Updating genCodeForBinary to be VEX aware by tannergooding · Pull Request #1344 · dotnet/runtime

tannergooding · 2020-01-07T01:20:12Z

This is a rough draft attempt at resolving #1342

This essentially just moves some of the genHWIntrinsic_R_R_RM logic into a inst_RV_RV_TT method and updates genCodeForBinary to call it for floating-point types if FEATURE_HW_INTRINSICS is defined.

As can be seen below, the numbers are overall fairly promising.

Summary of Code Size diffs:
(Lower is better)
Total bytes of diff: -3270 (-0.01% of base)
diff is an improvement.
Top file regressions (bytes):
39 : System.Drawing.Primitives.dasm (0.10% of base)
Top file improvements (bytes):
-2496 : System.Private.CoreLib.dasm (-0.05% of base)
-266 : System.Runtime.Numerics.dasm (-0.37% of base)
-178 : System.Linq.Parallel.dasm (-0.01% of base)
-113 : System.Data.Common.dasm (-0.01% of base)
-108 : Microsoft.Diagnostics.Tracing.TraceEvent.dasm (-0.00% of base)
-64 : System.Linq.dasm (-0.01% of base)
-40 : System.Private.Xml.dasm (-0.00% of base)
-16 : Microsoft.CodeAnalysis.VisualBasic.dasm (-0.00% of base)
-16 : Newtonsoft.Json.dasm (-0.00% of base)
-8 : Microsoft.CodeAnalysis.dasm (-0.00% of base)
-4 : Microsoft.CSharp.dasm (-0.00% of base)
12 total files with Code Size differences (11 improved, 1 regressed), 97 unchanged.
Top method regressions (bytes):
25 ( 6.60% of base) : System.Drawing.Primitives.dasm - RectangleF:Intersect(RectangleF,RectangleF):RectangleF
22 ( 3.72% of base) : Microsoft.Diagnostics.Tracing.TraceEvent.dasm - ThreadTimeStackComputer:AddUnkownAsyncDurationIfNeeded(StartStopActivity,double,TraceEvent):this
22 ( 7.61% of base) : System.Drawing.Primitives.dasm - RectangleF:Union(RectangleF,RectangleF):RectangleF
16 ( 2.86% of base) : Microsoft.Diagnostics.Tracing.TraceEvent.dasm - SampleProfilerThreadTimeComputer:AddUnkownAsyncDurationIfNeeded(StartStopActivity,double,TraceEvent):this
15 ( 1.86% of base) : Microsoft.CodeAnalysis.VisualBasic.dasm - OverloadResolution:FoldFloatingBinaryOperator(int,ConstantValue,ConstantValue,TypeSymbol,TypeSymbol):ConstantValue
12 ( 0.68% of base) : System.Private.CoreLib.dasm - DateTimeParse:ParseFormatO(ReadOnlySpan`1,byref):bool
12 ( 5.04% of base) : System.Private.Xml.dasm - XsdDateTime:.ctor(DateTimeOffset,int):this
4 ( 1.01% of base) : Microsoft.Diagnostics.Tracing.TraceEvent.dasm - StartStopActivityComputer:GetCurrentStartStopActivityStack(MutableTraceEventStackSource,TraceThread,TraceThread,bool):int:this
Top method improvements (bytes):
-384 (-13.55% of base) : System.Private.CoreLib.dasm - CalendricalCalculationsHelper:SumLongSequenceOfPeriodicTerms(double):double
-338 (-13.56% of base) : System.Private.CoreLib.dasm - Matrix4x4:Invert(Matrix4x4,byref):bool
-134 (-15.06% of base) : System.Private.CoreLib.dasm - Matrix4x4:Transform(Matrix4x4,Quaternion):Matrix4x4
-88 (-9.24% of base) : System.Runtime.Numerics.dasm - Complex:Asin_Internal(double,double,byref,byref,byref)
-65 (-13.32% of base) : System.Private.CoreLib.dasm - Matrix4x4:CreateFromQuaternion(Quaternion):Matrix4x4
-60 (-11.21% of base) : System.Private.CoreLib.dasm - Matrix4x4:CreateFromAxisAngle(Vector3,float):Matrix4x4
-59 (-10.42% of base) : System.Private.CoreLib.dasm - Quaternion:Slerp(Quaternion,Quaternion,float):Quaternion
-57 (-9.79% of base) : System.Private.CoreLib.dasm - Quaternion:Lerp(Quaternion,Quaternion,float):Quaternion
-52 (-10.00% of base) : System.Private.CoreLib.dasm - Matrix4x4:GetDeterminant():float:this
-52 (-11.98% of base) : System.Private.CoreLib.dasm - Quaternion:Divide(Quaternion,Quaternion):Quaternion
-52 (-11.98% of base) : System.Private.CoreLib.dasm - Quaternion:op_Division(Quaternion,Quaternion):Quaternion
-50 (-11.63% of base) : System.Private.CoreLib.dasm - Vector4:Transform(Vector3,Quaternion):Vector4
-48 (-13.75% of base) : System.Private.CoreLib.dasm - Plane:Transform(Plane,Matrix4x4):Plane
-48 (-16.55% of base) : System.Private.CoreLib.dasm - Vector4:Transform(Vector4,Matrix4x4):Vector4
