This is an archive of the discontinued LLVM Phabricator instance.

Differential D145621

[mlir][sparse] deduplicate non-unique coordinates unconditionally
ClosedPublic

Authored by Peiming on Mar 8 2023, 3:36 PM.

Details

Reviewers
aartbik
nicolasvasilache
bixia
wrengr
Commits
rGab99b5d1f6a2: [mlir][sparse] deduplicate non-unique coordinates unconditionally

Diff Detail

Event Timeline

Peiming created this revision.Mar 8 2023, 3:36 PM
Herald added a reviewer: aartbik. · View Herald TranscriptMar 8 2023, 3:36 PM
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: hanchung, jsetoain, Moerafaat and 23 others. · View Herald Transcript
Peiming requested review of this revision.Mar 8 2023, 3:36 PM
Herald added a reviewer: nicolasvasilache. · View Herald TranscriptMar 8 2023, 3:36 PM
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: stephenneuendorffer, nicolasvasilache. · View Herald Transcript
Peiming added inline comments.Mar 8 2023, 3:42 PM
mlir/test/Dialect/SparseTensor/sorted_coo.mlir
56–70

de-duplication while loop here

113–127

de-duplication while loop here

Peiming added reviewers: bixia, wrengr.Mar 8 2023, 3:48 PM
Harbormaster completed remote builds in B218227: Diff 503544.Mar 8 2023, 4:37 PM
aartbik accepted this revision.Mar 9 2023, 1:36 PM
This revision is now accepted and ready to land.Mar 9 2023, 1:36 PM
Closed by commit rGab99b5d1f6a2: [mlir][sparse] deduplicate non-unique coordinates unconditionally (authored by Peiming). · Explain WhyMar 9 2023, 2:00 PM
This revision was automatically updated to reflect the committed changes.
Peiming added a commit: rGab99b5d1f6a2: [mlir][sparse] deduplicate non-unique coordinates unconditionally.

Revision Contents

PathSize
mlir/
lib/
Dialect/
SparseTensor/
Transforms/
8 lines
test/
Dialect/
SparseTensor/
137 lines
106 lines

Diff 503929

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

Show First 20 Lines • Show All 1,281 Lines • ▼ Show 20 Lines
static bool translateBitsToTidDimPairs( static bool translateBitsToTidDimPairs(
CodegenEnv &env, unsigned li, unsigned idx, SmallVectorImpl<size_t> &tids, CodegenEnv &env, unsigned li, unsigned idx, SmallVectorImpl<size_t> &tids,
SmallVectorImpl<size_t> &dims, SmallVectorImpl<size_t> &affineTids, SmallVectorImpl<size_t> &dims, SmallVectorImpl<size_t> &affineTids,
SmallVectorImpl<size_t> &affineDims, SmallVectorImpl<AffineExpr> &exps) { SmallVectorImpl<size_t> &affineDims, SmallVectorImpl<AffineExpr> &exps) {
const BitVector &all = env.lat(li).bits; const BitVector &all = env.lat(li).bits;
const BitVector &simple = env.lat(li).simple; const BitVector &simple = env.lat(li).simple;
unsigned numloopCond = 0; unsigned numloopCond = 0;
bool hasNonUnique = false;
// Converts bits to array + dim pair // Converts bits to array + dim pair
env.merger().foreachTidDimPairInBits( env.merger().foreachTidDimPairInBits(
all, [&, idx](unsigned b, unsigned tid, std::optional<unsigned> dim, all, [&, idx](unsigned b, unsigned tid, std::optional<unsigned> dim,
DimLevelType dlt) { DimLevelType dlt) {
if (simple.test(b)) { if (simple.test(b)) {
if (isUndefDLT(dlt)) { if (isUndefDLT(dlt)) {
// An undefined dlt in the lattices, we probably mean to iterate // An undefined dlt in the lattices, we probably mean to iterate
// based on the dim of output tensor. // based on the dim of output tensor.
// E.g., this could be a synthetic tensor (for invariants and sparse // E.g., this could be a synthetic tensor (for invariants and sparse
// output tensor). // output tensor).
// out[i][j] = invariant; or a broadcast // out[i][j] = invariant; or a broadcast
// out[i][j] = in[i] (j is undef for input) // out[i][j] = in[i] (j is undef for input)
tid = env.merger().getOutTensorID(); tid = env.merger().getOutTensorID();
dim = env.merger().getDimNum(tid, idx); dim = env.merger().getDimNum(tid, idx);
// Skips invalid dim (e.g., when this is a zero ranked tensor). // Skips invalid dim (e.g., when this is a zero ranked tensor).
if (!dim) if (!dim)
return; return;
} }
hasNonUnique = !isUniqueDLT(dlt) || hasNonUnique;
tids.push_back(tid); tids.push_back(tid);
dims.push_back(*dim); dims.push_back(*dim);
numloopCond++; numloopCond++;
} else if (isDenseDLT(dlt)) { } else if (isDenseDLT(dlt)) {
tids.push_back(tid); tids.push_back(tid);
dims.push_back(*dim); dims.push_back(*dim);
} else { } else {
assert(isUndefDLT(dlt)); assert(isUndefDLT(dlt));
▲ Show 20 Lines • Show All 46 Lines • ▼ Show 20 Lines if (isDenseDLT(env.dlt(env.merger().getOutTensorID(), idx))) {
// unconditionally, since they may not appear in the lattice, but may be // unconditionally, since they may not appear in the lattice, but may be
// needed for linearized env. // needed for linearized env.
auto dim = *env.merger().getDimNum(env.merger().getOutTensorID(), idx); auto dim = *env.merger().getDimNum(env.merger().getOutTensorID(), idx);
tids.push_back(env.merger().getOutTensorID()); tids.push_back(env.merger().getOutTensorID());
dims.push_back(dim); dims.push_back(dim);
} }
assert(numloopCond > 0); assert(numloopCond > 0);
// If we just need to one loop conditions, the loop can be generated by a for // If we just need to one loop conditions and the conditions is not imposed on
// loop. // non-unique level, the loop can be generated by a for loop.
return numloopCond == 1; return numloopCond == 1 && !hasNonUnique;
} }
/// Starts a single loop in current sequence. /// Starts a single loop in current sequence.
static Operation *startLoop(CodegenEnv &env, OpBuilder &builder, unsigned at, static Operation *startLoop(CodegenEnv &env, OpBuilder &builder, unsigned at,
unsigned li, bool needsUniv) { unsigned li, bool needsUniv) {
// The set of tensors + dims to generate loops on // The set of tensors + dims to generate loops on
SmallVector<size_t> tids, dims; SmallVector<size_t> tids, dims;
// The set of dense tensors with non-trivial affine expression that just // The set of dense tensors with non-trivial affine expression that just
▲ Show 20 Lines • Show All 267 Lines • Show Last 20 Lines

