[SPIRV][Doc] Add arbitrary precision floating and fixed point specs #1934

ajaykumarkannan · 2020-06-19T15:57:56Z

These operations are to be leveraged by the DPC++ compiler for the Intel FPGA devices.

AlexeySachkov

Some partial review: haven't fully read SPV_INTEL_arbitrary_precision_floating_point spec yet

sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_fixed_point.asciidoc

sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_floating_point.asciidoc

jbrodman · 2020-06-26T16:40:17Z

Will approve after @AlexeySachkov does.

sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_fixed_point.asciidoc

sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_floating_point.asciidoc

This patch introduces SPV_INTEL_arbitrary_precision_fixed_point extension, which adds operations for arbitrary precision fixed point numbers. The ac_fixed datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - total width of the datatype. It's encoded in the width of the OpTypeInt. * I - determines the position of the decimal point. * S - determines if this is a signed or an unsigned number. Full specification (intel/llvm#1934): https://github.com/intel/llvm/blob/2f6e965e686354fbb25f9c177a667a646de302eb/sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_fixed_point.asciidoc

Add SPV_INTEL_arbitrary_precision_fixed_point extension, which introduces operations for arbitrary precision floating point numbers. The hls_float datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - width of the datatype. It's encoded in the width of the OpTypeInt. * E - represents the number of exponent bits. Exponent values are inferred from the width of corresponding variable. * M - represents the number of mantissa bits. The total width W is equal to `E+M+1` where the last bit is used to represent the sign. Full specification (intel/llvm#1934): https://github.com/intel/llvm/blob/bd86b218f749ea0e20ddc18c42db491faf54014a/sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_floating_point.asciidoc

Add SPV_INTEL_arbitrary_precision_fixed_point extension, which introduces operations for arbitrary precision floating point numbers. The hls_float datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - width of the datatype. It's encoded in the width of the OpTypeInt. * E - represents the number of exponent bits. Exponent values are inferred from the width of corresponding variable. * M - represents the number of mantissa bits. The total width W is equal to `E+M+1` where the last bit is used to represent the sign. Full specification (intel/llvm#1934): https://github.com/intel/llvm/blob/bd86b218f749ea0e20ddc18c42db491faf54014a/sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_floating_point.asciidoc Signed-off-by: Mikhail Lychkov <mikhail.lychkov@intel.com>

This patch introduces SPV_INTEL_arbitrary_precision_fixed_point extension, which adds operations for arbitrary precision fixed point numbers. The ac_fixed datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - total width of the datatype. It's encoded in the width of the OpTypeInt. * I - determines the position of the decimal point. * S - determines if this is a signed or an unsigned number. Full specification (intel/llvm#1934): https://github.com/intel/llvm/blob/2f6e965e686354fbb25f9c177a667a646de302eb/sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_fixed_point.asciidoc

Add SPV_INTEL_arbitrary_precision_fixed_point extension, which introduces operations for arbitrary precision floating point numbers. The hls_float datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - width of the datatype. It's encoded in the width of the OpTypeInt. * E - represents the number of exponent bits. Exponent values are inferred from the width of corresponding variable. * M - represents the number of mantissa bits. The total width W is equal to `E+M+1` where the last bit is used to represent the sign. Full specification (intel/llvm#1934): https://github.com/intel/llvm/blob/bd86b218f749ea0e20ddc18c42db491faf54014a/sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_floating_point.asciidoc Signed-off-by: Mikhail Lychkov <mikhail.lychkov@intel.com>

This patch introduces SPV_INTEL_arbitrary_precision_fixed_point extension, which adds operations for arbitrary precision fixed point numbers. The ac_fixed datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - total width of the datatype. It's encoded in the width of the OpTypeInt. * I - determines the position of the decimal point. * S - determines if this is a signed or an unsigned number. Full specification (intel/llvm#1934): https://github.com/intel/llvm/blob/2f6e965e686354fbb25f9c177a667a646de302eb/sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_fixed_point.asciidoc

Add SPV_INTEL_arbitrary_precision_fixed_point extension, which introduces operations for arbitrary precision floating point numbers. The hls_float datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - width of the datatype. It's encoded in the width of the OpTypeInt. * E - represents the number of exponent bits. Exponent values are inferred from the width of corresponding variable. * M - represents the number of mantissa bits. The total width W is equal to `E+M+1` where the last bit is used to represent the sign. Full specification (intel/llvm#1934): https://github.com/intel/llvm/blob/bd86b218f749ea0e20ddc18c42db491faf54014a/sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_floating_point.asciidoc Signed-off-by: Mikhail Lychkov <mikhail.lychkov@intel.com>

