diff --git a/src/bgfx.cpp b/src/bgfx.cpp
index 2a93d7d13..404f5227a 100644
--- a/src/bgfx.cpp
+++ b/src/bgfx.cpp
@@ -3495,7 +3495,7 @@ error:
 	void setVertexBuffer(VertexBufferHandle _handle, uint32_t _startVertex, uint32_t _numVertices)
 	{
 		BGFX_CHECK_MAIN_THREAD();
-		s_ctx->setVertexBuffer(_handle, _startVertex, _numVertices);
+		s_ctx->setVertexBuffer(0, _handle, _startVertex, _numVertices);
 	}
 
 	void setVertexBuffer(DynamicVertexBufferHandle _handle)
@@ -3506,7 +3506,7 @@ error:
 	void setVertexBuffer(DynamicVertexBufferHandle _handle, uint32_t _startVertex, uint32_t _numVertices)
 	{
 		BGFX_CHECK_MAIN_THREAD();
-		s_ctx->setVertexBuffer(_handle, _startVertex, _numVertices);
+		s_ctx->setVertexBuffer(0, _handle, _startVertex, _numVertices);
 	}
 
 	void setVertexBuffer(const TransientVertexBuffer* _tvb)
@@ -3518,7 +3518,7 @@ error:
 	{
 		BGFX_CHECK_MAIN_THREAD();
 		BX_CHECK(NULL != _tvb, "_tvb can't be NULL");
-		s_ctx->setVertexBuffer(_tvb, _startVertex, _numVertices);
+		s_ctx->setVertexBuffer(0, _tvb, _startVertex, _numVertices);
 	}
 
 	void setInstanceDataBuffer(const InstanceDataBuffer* _idb, uint32_t _num)
@@ -3543,12 +3543,15 @@ error:
 	void setTexture(uint8_t _stage, UniformHandle _sampler, TextureHandle _handle, uint32_t _flags)
 	{
 		BGFX_CHECK_MAIN_THREAD();
+		BX_CHECK(_stage < BGFX_CONFIG_MAX_TEXTURE_SAMPLERS, "Invalid stage %d (max %d).", _stage, BGFX_CONFIG_MAX_TEXTURE_SAMPLERS);
 		s_ctx->setTexture(_stage, _sampler, _handle, _flags);
 	}
 
 	void setTexture(uint8_t _stage, UniformHandle _sampler, FrameBufferHandle _handle, uint8_t _attachment, uint32_t _flags)
 	{
 		BGFX_CHECK_MAIN_THREAD();
+		BX_CHECK(_stage < BGFX_CONFIG_MAX_TEXTURE_SAMPLERS, "Invalid stage %d (max %d).", _stage, BGFX_CONFIG_MAX_TEXTURE_SAMPLERS);
+		BX_CHECK(_attachment < g_caps.maxFBAttachments, "Frame buffer attachment index %d is invalid.", _attachment);
 		s_ctx->setTexture(_stage, _sampler, _handle, _attachment, _flags);
 	}
 
@@ -3584,42 +3587,50 @@ error:
 	void setBuffer(uint8_t _stage, IndexBufferHandle _handle, Access::Enum _access)
 	{
 		BGFX_CHECK_MAIN_THREAD();
+		BX_CHECK(_stage < BGFX_CONFIG_MAX_TEXTURE_SAMPLERS, "Invalid stage %d (max %d).", _stage, BGFX_CONFIG_MAX_TEXTURE_SAMPLERS);
 		s_ctx->setBuffer(_stage, _handle, _access);
 	}
 
 	void setBuffer(uint8_t _stage, VertexBufferHandle _handle, Access::Enum _access)
 	{
 		BGFX_CHECK_MAIN_THREAD();
+		BX_CHECK(_stage < BGFX_CONFIG_MAX_TEXTURE_SAMPLERS, "Invalid stage %d (max %d).", _stage, BGFX_CONFIG_MAX_TEXTURE_SAMPLERS);
 		s_ctx->setBuffer(_stage, _handle, _access);
 	}
 
 	void setBuffer(uint8_t _stage, DynamicIndexBufferHandle _handle, Access::Enum _access)
 	{
 		BGFX_CHECK_MAIN_THREAD();
+		BX_CHECK(_stage < BGFX_CONFIG_MAX_TEXTURE_SAMPLERS, "Invalid stage %d (max %d).", _stage, BGFX_CONFIG_MAX_TEXTURE_SAMPLERS);
 		s_ctx->setBuffer(_stage, _handle, _access);
 	}
 
 	void setBuffer(uint8_t _stage, DynamicVertexBufferHandle _handle, Access::Enum _access)
 	{
 		BGFX_CHECK_MAIN_THREAD();
+		BX_CHECK(_stage < BGFX_CONFIG_MAX_TEXTURE_SAMPLERS, "Invalid stage %d (max %d).", _stage, BGFX_CONFIG_MAX_TEXTURE_SAMPLERS);
 		s_ctx->setBuffer(_stage, _handle, _access);
 	}
 
 	void setBuffer(uint8_t _stage, IndirectBufferHandle _handle, Access::Enum _access)
 	{
 		BGFX_CHECK_MAIN_THREAD();
+		BX_CHECK(_stage < BGFX_CONFIG_MAX_TEXTURE_SAMPLERS, "Invalid stage %d (max %d).", _stage, BGFX_CONFIG_MAX_TEXTURE_SAMPLERS);
 		s_ctx->setBuffer(_stage, _handle, _access);
 	}
 
