Skip to content

[RISCV] Update SpacemiT-X60 vector floating-point instructions latencies #150618

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
mikhailramalho merged 10 commits into from
Nov 25, 2025
Merged

[RISCV] Update SpacemiT-X60 vector floating-point instructions latencies #150618

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
17cd357
[RISCV] Update SpacemiT-X60 vector floating-point instructions
mikhailramalho Jul 25, 2025
a9a540b
Added helper class
mikhailramalho Jul 28, 2025
603a72c
Update ReleaseAtCycles from camel cdr data
mikhailramalho Jul 29, 2025
0e41d47
Merge remote-tracking branch 'origin/main' into x60-rvv-13
mikhailramalho Jul 31, 2025
976ba6a
Merge remote-tracking branch 'origin/main' into x60-rvv-13
mikhailramalho Aug 7, 2025
b206ab3
Update ReleaseAtCycles from camel cdr data
mikhailramalho Aug 7, 2025
b15d15e
Fix vfwmul occ
mikhailramalho Aug 8, 2025
07131e0
Update vfmul
mikhailramalho Nov 18, 2025
b86c593
Merge remote-tracking branch 'origin/main' into x60-rvv-13
mikhailramalho Nov 22, 2025
8ddd296
Update tests
mikhailramalho Nov 22, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
195 changes: 156 additions & 39 deletions llvm/lib/Target/RISCV/RISCVSchedSpacemitX60.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -104,6 +104,11 @@ class Get461018Latency<string mx> {
int c = GetLMULValue<[/*MF8=*/4, /*MF4=*/4, /*MF2=*/4, /*M1=*/4, /*M2=*/6, /*M4=*/10, /*M8=*/18], mx>.c;
}

// Used for: FP FMA operations, complex FP ops
class Get6678Latency<string mx> {
int c = GetLMULValue<[/*MF8=*/6, /*MF4=*/6, /*MF2=*/6, /*M1=*/6, /*M2=*/6, /*M4=*/7, /*M8=*/8], mx>.c;
}

//===----------------------------------------------------------------------===//

class SMX60IsWorstCaseMX<string mx, list<string> MxList> {
Expand Down Expand Up @@ -574,83 +579,169 @@ foreach mx = SchedMxListF in {
foreach sew = SchedSEWSet<mx, isF=1>.val in {
defvar IsWorstCase = SMX60IsWorstCaseMXSEW<mx, sew, SchedMxListF, isF=1>.c;

defm "" : LMULSEWWriteResMXSEW<"WriteVFALUV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFALUF", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFMulV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFMulF", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFMulAddV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFMulAddF", [SMX60_VFP], mx, sew, IsWorstCase>;
}
}
defvar VFALULat = Get4458Latency<mx>.c;
defvar VFALUOcc = ConstOneUntilM1ThenDouble<mx>.c;
let Latency = VFALULat, ReleaseAtCycles = [VFALUOcc] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFALUV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFALUF", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFMinMaxV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFMinMaxF", [SMX60_VFP], mx, sew, IsWorstCase>;
}

foreach mx = SchedMxListF in {
foreach sew = SchedSEWSet<mx, isF=1>.val in {
defvar IsWorstCase = SMX60IsWorstCaseMXSEW<mx, sew, SchedMxListF, isF=1>.c;
// Slightly increased latency for sew == 64
defvar VFMulVLat = !if(!eq(sew, 64), ConstValueUntilLMULThenDoubleBase<"M8", 5, 8, mx>.c,
Get4458Latency<mx>.c);
let Latency = VFMulVLat, ReleaseAtCycles = [ConstOneUntilM1ThenDouble<mx>.c] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFMulV", [SMX60_VFP], mx, sew, IsWorstCase>;
}
// VFMulF has the same latency as VFMulV, but slighlty lower ReleaseAtCycles
let Latency = VFMulVLat, ReleaseAtCycles = [ConstOneUntilM1ThenDouble<mx>.c] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFMulF", [SMX60_VFP], mx, sew, IsWorstCase>;
}

