[INFA][IR] Build and Evolve Low-level IR. Remove HalideIR dep. #3533
Conversation
| */ | ||
| static Range make_by_min_extent(Expr min, Expr extent); | ||
| // declare range. | ||
| TVM_DEFINE_NODE_REF_METHODS(Range, NodeRef, RangeNode); |
There was a problem hiding this comment.
Can we unify the RELAY and TVM macros into one?
There was a problem hiding this comment.
I agree, likely we will need to use these macros instead in relay and remove relay specific macros.
| }; | ||
|
|
||
| /*! \brief Create a vector where all the elements are value. */ | ||
| class Broadcast : public ExprNode { |
There was a problem hiding this comment.
This might be out of the scope for now, but I think we should use the XNode and X for the Node and NodeRef subclasses, it might be a good time to stop using raw pointers in the functor as well
| static constexpr const char* shift_right = "shift_right"; | ||
| static constexpr const char* popcount = "popcount"; | ||
| static constexpr const char* likely = "likely"; | ||
| static constexpr const char* glsl_texture_store = "glsl_texture_store"; |
There was a problem hiding this comment.
we do have a semi-functioning webgl backend that uses this intrinsic
| }); | ||
|
|
||
| TVM_STATIC_IR_FUNCTOR(IRPrinter, vtable) | ||
| .set_dispatch<Free>([](const Free* op, IRPrinter* p) { |
There was a problem hiding this comment.
I think we should unify both printers and start using a layout algorithm, we already have some lines which are incredibly ugly in current printing strategy, probably out of scope of this PR though.
There was a problem hiding this comment.
There is a lot of work in the printer. I am still trying to wrap up roundtripping.
| } | ||
|
|
||
| Stmt ProducerConsumer::make(FunctionRef func, bool is_producer, Stmt body) { | ||
| CHECK(body.defined()); |
There was a problem hiding this comment.
#define CHECK_DEFINE(x) CHECK(x.defined()) << #x << " is not defined";
There was a problem hiding this comment.
The error message of CHECK will automatically display the content, which is self-explained
There was a problem hiding this comment.
it will be CHECK: body.defined() failed
| }); | ||
|
|
||
| TVM_STATIC_IR_FUNCTOR(IRPrinter, vtable) | ||
| .set_dispatch<Free>([](const Free* op, IRPrinter* p) { |
There was a problem hiding this comment.
There is a lot of work in the printer. I am still trying to wrap up roundtripping.
tqchen
changed the title
Co-Authored-By: Jared Roesch <roeschinc@gmail.com>
Co-Authored-By: Jared Roesch <roeschinc@gmail.com>
|
Thanks @jroesch @MarisaKirisame , this PR is now merged |
|
Is this an effort to replace the Halide IR with something else? Is there a discussion somewhere? |
|
@kparzysz-quic see RFC #3474 |
…e#3533) * [INFA][IR] Build and Evolve Low-level IR. Remove dep from HalideIR. * Update include/tvm/node/ir_functor.h Co-Authored-By: Jared Roesch <roeschinc@gmail.com> * Update include/tvm/node/ir_functor.h Co-Authored-By: Jared Roesch <roeschinc@gmail.com>
…e#3533) * [INFA][IR] Build and Evolve Low-level IR. Remove dep from HalideIR. * Update include/tvm/node/ir_functor.h Co-Authored-By: Jared Roesch <roeschinc@gmail.com> * Update include/tvm/node/ir_functor.h Co-Authored-By: Jared Roesch <roeschinc@gmail.com>
This parameter is nullable for cases where the else block isn't present. Previously, it was represented as a `Stmt` holding `nullptr`, because `IfThenElse` (apache#3533) predates the `Optional` utility (apache#5314). This commit updates to use `Optional<Stmt>` instead, and updates all usages of `else_case`.
This parameter is nullable for cases where the else block isn't present. Previously, it was represented as a `Stmt` holding `nullptr`, because `IfThenElse` (apache#3533) predates the `Optional` utility (apache#5314). This commit updates to use `Optional<Stmt>` instead, and updates all usages of `else_case`.
This parameter is nullable for cases where the else block isn't present. Previously, it was represented as a `Stmt` holding `nullptr`, because `IfThenElse` (apache#3533) predates the `Optional` utility (apache#5314). This commit updates to use `Optional<Stmt>` instead, and updates all usages of `else_case`.
#3474