mlir/test/Dialect/SparseTensor/sorted_coo.mlir

// RUN: mlir-opt %s -sparsification | FileCheck %s // RUN: mlir-opt %s -sparsification --canonicalize | FileCheck %s
#SortedCOO = #sparse_tensor.encoding<{ #SortedCOO = #sparse_tensor.encoding<{
dimLevelType = [ "compressed-nu", "singleton" ] dimLevelType = [ "compressed-nu", "singleton" ]
}> }>
#trait_scale = { #trait_scale = {
indexing_maps = [ indexing_maps = [
affine_map<(i,j) -> (i,j)> // X (out) affine_map<(i,j) -> (i,j)> // X (out)
Show All 21 Lines#trait_mul = {
iterator_types = ["parallel", "parallel"], iterator_types = ["parallel", "parallel"],
doc = "Z(i,j) = X(i,j) * Y(i,j)" doc = "Z(i,j) = X(i,j) * Y(i,j)"
} }
// //
// Kernels that operate on SortedCOO format. // Kernels that operate on SortedCOO format.
// //
// CHECK-LABEL: func.func @sparse_scale( // CHECK-LABEL: func.func @sparse_scale(
// CHECK-SAME: %[[VAL_0:.*]]: tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>>) -> tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> { // CHECK-SAME: %[[VAL_0:.*]]: tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>>) -> tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> {
// CHECK-DAG: %[[VAL_1:.*]] = arith.constant 0 : index // CHECK-DAG: %[[VAL_1:.*]] = arith.constant false
// CHECK-DAG: %[[VAL_2:.*]] = arith.constant 1 : index // CHECK-DAG: %[[VAL_2:.*]] = arith.constant 0 : index
// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 2.000000e+00 : f32 // CHECK-DAG: %[[VAL_3:.*]] = arith.constant 1 : index
// CHECK-DAG: %[[VAL_4:.*]] = sparse_tensor.positions %[[VAL_0]] {level = 0 : index} : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex> // CHECK-DAG: %[[VAL_4:.*]] = arith.constant 2.000000e+00 : f32
// CHECK-DAG: %[[VAL_5:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xf32> // CHECK-DAG: %[[VAL_5:.*]] = sparse_tensor.positions %[[VAL_0]] {level = 0 : index} : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex>
// CHECK: %[[VAL_6:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_1]]] : memref<?xindex> // CHECK-DAG: %[[VAL_6:.*]] = sparse_tensor.coordinates %[[VAL_0]] {level = 0 : index} : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex, strided<[?], offset: ?>>
// CHECK: %[[VAL_7:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_2]]] : memref<?xindex> // CHECK-DAG: %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xf32>
// CHECK: scf.for %[[VAL_8:.*]] = %[[VAL_6]] to %[[VAL_7]] step %[[VAL_2]] { // CHECK-DAG: %[[VAL_8:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_2]]] : memref<?xindex>
// CHECK: %[[VAL_9:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_8]]] : memref<?xf32> // CHECK-DAG: %[[VAL_9:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex>
// CHECK: %[[VAL_10:.*]] = arith.mulf %[[VAL_9]], %[[VAL_3]] : f32 // CHECK: %[[VAL_10:.*]] = scf.while (%[[VAL_11:.*]] = %[[VAL_8]]) : (index) -> index {
// CHECK: memref.store %[[VAL_10]], %[[VAL_5]]{{\[}}%[[VAL_8]]] : memref<?xf32> // CHECK: %[[VAL_12:.*]] = arith.cmpi ult, %[[VAL_11]], %[[VAL_9]] : index
// CHECK: scf.condition(%[[VAL_12]]) %[[VAL_11]] : index
// CHECK: } do {
// CHECK: ^bb0(%[[VAL_13:.*]]: index):
// CHECK: %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_13]]] : memref<?xindex, strided<[?], offset: ?>>
// CHECK: %[[VAL_15:.*]] = scf.while (%[[VAL_16:.*]] = %[[VAL_13]]) : (index) -> index {
// CHECK: %[[VAL_17:.*]] = arith.cmpi ult, %[[VAL_16]], %[[VAL_9]] : index
// CHECK: %[[VAL_18:.*]] = scf.if %[[VAL_17]] -> (i1) {
// CHECK: %[[VAL_19:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_16]]] : memref<?xindex, strided<[?], offset: ?>>
// CHECK: %[[VAL_20:.*]] = arith.cmpi eq, %[[VAL_19]], %[[VAL_14]] : index
// CHECK: scf.yield %[[VAL_20]] : i1
// CHECK: } else {
// CHECK: scf.yield %[[VAL_1]] : i1
// CHECK: } // CHECK: }
// CHECK: %[[VAL_11:.*]] = sparse_tensor.load %[[VAL_0]] : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> // CHECK: scf.condition(%[[VAL_21:.*]]) %[[VAL_16]] : index
// CHECK: return %[[VAL_11]] : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> // CHECK: } do {
// CHECK: ^bb0(%[[VAL_22:.*]]: index):
// CHECK: %[[VAL_23:.*]] = arith.addi %[[VAL_22]], %[[VAL_3]] : index
// CHECK: scf.yield %[[VAL_23]] : index
// CHECK: }
PeimingAuthorUnsubmitted
Done
ReplyInline Actions

