thin3d: Add partial support for 32-bit index buffers. Add DrawIndexedUP.

Common/GPU/D3D11/thin3d_d3d11.cpp

@@ -131,6 +131,8 @@ public:
 	void Draw(int vertexCount, int offset) override;
 	void DrawIndexed(int indexCount, int offset) override;
 	void DrawUP(const void *vdata, int vertexCount) override;
+	void DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount, IndexFormat ifmt) override;
+
 	void Clear(int mask, uint32_t colorval, float depthVal, int stencilVal) override;
 
 	void BeginFrame(DebugFlags debugFlags) override;
@@ -218,6 +220,7 @@ private:
 
 	bool dirtyIndexBuffer_ = false;
 	ID3D11Buffer *nextIndexBuffer_ = nullptr;
+	DXGI_FORMAT nextIndexBufferFormat_ = DXGI_FORMAT_R16_UINT;
 	UINT nextIndexBufferOffset_ = 0;
 
 	InvalidationCallback invalidationCallback_;
@@ -235,6 +238,7 @@ private:
 	// Temporaries
 	ID3D11Texture2D *packTexture_ = nullptr;
 	Buffer *upBuffer_ = nullptr;
+	Buffer *upIBuffer_ = nullptr;
 
 	// System info
 	D3D_FEATURE_LEVEL featureLevel_;
@@ -351,6 +355,7 @@ D3D11DrawContext::D3D11DrawContext(ID3D11Device *device, ID3D11DeviceContext *de
 	const size_t UP_MAX_BYTES = 65536 * 24;
 
 	upBuffer_ = D3D11DrawContext::CreateBuffer(UP_MAX_BYTES, BufferUsageFlag::DYNAMIC | BufferUsageFlag::VERTEXDATA);
+	upIBuffer_ = D3D11DrawContext::CreateBuffer(UP_MAX_BYTES, BufferUsageFlag::DYNAMIC | BufferUsageFlag::INDEXDATA);
 
 	IDXGIDevice1 *dxgiDevice1 = nullptr;
 	hr = device_->QueryInterface(__uuidof(IDXGIDevice), reinterpret_cast<void **>(&dxgiDevice1));
@@ -364,6 +369,7 @@ D3D11DrawContext::~D3D11DrawContext() {
 	DestroyPresets();
 
 	upBuffer_->Release();
+	upIBuffer_->Release();
 	packTexture_->Release();
 
 	// Release references.
@@ -1255,7 +1261,7 @@ void D3D11DrawContext::ApplyCurrentState() {
 		context_->IASetVertexBuffers(0, 1, &nextVertexBuffer_, &curPipeline_->input->stride, &nextVertexBufferOffset_);
 	}
 	if (dirtyIndexBuffer_) {
-		context_->IASetIndexBuffer(nextIndexBuffer_, DXGI_FORMAT_R16_UINT, nextIndexBufferOffset_);
+		context_->IASetIndexBuffer(nextIndexBuffer_, nextIndexBufferFormat_, nextIndexBufferOffset_);
 		dirtyIndexBuffer_ = false;
 	}
 	if (curPipeline_->dynamicUniforms) {
@@ -1334,6 +1340,7 @@ void D3D11DrawContext::BindIndexBuffer(Buffer *indexBuffer, int offset) {
 	dirtyIndexBuffer_ = true;
 	nextIndexBuffer_ = buf ? buf->buf : 0;
 	nextIndexBufferOffset_ = buf ? offset : 0;
+	nextIndexBufferFormat_ = DXGI_FORMAT_R16_UINT;
 }
 
 void D3D11DrawContext::Draw(int vertexCount, int offset) {
@@ -1347,8 +1354,6 @@ void D3D11DrawContext::DrawIndexed(int indexCount, int offset) {
 }
 
 void D3D11DrawContext::DrawUP(const void *vdata, int vertexCount) {
-	ApplyCurrentState();
-
 	int byteSize = vertexCount * curPipeline_->input->stride;
 
