P1928 R12 std::simd - Merge data-parallel types from the Parallelism TS 2

wg21bot commented 5 years ago

P1928R0 Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

ogiroux commented 4 years ago

SG1 in Belfast.

wg21bot commented 2 years ago

P1928R1 Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

ogiroux commented 1 year ago

Poll #1: After significant experience with the TS, we recommend that the next version (the TS version with improvements) of std::simd target the IS (C++26) SF F N A SA 10 8 0 0 0

Unanimous consent

Poll #2: We like all of the recommended changes to std::simd proposed in p1928r1 (Includes making all of std::simd constexpr, and dropping an ABI stable type) SF F N A SA No objection to unanimous consent

Poll #4: Future papers and future revisions of existing papers that target std::simd should go directly to LEWG. (We do not believe there are SG1 issues with std::simd today.) SF F N A SA 9 8 0 0 0

wg21bot commented 1 year ago

P1928R2 Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

mattkretz commented 1 year ago

P1928R3 Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

brycelelbach commented 1 year ago

2023-02-08 Issaquah Library Evolution Meeting

P1928R3: std::simd

2023-02-08 UTC-8 Issaquah Library Evolution Minutes

Champion: Matthias Kretz (IP)

Chair: Bryce Adelstein Lelbach (IP) & Billy Baker (IP) & Robert Leahy (IP)

Minute Taker: Ben Craig (IP) & Inbal Levi (IP)

Start: 2023-02-08 08:32 UTC-8

Does this paper have:

Examples?
- Yes.
Field experience?
- Implementation experience?
  - Yes, multiple, including in GCC11.
- Usage experience?
  - Yes, multiple, including in GCC11.
- Deployment experience?
  - Yes, multiple, including in GCC11.
Performance considerations?
- Yes.
Discussion of prior art?
- Yes.
Changes Library Evolution previously requested?
- Yes.
Wording?
- Yes.
Breaking changes?
- No.
Feature test macro?
- No, but we will add one.
Freestanding?
- No, but it will be made freestanding in the next revision.
Format support?
- Discussed, should be considered for future revisions.
Allocator support?
- Discussed, N/A.
Format support?
- If it becomes a range, we get that for free.
Hash support?
- No, but it should have it.

Open Questions:

What should the default ABI tag be? In the TS it was the compatible tag (simd_abi::compatible), but feedback indicates it should be the performant tag (simd_abi::native).
If we change the default ABI tag to simd_abi::native, how do we retain the ability to interface with older code?
Should simd_abi::fixed_size be ABI incompatible?
What conversions happen when we have operations that take multiple SIMD objects, or SIMD objects and scalar types? What type is returned? Do we need a value preserving common type?
Should the single iterator constructor take a range or span instead?
Should the single iterator constructor be a factory function or static member function?
Should these facilities go in a nested namespace or have a simd_ prefix?
Are we satisfied with how simd::reference works?
Can and/or should we support variable-length SIMD, such as ARM SVE?

Future Work:

Scatter/gather support.
SIMD UDLs.

Poll Ideas:

Renaming of "generator" constructor to avoid clashing with stateful generators and generator.
Renaming of "vector_aligned" to avoid clashing with vector.
Renaming "reduce/hmin/hmax" to avoid clashing with reduce.
Renaming "copy_to/copy_from".
Renaming "_if" to avoid clashing with algorithms.
Renaming "conditional_operator".
Change default SIMD ABI tag native (performant) and remove compatible.
Change the fixed_size ABI tag to not recommend ABI compatibility.
Enable conversions between any simd specializations with equal element count, even if the conversion isn't portable between targets with different native SIMD widths.
Make conversions between simd specializations explicit if the element type conversion isn't value preserving.
Cast removals and changes.
Remove the load/store flags.
Add a default for load/store flags.
Make load/store flags typed.
Expose the vectorizable concept.
Remove where.
Replace the single iterator constructor with a range/span constructor.
Nested namespace/prefix for these facilities.
Add conditional_operator.
Upcasting/common type behavior.

POLL: Change the default SIMD ABI tag to simd_abi::native instead of simd_abi::compatible.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
16	12	0	0	1

Attendance: 23 (in person) + 20 (remote)

# of Authors: 1

Author Position: SF

Outcome: Consensus in favor.

POLL: Change simd_abi::fixed_size to not recommend implementations make it ABI compatible.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
16	7	1	0	1

Attendance: 23 (in person) + 18 (remote)

# of Authors: 1

Author Position: SF

Outcome: Consensus in favor.

Break: 10:14

Resume: 10:38

__POLL: Make simd::size an integral_constant instead of a static member function.__

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
9	8	7	1	0

Attendance: 21 (in person) + 14 (remote)

# of Authors: 1

Author Position: SF

Outcome: Consensus in favor.