de-duplication while loop here

Peiming: de-duplication while loop here
// CHECK: scf.for %[[VAL_24:.*]] = %[[VAL_13]] to %[[VAL_25:.*]] step %[[VAL_3]] {
// CHECK: %[[VAL_26:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_24]]] : memref<?xf32>
// CHECK: %[[VAL_27:.*]] = arith.mulf %[[VAL_26]], %[[VAL_4]] : f32
// CHECK: memref.store %[[VAL_27]], %[[VAL_7]]{{\[}}%[[VAL_24]]] : memref<?xf32>
// CHECK: } {"Emitted from" = "linalg.generic"}
// CHECK: scf.yield %[[VAL_28:.*]] : index
// CHECK: } attributes {"Emitted from" = "linalg.generic"}
// CHECK: %[[VAL_29:.*]] = sparse_tensor.load %[[VAL_0]] : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>>
// CHECK: return %[[VAL_29]] : tensor<?x?xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>>
// CHECK: } // CHECK: }
func.func @sparse_scale(%argx: tensor<?x?xf32, #SortedCOO>) -> tensor<?x?xf32, #SortedCOO> { func.func @sparse_scale(%argx: tensor<?x?xf32, #SortedCOO>) -> tensor<?x?xf32, #SortedCOO> {
%c = arith.constant 2.0 : f32 %c = arith.constant 2.0 : f32
%0 = linalg.generic #trait_scale %0 = linalg.generic #trait_scale
outs(%argx: tensor<?x?xf32, #SortedCOO>) { outs(%argx: tensor<?x?xf32, #SortedCOO>) {
^bb(%x: f32): ^bb(%x: f32):
%1 = arith.mulf %x, %c : f32 %1 = arith.mulf %x, %c : f32
linalg.yield %1 : f32 linalg.yield %1 : f32
} -> tensor<?x?xf32, #SortedCOO> } -> tensor<?x?xf32, #SortedCOO>
return %0 : tensor<?x?xf32, #SortedCOO> return %0 : tensor<?x?xf32, #SortedCOO>
} }
// CHECK-LABEL: func.func @matvec( // CHECK-LABEL: func.func @matvec(
// CHECK-SAME: %[[VAL_0:.*]]: tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>>, // CHECK-SAME: %[[VAL_0:.*]]: tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>>,
// CHECK-SAME: %[[VAL_1:.*]]: tensor<64xf64>, // CHECK-SAME: %[[VAL_1:.*]]: tensor<64xf64>,
// CHECK-SAME: %[[VAL_2:.*]]: tensor<32xf64>) -> tensor<32xf64> { // CHECK-SAME: %[[VAL_2:.*]]: tensor<32xf64>) -> tensor<32xf64> {
// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 0 : index // CHECK-DAG: %[[VAL_3:.*]] = arith.constant false
// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 1 : index // CHECK-DAG: %[[VAL_4:.*]] = arith.constant 0 : index
// CHECK-DAG: %[[VAL_5:.*]] = sparse_tensor.positions %[[VAL_0]] {level = 0 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex> // CHECK-DAG: %[[VAL_5:.*]] = arith.constant 1 : index
// CHECK-DAG: %[[VAL_6:.*]] = sparse_tensor.coordinates %[[VAL_0]] {level = 0 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex, strided<[?], offset: ?>> // CHECK-DAG: %[[VAL_6:.*]] = sparse_tensor.positions %[[VAL_0]] {level = 0 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex>
// CHECK-DAG: %[[VAL_7:.*]] = sparse_tensor.coordinates %[[VAL_0]] {level = 1 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex, strided<[?], offset: ?>> // CHECK-DAG: %[[VAL_7:.*]] = sparse_tensor.coordinates %[[VAL_0]] {level = 0 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex, strided<[?], offset: ?>>
// CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xf64> // CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.coordinates %[[VAL_0]] {level = 1 : index} : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xindex, strided<[?], offset: ?>>
// CHECK-DAG: %[[VAL_9:.*]] = bufferization.to_memref %[[VAL_1]] : memref<64xf64> // CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ] }>> to memref<?xf64>
// CHECK-DAG: %[[VAL_10:.*]] = bufferization.to_memref %[[VAL_2]] : memref<32xf64> // CHECK: %[[VAL_10:.*]] = bufferization.to_memref %[[VAL_2]] : memref<32xf64>
// CHECK-DAG: %[[VAL_11:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex> // CHECK: %[[VAL_11:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
// CHECK-DAG: %[[VAL_12:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex> // CHECK: %[[VAL_12:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
// CHECK: scf.for %[[VAL_13:.*]] = %[[VAL_11]] to %[[VAL_12]] step %[[VAL_4]] { // CHECK: %[[VAL_13:.*]] = scf.while (%[[VAL_14:.*]] = %[[VAL_11]]) : (index) -> index {
// CHECK: %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_13]]] : memref<?xindex, strided<[?], offset: ?>> // CHECK: %[[VAL_15:.*]] = arith.cmpi ult, %[[VAL_14]], %[[VAL_12]] : index
// CHECK: %[[VAL_15:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_14]]] : memref<32xf64> // CHECK: scf.condition(%[[VAL_15]]) %[[VAL_14]] : index
// CHECK: %[[VAL_16:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_13]]] : memref<?xindex, strided<[?], offset: ?>> // CHECK: } do {
// CHECK: %[[VAL_17:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_13]]] : memref<?xf64> // CHECK: ^bb0(%[[VAL_16:.*]]: index):
// CHECK: %[[VAL_18:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_16]]] : memref<64xf64> // CHECK: %[[VAL_17:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_16]]] : memref<?xindex, strided<[?], offset: ?>>
// CHECK: %[[VAL_19:.*]] = arith.mulf %[[VAL_17]], %[[VAL_18]] : f64 // CHECK: %[[VAL_18:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_16]]] : memref<?xindex, strided<[?], offset: ?>>
// CHECK: %[[VAL_20:.*]] = arith.addf %[[VAL_15]], %[[VAL_19]] : f64 // CHECK: %[[VAL_19:.*]] = scf.while (%[[VAL_20:.*]] = %[[VAL_16]]) : (index) -> index {
// CHECK: memref.store %[[VAL_20]], %[[VAL_10]]{{\[}}%[[VAL_14]]] : memref<32xf64> // CHECK: %[[VAL_21:.*]] = arith.cmpi ult, %[[VAL_20]], %[[VAL_12]] : index
// CHECK: %[[VAL_22:.*]] = scf.if %[[VAL_21]] -> (i1) {
// CHECK: %[[VAL_23:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_20]]] : memref<?xindex, strided<[?], offset: ?>>
// CHECK: %[[VAL_24:.*]] = arith.cmpi eq, %[[VAL_23]], %[[VAL_18]] : index
// CHECK: scf.yield %[[VAL_24]] : i1
// CHECK: } else {
// CHECK: scf.yield %[[VAL_3]] : i1
// CHECK: } // CHECK: }
// CHECK: %[[VAL_21:.*]] = bufferization.to_tensor %[[VAL_10]] : memref<32xf64> // CHECK: scf.condition(%[[VAL_25:.*]]) %[[VAL_20]] : index
// CHECK: return %[[VAL_21]] : tensor<32xf64> // CHECK: } do {
// CHECK: ^bb0(%[[VAL_26:.*]]: index):
// CHECK: %[[VAL_27:.*]] = arith.addi %[[VAL_26]], %[[VAL_5]] : index
// CHECK: scf.yield %[[VAL_27]] : index
// CHECK: }
PeimingAuthorUnsubmitted
Done
ReplyInline Actions

