[llvm] [RISCV][NFC] Factor out VLEN in the SiFive7 scheduling model (PR #143629)
Min-Yih Hsu via llvm-commits
llvm-commits at lists.llvm.org
Tue Jun 10 16:31:37 PDT 2025
https://github.com/mshockwave updated https://github.com/llvm/llvm-project/pull/143629
>From ceea6131f5cafbca35a2a2bf7a452bb881083c38 Mon Sep 17 00:00:00 2001
From: Min-Yih Hsu <min.hsu at sifive.com>
Date: Tue, 10 Jun 2025 16:09:11 -0700
Subject: [PATCH] [RISCV][NFC] Factor out VLEN in the SiFive7 scheduling model
Co-Authored-By: Michael Maitland <michaeltmaitland at gmail.com>
---
llvm/lib/Target/RISCV/RISCVSchedSiFive7.td | 64 ++++++++++------------
1 file changed, 30 insertions(+), 34 deletions(-)
diff --git a/llvm/lib/Target/RISCV/RISCVSchedSiFive7.td b/llvm/lib/Target/RISCV/RISCVSchedSiFive7.td
index af64a871a9292..c1d7cd4a716e7 100644
--- a/llvm/lib/Target/RISCV/RISCVSchedSiFive7.td
+++ b/llvm/lib/Target/RISCV/RISCVSchedSiFive7.td
@@ -88,9 +88,8 @@ class SiFive7GetCyclesSegmentedSeg2<string mx> {
// Cycles for segmented loads and stores are calculated using the
// formula vl * ceil((SEW * nf) / DLEN), where SEW * nf is the segment size.
-class SiFive7GetCyclesSegmented<string mx, int sew, int nf> {
- defvar VLEN = 512;
- defvar DLEN = 256;
+class SiFive7GetCyclesSegmented<string mx, int sew, int nf, int VLEN> {
+ defvar DLEN = !div(VLEN, 2);
// (VLEN * LMUL) / SEW
defvar VLUpperBound = !cond(
!eq(mx, "M1") : !div(VLEN, sew),
@@ -107,23 +106,20 @@ class SiFive7GetCyclesSegmented<string mx, int sew, int nf> {
int c = !mul(VLUpperBound, !div(!sub(!add(a, b), 1), b));
}
-class SiFive7GetCyclesOnePerElement<string mx, int sew> {
- // FIXME: On SiFive7, VLEN is 512. Although a user can request the compiler
- // to use a different VLEN, this model will not make scheduling decisions
- // based on the user specified VLEN.
+class SiFive7GetCyclesOnePerElement<string mx, int sew, int VLEN> {
// c = ceil(VLEN / SEW) * LMUL
// Note: c >= 1 since the smallest VLEN is 512 / 8 = 8, and the
// largest division performed on VLEN is in MF8 case with division
// by 8. Therefore, there is no need to ceil the result.
- int VLEN = !div(512, sew);
+ int numElements = !div(VLEN, sew);
int c = !cond(
- !eq(mx, "M1") : VLEN,
- !eq(mx, "M2") : !mul(VLEN, 2),
- !eq(mx, "M4") : !mul(VLEN, 4),
- !eq(mx, "M8") : !mul(VLEN, 8),
- !eq(mx, "MF2") : !div(VLEN, 2),
- !eq(mx, "MF4") : !div(VLEN, 4),
- !eq(mx, "MF8") : !div(VLEN, 8)
+ !eq(mx, "M1") : numElements,
+ !eq(mx, "M2") : !mul(numElements, 2),
+ !eq(mx, "M4") : !mul(numElements, 4),
+ !eq(mx, "M8") : !mul(numElements, 8),
+ !eq(mx, "MF2") : !div(numElements, 2),
+ !eq(mx, "MF4") : !div(numElements, 4),
+ !eq(mx, "MF8") : !div(numElements, 8)
);
}
@@ -139,10 +135,9 @@ class SiFive7GetDivOrSqrtFactor<int sew> {
/// Cycles for reductions take approximately VL*SEW/DLEN + 5(4 + log(DLEN/SEW))
/// cycles.