POLL: simd masked operations should look like (vote for as many options as you'd like):

Option	Votes
`where(u > 0, v).copy_from(ptr)`	12
`v.copy_from_if(u > 0, ptr)`	1
`v.copy_from_if(ptr, u > 0)`	2
`v.copy_from(ptr, u > 0)`	14
`v.copy_from(u > 0, ptr)`	3
`v.copy_from_where(u > 0, ptr)`	4
`v.copy_from_where(ptr, u > 0)`	11

Attendance: 24 (in person) + 13 (remote)

# of Authors: 1

POLL: simd masked operations should look like (vote once for your favorite):

Option	Votes
`where(u > 0, v).copy_from(ptr)`	5
`v.copy_from(ptr, u > 0)`	12
`v.copy_from_where(ptr, u > 0)`	6

Attendance: 24 (in person) + 13 (remote)

# of Authors: 1

Author Position: v.copy_from_where(ptr, u > 0)

Outcome: Consensus for v.copy_from(ptr, u > 0)

POLL: Make copy_to, copy_from, and the load constructor only do value-preserving conversions by default and require passing a flag to do non-value-preserving conversions.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
14	9	1	0	0

Attendance: 24 (in person) + 13 (remote)

# of Authors: 1

Author Position: SF

Outcome: Strong consensus in favor.

Attendance: 24 (in person) + 13 (remote)

Break: 11:58

Resume: 13:03

POLL: SIMD types and operations should be value preserving, even if that means they're inconsistent with the builtin numeric types.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
3	10	6	3	0

Attendance: 27 (in person) + 9 (remote)

# of Authors: 1

Author Position: WF

Outcome: Consensus in favor

WF: I generally like the direction, in the sense of safety, but I'm not sure of the consequences of such a design.

N: I would rather be consistently wrong than inconsistent.

WA: I don't want to be inconsistent with how things have been for the last 30 years.

WF: I'm not sure if we've evaluated the performance impact of this.

N: Just because we've made a mistake in the past doesn't mean we have to keep making it. I'm not convinced that the new approach is the right thing.

Which of these do we want?

float * simd<double> = simd<double>
float * simd<double> = simd<double>
int * simd<float> = simd<float>
int * simd<float> = simd<double>

POLL: 2 * simd<float> should produce simd<double> (status quo: simd<float>).

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
1	5	9	6	1

Attendance: 27 (in person) + 9 (remote)

# of Authors: 1

Author Position: SF

Outcome: No consensus.

N: I want more information. The author should tell me what to do.

N: To some degree I want the value preservation. On the other hand, the common case of using an int literal here leading to simd<double> would be surprising.

Names in this proposal:

simd
simd_mask
is_simd
is_simd_mask
is_simd_flag_type
simd_size
rebind_simd
resize_simd
*is_abi_tag -> is_simd_abi_tag
*element_aligned -> simd_element_aligned
*vector_aligned -> simd_vector_aligned
*overaligned -> simd_overaligned
*memory_alignment -> simd_memory_alignment
min
max
minmax
clamp
reduce
*split
*concat
hmin -> min_element
hmax -> max_element
(all|any)_of

POLL: Put SIMD types and operations into std:: and add the simd_ prefix to SIMD specific things (such as split and vector_aligned).

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
4	5	4	9	2

Attendance: 23 (in person) + 9 (remote)

# of Authors: 1

Author Position: WA

Outcome: No consensus.

POLL: Put SIMD types and operations into a nested namespace in std::.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
4	7	0	5	9

Attendance: 25 (in person) + 9 (remote)

# of Authors: 1

Author Position: SA

Outcome: No consensus.

POLL: simd should be a range.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
4	9	5	4	4

Attendance: 25 (in person) + 8 (remote)

# of Authors: 1

Author Position: WF

Outcome: No consensus.

POLL: There should be an explicit way to get a view to a simd.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
8	12	3	3	0

Attendance: 26 (in person) + 8 (remote)

# of Authors: 1

Author Position: WF

Outcome: Consensus in favor.

POLL: simd should have explicitly named functions for horizontal minimum and horizontal maximum.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
4	5	7	4	2

Attendance: 25 (in person) + 8 (remote)

# of Authors: 1

Author Position: SF

Outcome: No consensus.

Break: 15:19

Resume: 15:35

__POLL: Rename all_of/ any_of/none_of to reduce_all_of/reduce_any_of/reduce_none_of.__

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
2	1	1	8	5

Attendance: 22 (in person) + 2 (remote)

# of Authors: 1

Author Position: SA

Outcome: Consensus against.

POLL: Rename all_of/any_of/none_of to reduce_and/reduce_or/reduce_nand.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
2	6	2	4	3

Attendance: 22 (in person) + 2 (remote)

# of Authors: 1

Author Position: SA

Outcome: No consensus.