de-duplication while loop here

Peiming: de-duplication while loop here
// CHECK: %[[VAL_28:.*]] = tensor.extract %[[VAL_2]]{{\[}}%[[VAL_17]]] : tensor<32xf64>
// CHECK: %[[VAL_29:.*]] = scf.for %[[VAL_30:.*]] = %[[VAL_16]] to %[[VAL_31:.*]] step %[[VAL_5]] iter_args(%[[VAL_32:.*]] = %[[VAL_28]]) -> (f64) {
// CHECK: %[[VAL_33:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_30]]] : memref<?xindex, strided<[?], offset: ?>>
// CHECK: %[[VAL_34:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_30]]] : memref<?xf64>
// CHECK: %[[VAL_35:.*]] = tensor.extract %[[VAL_1]]{{\[}}%[[VAL_33]]] : tensor<64xf64>
// CHECK: %[[VAL_36:.*]] = arith.mulf %[[VAL_34]], %[[VAL_35]] : f64
// CHECK: %[[VAL_37:.*]] = arith.addf %[[VAL_32]], %[[VAL_36]] : f64
// CHECK: scf.yield %[[VAL_37]] : f64
// CHECK: } {"Emitted from" = "linalg.generic"}
// CHECK: memref.store %[[VAL_38:.*]], %[[VAL_10]]{{\[}}%[[VAL_17]]] : memref<32xf64>
// CHECK: scf.yield %[[VAL_39:.*]] : index
// CHECK: } attributes {"Emitted from" = "linalg.generic"}
// CHECK: %[[VAL_40:.*]] = bufferization.to_tensor %[[VAL_10]] : memref<32xf64>
// CHECK: return %[[VAL_40]] : tensor<32xf64>
// CHECK: } // CHECK: }
func.func @matvec(%arga: tensor<32x64xf64, #SortedCOO>, func.func @matvec(%arga: tensor<32x64xf64, #SortedCOO>,
%argb: tensor<64xf64>, %argb: tensor<64xf64>,
%argx: tensor<32xf64>) -> tensor<32xf64> { %argx: tensor<32xf64>) -> tensor<32xf64> {
%0 = linalg.generic #trait_matvec %0 = linalg.generic #trait_matvec
ins(%arga, %argb : tensor<32x64xf64, #SortedCOO>, tensor<64xf64>) ins(%arga, %argb : tensor<32x64xf64, #SortedCOO>, tensor<64xf64>)
outs(%argx: tensor<32xf64>) { outs(%argx: tensor<32xf64>) {
^bb(%A: f64, %b: f64, %x: f64): ^bb(%A: f64, %b: f64, %x: f64):
%0 = arith.mulf %A, %b : f64 %0 = arith.mulf %A, %b : f64
▲ Show 20 Lines • Show All 86 Lines • ▼ Show 20 Lines
// CHECK: %[[VAL_69:.*]] = arith.cmpi eq, %[[VAL_65]], %[[VAL_68]] : index // CHECK: %[[VAL_69:.*]] = arith.cmpi eq, %[[VAL_65]], %[[VAL_68]] : index
// CHECK: %[[VAL_70:.*]] = arith.cmpi eq, %[[VAL_66]], %[[VAL_68]] : index // CHECK: %[[VAL_70:.*]] = arith.cmpi eq, %[[VAL_66]], %[[VAL_68]] : index
// CHECK: %[[VAL_71:.*]] = arith.andi %[[VAL_69]], %[[VAL_70]] : i1 // CHECK: %[[VAL_71:.*]] = arith.andi %[[VAL_69]], %[[VAL_70]] : i1
// CHECK: scf.if %[[VAL_71]] { // CHECK: scf.if %[[VAL_71]] {
// CHECK: %[[VAL_72:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_63]]] : memref<?xf64> // CHECK: %[[VAL_72:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_63]]] : memref<?xf64>
// CHECK: %[[VAL_73:.*]] = memref.load %[[VAL_14]]{{\[}}%[[VAL_64]]] : memref<?xf64> // CHECK: %[[VAL_73:.*]] = memref.load %[[VAL_14]]{{\[}}%[[VAL_64]]] : memref<?xf64>
