len) = 0;
 		virtual void setName(Handle _handle, const char* _name, uint16_t _len) = 0;
 		virtual void submit(Frame* _render, ClearQuad& _clearQuad, TextVideoMemBlitter& _textVideoMemBlitter) = 0;
-		virtual void blitSetup(TextVideoMemBlitter& _blitter) = 0;
-		virtual void blitRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) = 0;
+		virtual void dbgTextRenderBegin(TextVideoMemBlitter& _blitter) = 0;
+		virtual void dbgTextRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) = 0;
+		virtual void dbgTextRenderEnd(TextVideoMemBlitter& _blitter) = 0;
 	};
 
 	inline RendererContextI::~RendererContextI()
@@ -4548,8 +4594,8 @@ namespace bgfx
 			BX_WARN(isValid(handle), "Failed to allocate draw indirect buffer handle.");
 			if (isValid(handle) )
 			{
-				uint32_t size  = _num * BGFX_CONFIG_DRAW_INDIRECT_STRIDE;
-				uint16_t flags = BGFX_BUFFER_DRAW_INDIRECT;
+				const uint32_t size  = _num * BGFX_CONFIG_DRAW_INDIRECT_STRIDE;
+				const uint16_t flags = BGFX_BUFFER_DRAW_INDIRECT;
 
 				CommandBuffer& cmdbuf = getCommandBuffer(CommandBuffer::CreateDynamicVertexBuffer);
 				cmdbuf.write(handle);

src/config.h

@@ -94,8 +94,8 @@
 #		define BGFX_CONFIG_RENDERER_OPENGLES (0 \
 					|| BX_PLATFORM_ANDROID      \
 					|| BX_PLATFORM_EMSCRIPTEN   \
-					|| BX_PLATFORM_RPI          \
 					|| BX_PLATFORM_NX           \
+					|| BX_PLATFORM_RPI          \
 					? BGFX_CONFIG_RENDERER_OPENGLES_MIN_VERSION : 0)
 #	endif // BGFX_CONFIG_RENDERER_OPENGLES
 
@@ -103,9 +103,9 @@
 #		define BGFX_CONFIG_RENDERER_VULKAN (0 \
 					|| BX_PLATFORM_ANDROID    \
 					|| BX_PLATFORM_LINUX      \
-					|| BX_PLATFORM_WINDOWS    \
 					|| BX_PLATFORM_NX         \
 					|| BX_PLATFORM_OSX        \
+					|| BX_PLATFORM_WINDOWS    \
 					? 1 : 0)
 #	endif // BGFX_CONFIG_RENDERER_VULKAN
 

src/renderer.h

@@ -503,14 +503,9 @@ namespace bgfx
 			return true;
 		}
 
-		for (uint32_t idx = 0, streamMask = _new.m_streamMask
-			; 0 != streamMask
-			; streamMask >>= 1, idx += 1
-			)
+		for (BitMaskToIndexIteratorT it(_new.m_streamMask); !it.isDone(); it.next() )
 		{
-			const uint32_t ntz = bx::uint32_cnttz(streamMask);
-			streamMask >>= ntz;
-			idx         += ntz;
+			const uint8_t idx = it.idx;
 
 			if (_current.m_stream[idx].m_handle.idx  != _new.m_stream[idx].m_handle.idx
 			||  _current.m_stream[idx].m_startVertex != _new.m_stream[idx].m_startVertex)
@@ -544,7 +539,7 @@ namespace bgfx
 		{
 			if (m_enabled)
 			{
-				ViewStats& viewStats = m_frame->m_perfStats.viewStats[m_numViews];
+				ViewStats& viewStats   = m_frame->m_perfStats.viewStats[m_numViews];
 				viewStats.cpuTimeBegin = bx::getHPCounter();
 
 				m_queryIdx = m_gpuTimer.begin(_view, m_frame->m_frameNum);
@@ -567,10 +562,10 @@ namespace bgfx
 				ViewStats& viewStats = m_frame->m_perfStats.viewStats[m_numViews];
 				const typename Ty::Result& result = m_gpuTimer.m_result[viewStats.view];
 
-				viewStats.cpuTimeEnd = bx::getHPCounter();
+				viewStats.cpuTimeEnd   = bx::getHPCounter();
 				viewStats.gpuTimeBegin = result.m_begin;
-				viewStats.gpuTimeEnd = result.m_end;
-				viewStats.gpuFrameNum = result.m_frameNum;
+				viewStats.gpuTimeEnd   = result.m_end;
+				viewStats.gpuFrameNum  = result.m_frameNum;
 
 				++m_numViews;
 				m_queryIdx = UINT32_MAX;

src/renderer_d3d11.cpp

@@ -1841,19 +1841,11 @@ namespace bgfx { namespace d3d11
 			return m_textures[_handle.idx].create(_mem, _flags, _skip);
 		}
 
