Expose AVX512 variants of SSE-SSE4.2 instructions #84909

tannergooding · 2023-04-17T00:26:36Z

This exposes 512-bit variants of all the relevant SSE-SSE4.2 instructions making progress towards completing:

For a few instructions, there is a new Q variant of the instruction that supports 64-bit elements. Where relevant, existing code paths were updated to utilize this instruction

The breakdown of the changes are:

JIT: +395, -44
- +84 of this is the main table, the rest is minor updates for specialized handling
LIB: +2386, -347
- +225 of this is the reference assembly
- +1302 lines is the "duplicated" entries for the *.PlatformNotSupported.cs version of the file
- +1712 lines of this is doc comments (split between the *.cs and *.PlatformNotSupported.cs)
TST: +10684, -570
- +392 lines of this is the main table used to generate the tests
- +10292 lines is csproj and manual tests where they aren't templated yet today

…lyHigh for AVX512F, AVX512BW, AVX512DQ, and AVX512DQ.VL

…etic, ShiftRightLogical, and ShiftRightLogical128BitLane for Avx512F, Avx512F.VL, and Avx512BW

dotnet-issue-labeler · 2023-04-17T00:26:44Z

Note regarding the new-api-needs-documentation label:

This serves as a reminder for when your PR is modifying a ref *.cs file and adding/modifying public APIs, please make sure the API implementation in the src *.cs file is documented with triple slash comments, so the PR reviewers can sign off that change.

ghost · 2023-04-17T00:26:48Z

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch, @kunalspathak
See info in area-owners.md if you want to be subscribed.

Issue Details

This exposes 512-bit variants of all the relevant SSE-SSE4.2 instructions making progress towards completing:

For a few instructions, there is a new Q variant of the instruction that supports 64-bit elements. Where relevant, existing code paths were updated to utilize this instruction

The breakdown of the changes are:

JIT: +395, -44
- +84 of this is the main table, the rest is minor updates for specialized handling
LIB: +2386, -347
- +225 of this is the reference assembly
- +1302 lines is the "duplicated" entries for the *.PlatformNotSupported.cs version of the file
- +1712 lines of this is doc comments (split between the *.cs and *.PlatformNotSupported.cs)
TST: +10684, -570
- +392 lines of this is the main table used to generate the tests
- +10292 lines is csproj and manual tests where they aren't templated yet today

Author:	tannergooding
Assignees:	tannergooding
Labels:	`area-CodeGen-coreclr`, `new-api-needs-documentation`
Milestone:	-

tannergooding · 2023-04-17T00:30:23Z

CC. @dotnet/avx512-contrib, @dotnet/jit-contrib, @dotnet/area-system-runtime-intrinsics

…t vectors

tannergooding · 2023-04-17T16:12:09Z

Diffs

We see a TP improvement of -0.02% in a couple contexts and codegen improvements for others. The codegen improvements come from:

containment for things like 256-bit pshufb
utilizing new the instructions that work with 64-bit elements (such as vpmaxsq)

Linux shows a couple regressions in some of the "recursive implementations" for the Avx.Shuffle and Avx2.AlignRight APIs.

Such diffs come from us supporting containment where we didn't previously and that hoisting the load into the jump table fallback.

We can/should probably avoid containment in the case we need to generate such a jump table, but its a pre-existing condition for all such nodes, so we can handle it in a follow up PR (#84933).

EgorBo

JIT and BCL changes LGTM, I skimmed through the tests and they also look ok

tannergooding · 2023-04-17T20:37:37Z

Tests are passing, but timing out. I have up #84948 to try and help solve some of that.

src/tests/JIT/HardwareIntrinsics/X86/Avx1/Avx1_handwritten_r.csproj

src/tests/JIT/HardwareIntrinsics/X86/Avx2/Avx2_handwritten_r.csproj

steveharter · 2023-04-17T21:04:17Z

src/tests/JIT/HardwareIntrinsics/X86/Avx512BW/Store.cs

+
+            if (Avx512BW.IsSupported)
+            {
+                using (TestTable<short> intTable = new TestTable<short>(new short[32] { 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4 }, new short[32]))


