Implement simple SIMD intrinsics for AVX/AVX2 by fiigii · Pull Request #16287 · dotnet/coreclr

fiigii · 2018-02-08T23:10:58Z

This is the first part of https://github.com/dotnet/coreclr/issues/16246

close https://github.com/dotnet/coreclr/issues/16244

fiigii · 2018-02-09T10:24:09Z

test Windows_NT x64 Checked jitincompletehwintrinsic
test Windows_NT x64 Checked jitx86hwintrinsicnoavx
test Windows_NT x64 Checked jitx86hwintrinsicnoavx2
test Windows_NT x64 Checked jitx86hwintrinsicnosimd
test Windows_NT x64 Checked jitnox86hwintrinsic

test Windows_NT x86 Checked jitincompletehwintrinsic
test Windows_NT x86 Checked jitx86hwintrinsicnoavx
test Windows_NT x86 Checked jitx86hwintrinsicnoavx2
test Windows_NT x86 Checked jitx86hwintrinsicnosimd
test Windows_NT x86 Checked jitnox86hwintrinsic

test Ubuntu x64 Checked jitincompletehwintrinsic
test Ubuntu x64 Checked jitx86hwintrinsicnoavx
test Ubuntu x64 Checked jitx86hwintrinsicnoavx2
test Ubuntu x64 Checked jitx86hwintrinsicnosimd
test Ubuntu x64 Checked jitnox86hwintrinsic

test OSX10.12 x64 Checked jitincompletehwintrinsic
test OSX10.12 x64 Checked jitx86hwintrinsicnoavx
test OSX10.12 x64 Checked jitx86hwintrinsicnoavx2
test OSX10.12 x64 Checked jitx86hwintrinsicnosimd
test OSX10.12 x64 Checked jitnox86hwintrinsic

fiigii · 2018-02-09T10:35:36Z

@CarolEidt @tannergooding PTAL

I removed the test cases of Avx2.CompareEqual/CompareGreaterThan.(U)Int64 due to the SSE4.1/SSE4.2 encoding issue. Logged at https://github.com/dotnet/coreclr/issues/16296

jkotas · 2018-02-09T17:48:29Z

test OSX10.12 x64 Checked jitincompletehwintrinsic
test OSX10.12 x64 Checked jitx86hwintrinsicnoavx
test OSX10.12 x64 Checked jitx86hwintrinsicnoavx2
test OSX10.12 x64 Checked jitx86hwintrinsicnosimd
test OSX10.12 x64 Checked jitnox86hwintrinsic

tannergooding · 2018-02-09T17:55:58Z

+    ("SimpleBinOpTest.template", new string[] { "Avx", "Avx",   "AndNot",     "Single",     "Vector256",    "32",          "(float)(random.NextDouble())",        "((~BitConverter.SingleToInt32Bits(left[0])) & BitConverter.SingleToInt32Bits(right[0])) != BitConverter.SingleToInt32Bits(result[0])",    "((~BitConverter.SingleToInt32Bits(left[i])) & BitConverter.SingleToInt32Bits(right[i])) != BitConverter.SingleToInt32Bits(result[i])"}),
+    ("SimpleTernOpTest.template",new string[] { "Avx", "Avx",   "BlendVariable", "Double",  "Vector256",    "32",          "(double)(random.NextDouble())", "(double)(((i % 2) == 0) ? -0.0E0 : 1.0E0)", "((BitConverter.DoubleToInt64Bits(thirdOp[0]) >> 63) & 1) == 1 ? BitConverter.DoubleToInt64Bits(secondOp[0]) != BitConverter.DoubleToInt64Bits(result[0]) : BitConverter.DoubleToInt64Bits(firstOp[0]) != BitConverter.DoubleToInt64Bits(result[0])", "((BitConverter.DoubleToInt64Bits(thirdOp[i]) >> 63) & 1) == 1 ? BitConverter.DoubleToInt64Bits(secondOp[i]) != BitConverter.DoubleToInt64Bits(result[i]) : BitConverter.DoubleToInt64Bits(firstOp[i]) != BitConverter.DoubleToInt64Bits(result[i])"}),
+    ("SimpleTernOpTest.template",new string[] { "Avx", "Avx",   "BlendVariable", "Single",  "Vector256",    "32",          "(float)(random.NextDouble())",  "(float)(((i % 2) == 0) ? -0.0E0 : 1.0E0)",  "((BitConverter.SingleToInt32Bits(thirdOp[0]) >> 31) & 1) == 1 ? BitConverter.SingleToInt32Bits(secondOp[0]) != BitConverter.SingleToInt32Bits(result[0]) : BitConverter.SingleToInt32Bits(firstOp[0]) != BitConverter.SingleToInt32Bits(result[0])", "((BitConverter.SingleToInt32Bits(thirdOp[i]) >> 31) & 1) == 1 ? BitConverter.SingleToInt32Bits(secondOp[i]) != BitConverter.SingleToInt32Bits(result[i]) : BitConverter.SingleToInt32Bits(firstOp[i]) != BitConverter.SingleToInt32Bits(result[i])"}),
+    ("SimpleUnOpTest.template",  new string[] { "Avx", "Avx",   "DuplicateEvenIndexed", "Double", "Vector256", "32",       "(double)(random.NextDouble())", "BitConverter.DoubleToInt64Bits(firstOp[0]) != BitConverter.DoubleToInt64Bits(result[0])", "(i % 2 == 0) ? (BitConverter.DoubleToInt64Bits(firstOp[i]) != BitConverter.DoubleToInt64Bits(result[i])) : (BitConverter.DoubleToInt64Bits(firstOp[i - 1]) != BitConverter.DoubleToInt64Bits(result[i]))"}),