__POLL: Rename popcount to reduce_count.__

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
4	9	2	1	2

Attendance: 22 (in person) + 2 (remote)

# of Authors: 1

Author Position: SF

Outcome: Weak consensus in favor.

POLL: Rename find_first_set/find_last_set to reduce_min_index/reduce_max_index.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
2	7	3	2	3

Attendance: 22 (in person) + 2 (remote)

# of Authors: 1

Author Position: SF

Outcome: Weak consensus in favor.

End: 16:04

Summary

We extensively reviewed P1928, which proposes adding SIMD types and operations from the C++ Parallelism Technical Specification version 2 to the C++ International Standard.

Some of the key topics discussed during our review:

What the default ABI tag should be and what ABI compatiblity should be recommended by ABI tags.
How conversions and common type computation should work for SIMD types and operations. We had consensus that we wanted these to be value preserving, even if that means they're inconsistent with the builtin numeric types.
The interface for masked/predicated SIMD operations. We had consensus to introduce a simpler approach than the where facility in the Technical Specification.
The names of many facilities and the nested namespace or prefix for the entire facility.

We intend to conduct another extensive review on the next revision at the 2023-06 Varna meeting, with the aim of approving the design and advancing the paper to Library Evolution for C++26.

Next Steps

Bring a revision of P1928R4 (std::simd), with the guidance below, to Library Evolution for further design review:

Produce wording that adds the entire simd feature with the suggested changes to the C++ International Standard.
Audit and update the wording taken from the Technical Specification to use the latest wording techniques.
Explore options for nested namespaces and prefixes for the proposed facility and add a section in the next revision that presents 2 to 4 options, including at least one option for nested namespaces and at least one option for prefixes.
Add a feature test macro __cpp_lib_simd.
Make all proposed facilities freestanding.
Change SIMD types and operations to be value preserving, even if that means they're inconsistent with the builtin numeric types.
Change copy_to, copy_from, and the load constructor to only do value-preserving conversions by default and require passing a flag to do non-value-preserving conversions.
Change the default SIMD ABI tag to simd_abi::native instead of simd_abi::compatible.
Change simd_abi::fixed_size to not recommend implementations make it ABI compatible.
Change iterator parameters to be passed by value, not const reference.
Change simd::size to be an integral_constant instead of a static member function.__
Remove where.
Add masked overloads of copy_to and copy_from that take a mask as the second parameter.
Rename popcount to reduce_count.
Rename find_first_set/find_last_set to reduce_min_index/reduce_max_index.
Prepare a list of alternative names for conditional_operator.
Write a separate paper that adds an explicit way to get a view to a simd.

ben-craig commented 1 year ago

Polling results in P2638 indicate that P1915 needs to remove split and concat

brycelelbach commented 1 year ago

https://wg21.link/P1928R4 will be available in the 2023-05 mailing.

wg21bot commented 1 year ago

P1928R4 std::simd - Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

inbal2l commented 1 year ago

2023-06-13 Library Evolution Varna

P1928R4: Merge data-parallel types from the Parallelism TS 2

2023-06-13 Library Evolution Minutes

2023-06-14 Library Evolution Minutes

2023-06-15 Library Evolution Minutes

Chair: Bryce Adelstein Lelbach, Fabio Fracassi

Champion: Matthias Kretz

Minute Taker(s): Inbal Levi / Mark Zeren / Robert Leahy

Summary

Few direction polls were taken during the discussion.

The paper was voted into C++ 26 on Thursday, 06-15:

POLL: Modify P1928D6 (“simd”) as described above, and then send the revised papre to library for C++26 (to be confirmed with a library evolution electronic poll).

SF	F	N	A	SA
16	3	1	0	0

Attendance: 22 (P) + 3 (R)

Author’s position: SF

# of Authors: 1

Outcome: Strong consensus in favor

N: There’s a lot of outstanding TOODs.

TODO: @brycelelbach

Next Steps

Forwarded to LWG for C++ 26 (to be confirmed with electronic poll).

TODO: @brycelelbach

wg21bot commented 1 year ago

P1928R5 std::simd - Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

wg21bot commented 1 year ago

P1928R6 std::simd - Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

wg21bot commented 1 year ago

P1928R7 std::simd - Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

mattkretz commented 11 months ago

P1928R8 std::simd - Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

inbal2l commented 11 months ago

Library Evolution Meeting Kona 2023-11-08

P1928R7: std::simd — merge data -parallel types from the Parallelism TS 2

2023-11-08 Library Evolution Kona 2023 Meeting Minutes

Chair: Billy Baker / Inbal Levi

Champion: Matthias Kretz

Minute Taker: Robert Leahy

Summary

Blog post on the topic of regularity: https://mattkretz.github.io/2023/11/16/preserving-regularity.html

No polls were taken.