 	UpdateBuffer(upBuffer_, (const uint8_t *)vdata, 0, byteSize, Draw::UPDATE_DISCARD);
@@ -1357,6 +1362,20 @@ void D3D11DrawContext::DrawUP(const void *vdata, int vertexCount) {
 	Draw(vertexCount, offset);
 }
 
+void D3D11DrawContext::DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount, IndexFormat ifmt) {
+	int vbyteSize = vertexCount * curPipeline_->input->stride;
+	int ibyteSize = indexCount * (ifmt == IndexFormat::U32 ? 4 : 2);
+
+	UpdateBuffer(upBuffer_, (const uint8_t *)vdata, 0, vbyteSize, Draw::UPDATE_DISCARD);
+	BindVertexBuffer(upBuffer_, 0);
+
+	UpdateBuffer(upIBuffer_, (const uint8_t *)idata, 0, ibyteSize, Draw::UPDATE_DISCARD);
+	BindIndexBuffer(upIBuffer_, 0);
+	// Override the index buffer format.
+	nextIndexBufferFormat_ = ifmt == IndexFormat::U32 ? DXGI_FORMAT_R32_UINT : DXGI_FORMAT_R16_UINT;
+	DrawIndexed(indexCount, 0);
+}
+
 uint32_t D3D11DrawContext::GetDataFormatSupport(DataFormat fmt) const {
 	DXGI_FORMAT giFmt = dataFormatToD3D11(fmt);
 	if (giFmt == DXGI_FORMAT_UNKNOWN)

Common/GPU/D3D9/thin3d_d3d9.cpp

@@ -573,6 +573,7 @@ public:
 	void Draw(int vertexCount, int offset) override;
 	void DrawIndexed(int vertexCount, int offset) override;
 	void DrawUP(const void *vdata, int vertexCount) override;
+	void DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount, IndexFormat ifmt) override;
 	void Clear(int mask, uint32_t colorval, float depthVal, int stencilVal) override;
 
 	uint64_t GetNativeObject(NativeObject obj, void *srcObject) override {
@@ -1172,6 +1173,16 @@ void D3D9Context::DrawUP(const void *vdata, int vertexCount) {
 	device_->DrawPrimitiveUP(curPipeline_->prim, D3DPrimCount(curPipeline_->prim, vertexCount), vdata, curPipeline_->inputLayout->GetStride());
 }
 
+void D3D9Context::DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount, IndexFormat ifmt) {
+	curPipeline_->inputLayout->Apply(device_);
+	curPipeline_->Apply(device_, stencilRef_, stencilWriteMask_, stencilCompareMask_);
+	ApplyDynamicState();
+
+	device_->DrawIndexedPrimitiveUP(curPipeline_->prim, 0, vertexCount, D3DPrimCount(curPipeline_->prim, indexCount),
+		idata, ifmt == IndexFormat::U32 ? D3DFMT_INDEX32 : D3DFMT_INDEX16,
+		vdata, curPipeline_->inputLayout->GetStride());
+}
+
 static uint32_t SwapRB(uint32_t c) {
 	return (c & 0xFF00FF00) | ((c >> 16) & 0xFF) | ((c << 16) & 0xFF0000);
 }

Common/GPU/OpenGL/thin3d_gl.cpp

@@ -443,6 +443,7 @@ public:
 	void Draw(int vertexCount, int offset) override;
 	void DrawIndexed(int vertexCount, int offset) override;
 	void DrawUP(const void *vdata, int vertexCount) override;
+	void DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount, IndexFormat ifmt) override;
 
 	void Clear(int mask, uint32_t colorval, float depthVal, int stencilVal) override;
 
@@ -1417,6 +1418,29 @@ void OpenGLContext::DrawUP(const void *vdata, int vertexCount) {
 	renderManager_.Draw(curPipeline_->inputLayout->inputLayout_, buf, offset, curPipeline_->prim, 0, vertexCount);
 }
 
