Added bgfx::setVertexCount for auto generated vertices with gl_VertexID.

include/bgfx/bgfx.h

@@ -1247,6 +1247,16 @@ namespace bgfx
 			, uint32_t _numVertices
 			);
 
+		/// Set number of vertices for auto generated vertices use in conjuction
+		/// with gl_VertexID.
+		///
+		/// @param[in] _numVertices Number of vertices.
+		///
+		/// @attention Availability depends on: `BGFX_CAPS_VERTEX_ID`.
+		/// @attention C99 equivalent is `bgfx_set_vertex_count`.
+		///
+		void setVertexCount(uint32_t _numVertices);
+
 		/// Set instance data buffer for draw primitive.
 		///
 		/// @param[in] _idb Transient instance data buffer.
@@ -3550,6 +3560,9 @@ namespace bgfx
 		, uint32_t _numVertices
 		);
 
+	///
+	void setVertexCount(uint32_t _numVertices);
+
 	/// Set instance data buffer for draw primitive.
 	///
 	/// @param[in] _idb Transient instance data buffer.

include/bgfx/c99/bgfx.h

@@ -983,6 +983,9 @@ BGFX_C_API void bgfx_set_dynamic_vertex_buffer(uint8_t _stream, bgfx_dynamic_ver
 /**/
 BGFX_C_API void bgfx_set_transient_vertex_buffer(uint8_t _stream, const bgfx_transient_vertex_buffer_t* _tvb, uint32_t _startVertex, uint32_t _numVertices);
 
+/**/
+BGFX_C_API void bgfx_set_vertex_count(uint32_t _numVertices);
+
 /**/
 BGFX_C_API void bgfx_set_instance_data_buffer(const bgfx_instance_data_buffer_t* _idb, uint32_t _start, uint32_t _num);
 
@@ -1085,6 +1088,9 @@ BGFX_C_API void bgfx_encoder_set_dynamic_vertex_buffer(struct bgfx_encoder_s* _e
 /**/
 BGFX_C_API void bgfx_encoder_set_transient_vertex_buffer(struct bgfx_encoder_s* _encoder, uint8_t _stream, const bgfx_transient_vertex_buffer_t* _tvb, uint32_t _startVertex, uint32_t _numVertices);
 
+/**/
+BGFX_C_API void bgfx_encoder_set_vertex_count(struct bgfx_encoder_s* _encoder, uint32_t _numVertices);
+
 /**/
 BGFX_C_API void bgfx_encoder_set_instance_data_buffer(struct bgfx_encoder_s* _encoder, const bgfx_instance_data_buffer_t* _idb, uint32_t _start, uint32_t _num);
 

include/bgfx/c99/platform.h

@@ -182,6 +182,7 @@ typedef struct bgfx_interface_vtbl
     void (*encoder_set_vertex_buffer)(struct bgfx_encoder_s* _encoder, uint8_t _stream, bgfx_vertex_buffer_handle_t _handle, uint32_t _startVertex, uint32_t _numVertices);
     void (*encoder_set_dynamic_vertex_buffer)(struct bgfx_encoder_s* _encoder, uint8_t _stream, bgfx_dynamic_vertex_buffer_handle_t _handle, uint32_t _startVertex, uint32_t _numVertices);
     void (*encoder_set_transient_vertex_buffer)(struct bgfx_encoder_s* _encoder, uint8_t _stream, const bgfx_transient_vertex_buffer_t* _tvb, uint32_t _startVertex, uint32_t _numVertices);
+    void (*encoder_set_vertex_count)(struct bgfx_encoder_s* _encoder, uint32_t _numVertices);
     void (*encoder_set_instance_data_buffer)(struct bgfx_encoder_s* _encoder, const bgfx_instance_data_buffer_t* _idb, uint32_t _start, uint32_t _num);
     void (*encoder_set_instance_data_from_vertex_buffer)(struct bgfx_encoder_s* _encoder, bgfx_vertex_buffer_handle_t _handle, uint32_t _startVertex, uint32_t _num);
     void (*encoder_set_instance_data_from_dynamic_vertex_buffer)(struct bgfx_encoder_s* _encoder, bgfx_dynamic_vertex_buffer_handle_t _handle, uint32_t _startVertex, uint32_t _num);

include/bgfx/defines.h

@@ -6,7 +6,7 @@
 #ifndef BGFX_DEFINES_H_HEADER_GUARD
 #define BGFX_DEFINES_H_HEADER_GUARD
 
