Jump to content

  • Log In with Google      Sign In   
  • Create Account

GPU Skinning problem (MD5 Doom)


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
2 replies to this topic

#1 DgekGD   Members   -  Reputation: 374

Like
0Likes
Like

Posted 13 November 2012 - 05:05 PM

Hello!

I'm trying to implement GPU skinning using md5 format. But after many hours of trying to figure out what's wrong I still get incorrect results for animation.

Firstly, I calculate Bind Pose and Inverse Bind Pose matrices for each bone:
[source lang="cpp"]Mat4x4 trans; trans.BuildTranslation(joints[i].posObjectSpace); //build translation from joint position in object spaceMat4x4 rot = joints[i].rot.BuildMatrix(); //joints[i].rot is unit quaternionMat4x4 bindPoseTransform = rot *trans;(**ppBones)[i].bindPoseTransform = rot * trans;(**ppBones)[i].inverseBindPoseTransform = bindPoseTransforms.Inverse();[/source]

After that I calculate vertex positions in object space and fill vertex buffer with positions, indices and weights
[source lang="cpp"]for (int j = 0; j < mesh.numVertices; j++) { //calculate position Vec3 pos(0.0f, 0.0f, 0.0f); md5_vertex vertex = mesh.vertices[j]; for (int k = 0; k < vertex.numWeights; k++) { md5_weight weight = mesh.weights[vertex.startWeight+k]; md5_joint joint = joints[weight.jointId]; //calculate vertex position in object space temp = Vec4(weight.posJointSpace); temp.w = 0.0f; Quaternion q = joint.rot; Quaternion revq = joint.rot.Inverse(); Vec3 temp3 = q * temp * revq; temp3 = temp3 + joint.posObjectSpace; temp3 *= weight.weight; pos += temp3;// * weight.weight; }//finally fill vertex structure with data v[j].pos = pos; v[j].texCoord = vertex.texCoords; for (int k = 0; k < vertex.numWeights; k++) { v[j].bonesIndices[k] = mesh.weights[vertex.startWeight+k].jointId; } v[j].bonesWeights = Vec4( vertex.numWeights > 0 ? mesh.weights[vertex.startWeight].weight : 0.0f, vertex.numWeights > 1 ? mesh.weights[vertex.startWeight+1].weight : 0.0f, vertex.numWeights > 2 ? mesh.weights[vertex.startWeight+2].weight : 0.0f, vertex.numWeights > 3 ? mesh.weights[vertex.startWeight+3].weight : 0.0f); assert( fabs(v[j].bonesWeights.x + v[j].bonesWeights.y + v[j].bonesWeights.z + v[j].bonesWeights.w - 1.0f) < 0.001f);[/source]
Next step is to calculate joints transforms in object space for each frame in animation:
[source lang="cpp"]//pos - current joint position relative to the parent (if have any)//rot - quaternion representng rotation of current joint//if we have parent joint... if (joint.id >= 0) { md5_joint parent = joints[joint.id]; //calculate joint pos Quaternion parentOrient = parent.rot; Vec4 temp( pos.x, pos.y, pos.z, 0.0f); pos = parentOrient * temp * parentOrient.Inverse(); pos = pos + parent.posObjectSpace; //concatenate rotation rot = parentOrient * rot'; rot.Normalize(); } //update joint position and orientation joints[j].posObjectSpace = pos; joints[j].rot = rot; Mat4x4 trans; trans.BuildTranslation(pos); Mat4x4 rotation = rot.BuildMatrix(); Mat4x4 transform = rotation * trans; newFrame.m_positions.push_back(pos); newFrame.m_rotations.push_back(rot); newFrame.m_transforms.push_back(transform); //use just fixed frames for debugging, no interpolation for now...[/source]
//Finally I render animation!
[source lang="cpp"]SkinningBuffer skinningBuffer;for (int j = 0; j < m_numBones; j++) { Mat4x4 world = pScene->GetWorld(); //world transform. identity for now Mat4x4 inverseBindPose = m_pBones[j].m_inverseBindPose; //inverse bind pose matrixMat4x4 curObjectSpace = m_pAnimations[0].GetTransform(j); //current object space transform for given bone skinningBuffer.bonesTransforms[j] = Transpose(inverseBindPose * curObjectSpace * world); } //next frame... m_pAnimations[0].NextFrame(); //update constant buffer for shader m_pSkinningBuffer->UpdateAndSet(&amp;amp;skinningBuffer, ST_Vertex, 3);[/source]

