PDA

View Full Version : Transform Feedback



DmitryM
03-13-2009, 02:38 AM
Hi all,

I'm trying to get GL_EXT_transform_feedback extension work in order to perform hardware skinning effectively. The problem appeared within the function glTransformFeedbackVaryingsEXT.

As described in OpenGL spec this funtion takes an array of strings representing varying names:
glTransformFeedbackVaryingsEXT (GLuint program, GLsizei count, const char ** varyings, GLenum bufferMode);

After reading it I considered this approach to be simple and easy. But in the actual glext.h header downloaded from this site (opengl.org) the prototype for glTransformFeedbackVaryingsEXT takes an array of integers representing varying *locations*:
glTransformFeedbackVaryingsEXT (GLuint program, GLsizei count, const GLint *locations, GLenum bufferMode);

The same prototype can be found in GLee. I'm not against using locations, but where can I get them? As far as I know, there is just a function glGetVaryingLocationNV, but what about general case (ATI, others)?

I'll be grateful if anybody points me to the working example of GL_EXT_transform_feedback. Thanks in advance.

Brolingstanz
03-14-2009, 06:19 AM
Make sure you have the latest headers (http://www.opengl.org/registry/) (updated recently).

Haven't used this extension yet myself but it seems that TF is part of the core in 3.0 and available right out of the box on a NV G80 or better with the latest drivers (182.08 by my reckoning).

DmitryM
03-15-2009, 06:18 AM
Of course I mean latest headers. I check almost every day.



Haven't used this extension yet myself but it seems that TF is part of the core in 3.0 and available right out of the box on a NV G80 or better with the latest drivers (182.08 by my reckoning).

The NV support is obvious as there is a complete GL_NV_transform_feedback extension. I'm actually talking about GL_EXT_* version.

DmitryM
03-30-2009, 01:03 AM
With new OpenGL 3.1 specification the glext.h header is correct and matches the spec. However, I still can't get transform feedback work. The code:

prog_id = glCreateProgramObject();
glAttachObject(prog_id, prog_vert);
glAttachObject(prog_id, prog_frag);

const char *vars[] = { "gl_Position" };
const unsigned ids[] = { buf_dst_id };

glTransformFeedbackVaryings(prog_id, 1, vars, GL_SEPARATE_ATTRIBS_EXT);
glBindBufferBase(GL_TRANSFORM_FEEDBACK_BUFFER_EXT, 0, ids[0]);

//!!! here get an error:
glLinkProgram(prog_id);
//!!! access violation

glUseProgramObject(prog_id)
glEnable( GL_RASTERIZER_DISCARD_EXT );
glBeginTransformFeedback(GL_POINTS);

glEnableClientState(GL_VERTEX_ARRAY);
glBindBuffer(GL_ARRAY_BUFFER, buf_src_id);
glVertexPointer(3, GL_FLOAT, 0, NULL);

glDrawArrays(GL_POINTS, 0, num_vertexes);
glEndTransformFeedback();

Jacques LIAO
03-30-2009, 03:29 AM
have you check if this EXTENSION is enabled in your platform and current driver? you can use GLEW or GLEE library to easily your process.

DmitryM
03-30-2009, 03:32 AM
Yes, the extension 'GL_EXT_transform_feedback' is supported.
I have a Radeon 2600HD video on board controlled by ATI Catalyst 9.3 driver.

S.Seegel
04-17-2009, 04:20 PM
I am quite certain transform feedback is not working in current ati drivers (I'm running some kind of catalyst 9.4 pre-release bundled with ubuntu 9.04RC on linux x86_64).

Whenever you call glTransformFeedbackVaryings with invalid varying names (e.g. a name not defined as an out variable in your vertex shader), the driver will segfault in glLinkProgram (which is the point when these settings are supposed to take effect).
In addition, build-in variables (such as gl_Position) are not accepted for feedback.

Obviously this bug can be worked around by using an user defined out variable for feedback.

But I still couldn't get transform feedback to work up to now, because I'm unable to bind a feedback buffer:
- glBindBufferBase results in an GL_INVALID_OPERATION
- glBindBufferOffset and glBindBufferRange cause a segfault

Has anybody had success in setting up transform feedback with ATIs opengl 3.0 drivers ?

Does anybody know which is the prefered way to report bugs to ati (I had no luck searching their website) ?

DmitryM
06-29-2009, 06:58 PM
Thank you, Seegel, for informative post.
Sorry for being so late with an answer - I just didn't have much to say until now.

I gave up trying to get it work in C++ with GL 2.1
Now I'm using Boo(.net) with OpenTK creating pure GL 3.0 context window.
And I faced the same problem you did...
http://www.opentk.com/node/980

If you ever get an idea how to solve it - let me know, please. I can't live without skinning performed on GPU using TF :)

yooyo
06-30-2009, 12:47 AM
I can't live without skinning performed on GPU

Use textures, FBO (with MRT), PBO and VBO.

Store all your vertices in rgb32f texture. Store additional attributes in other textures (indices, weights, normals, binormals, tangents).