-		void updateTextureBegin(TextureHandle /*_handle*/, uint8_t /*_side*/, uint8_t /*_mip*/) override
-		{
-		}
-
 		void updateTexture(TextureHandle _handle, uint8_t _side, uint8_t _mip, const Rect& _rect, uint16_t _z, uint16_t _depth, uint16_t _pitch, const Memory* _mem) override
 		{
 			m_textures[_handle.idx].update(_side, _mip, _rect, _z, _depth, _pitch, _mem);
 		}
 
-		void updateTextureEnd() override
-		{
-		}
-
 		void readTexture(TextureHandle _handle, void* _data, uint8_t _mip) override
 		{
 			const TextureD3D11& texture = m_textures[_handle.idx];
@@ -2131,7 +2123,7 @@ namespace bgfx { namespace d3d11
 
 		void submit(Frame* _render, ClearQuad& _clearQuad, TextVideoMemBlitter& _textVideoMemBlitter) override;
 
-		void blitSetup(TextVideoMemBlitter& _blitter) override
+		void dbgTextRenderBegin(TextVideoMemBlitter& _blitter) override
 		{
 			ID3D11DeviceContext* deviceCtx = m_deviceCtx;
 
@@ -2187,7 +2179,7 @@ namespace bgfx { namespace d3d11
 			commitTextureStage();
 		}
 
-		void blitRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) override
+		void dbgTextRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) override
 		{
 			const uint32_t numVertices = _numIndices*4/6;
 			if (0 < numVertices)
@@ -2202,6 +2194,10 @@ namespace bgfx { namespace d3d11
 			}
 		}
 
+		void dbgTextRenderEnd(TextVideoMemBlitter& /*_blitter*/) override
+		{
+		}
+
 		void preReset()
 		{
 			m_needPresent = false;
@@ -2898,7 +2894,7 @@ namespace bgfx { namespace d3d11
 
 			uint32_t fstencil = unpackStencil(0, _stencil);
 			uint32_t ref = (fstencil&BGFX_STENCIL_FUNC_REF_MASK)>>BGFX_STENCIL_FUNC_REF_SHIFT;
-			_stencil &= packStencil(~BGFX_STENCIL_FUNC_REF_MASK, ~BGFX_STENCIL_FUNC_REF_MASK);
+			_stencil &= kStencilNoRefMask;
 
 			bx::HashMurmur2A murmur;
 			murmur.begin();
@@ -3453,7 +3449,7 @@ namespace bgfx { namespace d3d11
 			}
 		}
 
-		void clearQuad(ClearQuad& _clearQuad, const Rect& _rect, const Clear& _clear, const float _palette[][4])
+		void clearQuad(const ClearQuad& _clearQuad, const Rect& _rect, const Clear& _clear, const float _palette[][4])
 		{
 			uint32_t width;
 			uint32_t height;
@@ -3470,10 +3466,9 @@ namespace bgfx { namespace d3d11
 				height = m_scd.height;
 			}
 
-			if (0      == _rect.m_x
-			&&  0      == _rect.m_y
-			&&  width  == _rect.m_width
-			&&  height == _rect.m_height)
+			const Rect fbRect(0, 0, bx::narrowCast<uint16_t>(width), bx::narrowCast<uint16_t>(height) );
+
+			if (_rect.isEqual(fbRect) )
 			{
 				clear(_clear, _palette);
 			}
@@ -3514,7 +3509,7 @@ namespace bgfx { namespace d3d11
 				const ShaderD3D11* vsh = program.m_vsh;
 				deviceCtx->VSSetShader(vsh->m_vertexShader, NULL, 0);
 				deviceCtx->VSSetConstantBuffers(0, 1, &vsh->m_buffer);
-				float mrtClearDepth[4] = { _clear.m_depth };
+				const float mrtClearDepth[4] = { _clear.m_depth };
 				deviceCtx->UpdateSubresource(vsh->m_buffer, 0, 0, mrtClearDepth, 0, 0);
 
 				if (NULL != m_currentColor)
@@ -3555,8 +3550,8 @@ namespace bgfx { namespace d3d11
 					deviceCtx->PSSetShader(NULL, NULL, 0);
 				}
 
-				VertexBufferD3D11& vb = m_vertexBuffers[_clearQuad.m_vb.idx];
-				const VertexLayout& layout = _clearQuad.m_layout;
+				const VertexBufferD3D11& vb = m_vertexBuffers[_clearQuad.m_vb.idx];
+				const VertexLayout&  layout = m_vertexLayouts[_clearQuad.m_layout.idx];
 
 				const uint32_t stride = layout.m_stride;
 				const uint32_t offset = 0;
@@ -6260,14 +6255,12 @@ namespace bgfx { namespace d3d11
 
 					if (UINT8_MAX != draw.m_streamMask)
 					{
-						for (uint32_t idx = 0, streamMask = draw.m_streamMask
-							; 0 != streamMask
-							; streamMask >>= 1, idx += 1, ++numStreams
+						for (BitMaskToIndexIteratorT it(draw.m_streamMask)
+							; !it.isDone()
+							; it.next(), numStreams++
 							)
 						{
-							const uint32_t ntz = bx::uint32_cnttz(streamMask);
-							streamMask >>= ntz;
-							idx         += ntz;
+							const uint8_t idx = it.idx;
 
 							currentState.m_stream[idx].m_layoutHandle = draw.m_stream[idx].m_layoutHandle;
 							currentState.m_stream[idx].m_handle       = draw.m_stream[idx].m_handle;
@@ -6476,11 +6469,11 @@ namespace bgfx { namespace d3d11
 									);
 							}
 						}
+					}
 
-						if (hasOcclusionQuery)
-						{
-							m_occlusionQuery.end();
-						}
+					if (hasOcclusionQuery)
+					{
+						m_occlusionQuery.end();
 					}
 
 					statsNumPrimsSubmitted[primIndex] += numPrimsSubmitted;
@@ -6694,7 +6687,7 @@ namespace bgfx { namespace d3d11
 				max = frameTime;
 			}
 
-			blit(this, _textVideoMemBlitter, tvm);
+			dbgTextSubmit(this, _textVideoMemBlitter, tvm);
 
 			BGFX_D3D11_PROFILER_END();
 		}
@@ -6702,7 +6695,7 @@ namespace bgfx { namespace d3d11
 		{
 			BGFX_D3D11_PROFILER_BEGIN_LITERAL("debugtext", kColorFrame);
 
-			blit(this, _textVideoMemBlitter, _render->m_textVideoMem);
+			dbgTextSubmit(this, _textVideoMemBlitter, _render->m_textVideoMem);
 
 			BGFX_D3D11_PROFILER_END();
 		}

src/renderer_d3d12.cpp

@@ -1868,7 +1868,7 @@ namespace bgfx { namespace d3d12
 					hr = m_swapChain->Present(syncInterval, presentFlags);
 				}
 
-				int64_t now = bx::getHPCounter();
+				const int64_t now = bx::getHPCounter();
 				m_presentElapsed = now - start;
 
 				m_lost = isLost(hr);
@@ -1969,20 +1969,12 @@ namespace bgfx { namespace d3d12
 			return m_textures[_handle.idx].create(_mem, _flags, _skip);
 		}
 
-		void updateTextureBegin(TextureHandle /*_handle*/, uint8_t /*_side*/, uint8_t /*_mip*/) override
-		{
-		}
-
 		void updateTexture(TextureHandle _handle, uint8_t _side, uint8_t _mip, const Rect& _rect, uint16_t _z, uint16_t _depth, uint16_t _pitch, const Memory* _mem) override
 		{
 			m_textures[_handle.idx].update(m_commandList, _side, _mip, _rect, _z, _depth, _pitch, _mem);
 		}
 
-		void updateTextureEnd() override
-		{
-		}
-
-		void readTexture(TextureHandle _handle, void* _data, uint8_t _mip ) override
+		void readTexture(TextureHandle _handle, void* _data, uint8_t _mip) override
 		{
 			const TextureD3D12& texture = m_textures[_handle.idx];
 
@@ -2005,8 +1997,8 @@ namespace bgfx { namespace d3d12
 
 			ID3D12Resource* readback = createCommittedResource(m_device, HeapProperty::ReadBack, total);
 
-			uint32_t srcWidth  = bx::uint32_max(1, texture.m_width >>_mip);
-			uint32_t srcHeight = bx::uint32_max(1, texture.m_height>>_mip);
+			const uint32_t srcWidth  = bx::max(1u, texture.m_width >>_mip);
+			const uint32_t srcHeight = bx::max(1u, texture.m_height>>_mip);
 
 			D3D12_BOX box;
 			box.left   = 0;
@@ -2287,7 +2279,7 @@ namespace bgfx { namespace d3d12
 
 		void submit(Frame* _render, ClearQuad& _clearQuad, TextVideoMemBlitter& _textVideoMemBlitter) override;
 
-		void blitSetup(TextVideoMemBlitter& _blitter) override
+		void dbgTextRenderBegin(TextVideoMemBlitter& _blitter) override
 		{
 			const uint32_t width  = m_scd.width;
 			const uint32_t height = m_scd.height;
@@ -2330,7 +2322,7 @@ namespace bgfx { namespace d3d12
 			float proj[16];
 			bx::mtxOrtho(proj, 0.0f, (float)width, (float)height, 0.0f, 0.0f, 1000.0f, 0.0f, false);
 
-			PredefinedUniform& predefined = m_program[_blitter.m_program.idx].m_predefined[0];
+			const PredefinedUniform& predefined = m_program[_blitter.m_program.idx].m_predefined[0];
 			uint8_t flags = predefined.m_type;
 			setShaderUniform(flags, predefined.m_loc, proj, 4);
 
@@ -2347,15 +2339,15 @@ namespace bgfx { namespace d3d12
 			m_commandList->SetGraphicsRootConstantBufferView(Rdt::CBV, gpuAddress);
 