+void OpenGLContext::DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount, IndexFormat ifmt) {
+	_assert_(curPipeline_->inputLayout != nullptr);
+	int stride = curPipeline_->inputLayout->stride;
+	uint32_t vdataSize = stride * vertexCount;
+	uint32_t idataSize = indexCount * (ifmt == IndexFormat::U32 ? 4 : 2);
+
+	FrameData &frameData = frameData_[renderManager_.GetCurFrame()];
+
+	GLRBuffer *vbuf;
+	uint32_t voffset;
+	uint8_t *dest = frameData.push->Allocate(vdataSize, 4, &vbuf, &voffset);
+	memcpy(dest, vdata, vdataSize);
+
+	GLRBuffer *ibuf;
+	uint32_t ioffset;
+	dest = frameData.push->Allocate(idataSize, 4, &ibuf, &ioffset);
+	memcpy(dest, idata, idataSize);
+
+	ApplySamplers();
+	_assert_(curPipeline_->inputLayout);
+	renderManager_.DrawIndexed(curPipeline_->inputLayout->inputLayout_, vbuf, voffset, ibuf, ioffset, curPipeline_->prim, 0, ifmt == IndexFormat::U32 ? GL_UNSIGNED_INT : GL_UNSIGNED_SHORT, vertexCount);
+}
+
 void OpenGLContext::Clear(int mask, uint32_t colorval, float depthVal, int stencilVal) {
 	float col[4];
 	Uint8x4ToFloat4(col, colorval);

Common/GPU/Vulkan/VulkanQueueRunner.cpp

@@ -1165,7 +1165,7 @@ void VulkanQueueRunner::PerformRenderPass(const VKRStep &step, VkCommandBuffer c
 				VkDescriptorSet set = (*descSets)[c.drawIndexed.descSetIndex].set;
 				_dbg_assert_(set != VK_NULL_HANDLE);
 				vkCmdBindDescriptorSets(cmd, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0, 1, &set, c.drawIndexed.numUboOffsets, c.drawIndexed.uboOffsets);
-				vkCmdBindIndexBuffer(cmd, c.drawIndexed.ibuffer, c.drawIndexed.ioffset, VK_INDEX_TYPE_UINT16);
+				vkCmdBindIndexBuffer(cmd, c.drawIndexed.ibuffer, c.drawIndexed.ioffset, c.drawIndexed.itype);
 				VkDeviceSize voffset = c.drawIndexed.voffset;
 				vkCmdBindVertexBuffers(cmd, 0, 1, &c.drawIndexed.vbuffer, &voffset);
 				vkCmdDrawIndexed(cmd, c.drawIndexed.count, c.drawIndexed.instances, 0, 0, 0);

Common/GPU/Vulkan/VulkanQueueRunner.h

@@ -80,6 +80,7 @@ struct VkRenderData {
 			uint32_t voffset;
 			uint32_t ioffset;
 			uint32_t count;
+			VkIndexType itype;  // TODO: pack this into one of the others, it's really a single bit
 		} drawIndexed;
 		struct {
 			uint32_t clearColor;

