Skip to content

[RISC-V] Update SpacemiT-X60 Vector Integer latencies #149207

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
merged 9 commits into from
Jul 24, 2025
Merged

[RISC-V] Update SpacemiT-X60 Vector Integer latencies #149207

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
18c2b1f
Added Vector Integer Instruction latencies
mikhailramalho Jul 16, 2025
d99a4b9
Added missing ReleaseAtCycles
mikhailramalho Jul 16, 2025
5cf6357
Updated helper classes
mikhailramalho Jul 17, 2025
c575ceb
Fixed typo
mikhailramalho Jul 17, 2025
8b9bd54
Merge remote-tracking branch 'origin/main' into x60-rvv-vi
mikhailramalho Jul 18, 2025
c62f52d
Changed latencies
mikhailramalho Jul 18, 2025
7e6e3d1
Address comments from reviewers
mikhailramalho Jul 22, 2025
b538a64
Renamed variables
mikhailramalho Jul 22, 2025
18c9edf
Swap Latency <-> ReleaseAtCycles
mikhailramalho Jul 23, 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
200 changes: 154 additions & 46 deletions llvm/lib/Target/RISCV/RISCVSchedSpacemitX60.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,67 @@ class SMX60IsWorstCaseMXSEW<string mx, int sew, list<string> MxList, bit isF = 0
bit c = !and(!eq(mx, LLMUL), !eq(sew, SSEW));
}

defvar SMX60VLEN = 256;
defvar SMX60DLEN = !div(SMX60VLEN, 2);

class Get1248Latency<string mx> {
int c = !cond(
!eq(mx, "M2") : 2,
!eq(mx, "M4") : 4,
!eq(mx, "M8") : 8,
true: 1
);
}

// Used for: logical opsz, shifts, sign ext, merge/move, FP sign/recip/convert, mask ops, slides
class Get4816Latency<string mx> {
int c = !cond(
!eq(mx, "M4") : 8,
!eq(mx, "M8") : 16,
true: 4
);
}

// Used for: arithmetic (add/sub/min/max), saturating/averaging, FP add/sub/min/max
class Get458Latency<string mx> {
int c = !cond(
!eq(mx, "M4") : 5,
!eq(mx, "M8") : 8,
true: 4
);
}

// Widening scaling pattern (4,4,4,4,5,8,8): plateaus at higher LMULs
// Used for: widening operations
class Get4588Latency<string mx> {
int c = !cond(
!eq(mx, "M2") : 5,
!eq(mx, "M4") : 8,
!eq(mx, "M8") : 8, // M8 not supported for most widening, fallback
true: 4
);
}

// Used for: mask-producing comparisons, carry ops with mask, FP comparisons
class Get461018Latency<string mx> {
int c = !cond(
!eq(mx, "M2") : 6,
!eq(mx, "M4") : 10,
!eq(mx, "M8") : 18,
true: 4
);
}

// Used for: e64 multiply pattern, complex ops
class Get781632Latency<string mx> {
int c = !cond(
!eq(mx, "M2") : 8,
!eq(mx, "M4") : 16,
!eq(mx, "M8") : 32,
true: 7
);
}

def SpacemitX60Model : SchedMachineModel {
let IssueWidth = 2; // dual-issue
let MicroOpBufferSize = 0; // in-order
Expand Down Expand Up @@ -322,71 +383,118 @@ foreach LMul = [1, 2, 4, 8] in {
foreach mx = SchedMxList in {
defvar IsWorstCase = SMX60IsWorstCaseMX<mx, SchedMxList>.c;

defm "" : LMULWriteResMX<"WriteVIALUV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIALUX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIALUI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVExtV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICALUV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICALUX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICALUI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICALUMV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICALUMX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICALUMI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICmpV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICmpX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICmpI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMinMaxV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMinMaxX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMergeV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMergeX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMergeI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMovV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMovX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMovI", [SMX60_VIEU], mx, IsWorstCase>;

defm "" : LMULWriteResMX<"WriteVShiftV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVShiftX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVShiftI", [SMX60_VIEU], mx, IsWorstCase>;

defm "" : LMULWriteResMX<"WriteVIMulV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMulX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMulAddV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMulAddX", [SMX60_VIEU], mx, IsWorstCase>;
let Latency = Get458Latency<mx>.c, ReleaseAtCycles = [4] in {
defm "" : LMULWriteResMX<"WriteVIMinMaxV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMinMaxX", [SMX60_VIEU], mx, IsWorstCase>;
}