Here is my vertex shader code:
[source lang="cpp"]//////////////////////////////////////////////////Constant Buffers////////////////////////////////////////////////cbuffer MatrixBuffer : register(b0){ float4x4 m_ViewProjection;}cbuffer LightBuffer : register(b1){ float4 ambientColor : packoffset(c0); float4 diffuseColor : packoffset(c1); float3 lightDir : packoffset(c2.x); float padb1 : packoffset(c2.w);};cbuffer CameraBuffer : register(b2){ float3 eyePos : packoffset(c0.x); float padb2 : packoffset(c0.w);};cbuffer SkinningBuffer : register(b3){ float4x4 m_BoneWorld[50];};/////////////////////////////////////////////////Texture Resources and Sampler States///////////////////////////////////////////////Texture2D objTexture : register(t0);Texture2D bumpMap : register(t1);Texture2D heightMap : register(t2);SamplerState LinearSampler : register(s0);/////////////////////////////////////////////////Shaders' input/output layouts///////////////////////////////////////////////struct vs_input{ float3 pos : POSITION; uint4 bones : BONEIDS; float4 weights : BONEWEIGHTS; float2 texCoord : TEXCOORD; float3 normal : NORMAL; float3 tangent : TANGENT; float3 binormal : BINORMAL;};struct ps_input{ float4 pos : SV_POSITION; float2 texCoord : TEXCOORD; float4 color : COLOR; float3 normal : NORMAL; float3 tangent : TANGENT; float3 binormal : BINORMAL;};////////////////////////////////////////////////Vertex Shader//////////////////////////////////////////////ps_input VS(vs_input input){ ps_input output; //apply bone transforms to position output.pos = mul(float4(input.pos, 1.0f), m_BoneWorld[input.bones.x] * input.weights.x); output.pos += mul(float4(input.pos, 1.0f), m_BoneWorld[input.bones.y] * input.weights.y); output.pos += mul(float4(input.pos, 1.0f), m_BoneWorld[input.bones.z] * input.weights.z); output.pos += mul(float4(input.pos, 1.0f), m_BoneWorld[input.bones.w] * input.weights.w); output.pos = mul(output.pos, m_ViewProjection); //apply bone transforms to normal output.normal = mul(float4(input.normal, 1.0f), m_BoneWorld[input.bones.x]) * input.weights.x; output.normal += mul(float4(input.normal, 1.0f), m_BoneWorld[input.bones.y]) * input.weights.y; output.normal += mul(float4(input.normal, 1.0f), m_BoneWorld[input.bones.z]) * input.weights.z; output.normal += mul(float4(input.normal, 1.0f), m_BoneWorld[input.bones.w]) * input.weights.w; //apply bone transforms to tangent output.tangent = mul(float4(input.tangent, 1.0f), m_BoneWorld[input.bones.x]) * input.weights.x; output.tangent += mul(float4(input.tangent, 1.0f), m_BoneWorld[input.bones.y]) * input.weights.y; output.tangent += mul(float4(input.tangent, 1.0f), m_BoneWorld[input.bones.z]) * input.weights.z; output.tangent += mul(float4(input.tangent, 1.0f), m_BoneWorld[input.bones.w]) * input.weights.w; //apply bone transforms to normal output.binormal = mul(float4(input.binormal, 1.0f), m_BoneWorld[input.bones.x]) * input.weights.x; output.binormal += mul(float4(input.binormal, 1.0f), m_BoneWorld[input.bones.y]) * input.weights.y; output.binormal += mul(float4(input.binormal, 1.0f), m_BoneWorld[input.bones.z]) * input.weights.z; output.binormal += mul(float4(input.binormal, 1.0f), m_BoneWorld[input.bones.w]) * input.weights.w; //transfer texture coordinates output.texCoord = input.texCoord; //output result return output;}[/source]
I'm pretty sure, that vertex object space positions are calculated correctly and shaders code works ok as I get correct results for Bind Pose. But when animation starts it doesn't work:
Posted Image

Posted Image

The model is taken from http://www.katsbits.com/download/models/ .

To simplify debugging, I created simple model from 2 joints and 4 vertices with two frames of animation and still it doesn't work, but it seems that i checked everything...
Posted Image

This is capture of the second frame(of two). The first frame is rendering correctly showing this black plane lies straight on the floor. In the second frame I just have left part lifted up a bit. Obvisously, I renders incorrectly for me. I attached file with this plane to my post.

Could you give me some ideas why I get wrong result? If you already have implemented GPU skinning could you type your matrices for inverseBindPose of each bone and curObjectTransform of bones for each frame, so that i can compare them with mine(for that black plane model)? I would really appreciate any help!

Thanks!

Attached Files


Edited by DgekGD, 13 November 2012 - 05:15 PM.


Sponsor:

#2 Vilem Otte   Crossbones+   -  Reputation: 1433

Like
1Likes
Like

Posted 13 November 2012 - 08:04 PM

As I've implemented MD5 skinning on GPU, I can give you few tips (I don't have time right now to move through your code as I'm being quite in time-pressure till thursday)...

1.) Start from simple MD5 file, that contains only translation and single bone. If it works, create only rotation, then combine them together, then build double-bone - e.g. trace where the problems lie
2.) Double check your math. I was sitting for few hours on problem with inverse matrices, then I found that VS actually *removes* few instructions from my SSE2 matrix inversion code (... forgot volatiles *huh* .. problem hadn't appeared on my main dev platform - linux & gcc)
3.) And triple check your quaternion <-> matrices. Often the problem lies only in rotations, and doing row-major matrix out of quaternion instead of column-major matrix is totally wrong (because of nature of 3x3 rotational matrix, you will actually get inverse rotation, and this can create mess you see).

And last point - try your skinning (e.g. inverse-bind-pose and matrix transformation) on CPU first, it's a lot easier to debug there, than on GPU. Posted Image

If you'll still have troubles implement it, I can post fragments of my code (with little how-to) here, although not till Thursday evening.

My current blog on programming, linux and stuff - http://gameprogrammerdiary.blogspot.com


#3 DgekGD   Members   -  Reputation: 374

Like
0Likes
Like

Posted 14 November 2012 - 07:58 AM

Thanks for your reply, Vilem Otte!

3.) And triple check your quaternion <-> matrices. Often the problem lies only in rotations, and doing row-major matrix out of quaternion instead of column-major matrix is totally wrong (because of nature of 3x3 rotational matrix, you will actually get inverse rotation, and this can create mess you see).


This is where the problem was coming from! So, i changed my code to transpose rotation matrices after they were built from quaternios and it works perfectly!
[source lang="cpp"]Mat4x4 rotation = rot.BuildMatrix();[/source]
to:
[source lang="cpp"]Mat4x4 rotation = rot.BuildMatrix();rotation.Transpose();[/source]
Thanks again for your help, you saved me from many hours of trying to solve this problem!




Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS