Skip to content

Commit 234a24f

Browse files
leonling-llleonling-ll
committed
CleanUp
1 parent 853c41f commit 234a24f

File tree

7 files changed

+29
-40
lines changed

7 files changed

+29
-40
lines changed

‎python/src/ir.cc‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1622,7 +1622,7 @@ void init_triton_ir(py::module &&m) {
16221622
if (haveDump) {
16231623
auto printingFlags = OpPrintingFlags();
16241624
printingFlags.elideLargeElementsAttrs(16);
1625-
// printingFlags.enableDebugInfo();
1625+
printingFlags.enableDebugInfo();
16261626
auto printAlways = [funcToDump](Pass *, Operation *op) -> bool {
16271627
if (funcToDump.empty())
16281628
return true;

‎third_party/intel/lib/Dialect/TritonIntelGPU/IR/Dialect.cpp‎

Lines changed: 0 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,5 @@
11
#include "triton/Dialect/Triton/IR/Dialect.h"
22

3-
#include <iostream>
43
#include <numeric>
54

65
#include "intel/include/Dialect/TritonIntelGPU/IR/LinearLayoutConversions.h"
@@ -153,8 +152,6 @@ DpasEncodingAttr::getShapePerCTATile(ArrayRef<int64_t> tensorShape) const {
153152
auto shapeC = getShapeC();
154153
SmallVector<unsigned> warpsPerCTA = getWarpsPerCTA();
155154
size_t rank = shapeC.size();
156-
assert(rank == shapeC.size() &&
157-
"ShapeC and WarpsPerCTA must have the same rank");
158155
SmallVector<unsigned> shapePerCTATile(rank);
159156
for (size_t i = 0; i < rank; ++i) {
160157
shapePerCTATile[i] = shapeC[i] * warpsPerCTA[i];
@@ -307,7 +304,6 @@ SmallVector<unsigned>
307304
DpasEncodingAttr::getShapePerCTATileForDotOperands(ArrayRef<int64_t> shape,
308305
int opIdx) const {
309306
auto parentShapePerCTATile = getShapePerCTATile(shape);
310-
// auto threadsPerWarp = getThreadsPerWarp();
311307
size_t rank = parentShapePerCTATile.size();
312308
if (opIdx == 0) {
313309
auto shapeA = getShapeA();

‎third_party/intel/lib/Dialect/TritonIntelGPU/IR/LinearLayoutConversions.cpp‎

Lines changed: 28 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -508,7 +508,7 @@ LinearLayout DPAStoLinearLayout(ArrayRef<int64_t> shape, Attribute layout,
508508
int systolicDepth = dpas.getSystolicDepth();
509509
int repeatCount = dpas.getRepeatCount();
510510
int executionSize = dpas.getExecutionSize();
511-
unsigned dimK, dimNonK;
511+
unsigned KDim, nonKDim;
512512
if (opIdx == 0) { // Operand A
513513
auto regBasesA = DPASRegBasesA(opsPerChannel, repeatCount, threadsPerWarp,
514514
systolicDepth);
@@ -517,16 +517,16 @@ LinearLayout DPAStoLinearLayout(ArrayRef<int64_t> shape, Attribute layout,
517517
tileLayout = LinearLayout({{kRegister, regBasesA}, {kLane, laneBasesA}},
518518
ArrayRef(outDimNames).take_back(2));
519519
// A only repeats by repCluster[rank - 2]
520-
dimNonK = rank - 2;
521-
dimK = rank - 1;
522-
tileLayout *= LinearLayout::identity1D(repCluster[dimNonK], kRegister,
523-
outDimNames[dimNonK]);
520+
nonKDim = rank - 2;
521+
KDim = rank - 1;
522+
tileLayout *= LinearLayout::identity1D(repCluster[nonKDim], kRegister,
523+
outDimNames[nonKDim]);
524524

