Skip to content

Commit d1c9203

Browse files
Muzammiluddin-Syed-ECEMuzammiluddin-Syed-ECE
committed
removing unnecessary files
Signed-off-by: Muzammiluddin Syed <[email protected]>
1 parent 0040779 commit d1c9203

File tree

2 files changed

+1
-624
lines changed

2 files changed

+1
-624
lines changed

‎mlir/lib/Dialect/GPU/Transforms/SubgroupReduceLowering.cpp‎

Lines changed: 1 addition & 161 deletions
Original file line numberDiff line numberDiff line change
@@ -11,12 +11,10 @@
1111
//===----------------------------------------------------------------------===//
1212

1313
#include "mlir/Dialect/Arith/IR/Arith.h"
14-
#include "mlir/Dialect/LLVMIR/ROCDLDialect.h"
1514
#include "mlir/Dialect/GPU/IR/GPUDialect.h"
1615
#include "mlir/Dialect/GPU/Transforms/Passes.h"
1716
#include "mlir/Dialect/GPU/Utils/GPUUtils.h"
1817
#include "mlir/Dialect/Vector/IR/VectorOps.h"
19-
#include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h"
2018
#include "mlir/IR/BuiltinTypes.h"
2119
#include "mlir/IR/Location.h"
2220
#include "mlir/IR/PatternMatch.h"
@@ -26,8 +24,6 @@
2624
#include <cassert>
2725
#include <cstdint>
2826

29-
#define DPP
30-
3127
using namespace mlir;
3228

3329
namespace {
@@ -192,8 +188,6 @@ Value createSubgroupShuffleReduction(OpBuilder &builder, Location loc,
192188
function_ref<Value(Value)> unpackFn) {
193189
// Lane value always stays in the original type. We use it to perform arith
194190
// reductions.
195-
llvm::errs() << "Cluster Stride: " << ci.clusterStride << "\n";
196-
llvm::errs() << "Cluster Size: " << ci.clusterSize << "\n";
197191
Value laneVal = input;
198192
// Parallel reduction using butterfly shuffles.
199193
for (unsigned i = ci.clusterStride; i < ci.clusterStride * ci.clusterSize;
@@ -212,146 +206,6 @@ Value createSubgroupShuffleReduction(OpBuilder &builder, Location loc,
212206
return laneVal;
213207
}
214208