defm "" : LMULSEWWriteResMXSEW<"WriteVFRecpV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFSgnjV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFSgnjF", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFMinMaxV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFMinMaxF", [SMX60_VFP], mx, sew, IsWorstCase>;
defvar VFSgnjLat = ConstValueUntilLMULThenDouble<"M2", 4, mx>.c;
defvar VFSgnjOcc = ConstOneUntilMF2ThenDouble<mx>.c;
let Latency = VFSgnjLat, ReleaseAtCycles = [VFSgnjOcc] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFRecpV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFSgnjV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFSgnjF", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFCvtIToFV", [SMX60_VFP], mx, sew, IsWorstCase>;
}

defm "" : LMULSEWWriteResMXSEW<"WriteVFCvtIToFV", [SMX60_VFP], mx, sew, IsWorstCase>;
// The following covers vfmacc, vfmsac, and their vfn* variants in the same group, but the
// ReleaseAtCycles takes one extra cycle for the vfn* variants.
// TODO: Should we split them?
// TODO: for some reason, the following cond is not working, and always use ConstValueUntilLMULThenDoubleBase<"M4", 5, 8, mx>.c
Copy link
Member

@mshockwave mshockwave Oct 28, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is this TODO still valid?

Copy link
Member Author

@mikhailramalho mikhailramalho Nov 22, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Unfortunately, yes. You can see in rvv-fp.s. For sew == 64, it should be picking 6678, but it stays in 5558 like the other condition.

I've tried to debug this some months ago, but got nowhere. The diff is small tho.

defvar VFMulAddLatency = !if(!eq(sew, 64),
Get6678Latency<mx>.c,
ConstValueUntilLMULThenDoubleBase<"M8", 5, 8, mx>.c
);
let Latency = VFMulAddLatency, ReleaseAtCycles = [ConstOneUntilM1ThenDouble<mx>.c] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFMulAddV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFMulAddF", [SMX60_VFP], mx, sew, IsWorstCase>;
}
}
}

foreach mx = SchedMxList in {
defvar IsWorstCase = SMX60IsWorstCaseMX<mx, SchedMxList>.c;

defm "" : LMULWriteResMX<"WriteVFCmpV", [SMX60_VFP], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVFCmpF", [SMX60_VFP], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVFClassV", [SMX60_VFP], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVFMergeV", [SMX60_VFP], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVFMovV", [SMX60_VFP], mx, IsWorstCase>;
// Slightly increased ReleaseAtCycles for M8: 18
defvar VFCmpOcc = !if(!eq(mx, "M8"),
!add(ConstOneUntilMF2ThenDouble<mx>.c, 2),
ConstOneUntilMF2ThenDouble<mx>.c
);
let Latency = Get461018Latency<mx>.c, ReleaseAtCycles = [VFCmpOcc] in {
defm "" : LMULWriteResMX<"WriteVFCmpV", [SMX60_VFP], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVFCmpF", [SMX60_VFP], mx, IsWorstCase>;
}

defm "" : LMULWriteResMX<"WriteVFCvtFToIV", [SMX60_VFP], mx, IsWorstCase>;
defvar VFClassLat = ConstValueUntilLMULThenDouble<"M2", 4, mx>.c;
defvar VFClassOcc = ConstOneUntilMF2ThenDouble<mx>.c;
let Latency = VFClassLat, ReleaseAtCycles = [VFClassOcc] in {
defm "" : LMULWriteResMX<"WriteVFClassV", [SMX60_VFP], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVFMergeV", [SMX60_VFP], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVFMovV", [SMX60_VFP], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVFCvtFToIV", [SMX60_VFP], mx, IsWorstCase>;
}
}

// Widening
foreach mx = SchedMxListW in {
foreach sew = SchedSEWSet<mx, isF=0, isWidening=1>.val in {
defvar IsWorstCase = SMX60IsWorstCaseMXSEW<mx, sew, SchedMxListW>.c;

defm "" : LMULSEWWriteResMXSEW<"WriteVFWCvtIToFV", [SMX60_VFP], mx, sew, IsWorstCase>;
defvar VFWCvtILat = ConstValueUntilLMULThenDouble<"M1", 4, mx>.c;
defvar VFWCvtIOcc = ConstOneUntilMF4ThenDouble<mx>.c;
let Latency = VFWCvtILat, ReleaseAtCycles = [VFWCvtIOcc] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFWCvtIToFV", [SMX60_VFP], mx, sew, IsWorstCase>;
}
}
}