Suggested change

using (TestTable<short> intTable = new TestTable<short>(new short[32] { 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4 }, new short[32]))

using (TestTable<short> intTable = new(new short[32] { 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4 }, new short[32]))

a bit more terse; here and elsewhere

This is setup in a way that makes it easier to switch to generated from a template in the future.

src/tests/JIT/HardwareIntrinsics/X86/Avx512DQ_VL_Vector128/Program.Avx512DQ_VL_Vector128.cs

src/tests/JIT/HardwareIntrinsics/X86/Avx512F/ConvertToVector512Double.Int32.cs

src/tests/JIT/HardwareIntrinsics/X86/Avx512F/ConvertToVector512Double.Single.cs

src/tests/JIT/HardwareIntrinsics/X86/Avx512F/ConvertToVector512Double.Int32.cs

src/tests/JIT/HardwareIntrinsics/X86/Avx512F_VL_Vector256/Program.Avx512F_VL_Vector256.cs

src/tests/JIT/HardwareIntrinsics/X86/Avx512F/LoadAlignedVector512NonTemporal.cs

src/coreclr/jit/gentree.cpp

tannergooding · 2023-04-17T23:30:00Z

Pulled in #84948 to help resolve test timeouts...

Edit: It did not resolve the timeouts, seems merged test wrappers are more special than we thought. I've pulled in #84959 instead after validating that CI is indeed filtering the tests down on the relevant platforms now.

steveharter

LGTM pending green CI

tannergooding · 2023-04-18T15:05:52Z

Resolved merge conflicts. Should pass now that the tests are properly getting filtered and last run showed OSX got past the timeouts, etc.

tannergooding added 15 commits April 16, 2023 07:38

Ensure already exposed AVX512 APIs are in the right ISAs

8636dab

Exposing Add and Subtract for AVX512F and AVX512BW

08fb613

Exposing Abs for AVX512F, AVX512F.VL, and AVX512BW

396abf1

Exposing AddSaturate and SubtractSaturate for AVX512BW

835fbad

Expose AlignRight and Average for AVX512BW

41d64d7

Expose ConvertToVector* for AVX512F and AVX512BW

3908c75

Expose Divide, Multiply, MultiplyAddAdjacent, MultiplyLow, and Multip…

d3b91bf

…lyHigh for AVX512F, AVX512BW, AVX512DQ, and AVX512DQ.VL

Expose DuplicateEvenIndexed and DuplicateOddIndexed in Avx512F

b1e8fa9

Expose LoadAlignedVector512NonTemporal in Avx512F

55cb66b

Expose Max and Min for Avx512F, Avx512F.VL, and Avx512BW

30ef5e3

Expose PackSignedSaturate and PackUnsignedSaturate for Avx512BW

ac0acd0

Expose ShiftLeftLogical, ShiftLeftLogical128BitLane, ShiftRightArithm…

03b3ecb

…etic, ShiftRightLogical, and ShiftRightLogical128BitLane for Avx512F, Avx512F.VL, and Avx512BW

Expose Shuffle, ShuffleLow, and ShuffleHigh for Avx512F and Avx512BW

b00635e

Expose Sqrt and SumAbsoluteDifferences for Avx512F and Avx512BW

aeffab3

Expose UnpackLow and UnpackHigh for Avx512F and Avx512BW

314605e

ghost assigned tannergooding Apr 17, 2023

dotnet-issue-labeler bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Apr 17, 2023

dotnet-issue-labeler bot added the new-api-needs-documentation label Apr 17, 2023

tannergooding added the avx512 Related to the AVX-512 architecture label Apr 17, 2023

tannergooding added 4 commits April 16, 2023 18:02

Ensure Avx2.Shuffle correctly handles containment for pshufb

f1d3013

Apply formatting patch

a3ec955

Fixing a couple of minor asserts

4a84b0c

Ensure TYP_DOUBLE is correctly handled for arithmetic right shift

7e6856f

This was referenced Apr 17, 2023

IOException running NuGet-Migrations during tests in dotnet CLI first run #80619

Closed

WasmTestOnBrowser-System.Text.Json.Tests.WorkItemExecution timing out #84434

Closed

tannergooding added 2 commits April 17, 2023 05:45

Work around an SPMI failure with the baseline

1b8bf9b

Fixing some test logic to account for "per lane" operations on 512-bi…

6bade9f

…t vectors

tannergooding marked this pull request as ready for review April 17, 2023 16:02

Ensure we don't fail for 64-bit vector multiplication on 32-bit

2afd909

tannergooding mentioned this pull request Apr 17, 2023

Don't contain operands for intrinsics when generating a jmp table fallback #84933

Merged

EgorBo approved these changes Apr 17, 2023

View reviewed changes

tannergooding mentioned this pull request Apr 17, 2023

Reconfigure the HWIntrinsic tests to run more selectively #84948

Merged

build-analysis bot mentioned this pull request Apr 17, 2023

[wasm] interpreter timeouts when WebSocket closes unexpectedly #84101

Closed

steveharter reviewed Apr 17, 2023

View reviewed changes

src/coreclr/jit/gentree.cpp Show resolved Hide resolved

Merge remote-tracking branch 'dotnet/main' into avx512-2

b4d9a35

tannergooding force-pushed the avx512-2 branch from 2af3423 to b4d9a35 Compare April 17, 2023 23:29

build-analysis bot mentioned this pull request Apr 18, 2023

Tracking issue for CI build timeouts #76454

Closed

tannergooding mentioned this pull request Apr 18, 2023

Split the MergedTestWrapper for the HWIntrinsic tests so we can actually filter per platform #84959

Merged

tannergooding force-pushed the avx512-2 branch from aa98aa3 to 30e37fa Compare April 18, 2023 11:54

steveharter approved these changes Apr 18, 2023

View reviewed changes

build-analysis bot mentioned this pull request Apr 18, 2023

nativeaot/SmokeTests/DwarfDump failing on linux-x64 Debug #84979

Closed

Merge remote-tracking branch 'dotnet/main' into avx512-2

55adfca

tannergooding force-pushed the avx512-2 branch from 30e37fa to 55adfca Compare April 18, 2023 15:13

tannergooding merged commit dbcffe7 into dotnet:main Apr 18, 2023

tannergooding deleted the avx512-2 branch April 18, 2023 19:29

ghost locked as resolved and limited conversation to collaborators May 19, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Expose AVX512 variants of SSE-SSE4.2 instructions #84909

Expose AVX512 variants of SSE-SSE4.2 instructions #84909

tannergooding commented Apr 17, 2023

dotnet-issue-labeler bot commented Apr 17, 2023

ghost commented Apr 17, 2023

tannergooding commented Apr 17, 2023

tannergooding commented Apr 17, 2023 •

edited

Loading

EgorBo left a comment

tannergooding commented Apr 17, 2023

steveharter Apr 17, 2023

tannergooding Apr 17, 2023

tannergooding commented Apr 17, 2023 •

edited

Loading

steveharter left a comment

tannergooding commented Apr 18, 2023

	using (TestTable<short> intTable = new TestTable<short>(new short[32] { 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4 }, new short[32]))
	using (TestTable<short> intTable = new(new short[32] { 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4, 1, -5, 100, 0, 1, 2, 3, 4 }, new short[32]))

Expose AVX512 variants of SSE-SSE4.2 instructions #84909

Expose AVX512 variants of SSE-SSE4.2 instructions #84909

Conversation

tannergooding commented Apr 17, 2023

dotnet-issue-labeler bot commented Apr 17, 2023

ghost commented Apr 17, 2023

tannergooding commented Apr 17, 2023

tannergooding commented Apr 17, 2023 • edited Loading

EgorBo left a comment

Choose a reason for hiding this comment

tannergooding commented Apr 17, 2023

steveharter Apr 17, 2023

Choose a reason for hiding this comment

tannergooding Apr 17, 2023

Choose a reason for hiding this comment

tannergooding commented Apr 17, 2023 • edited Loading

steveharter left a comment

Choose a reason for hiding this comment

tannergooding commented Apr 18, 2023

tannergooding commented Apr 17, 2023 •

edited

Loading

tannergooding commented Apr 17, 2023 •

edited

Loading