This patch introduces SPV_INTEL_arbitrary_precision_fixed_point extension, which adds operations for arbitrary precision fixed point numbers. The ac_fixed datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - total width of the datatype. It's encoded in the width of the OpTypeInt. * I - determines the position of the decimal point. * S - determines if this is a signed or an unsigned number. Full specification (intel/llvm#1934): https://github.com/intel/llvm/blob/2f6e965e686354fbb25f9c177a667a646de302eb/sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_fixed_point.asciidoc

Add SPV_INTEL_arbitrary_precision_fixed_point extension, which introduces operations for arbitrary precision floating point numbers. The hls_float datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - width of the datatype. It's encoded in the width of the OpTypeInt. * E - represents the number of exponent bits. Exponent values are inferred from the width of corresponding variable. * M - represents the number of mantissa bits. The total width W is equal to `E+M+1` where the last bit is used to represent the sign. Full specification (intel#1934): https://github.com/intel/llvm/blob/bd86b218f749ea0e20ddc18c42db491faf54014a/sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_floating_point.asciidoc Signed-off-by: Mikhail Lychkov <mikhail.lychkov@intel.com>

sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_fixed_point.asciidoc

AlexeySachkov · 2020-08-05T16:14:07Z

sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_floating_point.asciidoc

+
+_Result Type_ is an `OpTypeInt` of width `1+Eout+Mout` and is the type of _Result_.
+
+_M*_ contains the width of the mantissa of the floating point types within _Result_, _A_, and _B_.


Personally, I would prefer to have clear mapping between literal and argument which it corresponds to, but I guess it could be inferred with current wording as well: Mout for result, M1 for the first argument and M2 for the second

Probably would be good to hear more feedback from someone else

@AlexeySachkov any suggestions on what you would rename these to?

sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_floating_point.asciidoc

AlexeySachkov · 2020-08-05T16:25:31Z

sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_floating_point.asciidoc

+
+=== Validation Rules
+
+None.


Are you sure that there are no validation rules?

I think I can imagine at least a few:

M* literal arguments can't contains values that exceed width of OpTypeInt of corresponding argument minus 1

M* literal should contain value more than zero

It seems that there are no requirement to have arguments of the same width, but probably there are still some limitations related to this?

Note: the same applies to fixed-point extension

Good point - added for floating point. Based on the discussion about what values I can take, I don't know if this applies to the fixed_point case. What do you think?

AlexeySachkov · 2020-08-05T16:28:09Z

sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_floating_point.asciidoc

+
+After Section 3.16, add a new section "3.16d Rounding Accuracy" as follows:
+
+=== Rounding Accuracy


If I understand correctly, all instructions documented here have both "Subnormal Support" and "Rounding Accuracy" as literal arguments - this is a 8 bytes for storing 4 bits, which is kind of overkill.

This is not a major comment, but probably you will consider merging these two into a single bitmask and reducing amount of arguments of each new instructions.

The same applies to fixed-point extension

This one's on me - we could change this, but since the translator is already using the preliminary spec, we'll have to circle back and update it there as well. Can we table this one for a later changelist? Also, do internal SPIR-V API changes need a deprecation period as well?

This patch introduces SPV_INTEL_arbitrary_precision_fixed_point extension, which adds operations for arbitrary precision fixed point numbers. The ac_fixed datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - total width of the datatype. It's encoded in the width of the OpTypeInt. * I - determines the position of the decimal point. * S - determines if this is a signed or an unsigned number. Full specification (intel/llvm#1934): https://github.com/intel/llvm/blob/2f6e965e686354fbb25f9c177a667a646de302eb/sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_fixed_point.asciidoc

This patch introduces SPV_INTEL_arbitrary_precision_fixed_point extension, which adds operations for arbitrary precision fixed point numbers. The ac_fixed datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - total width of the datatype. It's encoded in the width of the OpTypeInt. * I - determines the position of the decimal point. * S - determines if this is a signed or an unsigned number. Full specification can be found in intel/llvm#1934 Co-authored-by: Viktoria Maksimova <viktoria.maksimova@intel.com>

This patch introduces SPV_INTEL_arbitrary_precision_fixed_point extension, which adds operations for arbitrary precision fixed point numbers. The ac_fixed datatype is represented as a pseudo type using OpTypeInt. Its parameters are: * W - total width of the datatype. It's encoded in the width of the OpTypeInt. * I - determines the position of the decimal point. * S - determines if this is a signed or an unsigned number. Full specification can be found in intel#1934 Co-authored-by: Viktoria Maksimova <viktoria.maksimova@intel.com>

MrSidims · 2021-10-12T08:14:26Z

sycl/doc/extensions/SPIRV/SPV_INTEL_arbitrary_precision_fixed_point.asciidoc

+|=====
+9+<|*OpFixedSinINTEL* +
+
+An `OpTypeInt` representing an arbitrary precision fixed point number (ac_fixed) is passed in as the _Input_ and the sine of the value is returned in _Result_.