 	void setImage(uint8_t _stage, UniformHandle _sampler, TextureHandle _handle, uint8_t _mip, Access::Enum _access, TextureFormat::Enum _format)
 	{
 		BGFX_CHECK_MAIN_THREAD();
+		BX_CHECK(_stage < BGFX_CONFIG_MAX_TEXTURE_SAMPLERS, "Invalid stage %d (max %d).", _stage, BGFX_CONFIG_MAX_TEXTURE_SAMPLERS);
 		s_ctx->setImage(_stage, _sampler, _handle, _mip, _access, _format);
 	}
 
 	void setImage(uint8_t _stage, UniformHandle _sampler, FrameBufferHandle _handle, uint8_t _attachment, Access::Enum _access, TextureFormat::Enum _format)
 	{
 		BGFX_CHECK_MAIN_THREAD();
+		BX_CHECK(_stage < BGFX_CONFIG_MAX_TEXTURE_SAMPLERS, "Invalid stage %d (max %d).", _stage, BGFX_CONFIG_MAX_TEXTURE_SAMPLERS);
+		BX_CHECK(_attachment < g_caps.maxFBAttachments, "Frame buffer attachment index %d is invalid.", _attachment);
 		s_ctx->setImage(_stage, _sampler, _handle, _attachment, _access, _format);
 	}
 
@@ -3662,6 +3673,7 @@ error:
 	{
 		BGFX_CHECK_MAIN_THREAD();
 		BGFX_CHECK_CAPS(BGFX_CAPS_TEXTURE_BLIT, "Texture blit is not supported!");
+		BX_CHECK(_attachment < g_caps.maxFBAttachments, "Frame buffer attachment index %d is invalid.", _attachment);
 		s_ctx->blit(_id, _dst, _dstMip, _dstX, _dstY, _dstZ, _src, _attachment, _srcMip, _srcX, _srcY, _srcZ, _width, _height, _depth);
 	}
 
diff --git a/src/bgfx_p.h b/src/bgfx_p.h
index a15d8d382..c0d673bd2 100644
--- a/src/bgfx_p.h
+++ b/src/bgfx_p.h
@@ -1216,6 +1216,20 @@ namespace bgfx
 		} m_un;
 	};
 
+	struct Stream
+	{
+		void clear()
+		{
+			m_startVertex = 0;
+			m_handle.idx  = invalidHandle;
+			m_decl.idx    = invalidHandle;
+		}
+
+		uint32_t           m_startVertex;
+		VertexBufferHandle m_handle;
+		VertexDeclHandle   m_decl;
+	};
+
 	struct RenderDraw
 	{
 		void clear()
@@ -1228,7 +1242,6 @@ namespace bgfx
 			m_matrix      = 0;
 			m_startIndex  = 0;
 			m_numIndices  = UINT32_MAX;
-			m_startVertex = 0;
 			m_numVertices = UINT32_MAX;
 			m_instanceDataOffset = 0;
 			m_instanceDataStride = 0;
@@ -1238,8 +1251,8 @@ namespace bgfx
 			m_num           = 1;
 			m_submitFlags   = BGFX_SUBMIT_EYE_FIRST;
 			m_scissor       = UINT16_MAX;
-			m_vertexBuffer.idx       = invalidHandle;
-			m_vertexDecl.idx         = invalidHandle;
+			m_streamMask    = 0;
+			m_stream[0].clear();
 			m_indexBuffer.idx        = invalidHandle;
 			m_instanceDataBuffer.idx = invalidHandle;
 			m_indirectBuffer.idx     = invalidHandle;
@@ -1255,6 +1268,7 @@ namespace bgfx
 		}
 
 		Binding  m_bind[BGFX_CONFIG_MAX_TEXTURE_SAMPLERS];
+		Stream   m_stream[BGFX_CONFIG_MAX_VERTEX_STREAMS];
 		uint64_t m_stateFlags;
 		uint64_t m_stencil;
 		uint32_t m_rgba;
@@ -1263,7 +1277,6 @@ namespace bgfx
 		uint32_t m_matrix;
 		uint32_t m_startIndex;
 		uint32_t m_numIndices;
-		uint32_t m_startVertex;
 		uint32_t m_numVertices;
 		uint32_t m_instanceDataOffset;
 		uint32_t m_numInstances;
@@ -1273,9 +1286,8 @@ namespace bgfx
 		uint16_t m_num;
 		uint16_t m_scissor;
 		uint8_t  m_submitFlags;
+		uint8_t  m_streamMask;
 
-		VertexBufferHandle   m_vertexBuffer;
-		VertexDeclHandle     m_vertexDecl;
 		IndexBufferHandle    m_indexBuffer;
 		VertexBufferHandle   m_instanceDataBuffer;
 		IndirectBufferHandle m_indirectBuffer;
@@ -1555,27 +1567,49 @@ namespace bgfx
 			m_discard = 0 == _numIndices;
 		}
 
-		void setVertexBuffer(VertexBufferHandle _handle, uint32_t _startVertex, uint32_t _numVertices)
+		void setVertexBuffer(uint8_t _stream, VertexBufferHandle _handle, uint32_t _startVertex, uint32_t _numVertices)
 		{
-			m_draw.m_startVertex  = _startVertex;
-			m_draw.m_numVertices  = _numVertices;
-			m_draw.m_vertexBuffer = _handle;
+			const uint8_t bit = 1<<_stream;
+			BX_CHECK(0 == (m_draw.m_streamMask & bit), "Vertex stream %d is already set.", _stream);
+			m_draw.m_streamMask |= bit;
+
+			Stream& stream = m_draw.m_stream[_stream];
+			stream.m_startVertex = _startVertex;
+			stream.m_handle      = _handle;
+			stream.m_decl.idx    = invalidHandle;
+			m_draw.m_numVertices = bx::uint32_min(m_draw.m_numVertices, _numVertices);
 		}
 