// CHECK: %[[VAL_74:.*]] = arith.mulf %[[VAL_72]], %[[VAL_73]] : f64 // CHECK: %[[VAL_74:.*]] = arith.mulf %[[VAL_72]], %[[VAL_73]] : f64
// CHECK: memref.store %[[VAL_74]], %[[VAL_15]]{{\[}}%[[VAL_31]], %[[VAL_68]]] : memref<32x64xf64> // CHECK: memref.store %[[VAL_74]], %[[VAL_15]]{{\[}}%[[VAL_31]], %[[VAL_68]]] : memref<32x64xf64>
// CHECK: } else {
// CHECK: } // CHECK: }
// CHECK: %[[VAL_75:.*]] = arith.cmpi eq, %[[VAL_65]], %[[VAL_68]] : index // CHECK: %[[VAL_75:.*]] = arith.cmpi eq, %[[VAL_65]], %[[VAL_68]] : index
// CHECK: %[[VAL_76:.*]] = arith.addi %[[VAL_63]], %[[VAL_6]] : index // CHECK: %[[VAL_76:.*]] = arith.addi %[[VAL_63]], %[[VAL_6]] : index
// CHECK: %[[VAL_77:.*]] = arith.select %[[VAL_75]], %[[VAL_76]], %[[VAL_63]] : index // CHECK: %[[VAL_77:.*]] = arith.select %[[VAL_75]], %[[VAL_76]], %[[VAL_63]] : index
// CHECK: %[[VAL_78:.*]] = arith.cmpi eq, %[[VAL_66]], %[[VAL_68]] : index // CHECK: %[[VAL_78:.*]] = arith.cmpi eq, %[[VAL_66]], %[[VAL_68]] : index
// CHECK: %[[VAL_79:.*]] = arith.addi %[[VAL_64]], %[[VAL_6]] : index // CHECK: %[[VAL_79:.*]] = arith.addi %[[VAL_64]], %[[VAL_6]] : index
// CHECK: %[[VAL_80:.*]] = arith.select %[[VAL_78]], %[[VAL_79]], %[[VAL_64]] : index // CHECK: %[[VAL_80:.*]] = arith.select %[[VAL_78]], %[[VAL_79]], %[[VAL_64]] : index
// CHECK: scf.yield %[[VAL_77]], %[[VAL_80]] : index, index // CHECK: scf.yield %[[VAL_77]], %[[VAL_80]] : index, index
// CHECK: } attributes {"Emitted from" = "linalg.generic"} // CHECK: } attributes {"Emitted from" = "linalg.generic"}
// CHECK: } else {
// CHECK: } // CHECK: }
// CHECK: %[[VAL_81:.*]] = arith.cmpi eq, %[[VAL_28]], %[[VAL_31]] : index // CHECK: %[[VAL_81:.*]] = arith.cmpi eq, %[[VAL_28]], %[[VAL_31]] : index
// CHECK: %[[VAL_82:.*]] = arith.select %[[VAL_81]], %[[VAL_83:.*]], %[[VAL_26]] : index // CHECK: %[[VAL_82:.*]] = arith.select %[[VAL_81]], %[[VAL_83:.*]], %[[VAL_26]] : index
// CHECK: %[[VAL_84:.*]] = arith.cmpi eq, %[[VAL_29]], %[[VAL_31]] : index // CHECK: %[[VAL_84:.*]] = arith.cmpi eq, %[[VAL_29]], %[[VAL_31]] : index
// CHECK: %[[VAL_85:.*]] = arith.select %[[VAL_84]], %[[VAL_86:.*]], %[[VAL_27]] : index // CHECK: %[[VAL_85:.*]] = arith.select %[[VAL_84]], %[[VAL_86:.*]], %[[VAL_27]] : index
// CHECK: scf.yield %[[VAL_82]], %[[VAL_85]] : index, index // CHECK: scf.yield %[[VAL_82]], %[[VAL_85]] : index, index
// CHECK: } attributes {"Emitted from" = "linalg.generic"} // CHECK: } attributes {"Emitted from" = "linalg.generic"}
// CHECK: %[[VAL_87:.*]] = bufferization.to_tensor %[[VAL_15]] : memref<32x64xf64> // CHECK: %[[VAL_87:.*]] = bufferization.to_tensor %[[VAL_15]] : memref<32x64xf64>
Show All 14 Lines

