[torch-mlir][sparse] add decomposition features to sparse compiler (#3505)

Fixes https://github.com/llvm/torch-mlir/issues/3499

python/torch_mlir/extras/fx_decomp_util.py

@@ -49,6 +49,7 @@ DEFAULT_DECOMPOSITIONS = [
     torch.ops.aten.nan_to_num.default,
     torch.ops.aten.unbind,
     torch.ops.aten._scaled_dot_product_flash_attention_for_cpu,
+    torch.ops.aten.diag,
 ]
 
 

test/python/fx_importer/sparse_test.py

@@ -12,6 +12,7 @@ import torch.export
 import torch.nn as nn
 import numpy as np
 
+from torch_mlir.extras.fx_decomp_util import get_decomposition_table
 from torch_mlir.extras.fx_importer import FxImporter
 from torch_mlir.extras.fx_importer import SparsityMeta
 from torch_mlir import ir
@@ -106,6 +107,9 @@ def sparse_export(
     # Build the regular FX traced graph with only dense arguments
     # (the current version would crash otherwise, see issue above).
     prog = torch.export.export(f, dargs, kwargs)
+    decomposition_table = get_decomposition_table()
+    if decomposition_table:
+        prog = prog.run_decompositions(decomposition_table)
     # Annotate sparse arguments in the graph and apply some very
     # basic propagation rules for sparsity.
     specs = prog.graph_signature.input_specs
@@ -120,7 +124,6 @@ def sparse_export(
                     node.meta["sparsity"] = sparse_metadata(args[k])
                 k = k + 1
         elif node.op == "call_function":
-            # TODO: use upstream _opname implementation when available
             opname = node.target._schema.name.split("::")[1]
             # Zero preserving elt-wise unary op.
             if opname in {"abs", "neg", "relu", "sin"}:
@@ -131,7 +134,7 @@ def sparse_export(
                     torch.sparse_coo, 0, dim, 0, None, torch.int64, torch.int64
                 )
             # TODO: Uncomment this to hack sparsity into the network.
-            # elif opname == "_to_dense":
+            # elif opname == "_to_dense" or opname == "to_dense":
             #     # hack (assumes we never really want the to_dense for now)
             #     node.meta["sparsity"] = node.args[0].meta.get("sparsity", None)
             elif opname == "select" and node.args[0].meta.get("sparsity", None):
@@ -176,8 +179,8 @@ def sparse_jit(f, *args, **kwargs):
     compiled = backend.compile(module)
     invoker = backend.load(compiled)
     xargs = []
-    # Prepare the buffer parameters (assume all dense).
+    # Prepare all the named buffer parameters (assume all dense).
-    # TODO: filters out scalar arguments, anything else?
+    # All scalar arguments are filtered out since they appear inline.
     params = dict(f.named_buffers(remove_duplicate=True))
     params_flat, params_spec = torch.utils._pytree.tree_flatten(params)
     for p in params_flat:
@@ -339,6 +342,7 @@ def test_sparse_SpMV():
 
 @run
 #
+# CHECK-LABEL: test_sparse_SpMM
 # CHECK:       #[[$COO:.*]] = #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 : compressed(nonunique), d1 : singleton(soa)), posWidth = 64, crdWidth = 64 }>
 # CHECK:       func.func @main(
 # CHECK-SAME:    %[[A:.*0]]: !torch.vtensor<[8,8],f32,#[[$COO]]>,
@@ -440,7 +444,7 @@ def test_sparse_eltwise():
     # Run it with PyTorch torch.sparse and with TORCH-MLIR sparse_jit.
     res1 = net(sparse_input)
     res2 = sparse_jit(net, sparse_input)
-    # TODO: make this work
+    # TODO: make this work in MLIR
     # res3 = sparse_jit(net, batch_input)
     print("torch.sparse")
     print(res1)
@@ -657,7 +661,14 @@ def test_sparse_network():
 # CHECK:                [0.1321, 0.2724, 0.2105, 0.3851],
 # CHECK:                [0.2478, 0.3439, 0.1898, 0.2185],
 # CHECK:                [0.0222, 0.1683, 0.2928, 0.5167]{{\]}})
+#
+# TODO: first row looks suspect...
+#
 # CHECK: torch.mlir
+# CHECK:   {{\[}}[0.         0.         0.         0.        ]
+# CHECK:         [0.13205223 0.27236593 0.21051763 0.38506418]
+# CHECK:         [0.24781987 0.34391665 0.18976606 0.2184974 ]
+# CHECK:         [0.02224578 0.16825409 0.29283574 0.51666445]{{\]}}
 #
 def test_sparse_feature_scaling():
     class Scale(nn.Module):
@@ -678,11 +689,11 @@ def test_sparse_feature_scaling():
 
     # Run it with PyTorch torch.sparse and with TORCH-MLIR sparse_jit.
     res1 = net(f)
-    # TODO: make this work
+    res2 = sparse_jit(net, f)
-    # res2 = sparse_jit(net, f)
     print("torch.sparse")
     print(res1)
     print("torch.mlir")
+    print(res2)
 
 
 @run