Store bone matrices in rgba32f texture. Make sure indices in texture points to correct bone texel.

Then render aligned quad with shader. Result should be stored in FBO with MRT. After that, readback result to PBO buffer, rebind PBO as VBO, setup pointers and render.

You can store multiple characters in one texture and perform skinning on all of them in single rendering pass.

DmitryM
06-30-2009, 04:39 AM
Thanks, yooyo

I'll consider using textures for skinning.
However, TF approach would be more natural (no need of PBO, MRT, all attribs are just VBO) and fast, so a working example is still needed...

Edit: current TF status:
upgraded to Catalyst 9.6
BindBufferBase still produces InvalidOperation error

Edit2:
BindBufferRange produces segfault as reported by S.Seegel
Seems like an old driver issue still presents in Catalyst

DmitryM
08-20-2009, 09:03 AM
I've made a small test example. The results are:
For ATI Catalyst 9.5: InvalidOperation on BindBufferBase call
For Nvidia 180.xx WHQL: (will be posted a little bit later)

If there is someone from ATI/Nvidia here, please, tell me if I'm doing something wrong or it's just your drivers bugs.



#include "glew.h"
#include "glfw.h"
#include <stdio.h>
#include <assert.h>

#pragma comment(lib,"GLFW.lib")
#pragma comment(lib,"glew32.lib")
#pragma comment(lib,"opengl32.lib")

static const char* const sourceVert = "varying vec4 out_pos; \
void main() { gl_Position = out_pos = gl_Vertex + vec4(0.5); } \
";
static const char* const sourceFrag = "varying vec4 out_pos; \
void main() {}";

static void checkShader(const GLuint id, const GLenum type, const char str[]) {
static char msg[240];
int size,aux;
glGetObjectParameterivARB(id, type, &amp;size);
if(!size) return;
glGetObjectParameterivARB(id, GL_OBJECT_INFO_LOG_LENGTH_ARB, &amp;size);
assert(size < sizeof(msg));
glGetInfoLogARB(id, size, &amp;aux, msg);
msg[size] = 0;
printf("%s: %s", str, msg);
}

int main() {
printf("Testing EXT_TRANSFORM_FEEDBACK\n");
glfwInit();
int ok = glfwOpenWindow(400, 300, 8,8,8, 0,0,0, GLFW_WINDOW);
if(!ok) return -1;
GLenum err = glewInit();
if(err != GLEW_OK) return -2;
printf("Context initialized\n");
if(!GLEW_EXT_transform_feedback || !GLEW_ARB_shader_objects || !GLEW_ARB_vertex_buffer_object)
return -3;
printf("Extensions supported\n");
GLuint shid, bid, tmp;

shid = glCreateProgramObjectARB();
tmp = glCreateShaderObjectARB(GL_VERTEX_SHADER_ARB);
glShaderSourceARB(tmp, 1, (const GLcharARB**)&amp;sourceVert, NULL);
glCompileShaderARB(tmp);
checkShader(tmp, GL_OBJECT_COMPILE_STATUS_ARB, "Compile vertex");
glAttachObjectARB(shid,tmp);

tmp = glCreateShaderObjectARB(GL_FRAGMENT_SHADER_ARB);
glShaderSourceARB(tmp, 1, (const GLcharARB**)&amp;sourceFrag, NULL);
glCompileShaderARB(tmp);
checkShader(tmp, GL_OBJECT_COMPILE_STATUS_ARB, "Compile fragment");
glAttachObjectARB(shid,tmp);

const char *vars[] = {"out_pos"};
glTransformFeedbackVaryingsEXT(shid, 1, vars, GL_SEPARATE_ATTRIBS_EXT);
glLinkProgramARB(shid);
checkShader(shid, GL_OBJECT_LINK_STATUS_ARB, "Link result");
printf("Shader loaded: %d\n", glGetError());

glGenBuffersARB(1, &amp;bid);
glBindBufferARB(GL_ARRAY_BUFFER, bid);
const float init[] = {0.1f,0.2f,0.3f,0.4f};
glBufferDataARB(GL_ARRAY_BUFFER, 4*sizeof(float)*1, init, GL_STREAM_DRAW);
printf("Buffer prepared: %d\n", glGetError());

glBindBufferBaseEXT(GL_TRANSFORM_FEEDBACK_BUFFER_E XT, 1, bid);
printf("Feedback assigned: %d\n", glGetError());

glUseProgramObjectARB(shid);
glBeginTransformFeedbackEXT(GL_POINTS);
glEnable(GL_RASTERIZER_DISCARD_EXT);
glBegin(GL_POINTS);
glVertex4fv(init);
glEnd();
glDisable(GL_RASTERIZER_DISCARD_EXT);
glEndTransformFeedbackEXT();
printf("Drawing completed: %d\n", glGetError());

float *const data = (float *)glMapBuffer(GL_ARRAY_BUFFER, GL_READ_ONLY);
for(int i=0; data &amp;&amp; i!=4; ++i)
printf("%.1f ", data[i]);
glUnmapBuffer(GL_ARRAY_BUFFER);
printf("\nData readed: %d\n", glGetError());
glfwTerminate();
getchar();
return 0;
}

GHotep
08-20-2009, 09:41 AM
Try changing this




glBindBufferBaseEXT(GL_TRANSFORM_FEEDBACK_BUFFER_E XT, 1, bid);



to




glBindBufferBaseEXT(GL_TRANSFORM_FEEDBACK_BUFFER_E XT, 0, bid);



since you're only binding one varying for transform feedback. The index parameter starts at 0.