let Latency = Get4816Latency<mx>.c, ReleaseAtCycles = [4] in {
// Pattern of vadd, vsub, vrsub: 4/4/5/8
// Pattern of vand, vor, vxor: 4/4/8/16
// They are grouped together, so we used the worst case 4/4/8/16
// TODO: use InstRW to override individual instructions' scheduling data
defm "" : LMULWriteResMX<"WriteVIALUV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIALUX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIALUI", [SMX60_VIEU], mx, IsWorstCase>;

defm "" : LMULWriteResMX<"WriteVExtV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMergeV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMergeX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMergeI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMovV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMovX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMovI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVShiftV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVShiftX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVShiftI", [SMX60_VIEU], mx, IsWorstCase>;

defm "" : LMULWriteResMX<"WriteVICALUV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICALUX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICALUI", [SMX60_VIEU], mx, IsWorstCase>;
}

let Latency = Get461018Latency<mx>.c, ReleaseAtCycles = [4] in {
defm "" : LMULWriteResMX<"WriteVICALUMV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICALUMX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICALUMI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICmpV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICmpX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVICmpI", [SMX60_VIEU], mx, IsWorstCase>;
}

// Pattern of vmacc, vmadd, vmul, vmulh, etc.: e8/e16 = 4/4/5/8, e32 = 5,5,5,8,
// e64 = 7,8,16,32. We use the worst-case until we can split the SEW.
// TODO: change WriteVIMulV, etc to be defined with LMULSEWSchedWrites
let Latency = Get781632Latency<mx>.c, ReleaseAtCycles = [7] in {
defm "" : LMULWriteResMX<"WriteVIMulV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMulX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMulAddV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIMulAddX", [SMX60_VIEU], mx, IsWorstCase>;
}
}

// Widening
// Pattern of vwmul, vwmacc, etc: e8/e16 = 4/4/5/8, e32 = 5,5,5,8
// We use the worst-case for all.
foreach mx = SchedMxListW in {
defvar IsWorstCase = SMX60IsWorstCaseMX<mx, SchedMxListW>.c;

defm "" : LMULWriteResMX<"WriteVIWALUV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWALUX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWALUI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWMulV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWMulX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWMulAddV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWMulAddX", [SMX60_VIEU], mx, IsWorstCase>;
let Latency = Get4588Latency<mx>.c, ReleaseAtCycles = [4] in {
defm "" : LMULWriteResMX<"WriteVIWALUV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWALUX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWALUI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWMulV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWMulX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWMulAddV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVIWMulAddX", [SMX60_VIEU], mx, IsWorstCase>;
}
}

// Vector Integer Division and Remainder
// Division and remainder operations
// Pattern of vdivu: 11/11/11/20/40/80/160
// Pattern of vdiv: 12/12/12/22/44/88/176
// Pattern of vremu: 12/12/12/22/44/88/176
// Pattern of vrem: 13/13/13/24/48/96/192
// We use for all: 12/12/12/24/48/96/192
// TODO: Create separate WriteVIRem to more closely match the latencies
foreach mx = SchedMxList in {
foreach sew = SchedSEWSet<mx>.val in {
defvar IsWorstCase = SMX60IsWorstCaseMXSEW<mx, sew, SchedMxList>.c;

defm "" : LMULSEWWriteResMXSEW<"WriteVIDivV", [SMX60_VIEU], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVIDivX", [SMX60_VIEU], mx, sew, IsWorstCase>;
// Slightly reduced for fractional LMULs
defvar Multiplier = !cond(
!eq(mx, "MF8") : 12,
!eq(mx, "MF4") : 12,
!eq(mx, "MF2") : 12,
true: 24
);

let Latency = !mul(Get1248Latency<mx>.c, Multiplier), ReleaseAtCycles = [12] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVIDivV", [SMX60_VIEU], mx, sew, IsWorstCase>;
defm "" : LMULSEWWriteResMXSEW<"WriteVIDivX", [SMX60_VIEU], mx, sew, IsWorstCase>;
}
}
}

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

defm "" : LMULWriteResMX<"WriteVNShiftV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVNShiftX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVNShiftI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVNClipV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVNClipX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVNClipI", [SMX60_VIEU], mx, IsWorstCase>;
// Slightly increased for integer LMULs
defvar Multiplier = !cond(
!eq(mx, "M2") : 2,
!eq(mx, "M4") : 2,
true: 1
);

let Latency = !mul(Get4816Latency<mx>.c, Multiplier), ReleaseAtCycles = [4] in {
defm "" : LMULWriteResMX<"WriteVNShiftV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVNShiftX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVNShiftI", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVNClipV", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVNClipX", [SMX60_VIEU], mx, IsWorstCase>;
defm "" : LMULWriteResMX<"WriteVNClipI", [SMX60_VIEU], mx, IsWorstCase>;
}
}

// 12. Vector Fixed-Point Arithmetic Instructions
Expand Down
Loading