@ajaykumarkannan @tiwaria1 if input is more than of 64bit width, clang for spir target will generate a byval pointer for Input parameter. With the current version of the spec we would need to create an extra copy/load to make the proper translation. Alternatively we can allow pointers as input. WDYT?

I don't see any issue with updating this spec to allow pointer inputs. Users will use our wrappers over the spirv_ops.hpp functions. Without such an update, we will have worse performance due to generation of those extra loads and stores.

MrSidims · 2022-10-12T20:57:55Z

@ajaykumarkannan @tiwaria1 hi, is it considered a finished extension or extra work is required?

tiwaria1 · 2022-10-13T11:25:08Z

@ajaykumarkannan @tiwaria1 hi, is it considered a finished extension or extra work is required?

We haven't looked at this in a while. There is that "using pointers as an input to avoid the extra load/stores" issue that you pointed to above that will require an update.

ajaykumarkannan · 2022-10-13T15:16:05Z

@ajaykumarkannan @tiwaria1 hi, is it considered a finished extension or extra work is required?

@MrSidims The actual change has been implemented so this PR is the current state of implementation. We can check this in and address the remaining gaps later if needed.

gmlueck · 2022-10-13T18:06:43Z

Can you please rename these files? SPIR-V extensions now live in "sycl/doc/design/spirv-extensions".

…bitrary_types * Move SPV_INTEL extensions for arbitrary types to sycl/doc/design/spirv-extensions

ajaykumarkannan · 2022-10-18T03:46:46Z

Can you please rename these files? SPIR-V extensions now live in "sycl/doc/design/spirv-extensions".

Thanks Greg - moved the files now (and did a pull down from main).

ajaykumarkannan · 2022-10-28T21:44:44Z

Ping @gmlueck @MrSidims can we merge this in, or is there anything further you guys would like to change before committing. We also have been requested to merge these specs into the Khronos repository here:

MrSidims · 2022-11-21T13:24:43Z

In general LGTM, but I have two suggestions/questions to address:
#1934 (comment)
and
#1934 (comment)

gmlueck · 2022-11-21T13:31:34Z

Ping @gmlueck @MrSidims can we merge this in, or is there anything further you guys would like to change before committing.

I suggest abandoning this PR since these extension specifications are being reviewed instead in the "intel-innersource" repo you list above. Can we move the review discussion there to keep it all in one place?

ajaykumarkannan · 2022-11-22T18:17:43Z

Good note - I'll move this out to the other repo.

ajaykumarkannan · 2022-11-22T18:22:10Z

@MrSidims pinged you on the two reviews here:

ajaykumarkannan requested review from jbrodman and mkinsner as code owners June 19, 2020 15:57

jbrodman previously approved these changes Jun 19, 2020

View reviewed changes

AlexeySachkov reviewed Jun 19, 2020

View reviewed changes

ajaykumarkannan dismissed jbrodman’s stale review via 2f6e965 June 23, 2020 03:46

AlexeySachkov reviewed Jun 29, 2020

View reviewed changes

mlychkov reviewed Jun 29, 2020

View reviewed changes

vmaksimo mentioned this pull request Jul 7, 2020

Enable translation of ac_fixed functions KhronosGroup/SPIRV-LLVM-Translator#614

Closed

AlexeySachkov mentioned this pull request Jul 8, 2020

[SYCL] Introduce external arbitrary fixed point functions #2062

Merged

mlychkov mentioned this pull request Jul 20, 2020

[SYCL][FPGA] Enable translation of hls_float functions KhronosGroup/SPIRV-LLVM-Translator#648

Merged

mlychkov mentioned this pull request Jul 27, 2020

[SYCL][FPGA] Enable translation of ac_fixed functions KhronosGroup/SPIRV-LLVM-Translator#653

Merged

AlexeySachkov reviewed Aug 5, 2020

View reviewed changes

ajaykumarkannan requested a review from a team as a code owner August 12, 2020 03:06

ajaykumarkannan added 2 commits May 19, 2021 02:33

Re-phrased what values indicate signedness in fixed_point

72d9680

Added validation rules for floating_point

f61dbcc

MrSidims reviewed Oct 12, 2021

View reviewed changes

MrSidims mentioned this pull request Oct 12, 2021

Add load to obtain input for ap/hls instructions in case if its type > 64bits KhronosGroup/SPIRV-LLVM-Translator#1246