-48 (-14.55% of base) : System.Runtime.Numerics.dasm - Complex:Tan(Complex):Complex
-47 (-10.63% of base) : System.Private.CoreLib.dasm - Plane:Transform(Plane,Quaternion):Plane
-43 (-10.17% of base) : System.Private.CoreLib.dasm - Vector3:Transform(Vector3,Quaternion):Vector3
-43 (-9.60% of base) : System.Private.CoreLib.dasm - Vector4:Transform(Vector4,Quaternion):Vector4
-40 (-7.84% of base) : System.Private.CoreLib.dasm - Matrix4x4:CreateShadow(Vector3,Plane):Matrix4x4
-40 (-8.62% of base) : System.Private.CoreLib.dasm - Matrix4x4:CreateReflection(Plane):Matrix4x4
Top method regressions (percentages):
22 ( 7.61% of base) : System.Drawing.Primitives.dasm - RectangleF:Union(RectangleF,RectangleF):RectangleF
25 ( 6.60% of base) : System.Drawing.Primitives.dasm - RectangleF:Intersect(RectangleF,RectangleF):RectangleF
12 ( 5.04% of base) : System.Private.Xml.dasm - XsdDateTime:.ctor(DateTimeOffset,int):this
22 ( 3.72% of base) : Microsoft.Diagnostics.Tracing.TraceEvent.dasm - ThreadTimeStackComputer:AddUnkownAsyncDurationIfNeeded(StartStopActivity,double,TraceEvent):this
16 ( 2.86% of base) : Microsoft.Diagnostics.Tracing.TraceEvent.dasm - SampleProfilerThreadTimeComputer:AddUnkownAsyncDurationIfNeeded(StartStopActivity,double,TraceEvent):this
15 ( 1.86% of base) : Microsoft.CodeAnalysis.VisualBasic.dasm - OverloadResolution:FoldFloatingBinaryOperator(int,ConstantValue,ConstantValue,TypeSymbol,TypeSymbol):ConstantValue
4 ( 1.01% of base) : Microsoft.Diagnostics.Tracing.TraceEvent.dasm - StartStopActivityComputer:GetCurrentStartStopActivityStack(MutableTraceEventStackSource,TraceThread,TraceThread,bool):int:this
12 ( 0.68% of base) : System.Private.CoreLib.dasm - DateTimeParse:ParseFormatO(ReadOnlySpan`1,byref):bool
Top method improvements (percentages):
-12 (-21.05% of base) : System.Private.CoreLib.dasm - Quaternion:Multiply(Quaternion,float):Quaternion
-12 (-21.05% of base) : System.Private.CoreLib.dasm - Quaternion:op_Multiply(Quaternion,float):Quaternion
-20 (-17.09% of base) : System.Private.CoreLib.dasm - Matrix3x2:Multiply(Matrix3x2,float):Matrix3x2
-20 (-17.09% of base) : System.Private.CoreLib.dasm - Matrix3x2:op_Multiply(Matrix3x2,float):Matrix3x2
-48 (-16.55% of base) : System.Private.CoreLib.dasm - Vector4:Transform(Vector4,Matrix4x4):Vector4
-16 (-15.53% of base) : System.Private.Xml.dasm - NumericExpr:GetValue(int,double,double):double
-4 (-15.38% of base) : Newtonsoft.Json.dasm - JsonValidatingReader:FloatingPointRemainder(double,double):double
-4 (-15.38% of base) : System.Private.CoreLib.dasm - CalendricalCalculationsHelper:Reminder(double,double):double
-134 (-15.06% of base) : System.Private.CoreLib.dasm - Matrix4x4:Transform(Matrix4x4,Quaternion):Matrix4x4
-36 (-14.69% of base) : System.Private.CoreLib.dasm - Vector4:Transform(Vector3,Matrix4x4):Vector4
-48 (-14.55% of base) : System.Runtime.Numerics.dasm - Complex:Tan(Complex):Complex
-24 (-14.46% of base) : System.Private.CoreLib.dasm - Vector3:TransformNormal(Vector3,Matrix4x4):Vector3
-8 (-14.04% of base) : System.Private.CoreLib.dasm - CalendricalCalculationsHelper:AsLocalTime(double,double):double
-4 (-13.79% of base) : System.Runtime.Numerics.dasm - Complex:Scale(Complex,double):Complex
-48 (-13.75% of base) : System.Private.CoreLib.dasm - Plane:Transform(Plane,Matrix4x4):Plane
-338 (-13.56% of base) : System.Private.CoreLib.dasm - Matrix4x4:Invert(Matrix4x4,byref):bool
-384 (-13.55% of base) : System.Private.CoreLib.dasm - CalendricalCalculationsHelper:SumLongSequenceOfPeriodicTerms(double):double
-24 (-13.48% of base) : System.Private.CoreLib.dasm - Vector4:Transform(Vector2,Matrix4x4):Vector4
-65 (-13.32% of base) : System.Private.CoreLib.dasm - Matrix4x4:CreateFromQuaternion(Quaternion):Matrix4x4
-16 (-13.22% of base) : System.Private.CoreLib.dasm - Quaternion:Normalize(Quaternion):Quaternion
171 total methods with Code Size differences (163 improved, 8 regressed), 201419 unchanged.

