gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix a pipeline issue

Browse Source
This commit is contained in:
Yan Da
2022-05-16 19:38:40 +08:00
parent c3c4ac3733
commit 79298d61bc
2 changed files with 11 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
lib/Dialect/TritonGPU/Transforms/Pipeline.cpp
View File

@@ -269,7 +269,9 @@ scf::ForOp LoopPipeliner::createNewForOp() {
Value newMask = builder.create<arith::AndIOp>(mask.getLoc(), Value newMask = builder.create<arith::AndIOp>(mask.getLoc(),
splatCond, splatCond,
nextMapping.lookupOrDefault(mask)); nextMapping.lookupOrDefault(mask));
nextMapping.map(mask, newMask); // if mask is defined outside the loop, don't update the map more than once
if (!(forOp.isDefinedOutsideOfLoop(mask) && nextMapping.contains(mask)))
nextMapping.map(mask, newMask);
} }
Operation *nextOp = builder.clone(*op, nextMapping); Operation *nextOp = builder.clone(*op, nextMapping);
// update mapping of results // update mapping of results

18
rewrite-test/jit/matmul/matmul.mlir
View File

@@ -204,16 +204,14 @@ module {
%91 = arith.addi %arg17, %c128 : index %91 = arith.addi %arg17, %c128 : index
%92 = arith.cmpi slt, %91, %47 : index %92 = arith.cmpi slt, %91, %47 : index
%93 = tt.broadcast %92 : (i1) -> tensor<128x128xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">> %93 = tt.broadcast %92 : (i1) -> tensor<128x128xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>
%94 = tt.broadcast %92 : (i1) -> tensor<128x128xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">> %94 = "triton_gpu.copy_async"(%arg15, %93, %cst_1) {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : (tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128xf16, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>) -> tensor<128x128xf16, #triton_gpu<"shared (memory) encoding<>">>
%95 = arith.andi %94, %93 : tensor<128x128xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">> %95 = "triton_gpu.copy_async"(%arg16, %93, %cst_1) {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : (tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128xf16, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>) -> tensor<128x128xf16, #triton_gpu<"shared (memory) encoding<>">>
%96 = "triton_gpu.copy_async"(%arg15, %93, %cst_1) {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : (tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128xf16, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>) -> tensor<128x128xf16, #triton_gpu<"shared (memory) encoding<>">> %96 = tt.broadcast %c128_i32 : (i32) -> tensor<128x128xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>
%97 = "triton_gpu.copy_async"(%arg16, %95, %cst_1) {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : (tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128xf16, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>) -> tensor<128x128xf16, #triton_gpu<"shared (memory) encoding<>">> %97 = tt.getelementptr %arg15, %96, : tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>
%98 = tt.broadcast %c128_i32 : (i32) -> tensor<128x128xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">> %98 = arith.muli %arg7, %c128_i32 : i32
%99 = tt.getelementptr %arg15, %98, : tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">> %99 = tt.broadcast %98 : (i32) -> tensor<128x128xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>
%100 = arith.muli %arg7, %c128_i32 : i32 %100 = tt.getelementptr %arg16, %99, : tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>
%101 = tt.broadcast %100 : (i32) -> tensor<128x128xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">> scf.yield %85, %87, %90, %94, %95, %97, %100, %91 : tensor<128x128xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128xf16, #triton_gpu<"shared (memory) encoding<>">>, tensor<128x128xf16, #triton_gpu<"shared (memory) encoding<>">>, tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, index
%102 = tt.getelementptr %arg16, %101, : tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>
scf.yield %85, %87, %90, %96, %97, %99, %102, %91 : tensor<128x128xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128xf16, #triton_gpu<"shared (memory) encoding<>">>, tensor<128x128xf16, #triton_gpu<"shared (memory) encoding<>">>, tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, tensor<128x128x!tt.ptr<f16>, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>, index
} }
%56 = arith.truncf %55#0 : tensor<128x128xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">> to tensor<128x128xf16, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">> %56 = arith.truncf %55#0 : tensor<128x128xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">> to tensor<128x128xf16, #triton_gpu<"coalesced encoding<threadTileSize = 1, 1, blockTileSize = 32, 1, order = 0, 1>">>
%57 = arith.muli %12, %c128_i32 : i32 %57 = arith.muli %12, %c128_i32 : i32