Outcome

The discussion brought up using “optional”, author will followup with more investigation before getting back to LEWG for additional discussion.

billy-baker commented 9 months ago

Library Evolution Meeting Kona 2024-01-16

P1928R8: std::simd — merge data -parallel types from the Parallelism TS 2

2024-01-16 Library Evolution Telecon Minutes

Champion: Matthias Kretz

Chair: Billy Baker

Minute Taker: Mark Hoemmen

Summary

Does this paper have:

Why this facility should go into the standard library? Discussed for SIMD in length before
Examples? Yes
Field experience? Yes
- Implementation experience? Yes
- Usage experience? Yes
- Deployment experience? Yes
Performance considerations? Yes
Discussion of prior art?
Changes Library Evolution previously requested? Yes
Support for "std::format", "hash", and allocators (if required)? No for hash, open question on formatting.
Wording? Yes
Feature test macro? Yes
Freestanding? No

Poll: Restore the precondition on reduce_min_index(empty_mask) and reduce_max_index(empty_mask) (TS status quo, UB).

SF	F	N	A	SA
7	6	1	0	0

Attendance: 18 # of Authors: 1 Author Position: SF Outcome: Strong consensus

Poll: Return an unspecified value on reduce_min_index(empty_mask) and reduce_max_index(empty_mask).

SF	F	N	A	SA
0	7	3	1	2

Attendance: 19 # of Authors: 1 Author Position: F Outcome: No consensus.

Poll: Return std::optional from reduce_min_index and reduce_max_index.

SF	F	N	A	SA
0	1	2	7	4

Attendance: 19 # of Authors: 1 Author Position: WA Outcome: Consensus against.

WA: Not a bad design. optional is not ergonomic. WF: optional is a vocabulary type. Seems reasonable.

POLL: Modify P1928R8 (Merge data-parallel types from the Parallelism TS 2) by restoring the TS specification for reduce_min_index/reduce_max_inde and adding the change to 16.4.2.3 to list the header, and then send the revised paper to LWG for C++26 to be confirmed with a Library Evolution electronic poll.

SF	F	N	A	SA
9	2	0	1	1

Attendance: 19 # of Authors: 1 Author Position: SF Outcome: Strong consensus to forward to electonic polling.

WA: Would be SF if load/store interface did not use single iterators (security) SA: Concerned about the comparison operators and constructors

Action Items:

Modify with the results of the polls.
Add to the wording to include in Table 21: C++ library headers [tab:headers.cpp].

ACTION: There was a general feeling that the specification of reduce_min_index and reduce_max_index would benefit from erroneous behavior. Using erroneous behavior was not polled since it is not yet in the working draft.

Next Steps

The paper will be updated according to the input and added to the Feb 2024 electronic poll.

wg21bot commented 5 months ago

P1928R9 std::simd - Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

ben-craig commented 4 months ago

2024-06-24 Library Evolution St. Louis Meeting

P1928R9: std::simd - Merge data-parallel types from the Parallelism TS 2

2024-06-24 Library Evolution St. Louis Minutes

Champion: Matthias Kretz

Chair: Ben Craig

Minute Taker: Khalil Estell

Summary

POLL: Skip electronic poll for P1928R9 (std::simd - Merge data-parallel types from the Parallelism TS 2) and send directly to LWG for C++26 classified as B3 (addition).

SF	WF	N	WA	SA
14	3	3	0	2

Attendance: 21 in person + 19 remote

# of Authors: 1

Author Position: SF

Outcome: Consensus in favor

Comments: SA: Ok with skipping electronic poll, but concerns on the proposal still hold due to incoherence with rest of the standard library.

SA: Not good with the forwarding process

Next Steps

Forwarded to LWG for C++26 (without an electronic poll). Will be reviewed in LWG.

ben-craig commented 4 months ago

2024-06-28 Library Evolution St. Louis Meeting

P1928R10: std::simd - Merge data-parallel types from the Parallelism TS 2

2024-06-28 Library Evolution St. Louis Minutes

Champion: Matthias Kretz

Chair: Ben Craig & Fabio Fracassi

Minute Taker: Steve Downey

Summary

POLL: Approve presented changes in P1928R10 made due to the resolution of P3275 (which concludes P3275).

No objection to unanimous consent

Attendance: 14 + 5

# of Authors: 1

Next Steps

Send P1928 back to LWG, to be included in C++26 (without an electronic poll).

wg21bot commented 3 months ago

P1928R10 std::simd - Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

wg21bot commented 3 months ago

P1928R11 std::simd - Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

wg21bot commented 2 weeks ago

P1928R12 std::simd - Merge data-parallel types from the Parallelism TS 2 (Matthias Kretz)

cplusplus / papers