-#define BGFX_API_VERSION UINT32_C(68)
+#define BGFX_API_VERSION UINT32_C(69)
 
 /// Color RGB/alpha/depth write. When it's not specified write will be disabled.
 #define BGFX_STATE_WRITE_R                 UINT64_C(0x0000000000000001) //!< Enable R write.
@@ -456,6 +456,7 @@
 #define BGFX_CAPS_TEXTURE_READ_BACK      UINT64_C(0x0000000000400000) //!< Read-back texture is supported.
 #define BGFX_CAPS_VERTEX_ATTRIB_HALF     UINT64_C(0x0000000000800000) //!< Vertex attribute half-float is supported.
 #define BGFX_CAPS_VERTEX_ATTRIB_UINT10   UINT64_C(0x0000000000800000) //!< Vertex attribute 10_10_10_2 is supported.
+#define BGFX_CAPS_VERTEX_ID              UINT64_C(0x0000000001000000) //!< Rendering with VertexID only is supported.
 
 ///
 #define BGFX_CAPS_FORMAT_TEXTURE_NONE             UINT16_C(0x0000) //!< Texture format is not supported.

src/bgfx.cpp

@@ -932,17 +932,25 @@ namespace bgfx
 		m_draw.m_uniformBegin = m_uniformBegin;
 		m_draw.m_uniformEnd   = m_uniformEnd;
 
-		uint32_t numVertices = UINT32_MAX;
-		for (uint32_t idx = 0, streamMask = m_draw.m_streamMask, ntz = bx::uint32_cnttz(streamMask)
-			; 0 != streamMask
-			; streamMask >>= 1, idx += 1, ntz = bx::uint32_cnttz(streamMask)
-			)
+		if (UINT8_MAX != m_draw.m_streamMask)
 		{
-			streamMask >>= ntz;
-			idx         += ntz;
-			numVertices = bx::min(numVertices, m_numVertices[idx]);
+			uint32_t numVertices = UINT32_MAX;
+			for (uint32_t idx = 0, streamMask = m_draw.m_streamMask, ntz = bx::uint32_cnttz(streamMask)
+				; 0 != streamMask
+				; streamMask >>= 1, idx += 1, ntz = bx::uint32_cnttz(streamMask)
+				)
+			{
+				streamMask >>= ntz;
+				idx         += ntz;
+				numVertices = bx::min(numVertices, m_numVertices[idx]);
+			}
+
+			m_draw.m_numVertices = numVertices;
+		}
+		else
+		{
+			m_draw.m_numVertices = m_numVertices[0];
 		}
-		m_draw.m_numVertices = numVertices;
 
 		if (isValid(_occlusionQuery) )
 		{
@@ -1167,6 +1175,7 @@ namespace bgfx
 		CAPS_FLAGS(BGFX_CAPS_TEXTURE_READ_BACK),
 		CAPS_FLAGS(BGFX_CAPS_VERTEX_ATTRIB_HALF),
 		CAPS_FLAGS(BGFX_CAPS_VERTEX_ATTRIB_UINT10),
+		CAPS_FLAGS(BGFX_CAPS_VERTEX_ID),
 #undef CAPS_FLAGS
 	};
 
@@ -3119,6 +3128,12 @@ error:
 		setVertexBuffer(_stream, _tvb, 0, UINT32_MAX);
 	}
 
+	void Encoder::setVertexCount(uint32_t _numVertices)
+	{
+		BGFX_CHECK_CAPS(BGFX_CAPS_VERTEX_ID, "Auto generated vertices are not supported!");
+		BGFX_ENCODER(setVertexCount(_numVertices) );
+	}
+
 	void Encoder::setInstanceDataBuffer(const InstanceDataBuffer* _idb)
 	{
 		setInstanceDataBuffer(_idb, 0, UINT32_MAX);
@@ -4286,6 +4301,12 @@ error:
 		setVertexBuffer(_stream, _tvb, 0, UINT32_MAX);
 	}
 
+	void setVertexCount(uint32_t _numVertices)
+	{
+		BGFX_CHECK_API_THREAD();
+		s_ctx->m_encoder0->setVertexCount(_numVertices);
+	}
+
 	void setInstanceDataBuffer(const InstanceDataBuffer* _idb)
 	{
 		BGFX_CHECK_API_THREAD();
@@ -5391,6 +5412,11 @@ BGFX_C_API void bgfx_set_transient_vertex_buffer(uint8_t _stream, const bgfx_tra
 	bgfx::setVertexBuffer(_stream, (const bgfx::TransientVertexBuffer*)_tvb, _startVertex, _numVertices);
 }
 