215-
#ifdef DPP
216-
Value createSubgroupDPPReduction(OpBuilder &b, Location loc,
217-
Value input, gpu::AllReduceOperation mode,
218-
const ClusterInfo &ci,
219-
function_ref<Value(Value)> packFn,
220-
function_ref<Value(Value)> unpackFn) {
221-
llvm::errs() << "createSubgroupDPPReduction" << "\n";
222-
Value result = input;
223-
if (ci.clusterSize >= 2) {
224-
auto permArg = b.getIntegerAttr(b.getIntegerType(32), 1);
225-
Value dppResult = b.create<amdgpu::DPPOp>(loc, result.getType(), result, result, amdgpu::DPPPerm::row_shr, permArg);
226-
llvm::errs() << dppResult << " c 2 \n";
227-
result = vector::makeArithReduction(b, loc,
228-
gpu::convertReductionKind(mode),
229-
result, dppResult);
230-
}
231-
232-
if (ci.clusterSize >= 4) {
233-
auto permArg = b.getIntegerAttr(b.getIntegerType(32), 2);
234-
Value dppResult = b.create<amdgpu::DPPOp>(loc, result.getType(), result, result, amdgpu::DPPPerm::row_shr, permArg);
235-
llvm::errs() << dppResult << " c 4 \n";
236-
result = vector::makeArithReduction(b, loc,
237-
gpu::convertReductionKind(mode),
238-
result, dppResult);
239-
}
240-
241-
if (ci.clusterSize >= 8) {
242-
243-
Value dppResult = b.create<amdgpu::DPPOp>(loc, result.getType(), result, result, amdgpu::DPPPerm::row_half_mirror, b.getUnitAttr());
244-
llvm::errs() << dppResult << " c 8 \n";
245-
result = vector::makeArithReduction(b, loc,
246-
gpu::convertReductionKind(mode),
247-
result, dppResult);
248-
}
249-
250-
if (ci.clusterSize >= 16) {
251-
Value dppResult = b.create<amdgpu::DPPOp>(loc, result.getType(), result, result, amdgpu::DPPPerm::row_mirror, b.getUnitAttr());
252-
llvm::errs() << dppResult << " c 16 \n";
253-
result = vector::makeArithReduction(b, loc,
254-
gpu::convertReductionKind(mode),
255-
result, dppResult);
256-
}
257-
258-
if (ci.clusterSize >= 32) {
259-
// auto permArg = builder.getInt32(15);
260-
// auto rowMask = builder.getInt32("0xa");
261-
// auto bankMask = builder.getInt32("0xf");
262-
// auto boundCtrl = builder.getBoolAttr(false);
263-
auto permArg = b.getIntegerAttr(b.getIntegerType(32), 15);
264-
Value dppResult = b.create<amdgpu::DPPOp>(loc, result.getType(), result, result, amdgpu::DPPPerm::row_bcast_15, b.getUnitAttr(), 10, 15, false);
265-
llvm::errs() << dppResult << " c 32 \n";
266-
result = vector::makeArithReduction(b, loc,
267-
gpu::convertReductionKind(mode),
268-
result, dppResult);
269-
}
270-
271-
if (ci.clusterSize == 64) {
272-
// auto permArg = builder.getInt32(31);
273-
// auto rowMask = builder.getInt32("0xc");
274-
// auto bankMask = builder.getInt32("0xf");
275-
// auto boundCtrl = builder.getBoolAttr(false);
276-
auto permArg = b.getIntegerAttr(b.getIntegerType(32), 31);
277-
Value dppResult = b.create<amdgpu::DPPOp>(loc, result.getType(), result, result, amdgpu::DPPPerm::row_bcast_31, b.getUnitAttr(), 12, 15, false);
278-
llvm::errs() << dppResult << " c 64 \n";
279-
result = vector::makeArithReduction(b, loc,
280-
gpu::convertReductionKind(mode),
281-
result, dppResult);
282-
}
283-
284-
// // read lane 63 with the final result.
285-
// auto lane = b.getIntegerAttr(b.getIntegerType(32), 63);
286-
// result = b.create<ROCDL::ReadLaneOp>(loc, input.getType(), result, lane);
287-
assert(result.getType() == input.getType());
288-
return result;
289-
}
290-
#endif
291-
292-
// Value createSubgroupDPPReduction(OpBuilder &b, Location loc,
293-
// Value input, gpu::AllReduceOperation mode,
294-
// const ClusterInfo &ci,
295-
// function_ref<Value(Value)> packFn,
296-
// function_ref<Value(Value)> unpackFn) {
297-
298-
// Value result = input;
299-
// if (ci.clusterSize >= 2) {
300-
// auto permArg = b.getInt32(1);
301-
// Value dppResult = builder.create<amdgpu::DPPOp>(packFn(result), packFn(result), amdgpu::DPPPerm::row_shr, permArg);
302-
// result = vector::makeArithReduction(builder, loc,
303-
// gpu::convertReductionKind(mode),
304-
// result, unpackFn(dppResult));
305-
// }
306-
307-
// if (ci.clusterSize >= 4) {
308-
// auto permArg = builder.getInt32(2);
309-
// Value dppResult = builder.create<amdgpu::DPPOp>(packFn(result), packFn(result), amdgpu::DPPPerm::row_shr, permArg);
310-
// result = vector::makeArithReduction(builder, loc,
311-
// gpu::convertReductionKind(mode),
312-
// result, unpackFn(dppResult));
313-
// }
314-
315-
// if (ci.clusterSize >= 8) {
316-
// Value dppResult = builder.create<amdgpu::DPPOp>(packFn(result), packFn(result), amdgpu::DPPPerm::row_half_mirror);
317-
// result = vector::makeArithReduction(builder, loc,
318-
// gpu::convertReductionKind(mode),
319-
// result, unpackFn(dppResult));
320-
// }
321-
322-
// if (ci.clusterSize >= 16) {
323-
// Value dppResult = builder.create<amdgpu::DPPOp>(packFn(result), packFn(result), amdgpu::DPPPerm::row_mirror);
324-
// result = vector::makeArithReduction(builder, loc,
325-
// gpu::convertReductionKind(mode),
326-
// result, unpackFn(dppResult));
327-
// }
328-
329-
// if (ci.clusterSize >= 32) {
330-
// auto permArg = builder.getInt32(15);
331-
// auto rowMask = builder.getInt32("0xa");
332-
// auto bankMask = builder.getInt32("0xf");
333-
// auto boundCtrl = builder.getBoolAttr(false);
334-
// Value dppResult = builder.create<amdgpu::DPPOp>(packFn(result), packFn(result), amdgpu::DPPPerm::row_bcast, permArg, rowMask, bankMask, boundCtrl);
335-
// result = vector::makeArithReduction(builder, loc,
336-
// gpu::convertReductionKind(mode),
337-
// result, unpackFn(dppResult));
338-
// }
339-
340-
// if (ci.clusterSize == 64) {
341-
// auto permArg = builder.getInt32(31);
342-
// auto rowMask = builder.getInt32("0xc");
343-
// auto bankMask = builder.getInt32("0xf");
344-
// auto boundCtrl = builder.getBoolAttr(false);
345-
// Value dppResult = builder.create<amdgpu::DPPOp>(packFn(result), packFn(result), amdgpu::DPPPerm::row_bcast, permArg, rowMask, bankMask, boundCtrl);
346-
// result = vector::makeArithReduction(builder, loc,
347-
// gpu::convertReductionKind(mode),
348-
// result, unpackFn(dppResult));
349-
// }
350-
351-
// assert(result.getType() == input.getType());
352-
// return result;
353-
// }
354-
355209
/// Lowers scalar gpu subgroup reductions to a series of shuffles.
356210
struct ScalarSubgroupReduceToShuffles final
357211
: OpRewritePattern<gpu::SubgroupReduceOp> {
@@ -363,7 +217,6 @@ struct ScalarSubgroupReduceToShuffles final
363217