DmitryM
08-20-2009, 10:53 AM
Thanks for reading the code, GHotep.
It was a simple mistake. Fixing it doesn't help. The same error 1282 (Invalid Operation).

GHotep
08-20-2009, 08:15 PM
Hmm. I don't see anything else obviously wrong - one guess, try adding a #version 120 or #version 130 as the first line of the shader. Another possibility, try binding gl_Position instead of out_pos, in case the latter is being eliminated during optimization.

DmitryM
08-21-2009, 06:59 AM
1) Tried to specify the version. 120 works in the same way as there's no version. 130 complains that version can't be specified in GL2 context...

2) Binding gl_Position results in the access violation crush on glLinkProgram. Looks like a driver bug for me...

Any ideas what to do next?

GHotep
08-21-2009, 02:23 PM
I'm just guessing at this point - I know I've used transform feedback recently on Nvidia hardware, but I don't have access to ATI cards. Maybe add



out_Color = out_pos;


to the fragment shader, in case the varying is getting optimized away. I agree that binding gl_Position shouldn't be causing a crash, since that's what I've done in the recent past.

DmitryM
08-21-2009, 02:29 PM
The last tests were performed with the following fragment shader:


#version 120
in vec4 out_pos;
void main() {
gl_FragColor = out_pos;
}

The result is the same... AFAIK, driver doesn't optimize across shader objects.

Did you try EXT_Transform_Feedback on NVidia or you just used NV-version of the extension? I'm not aware of *any* example of EXT version.

GHotep
08-23-2009, 08:21 AM
Did you try EXT_Transform_Feedback on NVidia or you just used NV-version of the extension? I'm not aware of *any* example of EXT version.



I'm writing currently in OpenGL 3.1, which promoted EXT_transform_feedback to the core with not too many changes. I'm at a bit of a loss with providing additional help - it may well be driver bugs, or there is something subtle that I am overlooking.

DmitryM
08-23-2009, 12:38 PM
So my question is: are you using Transform Feedback with GL 3.1? If you are and you don't see any mistakes in my code, could you please some working code of yours?

BionicBytes
08-24-2009, 05:31 AM
I can't say I have tried transform feedback - but have you enabled the extensions in your shaders ?

for example

#extension GL_EXT_texture_array : enable

DmitryM
08-24-2009, 05:47 AM
The shader used for TF doesn't need to know about the existence of TF. It just writes to the output varyings.

DmitryM
09-02-2009, 08:04 AM
Damn, I'm trying the way 'yooyo' suggested...

This ATI driver drives me crazy (9.5 at least). The error InvalidOperation is generated on BindBuffer(PixelUnpackBuffer,0) call...

Common, PBO was introduced in 2007. Is it so difficult to get it work after 2 years on ATI?..

BTW, I've filed a support request on the AMD developer central about TF issue. Still no answer. AMD seems not care much about developers, doesn't it?..

Heiko
09-02-2009, 08:33 AM
Have you tried whether the bug is still there in the 9.8 driver (that is the newest one)? There is not really a point in sending a bug report for an old driver I guess, unless you know the bug is still there in the newest driver.

I sent a bug report a while ago, together with sample code and system details. I got a great response that said they couldn't reproduce the bug with the newest drivers and asked whether I could verify if the bug was gone with the newest drivers. It was gone indeed, so I was satisfied.

DmitryM
09-02-2009, 08:45 AM
Lucky you, Heiko!

With a bug file about TransformFeedback I attached the simple C source (using GLew & GLFW) that reproduces it. I specified that it crashes on both 9.5 and 9.8. If there was something they need to know in addition, why not ask me?..

BTW, did you use AMD dev central support system? How have you posted a bug so that they noticed it? :)

Heiko
09-02-2009, 09:04 AM
I emailed everything to devrel[at]amd.com. System details were gpu, cpu, motherboard, amount of memory, OS and the driver I used. The code was a minimal example, well documented, not using any external libraries.

DmitryM
09-02-2009, 03:11 PM
Thanks, Heiko.
I've mailed them just now as well. Provided detailed SysInfo output. Hope to get *some* answer...

DmitryM
09-23-2009, 08:04 AM
Finally!

After some pure discussion with different AMD departments, I've got an advice to try 9.9 and check if the bug is still there (yes, they admitted it to be a bug).

Surprisingly, TF works in 9.9 under my GL-3.0 forward-compatible context. Congratulations for everyone who were waiting for this event!