+BGFX_C_API void bgfx_set_vertex_count(uint32_t _numVertices)
+{
+	bgfx::setVertexCount(_numVertices);
+}
+
 BGFX_C_API void bgfx_set_instance_data_buffer(const bgfx_instance_data_buffer_t* _idb, uint32_t _start, uint32_t _num)
 {
 	bgfx::setInstanceDataBuffer( (const bgfx::InstanceDataBuffer*)_idb, _start, _num);
@@ -5589,6 +5615,11 @@ BGFX_C_API void bgfx_encoder_set_transient_vertex_buffer(bgfx_encoder_s* _encode
 	BGFX_ENCODER(setVertexBuffer(_stream, (const bgfx::TransientVertexBuffer*)_tvb, _startVertex, _numVertices) );
 }
 
+BGFX_C_API void bgfx_encoder_set_vertex_count(bgfx_encoder_s* _encoder, uint32_t _numVertices)
+{
+	BGFX_ENCODER(setVertexCount(_numVertices) );
+}
+
 BGFX_C_API void bgfx_encoder_set_instance_data_buffer(bgfx_encoder_s* _encoder, const bgfx_instance_data_buffer_t* _idb, uint32_t _start, uint32_t _num)
 {
 	BGFX_ENCODER(setInstanceDataBuffer( (const bgfx::InstanceDataBuffer*)_idb, _start, _num) );
@@ -5855,6 +5886,7 @@ BGFX_C_API bgfx_interface_vtbl_t* bgfx_get_interface(uint32_t _version)
 	BGFX_IMPORT_FUNC(encoder_set_vertex_buffer)                            \
 	BGFX_IMPORT_FUNC(encoder_set_dynamic_vertex_buffer)                    \
 	BGFX_IMPORT_FUNC(encoder_set_transient_vertex_buffer)                  \
+	BGFX_IMPORT_FUNC(encoder_set_vertex_count)                             \
 	BGFX_IMPORT_FUNC(encoder_set_instance_data_buffer)                     \
 	BGFX_IMPORT_FUNC(encoder_set_instance_data_from_vertex_buffer)         \
 	BGFX_IMPORT_FUNC(encoder_set_instance_data_from_dynamic_vertex_buffer) \

src/bgfx_p.h

@@ -2194,6 +2194,7 @@ namespace bgfx
 
 		void setIndexBuffer(IndexBufferHandle _handle, uint32_t _firstIndex, uint32_t _numIndices)
 		{
+			BX_CHECK(UINT8_MAX != m_draw.m_streamMask, "");
 			m_draw.m_startIndex  = _firstIndex;
 			m_draw.m_numIndices  = _numIndices;
 			m_draw.m_indexBuffer = _handle;
@@ -2201,6 +2202,7 @@ namespace bgfx
 
 		void setIndexBuffer(const DynamicIndexBuffer& _dib, uint32_t _firstIndex, uint32_t _numIndices)
 		{
+			BX_CHECK(UINT8_MAX != m_draw.m_streamMask, "");
 			const uint32_t indexSize = 0 == (_dib.m_flags & BGFX_BUFFER_INDEX32) ? 2 : 4;
 			m_draw.m_startIndex  = _dib.m_startIndex + _firstIndex;
 			m_draw.m_numIndices  = bx::min(_numIndices, _dib.m_size/indexSize);
@@ -2209,6 +2211,7 @@ namespace bgfx
 
 		void setIndexBuffer(const TransientIndexBuffer* _tib, uint32_t _firstIndex, uint32_t _numIndices)
 		{
+			BX_CHECK(UINT8_MAX != m_draw.m_streamMask, "");
 			const uint32_t numIndices = bx::min(_numIndices, _tib->size/2);
 			m_draw.m_indexBuffer = _tib->handle;
 			m_draw.m_startIndex  = _tib->startIndex + _firstIndex;
@@ -2218,6 +2221,7 @@ namespace bgfx
 