Common/GPU/Vulkan/VulkanRenderManager.h

@@ -471,7 +471,7 @@ public:
 		curRenderStep_->render.numDraws++;
 	}
 
-	void DrawIndexed(int descSetIndex, int numUboOffsets, const uint32_t *uboOffsets, VkBuffer vbuffer, int voffset, VkBuffer ibuffer, int ioffset, int count, int numInstances) {
+	void DrawIndexed(int descSetIndex, int numUboOffsets, const uint32_t *uboOffsets, VkBuffer vbuffer, int voffset, VkBuffer ibuffer, int ioffset, int count, int numInstances, VkIndexType itype) {
 		_dbg_assert_(curRenderStep_ && curRenderStep_->stepType == VKRStepType::RENDER && curStepHasViewport_ && curStepHasScissor_);
 		VkRenderData &data = curRenderStep_->commands.push_uninitialized();
 		data.cmd = VKRRenderCommand::DRAW_INDEXED;
@@ -483,6 +483,7 @@ public:
 		data.drawIndexed.ibuffer = ibuffer;
 		data.drawIndexed.ioffset = ioffset;
 		data.drawIndexed.numUboOffsets = numUboOffsets;
+		data.drawIndexed.itype = itype;
 		_dbg_assert_(numUboOffsets <= ARRAY_SIZE(data.drawIndexed.uboOffsets));
 		for (int i = 0; i < numUboOffsets; i++)
 			data.drawIndexed.uboOffsets[i] = uboOffsets[i];

Common/GPU/Vulkan/thin3d_vulkan.cpp

@@ -491,6 +491,7 @@ public:
 	void Draw(int vertexCount, int offset) override;
 	void DrawIndexed(int vertexCount, int offset) override;
 	void DrawUP(const void *vdata, int vertexCount) override;
+	void DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount, IndexFormat ifmt) override;
 
 	void BindCurrentPipeline();
 	void ApplyDynamicState();
@@ -1497,7 +1498,7 @@ void VKContext::DrawIndexed(int vertexCount, int offset) {
 	int descSetIndex;
 	PackedDescriptor *descriptors = renderManager_.PushDescriptorSet(4, &descSetIndex);
 	BindDescriptors(vulkanUBObuf, descriptors);
-	renderManager_.DrawIndexed(descSetIndex, 1, &ubo_offset, vulkanVbuf, (int)vbBindOffset + curVBufferOffset_, vulkanIbuf, (int)ibBindOffset + offset * sizeof(uint32_t), vertexCount, 1);
+	renderManager_.DrawIndexed(descSetIndex, 1, &ubo_offset, vulkanVbuf, (int)vbBindOffset + curVBufferOffset_, vulkanIbuf, (int)ibBindOffset + offset * sizeof(uint32_t), vertexCount, 1, VK_INDEX_TYPE_UINT16);
 }
 
 void VKContext::DrawUP(const void *vdata, int vertexCount) {
@@ -1520,7 +1521,37 @@ void VKContext::DrawUP(const void *vdata, int vertexCount) {
 	int descSetIndex;
 	PackedDescriptor *descriptors = renderManager_.PushDescriptorSet(4, &descSetIndex);
 	BindDescriptors(vulkanUBObuf, descriptors);
-	renderManager_.Draw(descSetIndex, 1, &ubo_offset, vulkanVbuf, (int)vbBindOffset + curVBufferOffset_, vertexCount);
+	renderManager_.Draw(descSetIndex, 1, &ubo_offset, vulkanVbuf, (int)vbBindOffset, vertexCount);
+}
+
+void VKContext::DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount, IndexFormat ifmt) {
+	_dbg_assert_(vertexCount >= 0);
+	_dbg_assert_(indexCount >= 0);
+	if (vertexCount <= 0 || indexCount <= 0) {
+		return;
+	}
+
+	VkBuffer vulkanVbuf, vulkanIbuf, vulkanUBObuf;
+	size_t vdataSize = vertexCount * curPipeline_->stride;
+	uint32_t vbBindOffset;
+	uint8_t *vdataPtr = push_->Allocate(vdataSize, 4, &vulkanVbuf, &vbBindOffset);
+	_assert_(vdataPtr != nullptr);
+	memcpy(vdataPtr, vdata, vdataSize);
+
+	size_t idataSize = indexCount * (ifmt == IndexFormat::U32 ? 4 : 2);
+	uint32_t ibBindOffset;
+	uint8_t *idataPtr = push_->Allocate(idataSize, 4, &vulkanIbuf, &ibBindOffset);
+	_assert_(idataPtr != nullptr);
+	memcpy(idataPtr, vdata, idataSize);
+
+	uint32_t ubo_offset = (uint32_t)curPipeline_->PushUBO(push_, vulkan_, &vulkanUBObuf);
+
+	BindCurrentPipeline();
+	ApplyDynamicState();
+	int descSetIndex;
+	PackedDescriptor *descriptors = renderManager_.PushDescriptorSet(4, &descSetIndex);
+	BindDescriptors(vulkanUBObuf, descriptors);
+	renderManager_.DrawIndexed(descSetIndex, 1, &ubo_offset, vulkanVbuf, (int)vbBindOffset, vulkanIbuf, (int)ibBindOffset, indexCount, 1, ifmt == IndexFormat::U32 ? VK_INDEX_TYPE_UINT32 : VK_INDEX_TYPE_UINT16);
 }
 
