Differential D2938 Diff 15876 source/renderer/InstancingModelRenderer.cpp

Changeset View

Standalone View

source/renderer/InstancingModelRenderer.cpp

Show First 20 Lines • Show All 241 Lines • ▼ Show 20 Lines	struct InstancingModelRendererInternals

bool calculateTangents;		bool calculateTangents;

/// Previously prepared modeldef		/// Previously prepared modeldef
IModelDef* imodeldef;		IModelDef* imodeldef;

/// Index base for imodeldef		/// Index base for imodeldef
u8* imodeldefIndexBase;		u8* imodeldefIndexBase;
};

		GLuint buff;
		std::vector<CMatrix3D> transforms;
		};

// Construction and Destruction		// Construction and Destruction
InstancingModelRenderer::InstancingModelRenderer(bool gpuSkinning, bool calculateTangents)		InstancingModelRenderer::InstancingModelRenderer(bool gpuSkinning, bool calculateTangents)
{		{
m = new InstancingModelRendererInternals;		m = new InstancingModelRendererInternals;
m->gpuSkinning = gpuSkinning;		m->gpuSkinning = gpuSkinning;
m->calculateTangents = calculateTangents;		m->calculateTangents = calculateTangents;
m->imodeldef = 0;		m->imodeldef = 0;

		pglGenBuffersARB(1, &m->buff);
}		}
		vladislavbelovUnsubmitted Not Done Inline Actions If you don't use VertexBuffer then you need to wait some my refactorings. Because it's bad to introduce new low-level GL calls into high-level or middle-level code. vladislavbelov: If you don't use VertexBuffer then you need to wait some my refactorings. Because it's bad to…
		wraitiiAuthorUnsubmitted Done Inline Actions This really isn't high/mid-level code, to be honest. It's the code that actually draws stuff, it's as low-level as you get in our renderer architecture. I get your concern though, I don't really mind waiting, for now anyways. wraitii: This really isn't high/mid-level code, to be honest. It's the code that actually draws stuff…
		vladislavbelovUnsubmitted Not Done Inline Actions If a level knows about models and materials then it's not the low level, it's at least middle-level. Low-level includes direct GL calls and knowing about GPU structure. vladislavbelov: If a level knows about models and materials then it's not the low level, it's at least middle…

InstancingModelRenderer::~InstancingModelRenderer()		InstancingModelRenderer::~InstancingModelRenderer()
{		{
		pglDeleteBuffersARB(1, &m->buff);
		vladislavbelovUnsubmitted Not Done Inline Actions Should use `VertexBuffer` instead. vladislavbelov: Should use `VertexBuffer` instead.
delete m;		delete m;
}		}


// Build modeldef data if necessary - we have no per-CModel data		// Build modeldef data if necessary - we have no per-CModel data
CModelRData* InstancingModelRenderer::CreateModelData(const void* key, CModel* model)		CModelRData* InstancingModelRenderer::CreateModelData(const void* key, CModel* model)
{		{
CModelDefPtr mdef = model->GetModelDef();		CModelDefPtr mdef = model->GetModelDef();
Show All 14 Lines
}		}


void InstancingModelRenderer::UpdateModelData(CModel* UNUSED(model), CModelRData* UNUSED(data), int UNUSED(updateflags))		void InstancingModelRenderer::UpdateModelData(CModel* UNUSED(model), CModelRData* UNUSED(data), int UNUSED(updateflags))
{		{
// We have no per-CModel data		// We have no per-CModel data
}		}


// Setup one rendering pass.		// Setup one rendering pass.
void InstancingModelRenderer::BeginPass(int streamflags)		void InstancingModelRenderer::BeginPass(int streamflags)
{		{
ENSURE(streamflags == (streamflags & (STREAM_POS\|STREAM_NORMAL\|STREAM_UV0\|STREAM_UV1)));		ENSURE(streamflags == (streamflags & (STREAM_POS\|STREAM_NORMAL\|STREAM_UV0\|STREAM_UV1)));

		pglVertexAttribDivisorARB(10, 1);
		pglVertexAttribDivisorARB(11, 1);
		pglVertexAttribDivisorARB(12, 1);
		pglVertexAttribDivisorARB(13, 1);
}		}

// Cleanup rendering pass.		// Cleanup rendering pass.
void InstancingModelRenderer::EndPass(int UNUSED(streamflags))		void InstancingModelRenderer::EndPass(int UNUSED(streamflags))
{		{
CVertexBuffer::Unbind();		CVertexBuffer::Unbind();
}		}