525525
// K-dimension is shared among warps
526526
tileLayout *=
527-
LinearLayout::zeros1D(warpsPerCTA[dimK], kWarp, outDimNames[dimK]);
528-
tileLayout *= LinearLayout::identity1D(warpsPerCTA[dimNonK], kWarp,
529-
outDimNames[dimNonK]);
527+
LinearLayout::zeros1D(warpsPerCTA[KDim], kWarp, outDimNames[KDim]);
528+
tileLayout *= LinearLayout::identity1D(warpsPerCTA[nonKDim], kWarp,
529+
outDimNames[nonKDim]);
530530
if (rank == 3)
531531
tileLayout *=
532532
LinearLayout::identity1D(warpsPerCTA[0], kWarp, outDimNames[0]);
@@ -539,16 +539,16 @@ LinearLayout DPAStoLinearLayout(ArrayRef<int64_t> shape, Attribute layout,
539539
tileLayout = LinearLayout({{kRegister, regBasesB}, {kLane, laneBasesB}},
540540
ArrayRef(outDimNames).take_back(2));
541541
// B only repeats by repCluster[rank - 1]
542-
dimNonK = rank - 1;
543-
dimK = rank - 2;
544-
tileLayout *= LinearLayout::identity1D(repCluster[dimNonK], kRegister,
545-
outDimNames[dimNonK]);
542+
nonKDim = rank - 1;
543+
KDim = rank - 2;
544+
tileLayout *= LinearLayout::identity1D(repCluster[nonKDim], kRegister,
545+
outDimNames[nonKDim]);
546546

547547
// K-dimension is shared among warps
548-
tileLayout *= LinearLayout::identity1D(warpsPerCTA[dimNonK], kWarp,
549-
outDimNames[dimNonK]);
548+
tileLayout *= LinearLayout::identity1D(warpsPerCTA[nonKDim], kWarp,
549+
outDimNames[nonKDim]);
550550
tileLayout *=
551-
LinearLayout::zeros1D(warpsPerCTA[dimK], kWarp, outDimNames[dimK]);
551+
LinearLayout::zeros1D(warpsPerCTA[KDim], kWarp, outDimNames[KDim]);
552552
if (rank == 3)
553553
tileLayout *=
554554
LinearLayout::identity1D(warpsPerCTA[0], kWarp, outDimNames[0]);
@@ -561,18 +561,18 @@ LinearLayout DPAStoLinearLayout(ArrayRef<int64_t> shape, Attribute layout,
561561
// The per-inst layout is repeated at each repCluster.
562562
// Hence, multiply with the identity layouts starting from the
563563
// least significant dimension.
564-
dimNonK = rank - 2;
565-
dimK = rank - 1;
566-
tileLayout *= LinearLayout::identity1D(repCluster[dimK], kRegister,
567-
outDimNames[dimK]);
568-
tileLayout *= LinearLayout::identity1D(repCluster[dimNonK], kRegister,
569-
outDimNames[dimNonK]);
564+
nonKDim = rank - 2;
565+
KDim = rank - 1;
566+
tileLayout *= LinearLayout::identity1D(repCluster[KDim], kRegister,
567+
outDimNames[KDim]);
568+
tileLayout *= LinearLayout::identity1D(repCluster[nonKDim], kRegister,
569+
outDimNames[nonKDim]);
570570

571571
// // The identical layout is repeated among warps
572572
tileLayout *=
573-
LinearLayout::identity1D(warpsPerCTA[dimK], kWarp, outDimNames[dimK]);
574-
tileLayout *= LinearLayout::identity1D(warpsPerCTA[dimNonK], kWarp,
575-
outDimNames[dimNonK]);
573+
LinearLayout::identity1D(warpsPerCTA[KDim], kWarp, outDimNames[KDim]);
574+
tileLayout *= LinearLayout::identity1D(warpsPerCTA[nonKDim], kWarp,
575+
outDimNames[nonKDim]);
576576
if (rank == 3)
577577
tileLayout *=
578578
LinearLayout::identity1D(warpsPerCTA[0], kWarp, outDimNames[0]);
@@ -584,12 +584,12 @@ LinearLayout DPAStoLinearLayout(ArrayRef<int64_t> shape, Attribute layout,
584584
SmallVector<int64_t> numReps = dpas.getDPASRepetitions(shape, opIdx);
585585