-		void setVertexBuffer(const DynamicVertexBuffer& _dvb, uint32_t _startVertex, uint32_t _numVertices)
+		void setVertexBuffer(uint8_t _stream, const DynamicVertexBuffer& _dvb, uint32_t _startVertex, uint32_t _numVertices)
 		{
-			m_draw.m_startVertex  = _dvb.m_startVertex + _startVertex;
-			m_draw.m_numVertices  = bx::uint32_min(bx::uint32_imax(0, _dvb.m_numVertices - _startVertex), _numVertices);
-			m_draw.m_vertexBuffer = _dvb.m_handle;
-			m_draw.m_vertexDecl   = _dvb.m_decl;
+			const uint8_t bit = 1<<_stream;
+			BX_CHECK(0 == (m_draw.m_streamMask & bit), "Vertex stream %d is already set.", _stream);
+			m_draw.m_streamMask |= bit;
+
+			Stream& stream = m_draw.m_stream[_stream];
+			stream.m_startVertex = _dvb.m_startVertex + _startVertex;
+			stream.m_handle      = _dvb.m_handle;
+			stream.m_decl        = _dvb.m_decl;
+			m_draw.m_numVertices =
+				  bx::uint32_min(m_draw.m_numVertices
+				, bx::uint32_min(bx::uint32_imax(0, _dvb.m_numVertices - _startVertex), _numVertices)
+				);
 		}
 
-		void setVertexBuffer(const TransientVertexBuffer* _tvb, uint32_t _startVertex, uint32_t _numVertices)
+		void setVertexBuffer(uint8_t _stream, const TransientVertexBuffer* _tvb, uint32_t _startVertex, uint32_t _numVertices)
 		{
-			m_draw.m_startVertex  = _tvb->startVertex + _startVertex;
-			m_draw.m_numVertices  = bx::uint32_min(bx::uint32_imax(0, _tvb->size/_tvb->stride - _startVertex), _numVertices);
-			m_draw.m_vertexBuffer = _tvb->handle;
-			m_draw.m_vertexDecl   = _tvb->decl;
+			const uint8_t bit = 1<<_stream;
+			BX_CHECK(0 == (m_draw.m_streamMask & bit), "Vertex stream %d is already set.", _stream);
+			m_draw.m_streamMask |= bit;
+
+			Stream& stream = m_draw.m_stream[_stream];
+			stream.m_startVertex  = _tvb->startVertex + _startVertex;
+			stream.m_handle       = _tvb->handle;
+			stream.m_decl         = _tvb->decl;
+			m_draw.m_numVertices  =
+				  bx::uint32_min(m_draw.m_numVertices
+				, bx::uint32_min(bx::uint32_imax(0, _tvb->size/_tvb->stride - _startVertex), _numVertices)
+				);
 		}
 
 		void setInstanceDataBuffer(const InstanceDataBuffer* _idb, uint32_t _num)
@@ -3702,22 +3736,25 @@ namespace bgfx
 			m_submit->setIndexBuffer(_tib, _firstIndex, _numIndices);
 		}
 
-		BGFX_API_FUNC(void setVertexBuffer(VertexBufferHandle _handle, uint32_t _startVertex, uint32_t _numVertices) )
+		BGFX_API_FUNC(void setVertexBuffer(uint8_t _stream, VertexBufferHandle _handle, uint32_t _startVertex, uint32_t _numVertices) )
 		{
 			BGFX_CHECK_HANDLE("setVertexBuffer", m_vertexBufferHandle, _handle);
-			m_submit->setVertexBuffer(_handle, _startVertex, _numVertices);
+			BX_CHECK(_stream < BGFX_CONFIG_MAX_VERTEX_STREAMS, "Invalid stream %d (max %d).", _stream, BGFX_CONFIG_MAX_VERTEX_STREAMS);
+			m_submit->setVertexBuffer(_stream, _handle, _startVertex, _numVertices);
 		}
 
-		BGFX_API_FUNC(void setVertexBuffer(DynamicVertexBufferHandle _handle, uint32_t _startVertex, uint32_t _numVertices) )
+		BGFX_API_FUNC(void setVertexBuffer(uint8_t _stream, DynamicVertexBufferHandle _handle, uint32_t _startVertex, uint32_t _numVertices) )
 		{
 			BGFX_CHECK_HANDLE("setVertexBuffer", m_dynamicVertexBufferHandle, _handle);
-			m_submit->setVertexBuffer(m_dynamicVertexBuffers[_handle.idx], _startVertex, _numVertices);
+			BX_CHECK(_stream < BGFX_CONFIG_MAX_VERTEX_STREAMS, "Invalid stream %d (max %d).", _stream, BGFX_CONFIG_MAX_VERTEX_STREAMS);
+			m_submit->setVertexBuffer(_stream, m_dynamicVertexBuffers[_handle.idx], _startVertex, _numVertices);
 		}
 
-		BGFX_API_FUNC(void setVertexBuffer(const TransientVertexBuffer* _tvb, uint32_t _startVertex, uint32_t _numVertices) )
+		BGFX_API_FUNC(void setVertexBuffer(uint8_t _stream, const TransientVertexBuffer* _tvb, uint32_t _startVertex, uint32_t _numVertices) )
 		{
 			BGFX_CHECK_HANDLE("setVertexBuffer", m_vertexBufferHandle, _tvb->handle);
-			m_submit->setVertexBuffer(_tvb, _startVertex, _numVertices);
+			BX_CHECK(_stream < BGFX_CONFIG_MAX_VERTEX_STREAMS, "Invalid stream %d (max %d).", _stream, BGFX_CONFIG_MAX_VERTEX_STREAMS);
+			m_submit->setVertexBuffer(_stream, _tvb, _startVertex, _numVertices);
 		}
 