Show All 32 Lines	void InstancingModelRenderer::PrepareModelDef(const CShaderProgramPtr& shader, int streamflags, const CModelDef& def)
// GPU skinning requires extra attributes to compute positions/normals		// GPU skinning requires extra attributes to compute positions/normals
if (m->gpuSkinning)		if (m->gpuSkinning)
{		{
shader->VertexAttribPointer(str_a_skinJoints, 4, GL_UNSIGNED_BYTE, GL_FALSE, stride, base + m->imodeldef->m_BlendJoints.offset);		shader->VertexAttribPointer(str_a_skinJoints, 4, GL_UNSIGNED_BYTE, GL_FALSE, stride, base + m->imodeldef->m_BlendJoints.offset);
shader->VertexAttribPointer(str_a_skinWeights, 4, GL_UNSIGNED_BYTE, GL_TRUE, stride, base + m->imodeldef->m_BlendWeights.offset);		shader->VertexAttribPointer(str_a_skinWeights, 4, GL_UNSIGNED_BYTE, GL_TRUE, stride, base + m->imodeldef->m_BlendWeights.offset);
}		}

shader->AssertPointersBound();		shader->AssertPointersBound();

		pglBindBufferARB(GL_ARRAY_BUFFER, m->buff);
}		}


// Render one model		// Render one model
void InstancingModelRenderer::RenderModel(const CShaderProgramPtr& shader, int UNUSED(streamflags), CModel* model, CModelRData* UNUSED(data))		void InstancingModelRenderer::RenderModel(const CShaderProgramPtr& shader, int UNUSED(streamflags), CModel* model, CModelRData* UNUSED(data))
{		{
const CModelDefPtr& mdldef = model->GetModelDef();		const CModelDefPtr& mdldef = model->GetModelDef();

if (m->gpuSkinning)		if (m->gpuSkinning)
		wraitiiAuthorUnsubmitted Done Inline Actions This might actually make ARB performance slightly worse than it should be with the GetUniformBinding, need to check wraitii: This might actually make ARB performance slightly worse than it should be with the…
{		{
// Bind matrices for current animation state.		// Bind matrices for current animation state.
// Add 1 to NumBones because of the special 'root' bone.		// Add 1 to NumBones because of the special 'root' bone.
		vladislavbelovUnsubmitted Not Done Inline Actions Should return `false` in `CanInstance` if `m->gpuSkinning`. vladislavbelov: Should return `false` in `CanInstance` if `m->gpuSkinning`.
		wraitiiAuthorUnsubmitted Done Inline Actions I've actually fixed that so Instancing + GPU Skinning is working. I'll update the comment. wraitii: I've actually fixed that so Instancing + GPU Skinning is working. I'll update the comment.
// HACK: NVIDIA drivers return uniform name with "[0]", Intel Windows drivers without;		// HACK: NVIDIA drivers return uniform name with "[0]", Intel Windows drivers without;
// try uploading both names since one of them should work, and this is easier than		// try uploading both names since one of them should work, and this is easier than
// canonicalising the uniform names in CShaderProgramGLSL		// canonicalising the uniform names in CShaderProgramGLSL
shader->Uniform(str_skinBlendMatrices_0, mdldef->GetNumBones() + 1, model->GetAnimatedBoneMatrices());		shader->Uniform(str_skinBlendMatrices_0, mdldef->GetNumBones() + 1, model->GetAnimatedBoneMatrices());
shader->Uniform(str_skinBlendMatrices, mdldef->GetNumBones() + 1, model->GetAnimatedBoneMatrices());		shader->Uniform(str_skinBlendMatrices, mdldef->GetNumBones() + 1, model->GetAnimatedBoneMatrices());
}		}

// render the lot		// render the lot
size_t numFaces = mdldef->GetNumFaces();		size_t numFaces = mdldef->GetNumFaces();

if (!g_Renderer.m_SkipSubmit)		if (!g_Renderer.m_SkipSubmit)
{		{
// Draw with DrawRangeElements where available, since it might be more efficient		// Draw with DrawRangeElements where available, since it might be more efficient
#if CONFIG2_GLES		#if CONFIG2_GLES
		StanUnsubmitted Done Inline Actions Cast, magic numbers Stan: Cast, magic numbers
		vladislavbelovUnsubmitted Done Inline Actions Means possible reallocation on each different model. Also it might make sense to not draw instances if their number less than 2-4. vladislavbelov: Means possible reallocation on each different model. Also it might make sense to not draw…
		wraitiiAuthorUnsubmitted Done Inline Actions Means possible reallocation on each different model. That's "wanted" in that I don't want synchronisation between my draw calls, so in principle this orphans the buffer and may or may not reallocate. Also it might make sense to not draw instances if their number less than 2-4. I'd agree conceptually, since since we have a list of models to render already I feel like it cannot be slower. Empirically it didn't seem to change much on my system. wraitii: > Means possible reallocation on each different model. That's "wanted" in that I don't want…
		vladislavbelovUnsubmitted Done Inline Actions That's "wanted" in that I don't want synchronisation between my draw calls, so in principle this orphans the buffer and may or may not reallocate. I believe it works properly (without reallocating) only after a draw call is finished, else you'll have a new allocation for each new buffer. Since the draw call might be drawn or not. So it might make sense to collect transforms first. vladislavbelov: > That's "wanted" in that I don't want synchronisation between my draw calls, so in principle…
glDrawElements(GL_TRIANGLES, (GLsizei)numFaces*3, GL_UNSIGNED_SHORT, m->imodeldefIndexBase);		glDrawElements(GL_TRIANGLES, (GLsizei)numFaces*3, GL_UNSIGNED_SHORT, m->imodeldefIndexBase);
#else		#else
pglDrawRangeElementsEXT(GL_TRIANGLES, 0, (GLuint)m->imodeldef->m_Array.GetNumVertices()-1,		pglDrawRangeElementsEXT(GL_TRIANGLES, 0, (GLuint)m->imodeldef->m_Array.GetNumVertices()-1,
		vladislavbelovUnsubmitted Done Inline Actions Useless items copying. vladislavbelov: Useless items copying.
		wraitiiAuthorUnsubmitted Done Inline Actions What are you referring to here? I need to copy the model transform so I can glBufferSubData them? Or would you call glBufferSubData N times? wraitii: What are you referring to here? I need to copy the model transform so I can glBufferSubData…
		vladislavbelovUnsubmitted Done Inline Actions No, reserving before clear might cause copying of `m->transforms`. vladislavbelov: No, reserving before clear might cause copying of `m->transforms`.
(GLsizei)numFaces*3, GL_UNSIGNED_SHORT, m->imodeldefIndexBase);		(GLsizei)numFaces*3, GL_UNSIGNED_SHORT, m->imodeldefIndexBase);
#endif		#endif
}		}

