HALF_FLOAT in Compute shaders

[Evgeny Demidov]

I've seen https://esdiscuss.org/topic/float16array discussion.

May be you can help me before I sink in searching and experiments.

1. In https://www.ibiblio.org/e-notes/webgl/gpu/CSpetal.htm vertexes are generated by Compute Shader. All works with FLOAT data but I get broken picture with HALF_FLOAT data

   gl.bufferData(gl.SHADER_STORAGE_BUFFER, new Uint16Array(4*nr*nf), gl.STATIC_DRAW);

...

   gl.vertexAttribPointer(aPosLoc, 4, gl.HALF_FLOAT, gl.FALSE, 0, 0)

what shall I use instead Uint16Array or Float32Array?

2. can I convert f32 (Float32Array) into f16 buffer by simple Compute Shader for Matrix multiplications experiments?

Evgeny

[Ken Russell]

On Tue, Feb 19, 2019 at 7:48 AM Evgeny Demidov <demidov...@gmail.com> wrote:

I've seen https://esdiscuss.org/topic/float16array discussion.

May be you can help me before I sink in searching and experiments.

1. In https://www.ibiblio.org/e-notes/webgl/gpu/CSpetal.htm vertexes are generated by Compute Shader. All works with FLOAT data but I get broken picture with HALF_FLOAT data

   gl.bufferData(gl.SHADER_STORAGE_BUFFER, new Uint16Array(4*nr*nf), gl.STATIC_DRAW);

...

   gl.vertexAttribPointer(aPosLoc, 4, gl.HALF_FLOAT, gl.FALSE, 0, 0)

what shall I use instead Uint16Array or Float32Array?

Uint16Array is correct. The only thing that does is change how the CPU-side code interprets the bits. If you are uploading half-float data from the CPU to the GPU then you'll need to pack it yourself using a library like https://github.com/petamoriken/float16 .

 

2. can I convert f32 (Float32Array) into f16 buffer by simple Compute Shader for Matrix multiplications experiments?

If you are uploading data from the CPU then you will need to convert it per above. If you are just writing half-float values from your compute shader into the buffer, the GPU does the conversion so you don't need to do anything. (If you then want to read the half-float values in JavaScript you'll need to convert them on the way back using the same float16 library.)

-Ken

 

Evgeny

--
You received this message because you are subscribed to the Google Groups "WebGL Dev List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to webgl-dev-lis...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Evgeny Demidov]

On Wednesday, February 20, 2019 at 12:08:28 AM UTC+3, Kenneth Russell wrote:

On Tue, Feb 19, 2019 at 7:48 AM Evgeny Demidov <demidov...@gmail.com> wrote:

I've seen https://esdiscuss.org/topic/float16array discussion.

May be you can help me before I sink in searching and experiments.

1. In https://www.ibiblio.org/e-notes/webgl/gpu/CSpetal.htm vertexes are generated by Compute Shader. All works with FLOAT data but I get broken picture with HALF_FLOAT data

   gl.bufferData(gl.SHADER_STORAGE_BUFFER, new Uint16Array(4*nr*nf), gl.STATIC_DRAW);

...

   gl.vertexAttribPointer(aPosLoc, 4, gl.HALF_FLOAT, gl.FALSE, 0, 0)

what shall I use instead Uint16Array or Float32Array?

Uint16Array is correct. The only thing that does is change how the CPU-side code interprets the bits. If you are uploading half-float data from the CPU to the GPU then you'll need to pack it yourself using a library like https://github.com/petamoriken/float16 .

no all data are generated by compute shader (CS)!

But CS writes wrongly vec3 data (only float, vec2 and vec4 work correctly). I got similar broken meshes with vec3 vertex arrays. From the neighboring topic

"for this simple shader

const CSs = `#version 310 es

  layout (local_size_x = 2, local_size_y = 2, local_size_z = 1) in;

  layout (std430, binding = 0) buffer SSBO {

    vec3 data[];

  } ssbo;

void main() {

  ssbo.data[gl_LocalInvocationIndex] = vec3(gl_LocalInvocationID);

}

`;

https://www.ibiblio.org/e-notes/webgl/gpu/CDbuffer2x2.htm

I get strange result - output: [0,0,0, 0,1,0, 0,0,0, 1,0,0] (I inserted spaces by hand)" me and

"I suppose this is because of memory layout. An array of vec3 is rounded up to 4N even though std430.

If you allocate and get result with new Float32Array(4 * 4) by bufferData()/getBufferSubData(), you will get correct result - output: [0,0,0,(0), 1,0,0,(0), 0,1,0,(0), 1,1,0,(0)], (0) means auto inserted padding. The result you showed is first 12 components of this." Kentaro Kawakatsu.

May be "this is because of memory layout" again but complicated by f16.

and I think we needn't any library now. Let simple CS convert f32 <-> f16 data (at least for experiments) meanwhile we get Float16Array.

I'll try to make simple tests for writing into f16 buffers (Uint16Array 8-)

Evgeny

[Evgeny Demidov]

Malevich's CS 16f "black square" (is stolen) - f16 vertex array is written by CS

//   gl.bufferData(gl.SHADER_STORAGE_BUFFER, new Float32Array(8), gl.STATIC_DRAW);

     gl.bufferData(gl.SHADER_STORAGE_BUFFER, new Uint16Array(8), gl.STATIC_DRAW);

//   gl.vertexAttribPointer(aPosLoc, 2, gl.FLOAT, gl.FALSE, 0, 0)

     gl.vertexAttribPointer(aPosLoc, 2, gl.HALF_FLOAT, gl.FALSE, 0, 0)

https://www.ibiblio.org/e-notes/webgl/gpu/CShalf.htm

 f32 vertex array is written by CS - works fine

https://www.ibiblio.org/e-notes/webgl/gpu/CShalf32.htm

no CS, just hand made Uint16Array - works fine

https://www.ibiblio.org/e-notes/webgl/gpu/CShalf16.htm

how long does it takes to compile ANGLE (Chromium, Firefox) on Core2Duo?  :)

Evgeny

[Evgeny Demidov]

https://chromium.googlesource.com/angle/angle/+/c5af8ba6ebd2246b4c214f30bd50ba80fa8b1e3e/src/compiler/translator/blocklayout.cpp

72 void Std140BlockEncoder::getBlockLayoutInfo(GLenum type, ...

78 // We assume we are only dealing with 4 byte components (no doubles or half-words currently)

no f16 am I correct?

Evgeny

[Ken Russell]

Thanks for tracking that down Evgeny. It looks like a bug.

Filed https://bugs.chromium.org/p/angleproject/issues/detail?id=3160 to track it.

--

[Ken Russell]

Hi Evgeny,

Actually some clarification is needed on https://bugs.chromium.org/p/angleproject/issues/detail?id=3160 . Could you please follow up on it? Thanks.

-Ken