Merged

Add argument SignOfB to OpArbitraryFloatPowNINTEL

3155570

mlychkov requested review from AlexeySotkin, bashbaug and mbelicki as code owners December 17, 2021 11:39

github-actions bot added the Stale label Jun 16, 2022

github-actions bot closed this Jul 17, 2022

bader changed the title ~~[SYCL] Floating point and fixed point operations added to SPIR-V~~ [SPIRV][Doc] Add arbitrary precision floating and fixed point specs Jul 17, 2022

MrSidims reopened this Oct 12, 2022

github-actions bot removed the Stale label Oct 13, 2022

Merge branch 'sycl' of https://github.com/intel/llvm into kannanaj_ar…

71fa6a8

…bitrary_types * Move SPV_INTEL extensions for arbitrary types to sycl/doc/design/spirv-extensions

ajaykumarkannan requested a review from a team as a code owner October 18, 2022 03:45

MrSidims approved these changes Nov 21, 2022

View reviewed changes

MrSidims requested review from MrSidims and removed request for AlexeySotkin and mlychkov November 21, 2022 13:22

ajaykumarkannan closed this Nov 22, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[SPIRV][Doc] Add arbitrary precision floating and fixed point specs #1934

[SPIRV][Doc] Add arbitrary precision floating and fixed point specs #1934

ajaykumarkannan commented Jun 19, 2020

AlexeySachkov left a comment

jbrodman commented Jun 26, 2020

AlexeySachkov Aug 5, 2020

ajaykumarkannan May 19, 2021

AlexeySachkov Aug 5, 2020

ajaykumarkannan May 19, 2021

AlexeySachkov Aug 5, 2020

ajaykumarkannan May 19, 2021

MrSidims Oct 12, 2021 •

edited

Loading

tiwaria1 Oct 15, 2021 •

edited

Loading

MrSidims commented Oct 12, 2022

tiwaria1 commented Oct 13, 2022

ajaykumarkannan commented Oct 13, 2022 •

edited

Loading

gmlueck commented Oct 13, 2022

ajaykumarkannan commented Oct 18, 2022

ajaykumarkannan commented Oct 28, 2022

MrSidims commented Nov 21, 2022

gmlueck commented Nov 21, 2022

ajaykumarkannan commented Nov 22, 2022

ajaykumarkannan commented Nov 22, 2022


		_Result Type_ is an `OpTypeInt` of width `1+Eout+Mout` and is the type of _Result_.

		_M*_ contains the width of the mantissa of the floating point types within _Result_, _A_, and _B_.


		After Section 3.16, add a new section "3.16d Rounding Accuracy" as follows:

		=== Rounding Accuracy

[SPIRV][Doc] Add arbitrary precision floating and fixed point specs #1934

[SPIRV][Doc] Add arbitrary precision floating and fixed point specs #1934

Conversation

ajaykumarkannan commented Jun 19, 2020

AlexeySachkov left a comment

Choose a reason for hiding this comment

jbrodman commented Jun 26, 2020

AlexeySachkov Aug 5, 2020

Choose a reason for hiding this comment

ajaykumarkannan May 19, 2021

Choose a reason for hiding this comment

AlexeySachkov Aug 5, 2020

Choose a reason for hiding this comment

ajaykumarkannan May 19, 2021

Choose a reason for hiding this comment

AlexeySachkov Aug 5, 2020

Choose a reason for hiding this comment

ajaykumarkannan May 19, 2021

Choose a reason for hiding this comment

MrSidims Oct 12, 2021 • edited Loading

Choose a reason for hiding this comment

tiwaria1 Oct 15, 2021 • edited Loading

Choose a reason for hiding this comment

MrSidims commented Oct 12, 2022

tiwaria1 commented Oct 13, 2022

ajaykumarkannan commented Oct 13, 2022 • edited Loading

gmlueck commented Oct 13, 2022

ajaykumarkannan commented Oct 18, 2022

ajaykumarkannan commented Oct 28, 2022

MrSidims commented Nov 21, 2022

gmlueck commented Nov 21, 2022

ajaykumarkannan commented Nov 22, 2022

ajaykumarkannan commented Nov 22, 2022

MrSidims Oct 12, 2021 •

edited

Loading

tiwaria1 Oct 15, 2021 •

edited

Loading

ajaykumarkannan commented Oct 13, 2022 •

edited

Loading