 		void setVertexBuffer(uint8_t _stream, VertexBufferHandle _handle, uint32_t _startVertex, uint32_t _numVertices)
 		{
+			BX_CHECK(UINT8_MAX != m_draw.m_streamMask, "");
 			BX_CHECK(_stream < BGFX_CONFIG_MAX_VERTEX_STREAMS, "Invalid stream %d (max %d).", _stream, BGFX_CONFIG_MAX_VERTEX_STREAMS);
 			if (m_draw.setStreamBit(_stream, _handle) )
 			{
@@ -2231,6 +2235,7 @@ namespace bgfx
 
 		void setVertexBuffer(uint8_t _stream, const DynamicVertexBuffer& _dvb, uint32_t _startVertex, uint32_t _numVertices)
 		{
+			BX_CHECK(UINT8_MAX != m_draw.m_streamMask, "");
 			BX_CHECK(_stream < BGFX_CONFIG_MAX_VERTEX_STREAMS, "Invalid stream %d (max %d).", _stream, BGFX_CONFIG_MAX_VERTEX_STREAMS);
 			if (m_draw.setStreamBit(_stream, _dvb.m_handle) )
 			{
@@ -2246,6 +2251,7 @@ namespace bgfx
 
 		void setVertexBuffer(uint8_t _stream, const TransientVertexBuffer* _tvb, uint32_t _startVertex, uint32_t _numVertices)
 		{
+			BX_CHECK(UINT8_MAX != m_draw.m_streamMask, "");
 			BX_CHECK(_stream < BGFX_CONFIG_MAX_VERTEX_STREAMS, "Invalid stream %d (max %d).", _stream, BGFX_CONFIG_MAX_VERTEX_STREAMS);
 			if (m_draw.setStreamBit(_stream, _tvb->handle) )
 			{
@@ -2259,6 +2265,17 @@ namespace bgfx
 			}
 		}
 
+		void setVertexCount(uint32_t _numVertices)
+		{
+			BX_CHECK(0 == m_draw.m_streamMask, "Vertex buffer already set.");
+			m_draw.m_streamMask  = UINT8_MAX;
+			Stream& stream = m_draw.m_stream[0];
+			stream.m_startVertex = 0;
+			stream.m_handle.idx  = kInvalidHandle;
+			stream.m_decl.idx    = kInvalidHandle;
+			m_numVertices[0]     = _numVertices;
+		}
+
 		void setInstanceDataBuffer(const InstanceDataBuffer* _idb, uint32_t _start, uint32_t _num)
 		{
 			const uint32_t start = bx::min(_start, _idb->num);

src/renderer_d3d11.cpp

@@ -1124,6 +1124,7 @@ namespace bgfx { namespace d3d11
 					| BGFX_CAPS_TEXTURE_3D
 					| BGFX_CAPS_VERTEX_ATTRIB_HALF
 					| BGFX_CAPS_VERTEX_ATTRIB_UINT10
+					| BGFX_CAPS_VERTEX_ID
 					| BGFX_CAPS_FRAGMENT_DEPTH
 					| (getIntelExtensions(m_device)
 						? BGFX_CAPS_FRAGMENT_ORDERING
@@ -6093,34 +6094,38 @@ namespace bgfx { namespace d3d11
 