tannergooding · 2020-01-07T01:20:36Z

CC. @CarolEidt

tannergooding · 2020-01-07T01:22:58Z

The majority of diffs are positive changes, such as:

- vdivsd xmm0, qword ptr [reloc @RWD00]- vmovsd qword ptr [rbp-48H], xmm0+ vdivsd xmm6, xmm0, qword ptr [reloc @RWD00]

However, their are a few cases (such as DateTimeParse:ParseFormatO) where the epilogues (and there are multiple) have an additional instruction such as vmovaps xmm6, qword ptr [rsp+50H].

CarolEidt

This looks OK overall, but I'm not comfortable with continuing to have fundamental support for 3-operand encodings under FEATURE_HW_INTRINSICS. I think that's something that should be cleaned up.

tannergooding · 2020-01-07T19:50:35Z

This looks OK overall, but I'm not comfortable with continuing to have fundamental support for 3-operand encodings under FEATURE_HW_INTRINSICS

Right, and I called that out above. I'm working on updating the function to work without FEATURE_HW_INTRINSIC.

tannergooding · 2020-01-07T21:41:53Z

@@ -1336,6 +1336,7 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
 #if defined(_TARGET_XARCH_)
 void inst_RV_RV_IV(instruction ins, emitAttr size, regNumber reg1, regNumber reg2, unsigned ival);
 void inst_RV_TT_IV(instruction ins, emitAttr attr, regNumber reg1, GenTree* rmOp, int ival);


This was conditioned under FEATURE_HW_INTRINSICS for the implementation, but not the declaration here. So, I did the minimal fixup to allow it to be available without `FEATURE_HW_INTRINSICS.

CarolEidt

Thanks!

tannergooding · 2020-01-07T21:45:55Z

 regNumber op1reg = op1->isUsedFromReg() ? op1->GetRegNum() : REG_NA;
 regNumber op2reg = op2->isUsedFromReg() ? op2->GetRegNum() : REG_NA;

+ if (varTypeIsFloating(treeNode->TypeGet()))


This just hijacks all the floating-point types and forwards to inst_RV_RV_TT, the emit helpers (such as emitIns_SIMD_R_R_R) deal with the difference between VEX and non-VEX for dst and op1Reg.

Great - I think that's the right approach.

tannergooding · 2020-01-07T21:46:52Z

 }
 }

+#ifdef FEATURE_HW_INTRINSICS


I only made the minimum number of emitIns_SIMD_* methods needed available outside FEATURE_HW_INTRINSICS as the others can't be encountered during normal codegen (and likely never will be).

tannergooding · 2020-01-07T21:48:39Z

CC. @dotnet/jit-contrib

mikedn · 2020-01-07T22:50:41Z

+// isRMW -- true if the instruction is RMW; otherwise, false
+//
+void CodeGen::inst_RV_RV_TT(
+ instruction ins, emitAttr size, regNumber targetReg, regNumber op1Reg, GenTree* op2, bool isRMW)


isRMW should come before targetReg & operands. It's already pretty dubious that such information has to be communicated separately from instructions.

I put this at the end to keep parameter ordering with other overloads consistent and because this should likely be looked up a different way eventually (either based on the instruction or the node)

tannergooding · 2020-01-10T21:14:11Z

Any other feedback here? If not, I think it should be good to merge once CI passes

CarolEidt

LGTM - thanks for eliminating the FEATURE_HW_INTRINSICS dependencies.