 			TextureD3D12& texture = m_textures[_blitter.m_texture.idx];
-			uint32_t samplerFlags[] = { uint32_t(texture.m_flags & BGFX_SAMPLER_BITS_MASK) };
-			uint16_t samplerStateIdx = getSamplerState(samplerFlags, BX_COUNTOF(samplerFlags), NULL);
+			const uint32_t samplerFlags[] = { uint32_t(texture.m_flags & BGFX_SAMPLER_BITS_MASK) };
+			const uint16_t samplerStateIdx = getSamplerState(samplerFlags, BX_COUNTOF(samplerFlags), NULL);
 			m_commandList->SetGraphicsRootDescriptorTable(Rdt::Sampler, m_samplerAllocator.get(samplerStateIdx) );
 			D3D12_GPU_DESCRIPTOR_HANDLE srvHandle;
 			scratchBuffer.allocSrv(srvHandle, texture);
 			m_commandList->SetGraphicsRootDescriptorTable(Rdt::SRV, srvHandle);
 
-			VertexBufferD3D12&  vb     = m_vertexBuffers[_blitter.m_vb->handle.idx];
-			const VertexLayout& layout = m_vertexLayouts[_blitter.m_vb->layoutHandle.idx];
+			const VertexBufferD3D12& vb = m_vertexBuffers[_blitter.m_vb->handle.idx];
+			const VertexLayout& layout  = m_vertexLayouts[_blitter.m_vb->layoutHandle.idx];
 			D3D12_VERTEX_BUFFER_VIEW viewDesc;
 			viewDesc.BufferLocation = vb.m_gpuVA;
 			viewDesc.StrideInBytes  = layout.m_stride;
@@ -2372,7 +2364,7 @@ namespace bgfx { namespace d3d12
 			m_commandList->IASetPrimitiveTopology(D3D_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
 		}
 
-		void blitRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) override
+		void dbgTextRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) override
 		{
 			const uint32_t numVertices = _numIndices*4/6;
 			if (0 < numVertices)
@@ -2380,15 +2372,14 @@ namespace bgfx { namespace d3d12
 				m_indexBuffers [_blitter.m_ib->handle.idx].update(m_commandList, 0, _numIndices*2, _blitter.m_ib->data);
 				m_vertexBuffers[_blitter.m_vb->handle.idx].update(m_commandList, 0, numVertices*_blitter.m_layout.m_stride, _blitter.m_vb->data, true);
 
-				m_commandList->DrawIndexedInstanced(_numIndices
-					, 1
-					, 0
-					, 0
-					, 0
-					);
+				m_commandList->DrawIndexedInstanced(_numIndices, 1, 0, 0, 0);
 			}
 		}
 
+		void dbgTextRenderEnd(TextVideoMemBlitter& /*_blitter*/) override
+		{
+		}
+
 		void preReset()
 		{
 			finishAll();
@@ -3221,7 +3212,7 @@ namespace bgfx { namespace d3d12
 				| BGFX_STATE_PT_MASK
 				;
 
-			_stencil &= packStencil(~BGFX_STENCIL_FUNC_REF_MASK, ~BGFX_STENCIL_FUNC_REF_MASK);
+			_stencil &= kStencilNoRefMask;
 
 			VertexLayout layout;
 			if (0 < _numStreams)
@@ -4371,15 +4362,12 @@ namespace bgfx { namespace d3d12
 
 		if (UINT8_MAX != _draw.m_streamMask)
 		{
-			for (uint32_t idx = 0, streamMask = _draw.m_streamMask
-				; 0 != streamMask
-				; streamMask >>= 1, idx += 1, ++numStreams
+			for (BitMaskToIndexIteratorT it(_draw.m_streamMask)
+				; !it.isDone()
+				; it.next(), numStreams++
 				)
 			{
-				const uint32_t ntz = bx::uint32_cnttz(streamMask);
-				streamMask >>= ntz;
-				idx         += ntz;
-
+				const uint8_t idx = it.idx;
 				const Stream& stream = _draw.m_stream[idx];
 
 				uint16_t handle = stream.m_handle.idx;
@@ -7027,14 +7015,12 @@ namespace bgfx { namespace d3d12
 					uint8_t numStreams = 0;
 					if (UINT8_MAX != draw.m_streamMask)
 					{
-						for (uint32_t idx = 0, streamMask = draw.m_streamMask
-							; 0 != streamMask
-							; streamMask >>= 1, idx += 1, ++numStreams
+						for (BitMaskToIndexIteratorT it(draw.m_streamMask)
+							; !it.isDone()
+							; it.next(), numStreams++
 							)
 						{
-							const uint32_t ntz = bx::uint32_cnttz(streamMask);
-							streamMask >>= ntz;
-							idx         += ntz;
+							const uint8_t idx = it.idx;
 
 							currentState.m_stream[idx].m_layoutHandle = draw.m_stream[idx].m_layoutHandle;
 							currentState.m_stream[idx].m_handle       = draw.m_stream[idx].m_handle;
@@ -7577,7 +7563,7 @@ namespace bgfx { namespace d3d12
 				presentMax = m_presentElapsed;
 			}
 
-			blit(this, _textVideoMemBlitter, tvm);
+			dbgTextSubmit(this, _textVideoMemBlitter, tvm);
 