 		BGFX_API_FUNC(void setInstanceDataBuffer(const InstanceDataBuffer* _idb, uint32_t _num) )
@@ -3754,7 +3791,6 @@ namespace bgfx
 		BGFX_API_FUNC(void setTexture(uint8_t _stage, UniformHandle _sampler, FrameBufferHandle _handle, uint8_t _attachment, uint32_t _flags) )
 		{
 			BGFX_CHECK_HANDLE_INVALID_OK("setTexture/FrameBufferHandle", m_frameBufferHandle, _handle);
-			BX_CHECK(_attachment < g_caps.maxFBAttachments, "Frame buffer attachment index %d is invalid.", _attachment);
 			TextureHandle textureHandle = BGFX_INVALID_HANDLE;
 			if (isValid(_handle) )
 			{
@@ -3846,7 +3882,6 @@ namespace bgfx
 
 		BGFX_API_FUNC(void setImage(uint8_t _stage, UniformHandle _sampler, FrameBufferHandle _handle, uint8_t _attachment, Access::Enum _access, TextureFormat::Enum _format) )
 		{
-			BX_CHECK(_attachment < g_caps.maxFBAttachments, "Frame buffer attachment index %d is invalid.", _attachment);
 			TextureHandle textureHandle = BGFX_INVALID_HANDLE;
 			if (isValid(_handle) )
 			{
diff --git a/src/config.h b/src/config.h
index f9e4ddd24..dd80f94e4 100644
--- a/src/config.h
+++ b/src/config.h
@@ -237,6 +237,10 @@
 #	define BGFX_CONFIG_MAX_VERTEX_BUFFERS (4<<10)
 #endif // BGFX_CONFIG_MAX_VERTEX_BUFFERS
 
+#ifndef BGFX_CONFIG_MAX_VERTEX_STREAMS
+#	define BGFX_CONFIG_MAX_VERTEX_STREAMS 1
+#endif // BGFX_CONFIG_MAX_VERTEX_STREAMS
+
 #ifndef BGFX_CONFIG_MAX_DYNAMIC_INDEX_BUFFERS
 #	define BGFX_CONFIG_MAX_DYNAMIC_INDEX_BUFFERS (4<<10)
 #endif // BGFX_CONFIG_MAX_DYNAMIC_INDEX_BUFFERS
diff --git a/src/renderer_d3d11.cpp b/src/renderer_d3d11.cpp
index 605fc83ed..f58487b4b 100644
--- a/src/renderer_d3d11.cpp
+++ b/src/renderer_d3d11.cpp
@@ -5816,24 +5816,24 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 				}
 
 				if (programChanged
-				||  currentState.m_vertexDecl.idx         != draw.m_vertexDecl.idx
-				||  currentState.m_vertexBuffer.idx       != draw.m_vertexBuffer.idx
+				||  currentState.m_stream[0].m_decl.idx   != draw.m_stream[0].m_decl.idx
+				||  currentState.m_stream[0].m_handle.idx != draw.m_stream[0].m_handle.idx
 				||  currentState.m_instanceDataBuffer.idx != draw.m_instanceDataBuffer.idx
 				||  currentState.m_instanceDataOffset     != draw.m_instanceDataOffset
 				||  currentState.m_instanceDataStride     != draw.m_instanceDataStride)
 				{
-					currentState.m_vertexDecl             = draw.m_vertexDecl;
-					currentState.m_vertexBuffer           = draw.m_vertexBuffer;
+					currentState.m_stream[0].m_decl       = draw.m_stream[0].m_decl;
+					currentState.m_stream[0].m_handle     = draw.m_stream[0].m_handle;
 					currentState.m_instanceDataBuffer.idx = draw.m_instanceDataBuffer.idx;
 					currentState.m_instanceDataOffset     = draw.m_instanceDataOffset;
 					currentState.m_instanceDataStride     = draw.m_instanceDataStride;
 
-					uint16_t handle = draw.m_vertexBuffer.idx;
+					uint16_t handle = draw.m_stream[0].m_handle.idx;
 					if (invalidHandle != handle)
 					{
 						const VertexBufferD3D11& vb = m_vertexBuffers[handle];
 
-						uint16_t decl = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
+						uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[0].m_decl.idx : vb.m_decl.idx;
 						const VertexDecl& vertexDecl = m_vertexDecls[decl];
 						uint32_t stride = vertexDecl.m_stride;
 						uint32_t offset = 0;
@@ -5877,13 +5877,13 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 					}
 				}
 