-class SiFive7GetReductionCycles<string mx, int sew> {
+class SiFive7GetReductionCycles<string mx, int sew, int VLEN> {
// VLUpperBound*SEW/DLEN is equivalent to 2*LMUL since
// VLUpperBound=(VLEN*LMUL)/SEW.
- defvar VLEN = 512;
defvar DLEN = !div(VLEN, 2);
defvar TwoTimesLMUL = !cond(
!eq(mx, "M1") : 2,
@@ -160,8 +155,7 @@ class SiFive7GetReductionCycles<string mx, int sew> {
}
/// Cycles for ordered reductions take approximately 6*VL cycles
-class SiFive7GetOrderedReductionCycles<string mx, int sew> {
- defvar VLEN = 512;
+class SiFive7GetOrderedReductionCycles<string mx, int sew, int VLEN> {
// (VLEN * LMUL) / SEW
defvar VLUpperBound = !cond(
!eq(mx, "M1") : !div(VLEN, sew),
@@ -234,6 +228,8 @@ def SiFive7VCQ : ProcResource<1>; // Vector Command Queue
def SiFive7PipeAB : ProcResGroup<[SiFive7PipeA, SiFive7PipeB]>;
+defvar SiFive7VLEN = 512;
+
// Branching
let Latency = 3 in {
def : WriteRes<WriteJmp, [SiFive7PipeB]>;
@@ -481,7 +477,7 @@ foreach mx = SchedMxList in {
foreach mx = SchedMxList in {
defvar VLDSX0Cycles = SiFive7GetCyclesDefault<mx>.c;
- defvar Cycles = SiFive7GetCyclesOnePerElement<mx, 8>.c;
+ defvar Cycles = SiFive7GetCyclesOnePerElement<mx, 8, SiFive7VLEN>.c;
defvar IsWorstCase = SiFive7IsWorstCaseMX<mx, SchedMxList>.c;
defm SiFive7 : LMULWriteResMXVariant<"WriteVLDS8", VLDSX0Pred, [SiFive7VCQ, SiFive7VL],
4, [0, 1], [1, !add(1, VLDSX0Cycles)], !add(3, Cycles),
@@ -501,7 +497,7 @@ foreach mx = SchedMxList in {
// since LMUL >= 16/64.
foreach mx = ["MF4", "MF2", "M1", "M2", "M4", "M8"] in {
defvar VLDSX0Cycles = SiFive7GetCyclesDefault<mx>.c;
- defvar Cycles = SiFive7GetCyclesOnePerElement<mx, 16>.c;
+ defvar Cycles = SiFive7GetCyclesOnePerElement<mx, 16, SiFive7VLEN>.c;
defvar IsWorstCase = SiFive7IsWorstCaseMX<mx, SchedMxList>.c;
defm SiFive7 : LMULWriteResMXVariant<"WriteVLDS16", VLDSX0Pred, [SiFive7VCQ, SiFive7VL],
4, [0, 1], [1, !add(1, VLDSX0Cycles)], !add(3, Cycles),
@@ -518,7 +514,7 @@ foreach mx = ["MF4", "MF2", "M1", "M2", "M4", "M8"] in {
}
foreach mx = ["MF2", "M1", "M2", "M4", "M8"] in {
defvar VLDSX0Cycles = SiFive7GetCyclesDefault<mx>.c;
- defvar Cycles = SiFive7GetCyclesOnePerElement<mx, 32>.c;
+ defvar Cycles = SiFive7GetCyclesOnePerElement<mx, 32, SiFive7VLEN>.c;
defvar IsWorstCase = SiFive7IsWorstCaseMX<mx, SchedMxList>.c;
defm SiFive7 : LMULWriteResMXVariant<"WriteVLDS32", VLDSX0Pred, [SiFive7VCQ, SiFive7VL],
4, [0, 1], [1, !add(1, VLDSX0Cycles)], !add(3, Cycles),
@@ -535,7 +531,7 @@ foreach mx = ["MF2", "M1", "M2", "M4", "M8"] in {
}
foreach mx = ["M1", "M2", "M4", "M8"] in {
defvar VLDSX0Cycles = SiFive7GetCyclesDefault<mx>.c;
- defvar Cycles = SiFive7GetCyclesOnePerElement<mx, 64>.c;
+ defvar Cycles = SiFive7GetCyclesOnePerElement<mx, 64, SiFive7VLEN>.c;
defvar IsWorstCase = SiFive7IsWorstCaseMX<mx, SchedMxList>.c;
defm SiFive7 : LMULWriteResMXVariant<"WriteVLDS64", VLDSX0Pred, [SiFive7VCQ, SiFive7VL],
