FMA codegen

These two functions are doing the same thing. The only difference is the order of parameters of Fma.MultiplyAdd. However, the codegen of the second function is not as good as the first one. Maybe worth looking at this to see if the JIT can optimize it.

static unsafe void foo2(float scale, float[] src, float[] dst, int count)
{
    fixed (float* psrc = &src[0])
    fixed (float* pdst = &dst[0])
    {
        float* pSrcCurrent = psrc;
        float* pDstCurrent = pdst;
        float* pEnd = pdst + count;

        Vector256<float> scaleVector256 = Avx.SetAllVector256(scale);

        while (pDstCurrent + 8 <= pEnd)
        {
            Vector256<float> dstVector = Avx.LoadVector256(pDstCurrent);
            dstVector = Fma.MultiplyAdd(Avx.LoadVector256(pSrcCurrent), scaleVector256, dstVector);
            Avx.Store(pDstCurrent, dstVector);

            pSrcCurrent += 8;
            pDstCurrent += 8;
        }
    }
}

static unsafe void foo3(float scale, float[] src, float[] dst, int count)
{
    fixed (float* psrc = &src[0])
    fixed (float* pdst = &dst[0])
    {
        float* pSrcCurrent = psrc;
        float* pDstCurrent = pdst;
        float* pEnd = pdst + count;

        Vector256<float> scaleVector256 = Avx.SetAllVector256(scale);

        while (pDstCurrent + 8 <= pEnd)
        {
            Vector256<float> dstVector = Avx.LoadVector256(pDstCurrent);
            dstVector = Fma.MultiplyAdd(scaleVector256, Avx.LoadVector256(pSrcCurrent), dstVector);
            Avx.Store(pDstCurrent, dstVector);

            pSrcCurrent += 8;
            pDstCurrent += 8;
        }
    }
}

Codegen of foo2:

Codegen of foo3:

category:cq
theme:register-allocator
skill-level:expert
cost:medium

@fiigii

cc\ @fiigii @tannergooding

@CarolEidt

CC. @CarolEidt, it looks like there is some additional register shuffling going on.

it is unnecessarily shuffling ymm0 to ymm2 before the call and ymm2 to ymm1 after the call.

The first unnecessarily shuffling (ymm0 to ymm2) seems from emitter

void emitter::emitIns_SIMD_R_R_R_AR(
    instruction ins, emitAttr attr, regNumber targetReg, regNumber op1Reg, regNumber op2Reg, regNumber base)
{
    assert(IsFMAInstruction(ins));
    assert(UseVEXEncoding());

    if (op1Reg != targetReg)
    {
        // Ensure we aren't overwriting op2
        assert(op2Reg != targetReg);

        emitIns_R_R(INS_movaps, attr, targetReg, op1Reg);
    }

    emitIns_R_R_AR(ins, attr, targetReg, op2Reg, base, 0);
}

and the second one may be from delayfree?

@tannergooding

The current FMA instruction selection/RA is not optimal, but we need more scenarios to improve it.
@tannergooding As I know, you are going to use HW intrinsic for some math lib, which would be a great opportunity.

@fiigii

@fiigii, I don't think the issue is with the emitter, but rather the register allocator deciding that the source and destination cannot be the same here. Maybe I missed something in LSRA or in the containment checks for one of the FMA cases?

At least as of right now, we try to ensure at least one of the operands is contained, preferencing 3, then 2, then 1: https://github.com/dotnet/coreclr/blob/master/src/jit/lowerxarch.cpp#L3006

There is an existing issue with "reg-optional" in that we don't have a way to say that any of the three operand can be optional. Tracked by [LSRA][RyuJIT] Add support to allow either of op1 or op2 to be considered as reg optional rather than Lower fixing one of the operands as reg optional #6358

We then try to ensure that the correct delay free uses are setup in LSRA: https://github.com/dotnet/coreclr/blob/master/src/jit/lsraxarch.cpp#L2516

And finally just emit the appropriate encoding (based on which are is contained) in codegen: https://github.com/dotnet/coreclr/blob/master/src/jit/hwintrinsiccodegenxarch.cpp#L2410

@CarolEidt

Perhaps a possible fix here is telling the register allocator which operand ends up being the RMW operand.... @CarolEidt, do we have a good way to do that (is just setting tgtPref the right thing to do here?)