// bump stats		// bump stats
g_Renderer.m_Stats.m_DrawCalls++;		g_Renderer.m_Stats.m_DrawCalls++;
g_Renderer.m_Stats.m_ModelTris += numFaces;		g_Renderer.m_Stats.m_ModelTris += numFaces;

}		}

		void InstancingModelRenderer::RenderInstancedModel(const CShaderProgramPtr& shader, const std::vector<CModel*>& models)
		{
		if (g_Renderer.m_SkipSubmit)
		return;

		const CModelDefPtr& mdldef = models.front()->GetModelDef();

		if (m->gpuSkinning)
		{
		// HACK: this gives the same animation to all similar modeldefs,
		// which is somewhat obviously broken, but it renders something.

		StanUnsubmitted Not Done Inline Actions From https://github.com/0ad/0ad/blob/a4019a9d4986ab100d78e47f6696b4e61b8b2f79/source/renderer/InstancingModelRenderer.cpp#L357 Stan: From https://github.com/0ad/0ad/blob/a4019a9d4986ab100d78e47f6696b4e61b8b2f79/source/renderer/I…
		// Bind matrices for current animation state.
		// Add 1 to NumBones because of the special 'root' bone.
		// HACK: NVIDIA drivers return uniform name with "[0]", Intel Windows drivers without;
		// try uploading both names since one of them should work, and this is easier than
		// canonicalising the uniform names in CShaderProgramGLSL
		shader->Uniform(str_skinBlendMatrices_0, mdldef->GetNumBones() + 1, models.front()->GetAnimatedBoneMatrices());
		shader->Uniform(str_skinBlendMatrices, mdldef->GetNumBones() + 1, models.front()->GetAnimatedBoneMatrices());
		}

		size_t numFaces = mdldef->GetNumFaces();

		pglBufferDataARB(GL_ARRAY_BUFFER, std::max(64 * (size_t)50, 64 * models.size()), nullptr, GL_STREAM_DRAW);

		// Set up a uniform
		m->transforms.reserve(64);
		m->transforms.clear();

		for (CModel* model : models)
		m->transforms.emplace_back(model->GetTransform());

		/*
		if (idx + 4 * models.size() * 16 > 1024*1024)
		{
		glBufferData(GL_ARRAY_BUFFER, 1024*1024, nullptr, GL_STREAM_DRAW);
		idx = 0;
		}
		glBufferSubData(GL_ARRAY_BUFFER, idx, 4 * models.size() * 16, uniforms.data());
		idx += 4 * models.size() * 16;
		*/
		pglBufferSubDataARB(GL_ARRAY_BUFFER, 0, 64 * models.size(), m->transforms.data());

		/*
		for (int j = 0; j < 4; ++j)
		printf("%f %f %f %f\n", m->transforms.front()[j4], m->transforms.front()[j4+1], m->transforms.front()[j4+2], m->transforms.front()[j4+3]);
		printf("\n");
		*/

		shader->VertexAttribPointer(str_a_instancing0, 4, GL_FLOAT, GL_FALSE, 64, (void*)(0));
		shader->VertexAttribPointer(str_a_instancing1, 4, GL_FLOAT, GL_FALSE, 64, (void*)(0+16));
		shader->VertexAttribPointer(str_a_instancing2, 4, GL_FLOAT, GL_FALSE, 64, (void*)(0+32));
		shader->VertexAttribPointer(str_a_instancing3, 4, GL_FLOAT, GL_FALSE, 64, (void*)(0+48));

		pglDrawElementsInstancedARB(GL_TRIANGLES,
		(GLsizei)numFaces*3,
		GL_UNSIGNED_SHORT,
		m->imodeldefIndexBase, models.size());
		g_Renderer.m_Stats.m_DrawCalls++;
		g_Renderer.m_Stats.m_SavedDrawCalls += models.size() - 1;
		g_Renderer.m_Stats.m_ModelTris += numFaces * models.size();
		}