 void VKContext::BindCurrentPipeline() {

Common/GPU/thin3d.h

@@ -180,6 +180,11 @@ enum class Facing {
 	CW,
 };
 
+enum class IndexFormat {
+	U16,
+	U32,
+};
+
 enum BorderColor {
 	DONT_CARE,
 	TRANSPARENT_BLACK,
@@ -827,7 +832,8 @@ public:
 	virtual void Draw(int vertexCount, int offset) = 0;
 	virtual void DrawIndexed(int vertexCount, int offset) = 0;  // Always 16-bit indices.
 	virtual void DrawUP(const void *vdata, int vertexCount) = 0;
-	
+	virtual void DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount, IndexFormat ifmt) = 0;  // Supports 32-bit indices, for IMGUI use.
+
 	// Frame management (for the purposes of sync and resource management, necessary with modern APIs). Default implementations here.
 	virtual void BeginFrame(DebugFlags debugFlags) = 0;
 	virtual void EndFrame() = 0;

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -374,7 +374,7 @@ void DrawEngineVulkan::DoFlush() {
 			if (!ibuf) {
 				ibOffset = (uint32_t)pushIndex_->Push(decIndex_, sizeof(uint16_t) * vertexCount, 4, &ibuf);
 			}
-			renderManager->DrawIndexed(descSetIndex, ARRAY_SIZE(dynamicUBOOffsets), dynamicUBOOffsets, vbuf, vbOffset, ibuf, ibOffset, vertexCount, 1);
+			renderManager->DrawIndexed(descSetIndex, ARRAY_SIZE(dynamicUBOOffsets), dynamicUBOOffsets, vbuf, vbOffset, ibuf, ibOffset, vertexCount, 1, VK_INDEX_TYPE_UINT16);
 		} else {
 			renderManager->Draw(descSetIndex, ARRAY_SIZE(dynamicUBOOffsets), dynamicUBOOffsets, vbuf, vbOffset, vertexCount);
 		}
@@ -539,7 +539,7 @@ void DrawEngineVulkan::DoFlush() {
 			VkBuffer vbuf, ibuf;
 			vbOffset = (uint32_t)pushVertex_->Push(result.drawBuffer, numDecodedVerts_ * sizeof(TransformedVertex), 4, &vbuf);
 			ibOffset = (uint32_t)pushIndex_->Push(inds, sizeof(short) * result.drawNumTrans, 4, &ibuf);
-			renderManager->DrawIndexed(descSetIndex, ARRAY_SIZE(dynamicUBOOffsets), dynamicUBOOffsets, vbuf, vbOffset, ibuf, ibOffset, result.drawNumTrans, 1);
+			renderManager->DrawIndexed(descSetIndex, ARRAY_SIZE(dynamicUBOOffsets), dynamicUBOOffsets, vbuf, vbOffset, ibuf, ibOffset, result.drawNumTrans, 1, VK_INDEX_TYPE_UINT16);
 		} else if (result.action == SW_CLEAR) {
 			// Note: we won't get here if the clear is alpha but not color, or color but not alpha.
 			bool clearColor = gstate.isClearModeColorMask();