foreach mx = SchedMxListFW in {
defvar IsWorstCase = SMX60IsWorstCaseMX<mx, SchedMxListFW>.c;

defm "" : LMULWriteResMX<"WriteVFWCvtFToIV", [SMX60_VFP], mx, IsWorstCase>;
defvar VFWCvtFToIVLat = ConstValueUntilLMULThenDouble<"M1", 4, mx>.c;
defvar VFWCvtFToIVOcc = ConstOneUntilMF4ThenDouble<mx>.c;
let Latency = VFWCvtFToIVLat, ReleaseAtCycles = [VFWCvtFToIVOcc] in {
defm "" : LMULWriteResMX<"WriteVFWCvtFToIV", [SMX60_VFP], mx, IsWorstCase>;
}
}

foreach mx = SchedMxListFW in {
foreach sew = SchedSEWSet<mx, isF=1, isWidening=1>.val in {
defvar IsWorstCase = SMX60IsWorstCaseMXSEW<mx, sew, SchedMxListFW, isF=1>.c;

defm "" : LMULSEWWriteResMXSEW<"WriteVFWALUV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFWALUF", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFWMulV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFWMulF", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFWMulAddV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFWMulAddF", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFWCvtFToFV", [SMX60_VFP], mx, sew, IsWorstCase>;
defvar VFWCvtFToFVLat = ConstValueUntilLMULThenDouble<"M1", 4, mx>.c;
defvar VFWCvtFToFVOcc = ConstOneUntilMF4ThenDouble<mx>.c;
let Latency = VFWCvtFToFVLat, ReleaseAtCycles = [VFWCvtFToFVOcc] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFWCvtFToFV", [SMX60_VFP], mx, sew, IsWorstCase>;
}

// Latency for vfwsub/vfwadd.vv, vfwsub/vfwadd.vf: 4/4/4/5/8
// ReleaseAtCycles for vfwsub/vfwadd.vv, vfwsub/vfwadd.vf: 1/1/2/4/8
// Latency for vfwsub/vfwadd.wv, vfwsub/vfwadd.wf: 5/5/5/9/17
// ReleaseAtCycles for vfwsub/vfwadd.wv, vfwsub/vfwadd.wf: 1/2/4/8/17
// We use the worst-case
defvar VFWALULat = !add(ConstValueUntilLMULThenDouble<"M1", 4, mx>.c, 1); // 5/5/9/17
defvar VFWALUOcc = !if(!eq(mx, "M4"),
!add(ConstOneUntilMF4ThenDouble<mx>.c, 1), // 2/4/8/17
ConstOneUntilMF4ThenDouble<mx>.c
);
// TODO: Split .wf/.wv variants into separate scheduling classes
let Latency = VFWALULat, ReleaseAtCycles = [VFWALUOcc] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFWALUV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFWALUF", [SMX60_VFP], mx, sew, IsWorstCase>;
}

let Latency = Get4588Latency<mx>.c, ReleaseAtCycles = [ConstOneUntilMF2ThenDouble<mx>.c] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFWMulF", [SMX60_VFP], mx, sew, IsWorstCase>;
}

// Slightly increased latency for SEW == 32
defvar VFWMullOcc = !if(!eq(sew, 32),
GetLMULValue<[1, 1, 1, 3, 5, 9, 18], mx>.c,
ConstOneUntilMF2ThenDouble<mx>.c
);
defvar VFWMulVLat = ConstValueUntilLMULThenDoubleBase<"M8", 5, 8, mx>.c;
let Latency = VFWMulVLat, ReleaseAtCycles = [VFWMullOcc] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFWMulV", [SMX60_VFP], mx, sew, IsWorstCase>;
}

// Latency for vfwmacc, vfwnmacc, etc: e16 = 5/5/5/8; e32 = 6/6/7/8
defvar VFWMulAddVLat = !if(!eq(sew, 16),
ConstValueUntilLMULThenDoubleBase<"M4", 5, 8, mx>.c,
Get6678Latency<mx>.c
);
let Latency = VFWMulAddVLat, ReleaseAtCycles = [ConstOneUntilMF2ThenDouble<mx>.c] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFWMulAddV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFWMulAddF", [SMX60_VFP], mx, sew, IsWorstCase>;
}
}
}