-				if (isValid(currentState.m_vertexBuffer) )
+				if (isValid(currentState.m_stream[0].m_handle) )
 				{
 					uint32_t numVertices = draw.m_numVertices;
 					if (UINT32_MAX == numVertices)
 					{
-						const VertexBufferD3D11& vb = m_vertexBuffers[currentState.m_vertexBuffer.idx];
-						uint16_t decl = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
+						const VertexBufferD3D11& vb = m_vertexBuffers[currentState.m_stream[0].m_handle.idx];
+						uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[0].m_decl.idx : vb.m_decl.idx;
 						const VertexDecl& vertexDecl = m_vertexDecls[decl];
 						numVertices = vb.m_size/vertexDecl.m_stride;
 					}
@@ -5949,7 +5949,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 									deviceCtx->DrawIndexedInstanced(numIndices
 										, draw.m_numInstances
 										, 0
-										, draw.m_startVertex
+										, draw.m_stream[0].m_startVertex
 										, 0
 										);
 								}
@@ -5957,7 +5957,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 								{
 									deviceCtx->DrawIndexed(numIndices
 										, 0
-										, draw.m_startVertex
+										, draw.m_stream[0].m_startVertex
 										);
 								}
 							}
@@ -5973,7 +5973,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 									deviceCtx->DrawIndexedInstanced(numIndices
 										, draw.m_numInstances
 										, draw.m_startIndex
-										, draw.m_startVertex
+										, draw.m_stream[0].m_startVertex
 										, 0
 										);
 								}
@@ -5981,7 +5981,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 								{
 									deviceCtx->DrawIndexed(numIndices
 										, draw.m_startIndex
-										, draw.m_startVertex
+										, draw.m_stream[0].m_startVertex
 										);
 								}
 							}
@@ -5996,14 +5996,14 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 							{
 								deviceCtx->DrawInstanced(numVertices
 									, draw.m_numInstances
-									, draw.m_startVertex
+									, draw.m_stream[0].m_startVertex
 									, 0
 									);
 							}
 							else
 							{
 								deviceCtx->Draw(numVertices
-									, draw.m_startVertex
+									, draw.m_stream[0].m_startVertex
 									);
 							}
 						}
diff --git a/src/renderer_d3d12.cpp b/src/renderer_d3d12.cpp
index d60a5039d..2cd96fa2d 100644
--- a/src/renderer_d3d12.cpp
+++ b/src/renderer_d3d12.cpp
@@ -3431,10 +3431,10 @@ data.NumQualityLevels = 0;
 	{
 		Enum type = Enum(!!isValid(_draw.m_indexBuffer) );
 
-		VertexBufferD3D12& vb = s_renderD3D12->m_vertexBuffers[_draw.m_vertexBuffer.idx];
+		VertexBufferD3D12& vb = s_renderD3D12->m_vertexBuffers[_draw.m_stream[0].m_handle.idx];
 		vb.setState(_commandList, D3D12_RESOURCE_STATE_GENERIC_READ);
 
-		uint16_t declIdx = !isValid(vb.m_decl) ? _draw.m_vertexDecl.idx : vb.m_decl.idx;
+		uint16_t declIdx = !isValid(vb.m_decl) ? _draw.m_stream[0].m_decl.idx : vb.m_decl.idx;
 		const VertexDecl& vertexDecl = s_renderD3D12->m_vertexDecls[declIdx];
 		uint32_t numIndices = 0;
 
@@ -3464,7 +3464,7 @@ data.NumQualityLevels = 0;
 			}
 			cmd.draw.InstanceCount = _draw.m_numInstances;
 			cmd.draw.VertexCountPerInstance = numVertices;
-			cmd.draw.StartVertexLocation    = _draw.m_startVertex;
+			cmd.draw.StartVertexLocation    = _draw.m_stream[0].m_startVertex;
 			cmd.draw.StartInstanceLocation  = 0;
 		}
 		else
@@ -3506,7 +3506,7 @@ data.NumQualityLevels = 0;
 			cmd.drawIndexed.IndexCountPerInstance = numIndices;
 			cmd.drawIndexed.InstanceCount = _draw.m_numInstances;
 			cmd.drawIndexed.StartIndexLocation = _draw.m_startIndex;
-			cmd.drawIndexed.BaseVertexLocation = _draw.m_startVertex;
+			cmd.drawIndexed.BaseVertexLocation = _draw.m_stream[0].m_startVertex;
 			cmd.drawIndexed.StartInstanceLocation = 0;
 		}
 
@@ -5200,7 +5200,7 @@ data.NumQualityLevels = 0;
 
 				rendererUpdateUniforms(this, _render->m_uniformBuffer, draw.m_constBegin, draw.m_constEnd);
 
-				if (isValid(draw.m_vertexBuffer) )
+				if (isValid(draw.m_stream[0].m_handle) )
 				{
 					const uint64_t state = draw.m_stateFlags;
 					bool hasFactor = 0
@@ -5210,8 +5210,8 @@ data.NumQualityLevels = 0;
 						|| f3 == (state & f3)
 						;
 
-					const VertexBufferD3D12& vb = m_vertexBuffers[draw.m_vertexBuffer.idx];
-					uint16_t declIdx = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
+					const VertexBufferD3D12& vb = m_vertexBuffers[draw.m_stream[0].m_handle.idx];
+					uint16_t declIdx = !isValid(vb.m_decl) ? draw.m_stream[0].m_decl.idx : vb.m_decl.idx;
 
 					ID3D12PipelineState* pso =
 						getPipelineState(state
diff --git a/src/renderer_d3d9.cpp b/src/renderer_d3d9.cpp
index cc3555c4b..ebd8a5b9b 100644
--- a/src/renderer_d3d9.cpp
+++ b/src/renderer_d3d9.cpp
@@ -4021,22 +4021,22 @@ namespace bgfx { namespace d3d9
 				}
 
 				if (programChanged
-				||  currentState.m_vertexBuffer.idx != draw.m_vertexBuffer.idx
+				||  currentState.m_stream[0].m_handle.idx != draw.m_stream[0].m_handle.idx
 				||  currentState.m_instanceDataBuffer.idx != draw.m_instanceDataBuffer.idx
-				||  currentState.m_instanceDataOffset != draw.m_instanceDataOffset
-				||  currentState.m_instanceDataStride != draw.m_instanceDataStride)
+				||  currentState.m_instanceDataOffset     != draw.m_instanceDataOffset
+				||  currentState.m_instanceDataStride     != draw.m_instanceDataStride)
 				{
-					currentState.m_vertexBuffer = draw.m_vertexBuffer;
+					currentState.m_stream[0].m_handle     = draw.m_stream[0].m_handle;
 					currentState.m_instanceDataBuffer.idx = draw.m_instanceDataBuffer.idx;
-					currentState.m_instanceDataOffset = draw.m_instanceDataOffset;
-					currentState.m_instanceDataStride = draw.m_instanceDataStride;
+					currentState.m_instanceDataOffset     = draw.m_instanceDataOffset;
+					currentState.m_instanceDataStride     = draw.m_instanceDataStride;
 