 			BGFX_D3D12_PROFILER_END();
 		}
@@ -7585,7 +7571,7 @@ namespace bgfx { namespace d3d12
 		{
 			BGFX_D3D12_PROFILER_BEGIN_LITERAL("debugtext", kColorFrame);
 
-			blit(this, _textVideoMemBlitter, _render->m_textVideoMem);
+			dbgTextSubmit(this, _textVideoMemBlitter, _render->m_textVideoMem);
 
 			BGFX_D3D12_PROFILER_END();
 		}

src/renderer_gl.cpp

@@ -3399,19 +3399,11 @@ namespace bgfx { namespace gl
 			return NULL;
 		}
 
-		void updateTextureBegin(TextureHandle /*_handle*/, uint8_t /*_side*/, uint8_t /*_mip*/) override
-		{
-		}
-
 		void updateTexture(TextureHandle _handle, uint8_t _side, uint8_t _mip, const Rect& _rect, uint16_t _z, uint16_t _depth, uint16_t _pitch, const Memory* _mem) override
 		{
 			m_textures[_handle.idx].update(_side, _mip, _rect, _z, _depth, _pitch, _mem);
 		}
 
-		void updateTextureEnd() override
-		{
-		}
-
 		void readTexture(TextureHandle _handle, void* _data, uint8_t _mip) override
 		{
 			if (m_readBackSupported)
@@ -3690,7 +3682,7 @@ namespace bgfx { namespace gl
 
 		void submit(Frame* _render, ClearQuad& _clearQuad, TextVideoMemBlitter& _textVideoMemBlitter) override;
 
-		void blitSetup(TextVideoMemBlitter& _blitter) override
+		void dbgTextRenderBegin(TextVideoMemBlitter& _blitter) override
 		{
 			uint32_t width  = m_resolution.width;
 			uint32_t height = m_resolution.height;
@@ -3728,7 +3720,7 @@ namespace bgfx { namespace gl
 			}
 		}
 
-		void blitRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) override
+		void dbgTextRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) override
 		{
 			const uint32_t numVertices = _numIndices*4/6;
 			if (0 < numVertices)
@@ -3755,6 +3747,10 @@ namespace bgfx { namespace gl
 			}
 		}
 
+		void dbgTextRenderEnd(TextVideoMemBlitter& /*_blitter*/) override
+		{
+		}
+
 		void updateResolution(const Resolution& _resolution)
 		{
 			float maxAnisotropy = !!(_resolution.reset & BGFX_RESET_MAXANISOTROPY)
@@ -4613,8 +4609,8 @@ namespace bgfx { namespace gl
 					GL_CHECK(glDisable(GL_STENCIL_TEST) );
 				}
 
-				VertexBufferGL& vb = m_vertexBuffers[_clearQuad.m_vb.idx];
-				VertexLayout& layout = _clearQuad.m_layout;
+				const VertexBufferGL&   vb = m_vertexBuffers[_clearQuad.m_vb.idx];
+				const VertexLayout& layout = m_vertexLayouts[_clearQuad.m_layout.idx];
 
 				GL_CHECK(glBindBuffer(GL_ARRAY_BUFFER, vb.m_id) );
 
@@ -8281,14 +8277,9 @@ namespace bgfx { namespace gl
 					}
 
 					{
-						for (uint32_t idx = 0, streamMask = draw.m_streamMask
-							; 0 != streamMask
-							; streamMask >>= 1, idx += 1
-							)
+						for (BitMaskToIndexIteratorT it(draw.m_streamMask); !it.isDone(); it.next() )
 						{
-							const uint32_t ntz = bx::uint32_cnttz(streamMask);
-							streamMask >>= ntz;
-							idx         += ntz;
+							const uint8_t idx = it.idx;
 
 							if (currentState.m_stream[idx].m_handle.idx != draw.m_stream[idx].m_handle.idx)
 							{
@@ -8353,14 +8344,9 @@ namespace bgfx { namespace gl
 
 								if (UINT8_MAX != draw.m_streamMask)
 								{
-									for (uint32_t idx = 0, streamMask = draw.m_streamMask
-										; 0 != streamMask
-										; streamMask >>= 1, idx += 1
-										)
+									for (BitMaskToIndexIteratorT it(draw.m_streamMask); !it.isDone(); it.next() )
 									{
-										const uint32_t ntz = bx::uint32_cnttz(streamMask);
-										streamMask >>= ntz;
-										idx         += ntz;
+										const uint8_t idx = it.idx;
 
 										const VertexBufferGL& vb = m_vertexBuffers[draw.m_stream[idx].m_handle.idx];
 										const uint16_t decl = isValid(draw.m_stream[idx].m_layoutHandle)
@@ -8385,16 +8371,12 @@ namespace bgfx { namespace gl
 					if (0 != currentState.m_streamMask)
 					{
 						uint32_t numVertices = draw.m_numVertices;
+
 						if (UINT32_MAX == numVertices)
 						{
-							for (uint32_t idx = 0, streamMask = draw.m_streamMask
-								; 0 != streamMask
-								; streamMask >>= 1, idx += 1
-								)
+							for (BitMaskToIndexIteratorT it(draw.m_streamMask); !it.isDone(); it.next() )
 							{
-								const uint32_t ntz = bx::uint32_cnttz(streamMask);
-								streamMask >>= ntz;
-								idx         += ntz;
+								const uint8_t idx = it.idx;
 
 								const VertexBufferGL& vb = m_vertexBuffers[draw.m_stream[idx].m_handle.idx];
 								uint16_t decl = !isValid(vb.m_layoutHandle) ? draw.m_stream[idx].m_layoutHandle.idx : vb.m_layoutHandle.idx;
@@ -8840,7 +8822,7 @@ namespace bgfx { namespace gl
 				max = frameTime;
 			}
 