4, [0, 1], [1, !add(1, VLDSX0Cycles)], !add(3, Cycles),
@@ -588,7 +584,7 @@ foreach mx = SchedMxList in {
let Latency = 1, AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, Cycles)] in
defm "" : LMULWriteResMX<"WriteVSSEG2e" # eew, [SiFive7VCQ, SiFive7VS], mx, IsWorstCase>;
foreach nf=3-8 in {
- defvar Cycles = SiFive7GetCyclesSegmented<mx, eew, nf>.c;
+ defvar Cycles = SiFive7GetCyclesSegmented<mx, eew, nf, SiFive7VLEN>.c;
defvar IsWorstCase = SiFive7IsWorstCaseMX<mx, SchedMxList>.c;
// Does not chain so set latency high
let Latency = !add(3, Cycles), AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, Cycles)] in {
@@ -603,7 +599,7 @@ foreach mx = SchedMxList in {
foreach mx = SchedMxList in {
foreach nf=2-8 in {
foreach eew = [8, 16, 32, 64] in {
- defvar Cycles = SiFive7GetCyclesSegmented<mx, eew, nf>.c;
+ defvar Cycles = SiFive7GetCyclesSegmented<mx, eew, nf, SiFive7VLEN>.c;
defvar IsWorstCase = SiFive7IsWorstCaseMX<mx, SchedMxList>.c;
// Does not chain so set latency high
let Latency = !add(3, Cycles), AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, Cycles)] in {
@@ -669,7 +665,7 @@ foreach mx = SchedMxList in {
foreach mx = SchedMxList in {
foreach sew = SchedSEWSet<mx>.val in {
defvar Cycles = !mul(SiFive7GetDivOrSqrtFactor<sew>.c,
- !div(SiFive7GetCyclesOnePerElement<mx, sew>.c, 4));
+ !div(SiFive7GetCyclesOnePerElement<mx, sew, SiFive7VLEN>.c, 4));
defvar IsWorstCase = SiFive7IsWorstCaseMXSEW<mx, sew, SchedMxList>.c;
let Latency = Cycles, AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, Cycles)] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVIDivV", [SiFive7VCQ, SiFive7VA], mx, sew, IsWorstCase>;
@@ -774,7 +770,7 @@ foreach mx = SchedMxList in {
foreach mx = SchedMxListF in {
foreach sew = SchedSEWSet<mx, isF=1>.val in {
defvar Cycles = !mul(SiFive7GetDivOrSqrtFactor<sew>.c,
- !div(SiFive7GetCyclesOnePerElement<mx, sew>.c, 4));
+ !div(SiFive7GetCyclesOnePerElement<mx, sew, SiFive7VLEN>.c, 4));
defvar IsWorstCase = SiFive7IsWorstCaseMXSEW<mx, sew, SchedMxListF, 1>.c;
let Latency = Cycles, AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, Cycles)] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFSqrtV", [SiFive7VCQ, SiFive7VA], mx, sew, IsWorstCase>;
@@ -834,7 +830,7 @@ foreach mx = SchedMxListFW in {
// 14. Vector Reduction Operations
foreach mx = SchedMxList in {
foreach sew = SchedSEWSet<mx>.val in {
- defvar Cycles = SiFive7GetReductionCycles<mx, sew>.c;
+ defvar Cycles = SiFive7GetReductionCycles<mx, sew, SiFive7VLEN>.c;
defvar IsWorstCase = SiFive7IsWorstCaseMXSEW<mx, sew, SchedMxList>.c;
let Latency = Cycles, AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, Cycles)] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVIRedV_From", [SiFive7VCQ, SiFive7VA],
@@ -847,7 +843,7 @@ foreach mx = SchedMxList in {
foreach mx = SchedMxListWRed in {