364218
LogicalResult matchAndRewrite(gpu::SubgroupReduceOp op,
365219
PatternRewriter &rewriter) const override {
366-
llvm::errs() << "ScalarSubgroupReduceToShuffles" << "\n";
367220
if (op.getClusterSize().has_value() != matchClustered) {
368221
return rewriter.notifyMatchFailure(
369222
op, llvm::formatv("op is {0}clustered but pattern is configured to "
@@ -386,17 +239,10 @@ struct ScalarSubgroupReduceToShuffles final
386239
Location loc = op.getLoc();
387240
// Since this is already a native shuffle scalar, no packing is necessary.
388241
if (elemBitwidth == shuffleBitwidth) {
389-
llvm::errs() << "ScalarSubgroupReduceToShuffles - 1" << "\n";
390242
auto identityFn = [](Value v) { return v; };
391-
#ifndef DPP
392243
rewriter.replaceOp(op, createSubgroupShuffleReduction(
393244
rewriter, loc, op.getValue(), op.getOp(), *ci,
394245
identityFn, identityFn));
395-
#else
396-
rewriter.replaceOp(op, createSubgroupDPPReduction(
397-
rewriter, loc, op.getValue(), op.getOp(), *ci,
398-
identityFn, identityFn));
399-
#endif
400246
return success();
401247
}
402248

@@ -414,15 +260,10 @@ struct ScalarSubgroupReduceToShuffles final
414260
rewriter.create<arith::TruncIOp>(loc, equivIntType, packedVal);
415261
return rewriter.create<arith::BitcastOp>(loc, valueTy, asInt);
416262
};
417-
llvm::errs() << "ScalarSubgroupReduceToShuffles - 2" << "\n";
418-
#ifndef DPP
263+
419264
rewriter.replaceOp(
420265
op, createSubgroupShuffleReduction(rewriter, loc, op.getValue(),
421266
op.getOp(), *ci, packFn, unpackFn));
422-
#else
423-
rewriter.replaceOp(op, createSubgroupDPPReduction(rewriter, loc, op.getValue(),
424-
op.getOp(), *ci, packFn, unpackFn));
425-
#endif
426267
return success();
427268
}
428269

@@ -443,7 +284,6 @@ struct VectorSubgroupReduceToShuffles final
443284

444285
LogicalResult matchAndRewrite(gpu::SubgroupReduceOp op,
445286
PatternRewriter &rewriter) const override {
446-
llvm::errs() << "VectorSubgroupReduceToShuffles" << "\n";
447287
if (op.getClusterSize().has_value() != matchClustered) {
448288
return rewriter.notifyMatchFailure(
449289
op, llvm::formatv("op is {0}clustered but pattern is configured to "

0 commit comments

Comments
 (0)