-					uint16_t handle = draw.m_vertexBuffer.idx;
+					uint16_t handle = draw.m_stream[0].m_handle.idx;
 					if (invalidHandle != handle)
 					{
 						const VertexBufferD3D9& vb = m_vertexBuffers[handle];
 
-						uint16_t decl = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
+						uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[0].m_decl.idx : vb.m_decl.idx;
 						const VertexDeclD3D9& vertexDecl = m_vertexDecls[decl];
 						DX_CHECK(device->SetStreamSource(0, vb.m_ptr, 0, vertexDecl.m_decl.m_stride) );
 
@@ -4082,13 +4082,13 @@ namespace bgfx { namespace d3d9
 					}
 				}
 
-				if (isValid(currentState.m_vertexBuffer) )
+				if (isValid(currentState.m_stream[0].m_handle) )
 				{
 					uint32_t numVertices = draw.m_numVertices;
 					if (UINT32_MAX == numVertices)
 					{
-						const VertexBufferD3D9& vb = m_vertexBuffers[currentState.m_vertexBuffer.idx];
-						uint16_t decl = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
+						const VertexBufferD3D9& vb = m_vertexBuffers[currentState.m_stream[0].m_handle.idx];
+						uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[0].m_decl.idx : vb.m_decl.idx;
 						const VertexDeclD3D9& vertexDecl = m_vertexDecls[decl];
 						numVertices = vb.m_size/vertexDecl.m_decl.m_stride;
 					}
@@ -4115,7 +4115,7 @@ namespace bgfx { namespace d3d9
 							numPrimsRendered  = numPrimsSubmitted*draw.m_numInstances;
 
 							DX_CHECK(device->DrawIndexedPrimitive(prim.m_type
-								, draw.m_startVertex
+								, draw.m_stream[0].m_startVertex
 								, 0
 								, numVertices
 								, 0
@@ -4130,7 +4130,7 @@ namespace bgfx { namespace d3d9
 							numPrimsRendered  = numPrimsSubmitted*draw.m_numInstances;
 
 							DX_CHECK(device->DrawIndexedPrimitive(prim.m_type
-								, draw.m_startVertex
+								, draw.m_stream[0].m_startVertex
 								, 0
 								, numVertices
 								, draw.m_startIndex
@@ -4145,7 +4145,7 @@ namespace bgfx { namespace d3d9
 						numPrimsRendered  = numPrimsSubmitted*draw.m_numInstances;
 
 						DX_CHECK(device->DrawPrimitive(prim.m_type
-							, draw.m_startVertex
+							, draw.m_stream[0].m_startVertex
 							, numPrimsSubmitted
 							) );
 					}
diff --git a/src/renderer_gl.cpp b/src/renderer_gl.cpp
index b97ee35c9..ad5fd6d78 100644
--- a/src/renderer_gl.cpp
+++ b/src/renderer_gl.cpp
@@ -6779,25 +6779,27 @@ namespace bgfx { namespace gl
 					}
 
 					if (0 != defaultVao
-					&&  0 == draw.m_startVertex
+					&&  0 == draw.m_stream[0].m_startVertex
 					&&  0 == draw.m_instanceDataOffset)
 					{
 						if (programChanged
-						||  baseVertex                        != draw.m_startVertex
-						||  currentState.m_vertexBuffer.idx   != draw.m_vertexBuffer.idx
-						||  currentState.m_indexBuffer.idx    != draw.m_indexBuffer.idx
-						||  currentState.m_instanceDataOffset != draw.m_instanceDataOffset
-						||  currentState.m_instanceDataStride != draw.m_instanceDataStride
+						||  baseVertex                            != draw.m_stream[0].m_startVertex
+						||  currentState.m_stream[0].m_handle.idx != draw.m_stream[0].m_handle.idx
+						||  currentState.m_indexBuffer.idx        != draw.m_indexBuffer.idx
+						||  currentState.m_instanceDataOffset     != draw.m_instanceDataOffset
+						||  currentState.m_instanceDataStride     != draw.m_instanceDataStride
 						||  currentState.m_instanceDataBuffer.idx != draw.m_instanceDataBuffer.idx)
 						{
 							bx::HashMurmur2A murmur;
 							murmur.begin();
-							murmur.add(draw.m_vertexBuffer.idx);
 
-							if (isValid(draw.m_vertexBuffer) )
+							const Stream& stream = draw.m_stream[0];
+							murmur.add(stream.m_handle.idx);
+
+							if (isValid(stream.m_handle) )
 							{
-								const VertexBufferGL& vb = m_vertexBuffers[draw.m_vertexBuffer.idx];
-								uint16_t decl = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
+								const VertexBufferGL& vb = m_vertexBuffers[stream.m_handle.idx];
+								uint16_t decl = !isValid(vb.m_decl) ? stream.m_decl.idx : vb.m_decl.idx;
 								murmur.add(decl);
 							}
 
@@ -6808,11 +6810,12 @@ namespace bgfx { namespace gl
 							murmur.add(programIdx);
 							uint32_t hash = murmur.end();
 
-							currentState.m_vertexBuffer = draw.m_vertexBuffer;
+							currentState.m_stream[0].m_handle = stream.m_handle;
+							baseVertex                        = stream.m_startVertex;
+
 							currentState.m_indexBuffer = draw.m_indexBuffer;
 							currentState.m_instanceDataOffset = draw.m_instanceDataOffset;
 							currentState.m_instanceDataStride = draw.m_instanceDataStride;
-							baseVertex = draw.m_startVertex;
 