mlir/test/Dialect/SparseTensor/sparse_reshape_dot.mlir

// RUN: mlir-opt %s --linalg-generalize-named-ops --sparsification --cse --canonicalize | FileCheck %s // RUN: mlir-opt %s --linalg-generalize-named-ops --sparsification --cse --canonicalize | FileCheck %s
#COO_2D = #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ], posWidth = 32, crdWidth = 32 }> #COO_2D = #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton" ], posWidth = 32, crdWidth = 32 }>
#COO_3D = #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton-nu", "singleton" ], posWidth = 32, crdWidth = 32 }> #COO_3D = #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton-nu", "singleton" ], posWidth = 32, crdWidth = 32 }>
// CHECK-LABEL: func.func @sparse_reshape_fused( // CHECK-LABEL: func.func @sparse_reshape_fused(
// CHECK-SAME: %[[VAL_0:.*]]: tensor<5x6xf32>, // CHECK-SAME: %[[VAL_0:.*]]: tensor<5x6xf32>,
// CHECK-SAME: %[[VAL_1:.*]]: tensor<6x2x3xf32, // CHECK-SAME: %[[VAL_1:.*]]: tensor<6x2x3xf32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed-nu", "singleton-nu", "singleton" ], posWidth = 32, crdWidth = 32 }>>) -> tensor<?x?x?xf32> {
// CHECK-DAG: %[[VAL_2:.*]] = arith.constant 5 : index // CHECK-DAG: %[[VAL_2:.*]] = arith.constant false
// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 3 : index // CHECK-DAG: %[[VAL_3:.*]] = arith.constant 5 : index
// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 0 : index // CHECK-DAG: %[[VAL_4:.*]] = arith.constant 3 : index
// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 1 : index // CHECK-DAG: %[[VAL_5:.*]] = arith.constant 0 : index
// CHECK-DAG: %[[VAL_6:.*]] = tensor.empty() : tensor<5x6xf32> // CHECK-DAG: %[[VAL_6:.*]] = arith.constant 1 : index
// CHECK-DAG: %[[VAL_7:.*]] = sparse_tensor.positions %[[VAL_1]] {level = 0 : index} // CHECK-DAG: %[[VAL_7:.*]] = tensor.empty() : tensor<5x6xf32>
// CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.coordinates %[[VAL_1]] {level = 0 : index} // CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.positions %[[VAL_1]] {level = 0 : index}
// CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.coordinates %[[VAL_1]] {level = 1 : index} // CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.coordinates %[[VAL_1]] {level = 0 : index}
// CHECK-DAG: %[[VAL_10:.*]] = sparse_tensor.coordinates %[[VAL_1]] {level = 2 : index} // CHECK-DAG: %[[VAL_10:.*]] = sparse_tensor.coordinates %[[VAL_1]] {level = 1 : index}
// CHECK-DAG: %[[VAL_11:.*]] = sparse_tensor.values %[[VAL_1]] // CHECK-DAG: %[[VAL_11:.*]] = sparse_tensor.coordinates %[[VAL_1]] {level = 2 : index}
// CHECK-DAG: %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_6]] : memref<5x6xf32> // CHECK-DAG: %[[VAL_12:.*]] = sparse_tensor.values %[[VAL_1]]
// CHECK: scf.for %[[VAL_13:.*]] = %[[VAL_4]] to %[[VAL_2]] step %[[VAL_5]] { // CHECK-DAG: %[[VAL_13:.*]] = bufferization.to_memref %[[VAL_7]] : memref<5x6xf32>
// CHECK: %[[VAL_14:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_4]]] : memref<?xi32> // CHECK: scf.for %[[VAL_14:.*]] = %[[VAL_5]] to %[[VAL_3]] step %[[VAL_6]] {
// CHECK: %[[VAL_15:.*]] = arith.extui %[[VAL_14]] : i32 to i64 // CHECK: %[[VAL_15:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_5]]] : memref<?xi32>
// CHECK: %[[VAL_16:.*]] = arith.index_cast %[[VAL_15]] : i64 to index // CHECK: %[[VAL_16:.*]] = arith.extui %[[VAL_15]] : i32 to i64