-			blit(this, _textVideoMemBlitter, tvm);
+			dbgTextSubmit(this, _textVideoMemBlitter, tvm);
 
 			BGFX_GL_PROFILER_END();
 		}
@@ -8848,7 +8830,7 @@ namespace bgfx { namespace gl
 		{
 			BGFX_GL_PROFILER_BEGIN_LITERAL("debugtext", kColorFrame);
 
-			blit(this, _textVideoMemBlitter, _render->m_textVideoMem);
+			dbgTextSubmit(this, _textVideoMemBlitter, _render->m_textVideoMem);
 
 			BGFX_GL_PROFILER_END();
 		}

src/renderer_mtl.mm

@@ -1082,19 +1082,11 @@ static_assert(BX_COUNTOF(s_accessNames) == Access::Count, "Invalid s_accessNames
 			return NULL;
 		}
 
-		void updateTextureBegin(TextureHandle /*_handle*/, uint8_t /*_side*/, uint8_t /*_mip*/) override
-		{
-		}
-
 		void updateTexture(TextureHandle _handle, uint8_t _side, uint8_t _mip, const Rect& _rect, uint16_t _z, uint16_t _depth, uint16_t _pitch, const Memory* _mem) override
 		{
 			m_textures[_handle.idx].update(_side, _mip, _rect, _z, _depth, _pitch, _mem);
 		}
 
-		void updateTextureEnd() override
-		{
-		}
-
 		static MTLPixelFormat getSwapChainPixelFormat(SwapChainMtl* _swapChain)
 		{
 			return NULL != _swapChain
@@ -1346,12 +1338,11 @@ static_assert(BX_COUNTOF(s_accessNames) == Access::Count, "Invalid s_accessNames
 
 		void submit(Frame* _render, ClearQuad& _clearQuad, TextVideoMemBlitter& _textVideoMemBlitter) override;
 
-		void blitSetup(TextVideoMemBlitter& _blitter) override
+		void dbgTextRenderBegin(TextVideoMemBlitter& /*_blitter*/) override
 		{
-			BX_UNUSED(_blitter);
 		}
 
-		void blitRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) override
+		void dbgTextRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) override
 		{
 			const uint32_t numVertices = _numIndices*4/6;
 			if (0 < numVertices)
@@ -1466,6 +1457,10 @@ static_assert(BX_COUNTOF(s_accessNames) == Access::Count, "Invalid s_accessNames
 			}
 		}
 
+		void dbgTextRenderEnd(TextVideoMemBlitter& /*_blitter*/) override
+		{
+		}
+
 		bool isDeviceRemoved() override
 		{
 			return false;
@@ -1796,13 +1791,14 @@ static_assert(BX_COUNTOF(s_accessNames) == Access::Count, "Invalid s_accessNames
 
 			uint32_t numMrt = 1;
 			FrameBufferHandle fbh = m_fbh;
-			if (isValid(fbh) && m_frameBuffers[fbh.idx].m_swapChain == NULL)
+			if (isValid(fbh)
+			&&  NULL == m_frameBuffers[fbh.idx].m_swapChain)
 			{
 				const FrameBufferMtl& fb = m_frameBuffers[fbh.idx];
 				numMrt = bx::uint32_max(1, fb.m_num);
 			}
 
-			const VertexLayout* layout = &_clearQuad.m_layout;
+			const VertexLayout* layout  = &m_vertexLayouts[_clearQuad.m_layout.idx];
 			const PipelineStateMtl* pso = getPipelineState(
 				  state
 				, 0
@@ -1836,8 +1832,8 @@ static_assert(BX_COUNTOF(s_accessNames) == Access::Count, "Invalid s_accessNames
 				m_renderCommandEncoder.setFragmentBuffer(m_uniformBuffer, m_uniformBufferFragmentOffset, 0);
 			}
 