This is really more of a register allocation (and Lowering) issue. I've reclassified this as register-allocator theme, but it may be that it has been addressed already.

Another example I recently encountered:

unsafe static void MulKernel2x8(double* C,double* A,double* B,uint A_cols)
{
    Vector256<double> sum00, sum01, sum10, sum11;

    sum00 = sum01 = sum10 = sum11 = default;

    for (uint k = 0; k < A_cols; k += 1) {
        Vector256<double> va, vb0, vb1;
        vb0 = Avx.LoadVector256(B);
        vb1 = Avx.LoadVector256(B + 4);
        va = Avx.BroadcastScalarToVector256(A);
        sum00 = Fma.MultiplyAdd(va,vb0,sum00);
        sum01 = Fma.MultiplyAdd(va,vb1,sum01);
        va = Avx.BroadcastScalarToVector256(A + 1);
        sum10 = Fma.MultiplyAdd(va,vb0,sum10);
        sum11 = Fma.MultiplyAdd(va,vb1,sum11);
        A += 2;
        B += 8;
    }
    Avx.Store(C,sum00);
    Avx.Store(C + 4,sum01);
    Avx.Store(C + 8,sum10);
    Avx.Store(C + 12,sum11);
}

The compiled loop body:

vmovupd     ymm4,ymmword ptr [r8]  
vmovupd     ymm5,ymmword ptr [r8+20h]
vbroadcastsd ymm6,mmword ptr [rdx]  
vmovaps     ymm7,ymm6  
vfmadd213pd ymm7,ymm4,ymm3  
vmovaps     ymm3,ymm7  
vfmadd213pd ymm6,ymm5,ymm2  
vmovaps     ymm2,ymm6  
vbroadcastsd ymm6,mmword ptr [rdx+8] 
vfmadd213pd ymm4,ymm6,ymm1  
vmovaps     ymm1,ymm4  
vfmadd213pd ymm5,ymm6,ymm0  
vmovaps     ymm0,ymm5  
add         rdx,10h  
add         r8,40h

Ideal ouput:

vmovupd     ymm4,ymmword ptr [r8]  
vmovupd     ymm5,ymmword ptr [r8+20h]
vbroadcastsd ymm6,mmword ptr [rdx]  
vfmadd231pd ymm3,ymm6,ymm4  
vfmadd231pd ymm2,ymm6,ymm5  
vbroadcastsd ymm6,mmword ptr [rdx+8] 
vfmadd231pd ymm1,ymm6,ymm4  
vfmadd231pd ymm0,ymm6,ymm5  
add         rdx,10h  
add         r8,40h

It seems vfmadd231pd is hardly selected when both 3 operands are not from memory. I think this "mul and sum" pattern should be recognized.
BTW, why didn't we simply add Fma.MultiplyAdd231, etc?

This should be at least be analyzed; it may have been fixed e.g. by dotnet/coreclr/pull/19429

Moving to 7.0.0 since it's not a bug.

@BruceForstall

Un-assigning myself
cc @BruceForstall

This has been mostly resolved and we now generate identical code in most cases. There is still a possibility of an extra movaps in some cases, but that's a more general register allocator consideration and no longer specific to FMA

Navigation Menu

Repository navigation

FMA codegen #11260

1 remaining item

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

FMA codegen #11260

Description

Activity

helloguo commented on Oct 18, 2018

tannergooding commented on Oct 18, 2018

tannergooding commented on Oct 18, 2018

fiigii commented on Oct 18, 2018

fiigii commented on Oct 18, 2018

tannergooding commented on Oct 18, 2018

tannergooding commented on Oct 18, 2018

tannergooding commented on Oct 18, 2018

CarolEidt commented on Oct 12, 2020

kevinjwz commented on Oct 21, 2020

CarolEidt commented on Nov 14, 2020

1 remaining item

echesakov commented on Jul 26, 2021

echesakov commented on Mar 15, 2022

tannergooding commented on Feb 5, 2024

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions