Add check for float texture linear filtering support

[adamscott]

This PR makes it so that, on the web platform, the curve texture is rendered using an int- half-float-based image format instead of a float based one.

Thanks to @clayjohn #102089 (comment), this PR now adds a check for float texture linear filtering support.

Fixes #102088

[Calinou]

Did you try RGBH/RH instead? Using RGB8/R8 means we lose all HDR data and have much lower precision, which can be problematic for some use cases.

[adamscott]

Did you try RGBH/RH instead? Using RGB8/R8 means we lose all HDR data and have much lower precision, which can be problematic for some use cases.

@Calinou Fixed! And it works! And I greatly simplified the code, didn't know that Image::convert() existed.

[clayjohn]

Is this an iOS specific problem? AFAIK RGB32F textures should be supported by all WebGL implementations

edit: Just double checked, RGB32F is supported by WebGL 2.0 https://registry.khronos.org/webgl/specs/latest/2.0/. Therefore I suspect this is more a problem of safari on iOS not being compliant with WebGL 2

OOOOhhhh, According to https://developer.mozilla.org/en-US/docs/Web/API/WebGLRenderingContext/texImage2D the floating point formats are supported, but may not be texture filterable! Could you test on iOS with the linear filtering disabled? Edit: and that's confirmed by the OpenGL ES 3.0 spec https://registry.khronos.org/OpenGL/specs/es/3.0/es_spec_3.0.pdf

Okay, so, ideally what we would do is check if an implementation supports using the linear filter with a given texture, then convert the texture automatically. This is how we handle compressed textures:

godot/drivers/gles3/storage/texture_storage.cpp

Lines 462 to 471 in 6dc78c8

	case Image::FORMAT_DXT1: {
	if (config->s3tc_supported) {
	r_gl_internal_format = _EXT_COMPRESSED_RGBA_S3TC_DXT1_EXT;
	r_gl_format = GL_RGBA;
	r_gl_type = GL_UNSIGNED_BYTE;
	r_compressed = true;
	} else {
	need_decompress = true;
	}
	} break;

And here is what we did in 3.6 for GLES2 where float textures are only supported with an extension:

godot/drivers/gles2/rasterizer_storage_gles2.cpp

Lines 258 to 271 in d47e265

	if (!config.float_texture_supported) {
	ERR_PRINT("RGBA float texture not supported, converting to RGBA8.");
	if (image.is_valid()) {
	image->convert(Image::FORMAT_RGBA8);
	}
	r_real_format = Image::FORMAT_RGBA8;
	r_gl_internal_format = GL_RGBA;
	r_gl_format = GL_RGBA;
	r_gl_type = GL_UNSIGNED_BYTE;
	} else {
	r_gl_internal_format = GL_RGBA;
	r_gl_format = GL_RGBA;
	r_gl_type = GL_FLOAT;
	}

[clayjohn]

I found the missing piece!

We need to check if the OES_texture_float_linear extension is supported.

It is basically supported everywhere except iOS: https://caniuse.com/?search=OES_texture_float_linear

edit: Aaaaaaannd looks like we handled this correctly in 3.x

godot/drivers/gles3/rasterizer_storage_gles3.cpp

Line 8190 in d47e265

config.texture_float_linear_supported = config.extensions.has("GL_OES_texture_float_linear");

So, to summarize. We need to add config.texture_float_linear_supported = config.extensions.has("GL_OES_texture_float_linear"); to config.cpp. Then we need to check it when creating the texture in texture_storage.cpp. If unsupported we can just fall back to RGB16H, i.e:

This:

godot/drivers/gles3/storage/texture_storage.cpp

Lines 401 to 405 in 6dc78c8

	case Image::FORMAT_RGBF: {
	r_gl_internal_format = GL_RGB32F;
	r_gl_format = GL_RGB;
	r_gl_type = GL_FLOAT;
	} break;

Becomes:

 case Image::FORMAT_RGBF: { 
 	if (config.texture_float_linear_supported) {
 	 	r_gl_internal_format = GL_RGB32F; 
 	 	r_gl_format = GL_RGB; 
 	 	r_gl_type = GL_FLOAT; 
 	} else {
 	 	ERR_PRINT("RGB32 float texture not supported, converting to RGB16."); 
 	 	if (image.is_valid()) { 
 	 		image->convert(Image::FORMAT_RGBH); 
 	 	} 
 	 	r_real_format = Image::FORMAT_RGBH; 
 	 	r_gl_internal_format = GL_RGB16F; 
 	 	r_gl_format = GL_RGB; 
 	 	r_gl_type = GL_HALF_FLOAT;
 	}
 } break; 

[adamscott]

@clayjohn It works!

[clayjohn]

Looks great! This might even fix a few other iOS issues

[Calinou]

ERR_PRINT() might be a bit spammy in projects using lots of CurveTextures, so maybe we should use WARN_PRINT_ONCE() instead. The condition can't change at runtime anyway – if float textures aren't supported when the project starts, they won't be supported afterwards in the same session.

[Repiteo]

We can change this in a followup PR if it ends up being too intrusive

[Repiteo]

Thanks!