+			const float mrtClearDepth[4] = { _clear.m_depth };
 			float mrtClearColor[BGFX_CONFIG_MAX_FRAME_BUFFER_ATTACHMENTS][4];
-			float mrtClearDepth[4] = { _clear.m_depth };
 
 			if (BGFX_CLEAR_COLOR_USE_PALETTE & _clear.m_flags)
 			{
@@ -1921,6 +1917,7 @@ static_assert(BX_COUNTOF(s_accessNames) == Access::Count, "Invalid s_accessNames
 				FrameBufferMtl& frameBuffer = m_frameBuffers[m_fbh.idx];
 				frameBuffer.resolve();
 			}
+
 			if (!isValid(_fbh)
 			||  m_frameBuffers[_fbh.idx].m_swapChain)
 			{
@@ -2012,7 +2009,7 @@ static_assert(BX_COUNTOF(s_accessNames) == Access::Count, "Invalid s_accessNames
 			uint32_t fstencil = unpackStencil(0, _stencil);
 			uint32_t ref      = (fstencil&BGFX_STENCIL_FUNC_REF_MASK)>>BGFX_STENCIL_FUNC_REF_SHIFT;
 
-			_stencil &= packStencil(~BGFX_STENCIL_FUNC_REF_MASK, ~BGFX_STENCIL_FUNC_REF_MASK);
+			_stencil &= kStencilNoRefMask;
 
 			bx::HashMurmur2A murmur;
 			murmur.begin();
@@ -3250,7 +3247,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 					depth  >>= 1;
 				}
 			}
-			
+
 			MTL_RELEASE(desc, 0);
 
 			if (NULL != temp)
@@ -4916,14 +4913,13 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 
 					uint32_t numVertices = draw.m_numVertices;
 					uint8_t  numStreams  = 0;
-					for (uint32_t idx = 0, streamMask = draw.m_streamMask
-						; 0 != streamMask
-						; streamMask >>= 1, idx += 1, ++numStreams
+
+					for (BitMaskToIndexIteratorT it(draw.m_streamMask)
+						; !it.isDone()
+						; it.next(), numStreams++
 						)
 					{
-						const uint32_t ntz = bx::uint32_cnttz(streamMask);
-						streamMask >>= ntz;
-						idx         += ntz;
+						const uint8_t idx = it.idx;
 
 						currentState.m_stream[idx].m_layoutHandle   = draw.m_stream[idx].m_layoutHandle;
 						currentState.m_stream[idx].m_handle         = draw.m_stream[idx].m_handle;
@@ -5386,7 +5382,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 				max = frameTime;
 			}
 
-			blit(this, _textVideoMemBlitter, tvm);
+			dbgTextSubmit(this, _textVideoMemBlitter, tvm);
 			rce = m_renderCommandEncoder;
 
 			rce.popDebugGroup();
@@ -5395,7 +5391,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 		{
 			rce.pushDebugGroup("debugtext");
 
-			blit(this, _textVideoMemBlitter, _render->m_textVideoMem);
+			dbgTextSubmit(this, _textVideoMemBlitter, _render->m_textVideoMem);
 			rce = m_renderCommandEncoder;
 
 			rce.popDebugGroup();

src/renderer_noop.cpp

@@ -169,18 +169,10 @@ namespace bgfx { namespace noop
 			return NULL;
 		}
 
-		void updateTextureBegin(TextureHandle /*_handle*/, uint8_t /*_side*/, uint8_t /*_mip*/) override
-		{
-		}
-
 		void updateTexture(TextureHandle /*_handle*/, uint8_t /*_side*/, uint8_t /*_mip*/, const Rect& /*_rect*/, uint16_t /*_z*/, uint16_t /*_depth*/, uint16_t /*_pitch*/, const Memory* /*_mem*/) override
 		{
 		}
 
-		void updateTextureEnd() override
-		{
-		}
-
 		void readTexture(TextureHandle /*_handle*/, void* /*_data*/, uint8_t /*_mip*/) override
 		{
 		}
@@ -267,11 +259,15 @@ namespace bgfx { namespace noop
 			perfStats.gpuMemoryUsed = -INT64_MAX;
 		}
 
-		void blitSetup(TextVideoMemBlitter& /*_blitter*/) override
+		void dbgTextRenderBegin(TextVideoMemBlitter& /*_blitter*/) override
 		{
 		}
 
-		void blitRender(TextVideoMemBlitter& /*_blitter*/, uint32_t /*_numIndices*/) override
+		void dbgTextRender(TextVideoMemBlitter& /*_blitter*/, uint32_t /*_numIndices*/) override
+		{
+		}
+
+		void dbgTextRenderEnd(TextVideoMemBlitter& /*_blitter*/) override
 		{
 		}
 	};

src/renderer_vk.cpp

@@ -2311,7 +2311,7 @@ VK_IMPORT_DEVICE
 				fb.present();
 			}
 
-			int64_t now = bx::getHPCounter();
+			const int64_t now = bx::getHPCounter();
 
 			m_presentElapsed += now - start;
 		}
@@ -2403,19 +2403,11 @@ VK_IMPORT_DEVICE
 			return m_textures[_handle.idx].create(m_commandBuffer, _mem, _flags, _skip);
 		}
 
-		void updateTextureBegin(TextureHandle /*_handle*/, uint8_t /*_side*/, uint8_t /*_mip*/) override
-		{
-		}
-
 		void updateTexture(TextureHandle _handle, uint8_t _side, uint8_t _mip, const Rect& _rect, uint16_t _z, uint16_t _depth, uint16_t _pitch, const Memory* _mem) override
 		{
 			m_textures[_handle.idx].update(m_commandBuffer, _side, _mip, _rect, _z, _depth, _pitch, _mem);
 		}
 