// CHECK: %[[VAL_17:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_5]]] : memref<?xi32> // CHECK: %[[VAL_17:.*]] = arith.index_cast %[[VAL_16]] : i64 to index
// CHECK: %[[VAL_18:.*]] = arith.extui %[[VAL_17]] : i32 to i64 // CHECK: %[[VAL_18:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_6]]] : memref<?xi32>
// CHECK: %[[VAL_19:.*]] = arith.index_cast %[[VAL_18]] : i64 to index // CHECK: %[[VAL_19:.*]] = arith.extui %[[VAL_18]] : i32 to i64
// CHECK: scf.for %[[VAL_20:.*]] = %[[VAL_16]] to %[[VAL_19]] step %[[VAL_5]] { // CHECK: %[[VAL_20:.*]] = arith.index_cast %[[VAL_19]] : i64 to index
// CHECK: %[[VAL_21:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_20]]] : memref<?xi32, strided<[?], offset: ?>> // CHECK: %[[VAL_21:.*]] = scf.while (%[[VAL_22:.*]] = %[[VAL_17]]) : (index) -> index {
// CHECK: %[[VAL_22:.*]] = arith.extui %[[VAL_21]] : i32 to i64 // CHECK: %[[VAL_23:.*]] = arith.cmpi ult, %[[VAL_22]], %[[VAL_20]] : index
// CHECK: %[[VAL_23:.*]] = arith.index_cast %[[VAL_22]] : i64 to index // CHECK: scf.condition(%[[VAL_23]]) %[[VAL_22]] : index
// CHECK: %[[VAL_24:.*]] = tensor.extract %[[VAL_0]]{{\[}}%[[VAL_13]], %[[VAL_23]]] : tensor<5x6xf32> // CHECK: } do {
// CHECK: %[[VAL_25:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_20]]] : memref<?xi32, strided<[?], offset: ?>> // CHECK: ^bb0(%[[VAL_24:.*]]: index):
// CHECK: %[[VAL_25:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_24]]] : memref<?xi32, strided<[?], offset: ?>>
// CHECK: %[[VAL_26:.*]] = arith.extui %[[VAL_25]] : i32 to i64 // CHECK: %[[VAL_26:.*]] = arith.extui %[[VAL_25]] : i32 to i64
// CHECK: %[[VAL_27:.*]] = arith.index_cast %[[VAL_26]] : i64 to index // CHECK: %[[VAL_27:.*]] = arith.index_cast %[[VAL_26]] : i64 to index
// CHECK: %[[VAL_28:.*]] = arith.muli %[[VAL_27]], %[[VAL_3]] : index // CHECK: %[[VAL_28:.*]] = scf.while (%[[VAL_29:.*]] = %[[VAL_24]]) : (index) -> index {
// CHECK: %[[VAL_29:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_20]]] : memref<?xi32, strided<[?], offset: ?>> // CHECK: %[[VAL_30:.*]] = arith.cmpi ult, %[[VAL_29]], %[[VAL_20]] : index
// CHECK: %[[VAL_30:.*]] = arith.extui %[[VAL_29]] : i32 to i64 // CHECK: %[[VAL_31:.*]] = scf.if %[[VAL_30]] -> (i1) {
// CHECK: %[[VAL_31:.*]] = arith.index_cast %[[VAL_30]] : i64 to index // CHECK: %[[VAL_32:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_29]]] : memref<?xi32, strided<[?], offset: ?>>
// CHECK: %[[VAL_32:.*]] = arith.addi %[[VAL_28]], %[[VAL_31]] : index // CHECK: %[[VAL_33:.*]] = arith.extui %[[VAL_32]] : i32 to i64
// CHECK: %[[VAL_33:.*]] = tensor.extract %[[VAL_6]]{{\[}}%[[VAL_13]], %[[VAL_32]]] : tensor<5x6xf32> // CHECK: %[[VAL_34:.*]] = arith.index_cast %[[VAL_33]] : i64 to index
// CHECK: %[[VAL_34:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_20]]] : memref<?xf32> // CHECK: %[[VAL_35:.*]] = arith.cmpi eq, %[[VAL_34]], %[[VAL_27]] : index
// CHECK: %[[VAL_35:.*]] = arith.mulf %[[VAL_24]], %[[VAL_34]] : f32 // CHECK: scf.yield %[[VAL_35]] : i1
// CHECK: %[[VAL_36:.*]] = arith.addf %[[VAL_33]], %[[VAL_35]] : f32 // CHECK: } else {
// CHECK: memref.store %[[VAL_36]], %[[VAL_12]]{{\[}}%[[VAL_13]], %[[VAL_32]]] : memref<5x6xf32> // CHECK: scf.yield %[[VAL_2]] : i1
// CHECK: }
// CHECK: scf.condition(%[[VAL_36:.*]]) %[[VAL_29]] : index
// CHECK: } do {
// CHECK: ^bb0(%[[VAL_37:.*]]: index):
// CHECK: %[[VAL_38:.*]] = arith.addi %[[VAL_37]], %[[VAL_6]] : index
// CHECK: scf.yield %[[VAL_38]] : index
// CHECK: }