// Narrowing
foreach mx = SchedMxListW in {
defvar IsWorstCase = SMX60IsWorstCaseMX<mx, SchedMxListW>.c;

defm "" : LMULWriteResMX<"WriteVFNCvtFToIV", [SMX60_VFP], mx, IsWorstCase>;
defvar VFNCvtFToIVLat = ConstValueUntilLMULThenDouble<"M1", 4, mx>.c;
defvar VFNCvtFToIVOcc = ConstOneUntilMF4ThenDouble<mx>.c;
let Latency = VFNCvtFToIVLat, ReleaseAtCycles = [VFNCvtFToIVOcc] in {
defm "" : LMULWriteResMX<"WriteVFNCvtFToIV", [SMX60_VFP], mx, IsWorstCase>;
}
}

foreach mx = SchedMxListFW in {
foreach sew = SchedSEWSet<mx, isF=1, isWidening=1>.val in {

defvar IsWorstCase = SMX60IsWorstCaseMXSEW<mx, sew, SchedMxListFW, isF=1>.c;
defm "" : LMULSEWWriteResMXSEW<"WriteVFNCvtIToFV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFNCvtFToFV", [SMX60_VFP], mx, sew, IsWorstCase>;

defvar VFNCvtToFVLat = ConstValueUntilLMULThenDouble<"M1", 4, mx>.c;
defvar VFNCvtToFVOcc = ConstOneUntilMF4ThenDouble<mx>.c;
let Latency = VFNCvtToFVLat, ReleaseAtCycles = [VFNCvtToFVOcc] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFNCvtIToFV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFNCvtFToFV", [SMX60_VFP], mx, sew, IsWorstCase>;
}
}
}

Expand All @@ -659,9 +750,35 @@ foreach mx = SchedMxListF in {
foreach sew = SchedSEWSet<mx, 1>.val in {
defvar IsWorstCase = SMX60IsWorstCaseMXSEW<mx, sew, SchedMxListF, 1>.c;

defm "" : LMULSEWWriteResMXSEW<"WriteVFDivV", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFDivF", [SMX60_VFP], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVFSqrtV", [SMX60_VFP], mx, sew, IsWorstCase>;
// Compute ReleaseAtCycles based on SEW
// Latency for vfdiv.vf: e16/e32 = 12/24/48/96; e64 = 18/36/72/144
// Latency for vfrdiv.vf: e16/e32 = 12/24/48/96; e64 = 40/80/160/320
// We use the worst-case, vfdiv.vf is penalized in e64
// TODO: split vfdiv.vf and vfrdiv.vf into separate scheduling classes
defvar VFDivFFactor = !if(!eq(sew, 64), 40, 12);
defvar VFDivFLatAndOcc = !mul(ConstOneUntilM1ThenDouble<mx>.c, VFDivFFactor);
let Latency = VFDivFLatAndOcc, ReleaseAtCycles = [VFDivFLatAndOcc] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFDivF", [SMX60_VFP], mx, sew, IsWorstCase>;
}

defvar VFDivVFactor = !if(!eq(sew, 16), 12, 40);
defvar VFDivVLatAndOcc = !mul(ConstOneUntilM1ThenDouble<mx>.c, VFDivVFactor);
let Latency = VFDivVLatAndOcc, ReleaseAtCycles = [VFDivVLatAndOcc] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFDivV", [SMX60_VFP], mx, sew, IsWorstCase>;
}
}
}

// Pattern for vfsqrt.v: e16 = 18/36/72/144; e32 = 38/76/152/304; e64 = 40/80/160/320
foreach mx = SchedMxListF in {
foreach sew = SchedSEWSet<mx, 1>.val in {
defvar IsWorstCase = SMX60IsWorstCaseMXSEW<mx, sew, SchedMxListF, 1>.c;

defvar VFSqrtVFactor = !if(!eq(sew, 16), 12, 40);
defvar VFSqrtVLatAndOcc = !mul(ConstOneUntilM1ThenDouble<mx>.c, VFSqrtVFactor);
let Latency = VFSqrtVLatAndOcc, ReleaseAtCycles = [VFSqrtVLatAndOcc] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFSqrtV", [SMX60_VFP], mx, sew, IsWorstCase>;
}
}
}

Expand Down
Loading