nit: It would be useful to treat this code as a "table" and keep the individual columns aligned, where possible.

Will fix. Thanks.

jkotas · 2018-02-09T17:57:30Z

test OSX10.12 x64 Checked Innerloop Build and Test

tannergooding · 2018-02-09T18:02:32Z

+    ("SimpleBinOpTest.template", new string[] { "Avx", "Avx",   "AndNot",     "Double",     "Vector256",    "32",          "(double)(random.NextDouble())",       "((~BitConverter.DoubleToInt64Bits(left[0])) & BitConverter.DoubleToInt64Bits(right[0])) != BitConverter.DoubleToInt64Bits(result[0])",    "((~BitConverter.DoubleToInt64Bits(left[i])) & BitConverter.DoubleToInt64Bits(right[i])) != BitConverter.DoubleToInt64Bits(result[i])"}),
+    ("SimpleBinOpTest.template", new string[] { "Avx", "Avx",   "AndNot",     "Single",     "Vector256",    "32",          "(float)(random.NextDouble())",        "((~BitConverter.SingleToInt32Bits(left[0])) & BitConverter.SingleToInt32Bits(right[0])) != BitConverter.SingleToInt32Bits(result[0])",    "((~BitConverter.SingleToInt32Bits(left[i])) & BitConverter.SingleToInt32Bits(right[i])) != BitConverter.SingleToInt32Bits(result[i])"}),
+    ("SimpleTernOpTest.template",new string[] { "Avx", "Avx",   "BlendVariable", "Double",  "Vector256",    "32",          "(double)(random.NextDouble())", "(double)(((i % 2) == 0) ? -0.0E0 : 1.0E0)", "((BitConverter.DoubleToInt64Bits(thirdOp[0]) >> 63) & 1) == 1 ? BitConverter.DoubleToInt64Bits(secondOp[0]) != BitConverter.DoubleToInt64Bits(result[0]) : BitConverter.DoubleToInt64Bits(firstOp[0]) != BitConverter.DoubleToInt64Bits(result[0])", "((BitConverter.DoubleToInt64Bits(thirdOp[i]) >> 63) & 1) == 1 ? BitConverter.DoubleToInt64Bits(secondOp[i]) != BitConverter.DoubleToInt64Bits(result[i]) : BitConverter.DoubleToInt64Bits(firstOp[i]) != BitConverter.DoubleToInt64Bits(result[i])"}),
+    ("SimpleTernOpTest.template",new string[] { "Avx", "Avx",   "BlendVariable", "Single",  "Vector256",    "32",          "(float)(random.NextDouble())",  "(float)(((i % 2) == 0) ? -0.0E0 : 1.0E0)",  "((BitConverter.SingleToInt32Bits(thirdOp[0]) >> 31) & 1) == 1 ? BitConverter.SingleToInt32Bits(secondOp[0]) != BitConverter.SingleToInt32Bits(result[0]) : BitConverter.SingleToInt32Bits(firstOp[0]) != BitConverter.SingleToInt32Bits(result[0])", "((BitConverter.SingleToInt32Bits(thirdOp[i]) >> 31) & 1) == 1 ? BitConverter.SingleToInt32Bits(secondOp[i]) != BitConverter.SingleToInt32Bits(result[i]) : BitConverter.SingleToInt32Bits(firstOp[i]) != BitConverter.SingleToInt32Bits(result[i])"}),


Why -0.0E0 and -1.0E0 instead of just -0.0 and 1.0?

tannergooding · 2018-02-09T18:04:58Z

+
+namespace JIT.HardwareIntrinsics.X86
+{
+    public unsafe struct SimpleTernaryOpTest__DataTable<T> : IDisposable where T : struct


I had originally kept the unaligned data tables around because not all the tests had been moved over to use the templates (and therefore couldn't be easily switch to using he aligned data tables).

