Confirm R16G16_S10_5 support

[lexaknyazev]

The VK_NV_optical_flow extension adds a new VK_FORMAT_R16G16_S10_5_NV format:

VK_FORMAT_R16G16_S10_5_NV specifies a two-component, fixed-point format where most significant bit specifies the sign bit, next 10 bits specify the integer value and last 5 bits represent the fractional value.

It should be checked whether it's representable with DFD.

[MarkCallow]

@fluppeteer please educate us whether this format is representable with DFD.

[fluppeteer]

Fixed-point sign-magnitude? Interesting. When I've got updates in place I'll try to make that clearer, since sign-magnitude is only really discussed under the floating point formats section.

I'd write it as this:

vendorId = 0
descriptorType = 0
versionNumber = 2
descriptorBlockSize = 88
colorModel: KHR_DF_MODEL_RGBSDA
colorPrimaries: KHR_DF_PRIMARIES_BT709 (or KHR_DF_PRIMARIES_UNSPECIFIED)
transfer = KHR_DF_TRANSFER_LINEAR (presumably)
flags = KHR_DF_FLAG_ALPHA_STRAIGHT
texelBlockDimension0: 0
texelBlockDimension1: 0
texelBlockDimension2: 0
texelBlockDimension3: 0
bytesPlane0: 2
bytesPlane1: 0
bytesPlane2: 0
bytesPlane3: 0
bytesPlane4: 0
bytesPlane5: 0
bytesPlane6: 0
bytesPlane7: 0
First sample: G mantissa
bitOffset: 0
bitLength: 14 (= "15")
channelType: GREEN
F = S = E = L = 0
samplePosition 0: 0
samplePosition 1: 0
samplePosition 2: 0
samplePosition 3: 0
sampleLower: 0
sampleUpper: 1 << 5 ("1.0" is represented as 32)
Second sample: G sign bit
bitOffset: 15
bitLength: 0 (= "1")
channelType: GREEN
F = E = L = 0
S = 1
samplePosition 0: 0
samplePosition 1: 0
samplePosition 2: 0
samplePosition 3: 0
sampleLower: 1 - 1-bit "-1"
sampleUpper: 0 - 1-bit "0"
Third sample: R mantissa
bitOffset: 16
bitLength: 14 (= "15")
channelType: RED
F = S = E = L = 0
samplePosition 0: 0
samplePosition 1: 0
samplePosition 2: 0
samplePosition 3: 0
sampleLower: 0
sampleUpper: 1 << 5 ("1.0" is represented as 32)
Second sample: R sign bit
bitOffset: 31
bitLength: 0 (= "1")
channelType: RED
F = E = L = 0
S = 1
samplePosition 0: 0
samplePosition 1: 0
samplePosition 2: 0
samplePosition 3: 0
sampleLower: 1 - 1-bit "-1"
sampleUpper: 0 - 1-bit "0"

I think that's consistent with, e.g., table 103 in the KDFS. I'm rusty, so please check it looks sensible, don't just trust me. :-)

In the above I'm assuming sign-magnitude, so "-1.0" is 0x8020". If it's fully-signed and "-1.0" is 0xFFE0, it's simpler.

I'm not a huge fan of the VK_FORMAT_R16G16_S10_5_NV naming (especially since S is used for stencil in depth-stencil format), but it's going to have to be new no matter what is done, and I don't have a firm suggestion. There are already the INT formats that are effectively the same as NORM but with a different definition of "1.0", and this feels like a variant.

[MarkCallow]

Thanks Andrew. Looks okay to me except that bytesPlane0 should be 4. I'm not fond of the naming either.

[fluppeteer]

Thanks Andrew. Looks okay to me except that bytesPlane0 should be 4. I'm not fond of the naming either.

D'oh. Sorry, I realised I was thinking of it as a single-channel format and then corrected myself and updated the number of samples, but forgot to fix bytesPlane0.

[MarkCallow]

@lexaknyazev how can we confirm if this is sign-magnitude or fully signed?