// CHECK: %[[VAL_39:.*]] = tensor.extract %[[VAL_0]]{{\[}}%[[VAL_14]], %[[VAL_27]]] : tensor<5x6xf32>
// CHECK: scf.for %[[VAL_40:.*]] = %[[VAL_24]] to %[[VAL_41:.*]] step %[[VAL_6]] {
// CHECK: %[[VAL_42:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_40]]] : memref<?xi32, strided<[?], offset: ?>>
// CHECK: %[[VAL_43:.*]] = arith.extui %[[VAL_42]] : i32 to i64
// CHECK: %[[VAL_44:.*]] = arith.index_cast %[[VAL_43]] : i64 to index
// CHECK: %[[VAL_45:.*]] = arith.muli %[[VAL_44]], %[[VAL_4]] : index
// CHECK: %[[VAL_46:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_40]]] : memref<?xi32, strided<[?], offset: ?>>
// CHECK: %[[VAL_47:.*]] = arith.extui %[[VAL_46]] : i32 to i64
// CHECK: %[[VAL_48:.*]] = arith.index_cast %[[VAL_47]] : i64 to index
// CHECK: %[[VAL_49:.*]] = arith.addi %[[VAL_45]], %[[VAL_48]] : index
// CHECK: %[[VAL_50:.*]] = tensor.extract %[[VAL_7]]{{\[}}%[[VAL_14]], %[[VAL_49]]] : tensor<5x6xf32>
// CHECK: %[[VAL_51:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_40]]] : memref<?xf32>
// CHECK: %[[VAL_52:.*]] = arith.mulf %[[VAL_39]], %[[VAL_51]] : f32
// CHECK: %[[VAL_53:.*]] = arith.addf %[[VAL_50]], %[[VAL_52]] : f32
// CHECK: memref.store %[[VAL_53]], %[[VAL_13]]{{\[}}%[[VAL_14]], %[[VAL_49]]] : memref<5x6xf32>
// CHECK: } {"Emitted from" = "linalg.generic"} // CHECK: } {"Emitted from" = "linalg.generic"}
// CHECK: scf.yield %[[VAL_54:.*]] : index
// CHECK: } attributes {"Emitted from" = "linalg.generic"}
// CHECK: } {"Emitted from" = "linalg.generic"} // CHECK: } {"Emitted from" = "linalg.generic"}
// CHECK: %[[VAL_37:.*]] = bufferization.to_tensor %[[VAL_12]] : memref<5x6xf32> // CHECK: %[[VAL_55:.*]] = bufferization.to_tensor %[[VAL_13]] : memref<5x6xf32>
// CHECK: %[[VAL_38:.*]] = tensor.expand_shape %[[VAL_37]] {{\[\[}}0], [1, 2]] : tensor<5x6xf32> into tensor<5x2x3xf32> // CHECK: %[[VAL_56:.*]] = tensor.expand_shape %[[VAL_55]] {{\[\[}}0], [1, 2]] : tensor<5x6xf32> into tensor<5x2x3xf32>
// CHECK: %[[VAL_39:.*]] = tensor.cast %[[VAL_38]] : tensor<5x2x3xf32> to tensor<?x?x?xf32> // CHECK: %[[VAL_57:.*]] = tensor.cast %[[VAL_56]] : tensor<5x2x3xf32> to tensor<?x?x?xf32>
// CHECK: return %[[VAL_39]] : tensor<?x?x?xf32> // CHECK: return %[[VAL_57]] : tensor<?x?x?xf32>
// CHECK: } // CHECK: }
func.func @sparse_reshape_fused(%arg0: tensor<5x6xf32>, %arg1: tensor<6x2x3xf32, #COO_3D>) -> tensor<?x?x?xf32> { func.func @sparse_reshape_fused(%arg0: tensor<5x6xf32>, %arg1: tensor<6x2x3xf32, #COO_3D>) -> tensor<?x?x?xf32> {
%collapsed = tensor.collapse_shape %arg1 [[0], [1, 2]] : tensor<6x2x3xf32, #COO_3D> into tensor<6x6xf32, #COO_2D> %collapsed = tensor.collapse_shape %arg1 [[0], [1, 2]] : tensor<6x2x3xf32, #COO_3D> into tensor<6x6xf32, #COO_2D>
%0 = tensor.empty() : tensor<5x6xf32> %0 = tensor.empty() : tensor<5x6xf32>
%2 = linalg.matmul ins(%arg0, %collapsed : tensor<5x6xf32>, tensor<6x6xf32, #COO_2D>) outs(%0 : tensor<5x6xf32>) -> tensor<5x6xf32> %2 = linalg.matmul ins(%arg0, %collapsed : tensor<5x6xf32>, tensor<6x6xf32, #COO_2D>) outs(%0 : tensor<5x6xf32>) -> tensor<5x6xf32>
%expanded = tensor.expand_shape %2 [[0], [1, 2]] : tensor<5x6xf32> into tensor<5x2x3xf32> %expanded = tensor.expand_shape %2 [[0], [1, 2]] : tensor<5x6xf32> into tensor<5x2x3xf32>
%ret1 = tensor.cast %expanded : tensor<5x2x3xf32> to tensor<?x?x?xf32> %ret1 = tensor.cast %expanded : tensor<5x2x3xf32> to tensor<?x?x?xf32>
return %ret1 : tensor<?x?x?xf32> return %ret1 : tensor<?x?x?xf32>
} }