We might not need/want to add new unaligned data tables and instead should probably just add new properties to the aligned table to get an unaligned address (for the scenarios where we need to explicitly test it).

We might not need/want to add new unaligned data tables and instead should probably just add new properties to the aligned table to get an unaligned address (for the scenarios where we need to explicitly test it).

In which case it might make sense to remove the "aligned" from the name, as it would then just become the "data table"?

as it would then just become the "data table"?

Agree.

Yes. As said above, the primary reason the original ones were kept is because not all tests could be moved to the template quite yet (my PR originally just modified DataTable to support alignment and then later had to be split into two separate tables).

tannergooding

This looks good to me, overall. Just a couple of questions/nits.

CarolEidt

LGTM

CarolEidt · 2018-02-09T18:43:31Z

+
+namespace JIT.HardwareIntrinsics.X86
+{
+    public unsafe struct SimpleTernaryOpTest__DataTable<T> : IDisposable where T : struct


We might not need/want to add new unaligned data tables and instead should probably just add new properties to the aligned table to get an unaligned address (for the scenarios where we need to explicitly test it).

In which case it might make sense to remove the "aligned" from the name, as it would then just become the "data table"?

fiigii · 2018-02-09T23:04:29Z

test Windows_NT x64 Checked jitincompletehwintrinsic
test Windows_NT x64 Checked jitx86hwintrinsicnoavx
test Windows_NT x64 Checked jitx86hwintrinsicnoavx2
test Windows_NT x64 Checked jitx86hwintrinsicnosimd
test Windows_NT x64 Checked jitnox86hwintrinsic

test Windows_NT x86 Checked jitincompletehwintrinsic
test Windows_NT x86 Checked jitx86hwintrinsicnoavx
test Windows_NT x86 Checked jitx86hwintrinsicnoavx2
test Windows_NT x86 Checked jitx86hwintrinsicnosimd
test Windows_NT x86 Checked jitnox86hwintrinsic

test Ubuntu x64 Checked jitincompletehwintrinsic
test Ubuntu x64 Checked jitx86hwintrinsicnoavx
test Ubuntu x64 Checked jitx86hwintrinsicnoavx2
test Ubuntu x64 Checked jitx86hwintrinsicnosimd
test Ubuntu x64 Checked jitnox86hwintrinsic

test OSX10.12 x64 Checked jitincompletehwintrinsic
test OSX10.12 x64 Checked jitx86hwintrinsicnoavx
test OSX10.12 x64 Checked jitx86hwintrinsicnoavx2
test OSX10.12 x64 Checked jitx86hwintrinsicnosimd
test OSX10.12 x64 Checked jitnox86hwintrinsic

fiigii · 2018-02-10T03:36:43Z

test Windows_NT x86 Checked jitx86hwintrinsicnoavx
test Windows_NT x86 Checked jitx86hwintrinsicnosimd

fiigii · 2018-02-10T10:29:39Z

All feedback addressed and all tests passed.

@tannergooding Could you please merge this PR?

fiigii force-pushed the avxsimd branch from d879b51 to d5ad46b Compare February 9, 2018 10:21

tannergooding reviewed Feb 9, 2018

View reviewed changes

tannergooding approved these changes Feb 9, 2018

View reviewed changes

FeiPengIntel added 2 commits February 9, 2018 10:41

Add test cases for some AVX/AVX2 intrinsics

2df4117

Add new test template

5a93c08

CarolEidt approved these changes Feb 9, 2018

View reviewed changes

FeiPengIntel added 2 commits February 9, 2018 15:03

Add AVX/AVX2 templete-drien tests and enable containment tests

4b17916

Implement certain simple SIMD intrinsics for AVX/AVX2

d663ffa

fiigii force-pushed the avxsimd branch from d5ad46b to d663ffa Compare February 9, 2018 23:04

tannergooding merged commit ff510db into dotnet:master Feb 10, 2018

fiigii mentioned this pull request Jan 31, 2020

[RyuJIT] Update the emitter for VEX-encoded SSE4.1/4.2 instructions with containment dotnet/runtime#9697

Closed

Conversation

fiigii commented Feb 8, 2018

Uh oh!

fiigii commented Feb 9, 2018

Uh oh!

fiigii commented Feb 9, 2018

Uh oh!

jkotas commented Feb 9, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jkotas commented Feb 9, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tannergooding left a comment

Choose a reason for hiding this comment

Uh oh!

CarolEidt left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

fiigii commented Feb 9, 2018

Uh oh!

fiigii commented Feb 10, 2018

Uh oh!

fiigii commented Feb 10, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

fiigii commented Feb 10, 2018 •

edited

Loading