 							GLuint id = m_vaoStateCache.find(hash);
 							if (UINT32_MAX != id)
@@ -6828,14 +6831,14 @@ namespace bgfx { namespace gl
 
 								program.add(hash);
 
-								if (isValid(draw.m_vertexBuffer) )
+								if (isValid(stream.m_handle) )
 								{
-									VertexBufferGL& vb = m_vertexBuffers[draw.m_vertexBuffer.idx];
+									VertexBufferGL& vb = m_vertexBuffers[stream.m_handle.idx];
 									vb.add(hash);
 									GL_CHECK(glBindBuffer(GL_ARRAY_BUFFER, vb.m_id) );
 
-									uint16_t decl = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
-									program.bindAttributes(m_vertexDecls[decl], draw.m_startVertex);
+									uint16_t decl = !isValid(vb.m_decl) ? stream.m_decl.idx : vb.m_decl.idx;
+									program.bindAttributes(m_vertexDecls[decl], stream.m_startVertex);
 
 									if (isValid(draw.m_instanceDataBuffer) )
 									{
@@ -6869,24 +6872,24 @@ namespace bgfx { namespace gl
 						&&  0 != currentVao)
 						{
 							GL_CHECK(glBindVertexArray(defaultVao) );
-							currentState.m_vertexBuffer.idx = invalidHandle;
-							currentState.m_indexBuffer.idx = invalidHandle;
+							currentState.m_stream[0].m_handle.idx = invalidHandle;
+							currentState.m_indexBuffer.idx        = invalidHandle;
 							bindAttribs = true;
 							currentVao = 0;
 						}
 
 						if (programChanged
-						||  currentState.m_vertexBuffer.idx != draw.m_vertexBuffer.idx
+						||  currentState.m_stream[0].m_handle.idx != draw.m_stream[0].m_handle.idx
 						||  currentState.m_instanceDataBuffer.idx != draw.m_instanceDataBuffer.idx
-						||  currentState.m_instanceDataOffset != draw.m_instanceDataOffset
-						||  currentState.m_instanceDataStride != draw.m_instanceDataStride)
+						||  currentState.m_instanceDataOffset     != draw.m_instanceDataOffset
+						||  currentState.m_instanceDataStride     != draw.m_instanceDataStride)
 						{
-							currentState.m_vertexBuffer = draw.m_vertexBuffer;
+							currentState.m_stream[0].m_handle     = draw.m_stream[0].m_handle;
 							currentState.m_instanceDataBuffer.idx = draw.m_instanceDataBuffer.idx;
-							currentState.m_instanceDataOffset = draw.m_instanceDataOffset;
-							currentState.m_instanceDataStride = draw.m_instanceDataStride;
+							currentState.m_instanceDataOffset     = draw.m_instanceDataOffset;
+							currentState.m_instanceDataStride     = draw.m_instanceDataStride;
 
-							uint16_t handle = draw.m_vertexBuffer.idx;
+							uint16_t handle = draw.m_stream[0].m_handle.idx;
 							if (invalidHandle != handle)
 							{
 								VertexBufferGL& vb = m_vertexBuffers[handle];
@@ -6915,15 +6918,15 @@ namespace bgfx { namespace gl
 							}
 						}
 
-						if (isValid(currentState.m_vertexBuffer) )
+						if (isValid(currentState.m_stream[0].m_handle) )
 						{
-							if (baseVertex != draw.m_startVertex
+							if (baseVertex != draw.m_stream[0].m_startVertex
 							||  bindAttribs)
 							{
-								baseVertex = draw.m_startVertex;
-								const VertexBufferGL& vb = m_vertexBuffers[draw.m_vertexBuffer.idx];
-								uint16_t decl = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
-								program.bindAttributes(m_vertexDecls[decl], draw.m_startVertex);
+								baseVertex = draw.m_stream[0].m_startVertex;
+								const VertexBufferGL& vb = m_vertexBuffers[draw.m_stream[0].m_handle.idx];
+								uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[0].m_decl.idx : vb.m_decl.idx;
+								program.bindAttributes(m_vertexDecls[decl], draw.m_stream[0].m_startVertex);
 
 								if (isValid(draw.m_instanceDataBuffer) )
 								{
@@ -6934,13 +6937,13 @@ namespace bgfx { namespace gl
 						}
 					}
 
-					if (isValid(currentState.m_vertexBuffer) )
+					if (isValid(currentState.m_stream[0].m_handle) )
 					{
 						uint32_t numVertices = draw.m_numVertices;
 						if (UINT32_MAX == numVertices)
 						{
-							const VertexBufferGL& vb = m_vertexBuffers[currentState.m_vertexBuffer.idx];
-							uint16_t decl = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
+							const VertexBufferGL& vb = m_vertexBuffers[currentState.m_stream[0].m_handle.idx];
+							uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[0].m_decl.idx : vb.m_decl.idx;
 							const VertexDecl& vertexDecl = m_vertexDecls[decl];
 							numVertices = vb.m_size/vertexDecl.m_stride;
 						}
diff --git a/src/renderer_mtl.mm b/src/renderer_mtl.mm
index d04ff0b0f..3c44dd832 100644
--- a/src/renderer_mtl.mm
+++ b/src/renderer_mtl.mm
@@ -3382,13 +3382,13 @@ namespace bgfx { namespace mtl
 