foreach sew = SchedSEWSet<mx, 0, 1>.val in {
- defvar Cycles = SiFive7GetReductionCycles<mx, sew>.c;
+ defvar Cycles = SiFive7GetReductionCycles<mx, sew, SiFive7VLEN>.c;
defvar IsWorstCase = SiFive7IsWorstCaseMXSEW<mx, sew, SchedMxListWRed>.c;
let Latency = Cycles, AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, Cycles)] in
defm "" : LMULSEWWriteResMXSEW<"WriteVIWRedV_From", [SiFive7VCQ, SiFive7VA],
@@ -857,7 +853,7 @@ foreach mx = SchedMxListWRed in {
foreach mx = SchedMxListF in {
foreach sew = SchedSEWSet<mx, 1>.val in {
- defvar RedCycles = SiFive7GetReductionCycles<mx, sew>.c;
+ defvar RedCycles = SiFive7GetReductionCycles<mx, sew, SiFive7VLEN>.c;
defvar IsWorstCase = SiFive7IsWorstCaseMXSEW<mx, sew, SchedMxListF, 1>.c;
let Latency = RedCycles, AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, RedCycles)] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFRedV_From", [SiFive7VCQ, SiFive7VA],
@@ -865,7 +861,7 @@ foreach mx = SchedMxListF in {
defm "" : LMULSEWWriteResMXSEW<"WriteVFRedMinMaxV_From", [SiFive7VCQ, SiFive7VA],
mx, sew, IsWorstCase>;
}
- defvar OrdRedCycles = SiFive7GetOrderedReductionCycles<mx, sew>.c;
+ defvar OrdRedCycles = SiFive7GetOrderedReductionCycles<mx, sew, SiFive7VLEN>.c;
let Latency = OrdRedCycles, AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, OrdRedCycles)] in
defm "" : LMULSEWWriteResMXSEW<"WriteVFRedOV_From", [SiFive7VCQ, SiFive7VA],
mx, sew, IsWorstCase>;
@@ -874,12 +870,12 @@ foreach mx = SchedMxListF in {
foreach mx = SchedMxListFWRed in {
foreach sew = SchedSEWSet<mx, 1, 1>.val in {
- defvar RedCycles = SiFive7GetReductionCycles<mx, sew>.c;
+ defvar RedCycles = SiFive7GetReductionCycles<mx, sew, SiFive7VLEN>.c;
defvar IsWorstCase = SiFive7IsWorstCaseMXSEW<mx, sew, SchedMxListFWRed, 1>.c;
let Latency = RedCycles, AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, RedCycles)] in
defm "" : LMULSEWWriteResMXSEW<"WriteVFWRedV_From", [SiFive7VCQ, SiFive7VA],
mx, sew, IsWorstCase>;
- defvar OrdRedCycles = SiFive7GetOrderedReductionCycles<mx, sew>.c;
+ defvar OrdRedCycles = SiFive7GetOrderedReductionCycles<mx, sew, SiFive7VLEN>.c;
let Latency = OrdRedCycles, AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, OrdRedCycles)] in
defm "" : LMULSEWWriteResMXSEW<"WriteVFWRedOV_From", [SiFive7VCQ, SiFive7VA],
mx, sew, IsWorstCase>;
@@ -924,7 +920,7 @@ foreach mx = SchedMxList in {
foreach mx = SchedMxList in {
foreach sew = SchedSEWSet<mx>.val in {
- defvar Cycles = SiFive7GetCyclesOnePerElement<mx, sew>.c;
+ defvar Cycles = SiFive7GetCyclesOnePerElement<mx, sew, SiFive7VLEN>.c;
defvar IsWorstCase = SiFive7IsWorstCaseMXSEW<mx, sew, SchedMxList>.c;
let Latency = !add(Cycles, 3), AcquireAtCycles = [0, 1], ReleaseAtCycles = [1, !add(1, Cycles)] in {
defm "" : LMULSEWWriteResMXSEW<"WriteVRGatherVV", [SiFive7VCQ, SiFive7VA], mx, sew, IsWorstCase>;
More information about the llvm-commits
mailing list