 					uint32_t numVertices = draw.m_numVertices;
 					uint8_t  numStreams  = 0;
-					for (uint32_t idx = 0, streamMask = draw.m_streamMask, ntz = bx::uint32_cnttz(streamMask)
-						; 0 != streamMask
-						; streamMask >>= 1, idx += 1, ntz = bx::uint32_cnttz(streamMask), ++numStreams
-						)
+
+					if (UINT8_MAX != draw.m_streamMask)
 					{
-						streamMask >>= ntz;
-						idx         += ntz;
-
-						currentState.m_stream[idx].m_decl        = draw.m_stream[idx].m_decl;
-						currentState.m_stream[idx].m_handle      = draw.m_stream[idx].m_handle;
-						currentState.m_stream[idx].m_startVertex = draw.m_stream[idx].m_startVertex;
-
-						const uint16_t handle = draw.m_stream[idx].m_handle.idx;
-						const VertexBufferD3D11& vb = m_vertexBuffers[handle];
-						const uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[idx].m_decl.idx : vb.m_decl.idx;
-						const VertexDecl& vertexDecl = m_vertexDecls[decl];
-						const uint32_t stride = vertexDecl.m_stride;
-
-						buffers[numStreams] = vb.m_ptr;
-						strides[numStreams] = stride;
-						offsets[numStreams] = draw.m_stream[idx].m_startVertex * stride;
-						decls[numStreams]   = &vertexDecl;
-
-						numVertices = bx::uint32_min(UINT32_MAX == draw.m_numVertices
-							? vb.m_size/stride
-							: draw.m_numVertices
-							, numVertices
-							);
+						for (uint32_t idx = 0, streamMask = draw.m_streamMask, ntz = bx::uint32_cnttz(streamMask)
+							; 0 != streamMask
+							; streamMask >>= 1, idx += 1, ntz = bx::uint32_cnttz(streamMask), ++numStreams
+							)
+						{
+							streamMask >>= ntz;
+							idx         += ntz;
+
+							currentState.m_stream[idx].m_decl        = draw.m_stream[idx].m_decl;
+							currentState.m_stream[idx].m_handle      = draw.m_stream[idx].m_handle;
+							currentState.m_stream[idx].m_startVertex = draw.m_stream[idx].m_startVertex;
+
+							const uint16_t handle = draw.m_stream[idx].m_handle.idx;
+							const VertexBufferD3D11& vb = m_vertexBuffers[handle];
+							const uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[idx].m_decl.idx : vb.m_decl.idx;
+							const VertexDecl& vertexDecl = m_vertexDecls[decl];
+							const uint32_t stride = vertexDecl.m_stride;
+
+							buffers[numStreams] = vb.m_ptr;
+							strides[numStreams] = stride;
+							offsets[numStreams] = draw.m_stream[idx].m_startVertex * stride;
+							decls[numStreams]   = &vertexDecl;
+
+							numVertices = bx::uint32_min(UINT32_MAX == draw.m_numVertices
+								? vb.m_size/stride
+								: draw.m_numVertices
+								, numVertices
+								);
+						}
 					}
 
 					currentState.m_numVertices = numVertices;
@@ -6132,9 +6137,9 @@ namespace bgfx { namespace d3d11
 						if (isValid(draw.m_instanceDataBuffer) )
 						{
 							const VertexBufferD3D11& inst = m_vertexBuffers[draw.m_instanceDataBuffer.idx];
-							uint32_t instStride = draw.m_instanceDataStride;
+							const uint32_t instStride = draw.m_instanceDataStride;
 							deviceCtx->IASetVertexBuffers(numStreams, 1, &inst.m_ptr, &instStride, &draw.m_instanceDataOffset);
-							setInputLayout(numStreams, decls, m_program[programIdx], draw.m_instanceDataStride/16);
+							setInputLayout(numStreams, decls, m_program[programIdx], uint16_t(instStride/16) );
 						}
 						else
 						{
@@ -6145,6 +6150,14 @@ namespace bgfx { namespace d3d11
 					else
 					{
 						deviceCtx->IASetVertexBuffers(0, 1, s_zero.m_buffer, s_zero.m_zero, s_zero.m_zero);
+
+						if (isValid(draw.m_instanceDataBuffer) )
+						{
+							const VertexBufferD3D11& inst = m_vertexBuffers[draw.m_instanceDataBuffer.idx];
+							const uint32_t instStride = draw.m_instanceDataStride;
+							deviceCtx->IASetVertexBuffers(0, 1, &inst.m_ptr, &instStride, &draw.m_instanceDataOffset);
+							setInputLayout(0, NULL, m_program[programIdx], uint16_t(instStride/16) );
+						}
 					}
 				}
 

src/renderer_d3d12.cpp

@@ -1045,6 +1045,7 @@ namespace bgfx { namespace d3d12
 					| BGFX_CAPS_DRAW_INDIRECT
 					| BGFX_CAPS_VERTEX_ATTRIB_HALF
 					| BGFX_CAPS_VERTEX_ATTRIB_UINT10
+					| BGFX_CAPS_VERTEX_ID
 					| BGFX_CAPS_FRAGMENT_DEPTH
 					| BGFX_CAPS_BLEND_INDEPENDENT
 					| BGFX_CAPS_COMPUTE
@@ -2643,14 +2644,17 @@ namespace bgfx { namespace d3d12
 			_stencil &= packStencil(~BGFX_STENCIL_FUNC_REF_MASK, ~BGFX_STENCIL_FUNC_REF_MASK);
 