 				if (key.m_program != programIdx
 				|| (BGFX_STATE_BLEND_MASK|BGFX_STATE_BLEND_EQUATION_MASK|BGFX_STATE_ALPHA_WRITE|BGFX_STATE_RGB_WRITE|BGFX_STATE_BLEND_INDEPENDENT|BGFX_STATE_MSAA|BGFX_STATE_BLEND_ALPHA_TO_COVERAGE) & changedFlags
-				||  currentState.m_vertexBuffer.idx != draw.m_vertexBuffer.idx
-				||  currentState.m_vertexDecl.idx != draw.m_vertexDecl.idx
-				||  currentState.m_instanceDataStride != draw.m_instanceDataStride
+				||  currentState.m_stream[0].m_handle.idx != draw.m_stream[0].m_handle.idx
+				||  currentState.m_stream[0].m_decl.idx   != draw.m_stream[0].m_decl.idx
+				||  currentState.m_instanceDataStride     != draw.m_instanceDataStride
 				|| ( (blendFactor != draw.m_rgba) && !!(newFlags & BGFX_STATE_BLEND_INDEPENDENT) ) )
 				{
 					programIdx = key.m_program;
-					currentState.m_vertexDecl         = draw.m_vertexDecl;
+					currentState.m_stream[0].m_decl   = draw.m_stream[0].m_decl;
 					currentState.m_instanceDataStride = draw.m_instanceDataStride;
 
 					if (invalidHandle == programIdx)
@@ -3403,18 +3403,18 @@ namespace bgfx { namespace mtl
 
 						RenderPipelineState pipelineState = NULL;
 
-						if ( isValid(draw.m_vertexBuffer) )
+						if (isValid(draw.m_stream[0].m_handle) )
 						{
-							uint16_t handle = draw.m_vertexBuffer.idx;
+							uint16_t handle = draw.m_stream[0].m_handle.idx;
 							const VertexBufferMtl& vb = m_vertexBuffers[handle];
 							VertexDeclHandle decl;
-							decl.idx = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
+							decl.idx = !isValid(vb.m_decl) ? draw.m_stream[0].m_decl.idx : vb.m_decl.idx;
 
 							pipelineState = program.getRenderPipelineState(newFlags, draw.m_rgba, fbh, decl, draw.m_instanceDataStride/16);
 						}
 
-						if (NULL == pipelineState )
-						{  //call with invalid program
+						if (NULL == pipelineState)
+						{
 							currentProgram = NULL;
 							programIdx = invalidHandle;
 							continue;
@@ -3500,24 +3500,24 @@ namespace bgfx { namespace mtl
 					}
 				}
 
-				if (currentState.m_vertexBuffer.idx       != draw.m_vertexBuffer.idx
-				||  currentState.m_startVertex            != draw.m_startVertex
-				||  currentState.m_instanceDataBuffer.idx != draw.m_instanceDataBuffer.idx
-				||  currentState.m_instanceDataOffset     != draw.m_instanceDataOffset)
+				if (currentState.m_stream[0].m_handle.idx  != draw.m_stream[0].m_handle.idx
+				||  currentState.m_stream[0].m_startVertex != draw.m_stream[0].m_startVertex
+				||  currentState.m_instanceDataBuffer.idx  != draw.m_instanceDataBuffer.idx
+				||  currentState.m_instanceDataOffset      != draw.m_instanceDataOffset)
 				{
-					currentState.m_vertexBuffer           = draw.m_vertexBuffer;
-					currentState.m_startVertex            = draw.m_startVertex;
-					currentState.m_instanceDataBuffer.idx = draw.m_instanceDataBuffer.idx;
-					currentState.m_instanceDataOffset     = draw.m_instanceDataOffset;
+					currentState.m_stream[0].m_handle       = draw.m_stream[0].m_handle;
+					currentState.m_stream[0].m_startVertex  = draw.m_stream[0].m_startVertex;
+					currentState.m_instanceDataBuffer.idx   = draw.m_instanceDataBuffer.idx;
+					currentState.m_instanceDataOffset       = draw.m_instanceDataOffset;
 
-					uint16_t handle = draw.m_vertexBuffer.idx;
+					uint16_t handle = draw.m_stream[0].m_handle.idx;
 					if (invalidHandle != handle)
 					{
 						const VertexBufferMtl& vb = m_vertexBuffers[handle];
 
-						uint16_t decl = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
+						uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[0].m_decl.idx : vb.m_decl.idx;
 						const VertexDecl& vertexDecl = m_vertexDecls[decl];
-						uint32_t offset = draw.m_startVertex  * vertexDecl.getStride();
+						uint32_t offset = draw.m_stream[0].m_startVertex  * vertexDecl.getStride();
 
 						rce.setVertexBuffer(vb.getBuffer(), offset, 1);
 
@@ -3529,13 +3529,13 @@ namespace bgfx { namespace mtl
 					}
 				}
 
-				if (isValid(currentState.m_vertexBuffer) )
+				if (isValid(currentState.m_stream[0].m_handle) )
 				{
 					uint32_t numVertices = draw.m_numVertices;
 					if (UINT32_MAX == numVertices)
 					{
-						const VertexBufferMtl& vb = m_vertexBuffers[currentState.m_vertexBuffer.idx];
-						uint16_t decl = !isValid(vb.m_decl) ? draw.m_vertexDecl.idx : vb.m_decl.idx;
+						const VertexBufferMtl& vb = m_vertexBuffers[currentState.m_stream[0].m_handle.idx];
+						uint16_t decl = !isValid(vb.m_decl) ? draw.m_stream[0].m_decl.idx : vb.m_decl.idx;
 						const VertexDecl& vertexDecl = m_vertexDecls[decl];
 						numVertices = vb.m_size/vertexDecl.m_stride;
 					}