CarolEidt · 2020-01-10T22:37:33Z

 regNumber op1reg = op1->isUsedFromReg() ? op1->GetRegNum() : REG_NA;
 regNumber op2reg = op2->isUsedFromReg() ? op2->GetRegNum() : REG_NA;

+ if (varTypeIsFloating(treeNode->TypeGet()))


Great - I think that's the right approach.

tannergooding · 2020-08-28T19:42:58Z

SIMD.ConsoleMandel.ScalarFloatSinglethreadADT directly benefited from this PR and is ~7.3x faster because of it:

sandreenko · 2020-08-28T20:01:46Z

SIMD.ConsoleMandel.ScalarFloatSinglethreadADT directly benefited from this PR and is ~7.3x faster because of it.

not sure what you mean and why it is important now but on the left part the main problem was that these ADT were not put in registers. This problem was fixed by #37745.

tannergooding · 2020-08-28T21:02:00Z

Hmm, I'm seeing 4.4s in 3.1 and 0.6s in 5.0; with it being 2.1s if just this PR is reverted; the additional moves inserted from the ops being interpreted as RMW were making a big difference in the pipelining of the code (at least from what I was able to discern from uProf).

sandreenko · 2020-08-28T21:51:00Z

Hmm, I'm seeing 4.4s in 3.1 and 0.6s in 5.0; with it being 2.1s if just this PR is reverted; the additional moves inserted from the ops being interpreted as RMW were making a big difference in the pipelining of the code (at least from what I was able to discern from uProf).

yeah, maybe it will be 3x time slower without this PR, I am just saying that these diffs that you have shown are not all caused by this PR and this PR, applied to 3.1 alone, won't produce any diffs on this test.

tannergooding added the NO-MERGE The PR is not ready for merge yet (see discussion for detailed reasons) label Jan 7, 2020

This comment has been minimized.
Sign in to view

CarolEidt reviewed Jan 7, 2020
View reviewed changes

Comment threadsrc/coreclr/src/jit/hwintrinsiccodegenxarch.cpp Outdated

Updating genCodeForBinary to be VEX aware
628ce55

tannergooding force-pushed the float-rmw branch from 25026ae to 628ce55Compare January 7, 2020 21:40

tannergooding commented Jan 7, 2020
View reviewed changes

CarolEidt approved these changes Jan 7, 2020
View reviewed changes

tannergooding commented Jan 7, 2020
View reviewed changes

tannergooding changed the title ~~[WIP] Updating genCodeForBinary to be VEX aware~~Updating genCodeForBinary to be VEX awareJan 7, 2020

tannergooding removed the NO-MERGE The PR is not ready for merge yet (see discussion for detailed reasons) label Jan 7, 2020

mikedn reviewed Jan 7, 2020
View reviewed changes

Comment threadsrc/coreclr/src/jit/instr.cpp Outdated

mikedn reviewed Jan 7, 2020
View reviewed changes

Comment threadsrc/coreclr/src/jit/instr.cpp Outdated

mikedn reviewed Jan 7, 2020
View reviewed changes

Comment threadsrc/coreclr/src/jit/instr.cpp Outdated

Resolving some PR feedback
dedbaf5

tannergooding merged commit ef27a17 into dotnet:masterJan 10, 2020

CarolEidt reviewed Jan 10, 2020
View reviewed changes

ghost locked as resolved and limited conversation to collaborators Dec 11, 2020

Uh oh!

Conversation

tannergooding commented Jan 7, 2020

Uh oh!

tannergooding commented Jan 7, 2020

Uh oh!

This comment has been minimized.

tannergooding commented Jan 7, 2020

Uh oh!

CarolEidt left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

tannergooding commented Jan 7, 2020

Uh oh!

tannergoodingJan 7, 2020

Choose a reason for hiding this comment

Uh oh!

CarolEidt left a comment

Choose a reason for hiding this comment

Uh oh!

tannergoodingJan 7, 2020

Choose a reason for hiding this comment

Uh oh!

CarolEidtJan 10, 2020

Choose a reason for hiding this comment

Uh oh!

tannergoodingJan 7, 2020

Choose a reason for hiding this comment

Uh oh!

tannergooding commented Jan 7, 2020

Uh oh!

Uh oh!

mikednJan 7, 2020

Choose a reason for hiding this comment

Uh oh!

tannergoodingJan 8, 2020

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

tannergooding commented Jan 10, 2020

Uh oh!

CarolEidt left a comment

Choose a reason for hiding this comment

Uh oh!

CarolEidtJan 10, 2020

Choose a reason for hiding this comment

Uh oh!

tannergooding commented Aug 28, 2020

Uh oh!

sandreenko commented Aug 28, 2020

Uh oh!

tannergooding commented Aug 28, 2020

Uh oh!

sandreenko commented Aug 28, 2020

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants