| name | add-ttir-d2m-lowering |
| description | Elementwise TTIR→D2M→TTMetal path: tablegen, TTIRToD2M.cpp, D2MToTTKernel.cpp, and — only when the kernel API callee is new — TTKernelIncludesMap.h (per-op api/compute/eltwise_unary/*.h mapping for JIT). Does not edit D2MGenericRegionOps.cpp or TTKernelToCpp.cpp. Not for reductions, matmul, views, or CCL. |
TTIR elementwise → D2M (TTMetal path)
Allowed edits (these layers):
-
Tablegen — e.g. include/ttmlir/Dialect/D2M/IR/D2MGenericRegionOps.td (and any other .td you
already own for the op). Pick the same base class as the nearest op (unary:
D2M_GenericRegionComputeUnaryDstOp; typical binary: …FPUOrSFPUBinary; ternary: …TernaryDstOp).
Prefer ops that need no hand-written C++ in D2MGenericRegionOps.cpp; that file is out of
scope for this workflow.
-
lib/Conversion/TTIRToD2M/TTIRToD2M.cpp — in populateTTIRToD2MPatterns, add one line to the
big patterns.add< … > list with the other elementwise rewriters, e.g.
D2MNamedElementwiseRewriter<ttir::YourOp, d2m::TileYourOp>, (keep ordering consistent with
neighbors). Use notifyMatchFailure inside patterns, not emitOpError.
-
lib/Conversion/D2MToTTKernel/D2MToTTKernel.cpp — extend ComputeOpMap / IntComputeOpMap and
the patterns.add<…D2MSFPUOpsRewriter…> list to match the nearest unary/binary tile op.
If the TTKernel op takes i32-encoded scalar params (float attrs bit-reinterpreted, or int
attrs, or a runtime scalar Value), reuse the shared helpers defined at the top of the anonymous
namespace rather than re-inlining a lambda:
floatAttrToI32Bits(rewriter, loc, attr) — FloatAttr → i32 bits (e.g. selu scale/alpha,
clamp_scalar float min/max).intAttrToI32(rewriter, loc, attr) — IntegerAttr → sign-extended i32 (e.g. clamp_scalar int
min/max).scalarToI32Bits(rewriter, loc, value) — runtime scalar Value → i32 (float widened+bitcast,
int sign-extended/truncated). Used by binop_with_scalar-style scalar rhs lowerings.
Ops with scalar attributes typically need a dedicated else if constexpr (std::is_same_v<SFPUOp, ttkernel::FooTileOp>) branch in the D2MSFPUOpsRewriter body that pulls
attrs off op and calls the shared helper — see the SeluTileOp / ClampScalarTileOp branches
as templates.
-
include/ttmlir/Target/TTKernel/TTKernelIncludesMap.h (only if the kernel API callee is new) —
the ScopedModuleHelper in lib/Target/TTKernel/TTKernelToCpp.cpp no longer hardcodes
api/compute/eltwise_unary/*.h. It walks the region and looks up each emitc.call_opaque
callee in getCalleeToHeadersMap(). If your op lowers to a tt-metal SFPU helper
(foo_tile / foo_tile_init) that isn't already in that map, add entries like:
{"foo_tile", {"api/compute/eltwise_unary/foo.h", ""}},
{"foo_tile_init", {"api/compute/eltwise_unary/foo.h", ""}},
The callee string must match the TTKernel_SFPUOp<"foo_tile", …> / TTKernel_InitOp<"foo_tile_init">
name in TTKernelOps.td exactly. Do not edit TTKernelToCpp.cpp to add includes directly —
the old unconditional emitc::IncludeOp block was removed. Without a map entry, wormhole JIT can
fail with "foo_tile was not declared in this scope" in chlkc_unpack.cpp.
Out of scope here: D2MGenericRegionOps.cpp, TTKernelToCpp.cpp. For TTNN / flatbuffer / full
builder parity across all targets, use .claude/skills/add-op/SKILL.md.
Tests (minimal): extend existing TTIR→D2M lit at
test/ttmlir/Conversion/TTIRToD2M/named_to_generic.mlir. Chain the new op into the SSA dataflow of
the existing named_elementwise function (bump the %N numbering and add a
// CHECK: d2m.tile_<op> + the ttir.<op> call) — do not create a separate
named_elementwise_* func for the new op. No lit under test/ttmlir/Conversion/D2MToTTKernel/
is required.
Golden (TTMetal-only, no TTNN): add ttir_<op>.mlir under mlir_snippets/ttir/ — one
snippet per new op so test_parse_split_ops.py exercises parse/split for each. Add the golden
in tools/golden/mapping.py and the matching @tag / @parse / @split in
tools/builder/ttir/ttir_builder.py (same pattern as square / exp: pass output_type_mlir
into the golden, no _op_proxy).
For ops that carry MLIR attributes (e.g. SELU's scale / alpha, clamp's min / max), the
golden function should accept the MLIR attr types (FloatAttr, IntegerAttr, …) as
positional arguments and unpack them internally with unpack_mlir_attr — do not give the
golden Python-level defaults that duplicate the tablegen DefaultValuedAttr. The builder
@tag method is allowed to keep Python-float defaults as a caller convenience; just convert them
to FloatAttr.get_f32(...) and pass the FloatAttr directly into the golden (both from @tag
and from @parse, where you already have the attr off old_op). Mirror the
ttnn_clamp_scalar_golden / ttnn_leaky_relu_golden shape for this.
In test/python/golden/ttir_ops/eltwise/test_ttir_unary.py (or sibling), mark the op with
SkipIf("ttnn", "emitc", "emitpy", "sim") so it runs only on ttmetal on silicon until TTNN
lowering exists. SkipIf is already imported from test_utils; prefer it over the more verbose
Marks(pytest.mark.skip_config([...]), …) form.
Run cmake --build build after changes.
Checklist