 			VertexDecl decl;
-			bx::memCopy(&decl, _vertexDecls[0], sizeof(VertexDecl) );
-			const uint16_t* attrMask = program.m_vsh->m_attrMask;
-
-			for (uint32_t ii = 0; ii < Attrib::Count; ++ii)
+			if (0 < _numStreams)
 			{
-				uint16_t mask = attrMask[ii];
-				uint16_t attr = (decl.m_attributes[ii] & mask);
-				decl.m_attributes[ii] = attr == 0 ? UINT16_MAX : attr == UINT16_MAX ? 0 : attr;
+				bx::memCopy(&decl, _vertexDecls[0], sizeof(VertexDecl) );
+				const uint16_t* attrMask = program.m_vsh->m_attrMask;
+
+				for (uint32_t ii = 0; ii < Attrib::Count; ++ii)
+				{
+					uint16_t mask = attrMask[ii];
+					uint16_t attr = (decl.m_attributes[ii] & mask);
+					decl.m_attributes[ii] = attr == 0 ? UINT16_MAX : attr == UINT16_MAX ? 0 : attr;
+				}
 			}
 
 			bx::HashMurmur2A murmur;
@@ -2659,6 +2663,7 @@ namespace bgfx { namespace d3d12
 			murmur.add(_stencil);
 			murmur.add(program.m_vsh->m_hash);
 			murmur.add(program.m_vsh->m_attrMask, sizeof(program.m_vsh->m_attrMask) );
+
 			if (NULL != program.m_fsh)
 			{
 				murmur.add(program.m_fsh->m_hash);
@@ -2668,6 +2673,7 @@ namespace bgfx { namespace d3d12
 			{
 				murmur.add(_vertexDecls[ii]->m_hash);
 			}
+
 			murmur.add(decl.m_attributes, sizeof(decl.m_attributes) );
 			murmur.add(m_fbh.idx);
 			murmur.add(_numInstanceData);
@@ -2785,7 +2791,10 @@ namespace bgfx { namespace d3d12
 
 			D3D12_INPUT_ELEMENT_DESC vertexElements[Attrib::Count + 1 + BGFX_CONFIG_MAX_INSTANCE_DATA_COUNT];
 			desc.InputLayout.NumElements = setInputLayout(vertexElements, _numStreams, _vertexDecls, program, _numInstanceData);
-			desc.InputLayout.pInputElementDescs = vertexElements;
+			desc.InputLayout.pInputElementDescs = 0 == desc.InputLayout.NumElements
+				? NULL
+				: vertexElements
+				;
 
 			uint8_t primIndex = uint8_t( (_state&BGFX_STATE_PT_MASK) >> BGFX_STATE_PT_SHIFT);
 			desc.PrimitiveTopologyType = s_primInfo[primIndex].m_topologyType;
@@ -3763,38 +3772,42 @@ namespace bgfx { namespace d3d12
 		return *cmd;
 	}
 
-	uint8_t fill(ID3D12GraphicsCommandList* _commandList, D3D12_VERTEX_BUFFER_VIEW* _vbv, const RenderDraw& _draw, uint32_t& numVertices)
+	uint8_t fill(ID3D12GraphicsCommandList* _commandList, D3D12_VERTEX_BUFFER_VIEW* _vbv, const RenderDraw& _draw, uint32_t& _outNumVertices)
 	{
-		uint8_t  numStreams = 0;
-		numVertices = _draw.m_numVertices;
-		for (uint32_t idx = 0, streamMask = _draw.m_streamMask, ntz = bx::uint32_cnttz(streamMask)
-			; 0 != streamMask
-			; streamMask >>= 1, idx += 1, ntz = bx::uint32_cnttz(streamMask), ++numStreams
-			)
+		uint8_t numStreams = 0;
+		_outNumVertices = _draw.m_numVertices;
+
+		if (UINT8_MAX != _draw.m_streamMask)
 		{
-			streamMask >>= ntz;
-			idx += ntz;
+			for (uint32_t idx = 0, streamMask = _draw.m_streamMask, ntz = bx::uint32_cnttz(streamMask)
+				; 0 != streamMask
+				; streamMask >>= 1, idx += 1, ntz = bx::uint32_cnttz(streamMask), ++numStreams
+				)
+			{
+				streamMask >>= ntz;
+				idx += ntz;
 
-			const Stream& stream = _draw.m_stream[idx];
+				const Stream& stream = _draw.m_stream[idx];
 