-		void updateTextureEnd() override
-		{
-		}
-
 		void readTexture(TextureHandle _handle, void* _data, uint8_t _mip) override
 		{
 			TextureVK& texture = m_textures[_handle.idx];
@@ -2684,7 +2676,7 @@ VK_IMPORT_DEVICE
 
 		void submit(Frame* _render, ClearQuad& _clearQuad, TextVideoMemBlitter& _textVideoMemBlitter) override;
 
-		void blitSetup(TextVideoMemBlitter& _blitter) override
+		void dbgTextRenderBegin(TextVideoMemBlitter& _blitter) override
 		{
 			const uint32_t width  = m_backBuffer.m_width;
 			const uint32_t height = m_backBuffer.m_height;
@@ -2778,7 +2770,7 @@ VK_IMPORT_DEVICE
 				);
 		}
 
-		void blitRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) override
+		void dbgTextRender(TextVideoMemBlitter& _blitter, uint32_t _numIndices) override
 		{
 			const uint32_t numVertices = _numIndices*4/6;
 
@@ -2806,6 +2798,10 @@ VK_IMPORT_DEVICE
 			}
 		}
 
+		void dbgTextRenderEnd(TextVideoMemBlitter& /*_blitter*/) override
+		{
+		}
+
 		void preReset()
 		{
 			for (uint32_t ii = 0; ii < BX_COUNTOF(m_frameBuffers); ++ii)
@@ -3021,7 +3017,7 @@ VK_IMPORT_DEVICE
 
 				newFrameBuffer.acquire(m_commandBuffer);
 
-				int64_t now = bx::getHPCounter();
+				const int64_t now = bx::getHPCounter();
 
 				if (NULL != newFrameBuffer.m_nwh)
 				{
@@ -3690,7 +3686,7 @@ VK_IMPORT_DEVICE
 				| BGFX_STATE_PT_MASK
 				;
 
-			_stencil &= packStencil(~BGFX_STENCIL_FUNC_REF_MASK, ~BGFX_STENCIL_FUNC_REF_MASK);
+			_stencil &= kStencilNoRefMask;
 
 			VertexLayout layout;
 			if (0 < _numStreams)
@@ -5580,6 +5576,7 @@ retry:
 		smci.codeSize = m_code->size;
 		smci.pCode    = (const uint32_t*)m_code->data;
 
+		BX_TRACE("%x", bx::hash<bx::HashMurmur3>(code, shaderSize) );
 		VK_CHECK(vkCreateShaderModule(
 			  s_renderVK->m_device
 			, &smci
@@ -9150,11 +9147,8 @@ retry:
 					}
 
 					if (beginRenderPass && (false
-					||  _render->m_view[view].m_fbh.idx       != fbh.idx
-					||  _render->m_view[view].m_rect.m_x      != viewState.m_rect.m_x
-					||  _render->m_view[view].m_rect.m_y      != viewState.m_rect.m_y
-					||  _render->m_view[view].m_rect.m_width  != viewState.m_rect.m_width
-					||  _render->m_view[view].m_rect.m_height != viewState.m_rect.m_height
+					||  _render->m_view[view].m_fbh.idx != fbh.idx
+					|| !_render->m_view[view].m_rect.isEqual(viewState.m_rect)
 					   ) )
 					{
 						vkCmdEndRenderPass(m_commandBuffer);
@@ -9540,14 +9534,12 @@ retry:
 					uint32_t numVertices = draw.m_numVertices;
 					if (UINT8_MAX != draw.m_streamMask)
 					{
-						for (uint32_t idx = 0, streamMask = draw.m_streamMask
-							; 0 != streamMask
-							; streamMask >>= 1, idx += 1, ++numStreams
+						for (BitMaskToIndexIteratorT it(draw.m_streamMask)
+							; !it.isDone()
+							; it.next(), numStreams++
 							)
 						{
-							const uint32_t ntz = bx::uint32_cnttz(streamMask);
-							streamMask >>= ntz;
-							idx         += ntz;
+							const uint8_t idx = it.idx;
 
 							currentState.m_stream[idx] = draw.m_stream[idx];
 
@@ -10168,7 +10160,7 @@ retry:
 				presentMax = m_presentElapsed;
 			}
 
-			blit(this, _textVideoMemBlitter, tvm);
+			dbgTextSubmit(this, _textVideoMemBlitter, tvm);
 
 			BGFX_VK_PROFILER_END();
 		}
@@ -10176,7 +10168,7 @@ retry:
 		{
 			BGFX_VK_PROFILER_BEGIN_LITERAL("debugtext", kColorFrame);
 
-			blit(this, _textVideoMemBlitter, _render->m_textVideoMem);
+			dbgTextSubmit(this, _textVideoMemBlitter, _render->m_textVideoMem);
 
 			BGFX_VK_PROFILER_END();
 		}