586586
// numReps is always 3D, we should add 1 to dim id when rank is 2
587-
int repDimK = rank == 2 ? dimK + 1 : dimK;
588-
int repDimNonK = rank == 2 ? dimNonK + 1 : dimNonK;
587+
int repDimK = rank == 2 ? KDim + 1 : KDim;
588+
int repDimNonK = rank == 2 ? nonKDim + 1 : nonKDim;
589589
tileLayout *=
590-
LinearLayout::identity1D(numReps[repDimK], kRegister, outDimNames[dimK]);
590+
LinearLayout::identity1D(numReps[repDimK], kRegister, outDimNames[KDim]);
591591
tileLayout *= LinearLayout::identity1D(numReps[repDimNonK], kRegister,
592-
outDimNames[dimNonK]);
592+
outDimNames[nonKDim]);
593593
if (rank == 3)
594594
tileLayout *=
595595
LinearLayout::identity1D(numReps[0], kRegister, outDimNames[0]);

‎third_party/intel/lib/TritonIntelGPUToLLVM/ConvertLayoutOpToLLVM.cpp‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,6 @@
11
#include "PatternTritonGPUOpToLLVM.h"
22
#include "TargetInfo.h"
33
#include "Utility.h"
4-
#include <iostream>
54

65
#include "intel/include/Analysis/Utility.h"
76
#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h"

‎third_party/intel/lib/TritonIntelGPUToLLVM/ConvertLayoutOpToLLVM/SharedToDotOperandDPAS.cpp‎

Lines changed: 0 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,8 +1,6 @@
11
#include "../TritonGPUToLLVMBase.h"
22
#include "../Utility.h"
33
#include "mlir/Dialect/LLVMIR/LLVMTypes.h"
4-
#include "mlir/Support/LLVM.h"
5-
#include "triton/Dialect/TritonGPU/IR/Dialect.h"
64
#include "llvm/Support/ErrorHandling.h"
75

86
using ValueTable = std::map<std::array<int, 3>, Value>;

‎third_party/intel/lib/TritonIntelGPUToLLVM/LoadStoreOpToLLVM.cpp‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -349,7 +349,6 @@ struct PrefetchOpConversion
349349
Type eltTy = tensorType.getElementType();
350350
const ArrayRef<int64_t> shapeRef = tensorType.getShape();
351351
SmallVector<int64_t> tensorShape{shapeRef.begin(), shapeRef.end()};
352-
assert(tensorShape.size() == 2 && "Only 2D tensors are prefetch supported");
353352

354353
if (!memoryRowMajor) {
355354
// Swap the shape to make it row major and then get the tiling

‎third_party/intel/lib/TritonIntelGPUToLLVM/Utility.h‎

Lines changed: 0 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -16,7 +16,6 @@
1616
#include "triton/Conversion/TritonGPUToLLVM/Utility.h"
1717
#include "triton/Dialect/Triton/IR/Utility.h"
1818
#include "llvm/Support/ErrorHandling.h"
19-
#include <iostream>
2019

2120
#define DEBUG_TYPE "ttgpu_to_llvm"
2221

@@ -656,9 +655,7 @@ inline DenseMap<unsigned, Value> getSwizzledSharedPtrs(
656655
// Order
657656
auto inOrder = triton::gpu::getOrder(srcEncoding);
658657
auto outOrder = triton::gpu::getOrder(resSharedLayout);
659-
unsigned rank = outOrder.size();
660658
assert(maxPhase == 1 ||
661-
// outVec * maxPhase <= srcShape[outOrder[rank-2]] &&
662659
outVec * maxPhase <= srcShape[outOrder[0]] &&
663660
"Swizzling would generate out of bounds memory accesses");
664661
// Tensor indices held by the current thread, as LLVM values

0 commit comments

Comments
 (0)