-			uint16_t handle = stream.m_handle.idx;
-			VertexBufferD3D12& vb = s_renderD3D12->m_vertexBuffers[handle];
-			vb.setState(_commandList, D3D12_RESOURCE_STATE_GENERIC_READ);
+				uint16_t handle = stream.m_handle.idx;
+				VertexBufferD3D12& vb = s_renderD3D12->m_vertexBuffers[handle];
+				vb.setState(_commandList, D3D12_RESOURCE_STATE_GENERIC_READ);
 
-			uint16_t decl = !isValid(vb.m_decl) ? stream.m_decl.idx : vb.m_decl.idx;
-			const VertexDecl& vertexDecl = s_renderD3D12->m_vertexDecls[decl];
-			uint32_t stride = vertexDecl.m_stride;
+				uint16_t decl = !isValid(vb.m_decl) ? stream.m_decl.idx : vb.m_decl.idx;
+				const VertexDecl& vertexDecl = s_renderD3D12->m_vertexDecls[decl];
+				uint32_t stride = vertexDecl.m_stride;
 
-			D3D12_VERTEX_BUFFER_VIEW& vbv = _vbv[numStreams];
-			vbv.BufferLocation = vb.m_gpuVA + stream.m_startVertex * stride;
-			vbv.StrideInBytes  = vertexDecl.m_stride;
-			vbv.SizeInBytes    = vb.m_size;
+				D3D12_VERTEX_BUFFER_VIEW& vbv = _vbv[numStreams];
+				vbv.BufferLocation = vb.m_gpuVA + stream.m_startVertex * stride;
+				vbv.StrideInBytes  = vertexDecl.m_stride;
+				vbv.SizeInBytes    = vb.m_size;
 
-			numVertices = bx::uint32_min(UINT32_MAX == _draw.m_numVertices
-				? vb.m_size/stride
-				: _draw.m_numVertices
-				, numVertices
-				);
+				_outNumVertices = bx::uint32_min(UINT32_MAX == _draw.m_numVertices
+					? vb.m_size/stride
+					: _draw.m_numVertices
+					, _outNumVertices
+					);
+			}
 		}
 
 		return numStreams;
@@ -6005,25 +6018,28 @@ namespace bgfx { namespace d3d12
 
 					const VertexDecl* decls[BGFX_CONFIG_MAX_VERTEX_STREAMS];
 
-					uint8_t  numStreams = 0;
-					for (uint32_t idx = 0, streamMask = draw.m_streamMask, ntz = bx::uint32_cnttz(streamMask)
-						; 0 != streamMask
-						; streamMask >>= 1, idx += 1, ntz = bx::uint32_cnttz(streamMask), ++numStreams
-						)
+					uint8_t numStreams = 0;
+					if (UINT8_MAX != draw.m_streamMask)
 					{
-						streamMask >>= ntz;
-						idx += ntz;
+						for (uint32_t idx = 0, streamMask = draw.m_streamMask, ntz = bx::uint32_cnttz(streamMask)
+							; 0 != streamMask
+							; streamMask >>= 1, idx += 1, ntz = bx::uint32_cnttz(streamMask), ++numStreams
+							)
+						{
+							streamMask >>= ntz;
+							idx += ntz;
 
-						currentState.m_stream[idx].m_decl        = draw.m_stream[idx].m_decl;
-						currentState.m_stream[idx].m_handle      = draw.m_stream[idx].m_handle;
-						currentState.m_stream[idx].m_startVertex = draw.m_stream[idx].m_startVertex;
+							currentState.m_stream[idx].m_decl        = draw.m_stream[idx].m_decl;
+							currentState.m_stream[idx].m_handle      = draw.m_stream[idx].m_handle;
+							currentState.m_stream[idx].m_startVertex = draw.m_stream[idx].m_startVertex;
 
-						uint16_t handle = draw.m_stream[idx].m_handle.idx;
-						const VertexBufferD3D12& vb = m_vertexBuffers[handle];
-						uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[idx].m_decl.idx : vb.m_decl.idx;
-						const VertexDecl& vertexDecl = m_vertexDecls[decl];
+							uint16_t handle = draw.m_stream[idx].m_handle.idx;
+							const VertexBufferD3D12& vb = m_vertexBuffers[handle];
+							uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[idx].m_decl.idx : vb.m_decl.idx;
+							const VertexDecl& vertexDecl = m_vertexDecls[decl];
 
-						decls[numStreams] = &vertexDecl;
+							decls[numStreams] = &vertexDecl;
+						}
 					}
 
 					ID3D12PipelineState* pso =