This format is already in the valid list in libktx's vkformat_check.c but no DFD is generated for it. If we want to support it, I'll open a bug in KTX-Software to fix the DFD generation once we know the answer about sign-magnitude. If we don't want to support it, we need to list it in the prohibited formats section.

[lexaknyazev]

Asked internally.

[MarkCallow]

Asked internally.

Thanks. While you are at it ask about the transfer function.

[lexaknyazev]

Got an answer:

• It's two's complement (like GL_FIXED)
• You convert the int16 values to float by dividing them by 32
• Range is [-1024 ... 1023.96875]
• The values are linear

[fluppeteer]

Thanks. In that case it's simpler, and each channel is just one 16-bit signed integer sample with a sample_upper of 32 and sample_lower of -32, I think.

[MarkCallow]

Thanks for the answers @lexaknyazev.

I see no reason not to support this in KTX. The DFD is straightforward. If we do, no spec. change is needed, right?

I don't think we'll want to support it in ktx create as none of the currently supported input formats seems appropriate and I don't think the image processing we have in KTX-Software supports the kind of bid-twiddling need to convert, e.g, a floating point input to this fixed point format. It could of course be added but is it worth it, given this is a vendor extension?

We'll need to update vk2dfd, dfd2vk, validate and info at a minimum.

[lexaknyazev]

no spec. change is needed, right?

Right.

Once DFD tools, ktx info, and ktx validate support this format, ktx create and ktx extract should work as well with the --raw mode.

[MarkCallow]

I've opened KhronosGroup/KTX-Software#859 to complete support in KTX-Software. Closing this.

[MarkCallow]

@fluppeteer I'm working on adding support in KTX-Software. I have vk2dfd working and I'm now making dfd2vk work. This is based around using interpretDFD. What is your recommendation for handling this in that function? I am thinking of adding an i_FIXED_POINT_FORMAT_BIT to InterpretDFDResult which I would set if the sampleUpper = -sampleLower and the value is somewhere in the middle of the range representable by the number of bits in the channel. Does that sound reasonable?

[MarkCallow]

I am thinking of adding an i_FIXED_POINT_FORMAT_BIT to InterpretDFDResult which I would set if the sampleUpper = -sampleLower and the value is somewhere in the middle of the range representable by the number of bits in the channel. Does that sound reasonable?

ping @fluppeteer.

[fluppeteer]

Sorry Mark. Yes, I think that works for fixed-point (basically anything other than "maximum number representable-ish = NORM, 1 = INT/SCALED"). If it's really a fixed-point format and doesn't have a separate signed bit, though, I ought to push back strongly on the Vulkan format naming, because it's hopelessly misleading; I'll try to raise an internal Khronos issue.

[lexaknyazev]

The format has been renamed to VK_FORMAT_R16G16_SFIXED5_NV in the upstream specs to better reflect its layout.

[MarkCallow]

It is KTX-Software that needs updating. Nothing to do here. I should have closed this when I merged KTX-Software PR #864.

There is no mapping for this to formats in other APIs. @lexaknyazev are there any possible mappings? If so, please add then to formats.json. If not, please close this. Feel free to open an issue in KTX-Software to remind me to change the name there.

[MarkCallow]

The format has been renamed to VK_FORMAT_R16G16_SFIXED5_NV in the upstream specs to better reflect its layout.

I went to start the update and found that this name is not in the vk.xml in GitLab's vulkan/vulkan/xml. It is in the vk.xml in GitHub's, KhronosGroup/Vulkan-Docs/xml. What gives @lexaknyazev? Has it been withdrawn?

[lexaknyazev]

this name is not in the vk.xml in GitLab's vulkan/vulkan/xml

It's there, in the main branch.

[MarkCallow]

It's there, in the main branch.

Oh! Silly me. I didn't realize my workarea didn't have main checked out.

[MarkCallow]

The name has been updated in KTX-Software PR #921.