[Datatypes] Custom datatypes #2900

gussmith23 · 2019-03-26T20:21:26Z

No description provided.

include/tvm/runtime/packed_func.h

src/codegen/datatype/datatype_registry.cc

src/codegen/llvm/codegen_llvm.cc

src/pass/lower_datatypes.cc

src/codegen/datatype/datatype_registry.cc

src/codegen/datatype/datatype_registry.h

src/codegen/llvm/codegen_llvm.cc

include/tvm/ir_operator.h

include/tvm/ir_pass.h

src/runtime/type_conversions.cc

python/tvm/datatype.py

CMakeLists.txt

gussmith23 · 2019-04-21T01:52:43Z

Revert accidental submodule edits.
Get it building (master seems to be broken currently)

tqchen · 2019-04-22T00:30:36Z

Some high-level comments:

Let us start to think about a proper namespace for this feature
- custom_datatypes vs datatype: the later one might be better to indicate this is a centralized utility to handle data types(including providing into for build-in ones)
Let us put bfloat's impl to 3rdparty, as it is taken from TF codebase, see https://github.com/dmlc/tvm/tree/master/3rdparty/compiler-rt

Please consider starting to request code reviews from reviewers when we have a ready to go testcase

gussmith23 · 2019-04-23T22:39:35Z

@tqchen Should the change from custom_datatypes to datatype be universal? based on some comments you had made previously, I changed some things from "datatype" to "custom datatypes", and then I decided to just make everything 100% consistent across the whole change. Do you have any feeling for where we should be sure to mention "custom"?

gussmith23 · 2019-04-25T23:01:30Z

mention Universal library

tqchen · 2019-04-26T03:18:01Z

I think most cases we only need to mention datatype

In the past, e.g. when lowering custom datatypes for an Add, we would lower a and b first, and then decide whether the resulting new Add needed to be lowered based on the (new) types of a and b. Now, instead, we need to check the types of a and b first (to see if they're custom types), and then lower them (so they'll become non-custom types), and then lower the new Add.

This reverts commit d554a58. This broke parsing. Will figure this out later. There isn't a really clean way to separate this out given how the rest of the function is written.

src/codegen/datatype/registry.h

tests/python/unittest/test_custom_datatypes_mybfloat16.py

tqchen · 2019-05-13T21:16:27Z

@weberlo @jroesch @tmoreau89 @Ravenwater please take another round of look and https://docs.tvm.ai/contribute/code_review.html#approve-and-request-changes-explicitly

gussmith23 · 2019-05-14T21:23:15Z

@weberlo thanks for comments; those are now resolved.

@jroesch @tmoreau89 @Ravenwater when you can: https://docs.tvm.ai/contribute/code_review.html#approve-and-request-changes-explicitly

gussmith23 · 2019-05-15T18:28:04Z

Closed and reopened to re-run CI.

* Register and use custom datatypes in TVM This patch adds the ability to register and use a custom datatype from Python, using the `register_datatype` call. The datatype can then be passed as the `dtype` parameter using the syntax `dtype="custom[<type_name>]bitsxlanes"`. * Removes extra file * Register custom datatypes with TVM; specify Cast and Add lowering This commit adds functionality for registering custom datatypes with TVM, and furthermore adding custom lowering functions to lower those custom datatypes. This commit only adds lowering for the Cast and Add ops; more ops will be added soon. Check out some custom datatype samples in my repository of samples: https://github.com/gussmith23/tvm-custom-datatype-samples * Register and lower casts from Python * Formatting * Fix include; was including too much * Add comment * Add DatatypeRegistered * Add storage size field to custom datatypes This field indicates the bitwidth of the opaque block of data into which instances of the datatype will be stored, when TVM compiles. For example, if I create a datatype with a storage size of 16, then - Constants of that datatype will be created as unsigned 16-bit ints - Calls to external functions taking that datatype will pass the data as unsigned 16-bit ints - External functions returning that datatype will be assumed to return unsigned 16-bit ints. * Change how lowering funcs (Cast and other ops) are named in registry tvm.datatypes.lower.<target>.cast.<dst-type>.<src-type> becomes tvm.datatypes.lower.<target>.Cast.<dst-type>.<src-type> And fixes some sloppy code around how the other ops were being formatted. * Update Python register_datatype to accept storage size * Oops, left out one cast->Cast change * Look up storage size when parsing `custom[typename]` When we encounter this type string in Python, it will be parsed into a Halide type object in C++. Some of my original code supported this parsing, but we now have to attach the storage type to the type (by setting the bits field). * Change how external calls for casting/other ops are done Firstly, we now use the storage size of the custom type when determining input/output types; e.g. a cast to a custom type with storage size 16 is seen as a call to an external function returning an opaque uint of size 16. Secondly, write a macro to handle the other ops. Originally I thought I could handle these at runtime, with a single `_register_op` global. I transitioned instead to using individual `_register_Add` etc. calls generated with a macro, but I don't remember why. * When encountering a custom type immediate, generate UIntImm * Translate custom types to LLVM type * Generate correct return type in Casts Originally I was assuming that the result type from casts was always a custom datatype, and so I was making the Call return a UInt type. * Use TVM-idiomatic recursion style in DatatypesLowerer This was actually a bug, I'm pretty sure; we wouldn't have recursed deep on any complex programs. As a result of making this change, I also uncovered another potential bug, where the datatypes lowering pass would attempt to lower a Load of a custom type. By commenting out the `Mutate_` for Load, I was able to stop the error from cropping up, but frankly, I'm not satisfied with the solution; how is it that we are able to run codegen when Loads of custom datatypes are present in the IR? I have not written any code, to my knowledge, that will support this. Perhaps Load does not care about the underlying datatype? * Use CHECK * Add comment about which Mutate_s are needed * Add comments * Add GetCustomDatatypeRegistered as an extern C function * Formatting, comments, casting * Change how datatype string is formatted * Use bits() instead of GetStorageSize Use bits() instead of GetStorageSize * Change comment * Add datatype.py * Change registered function name (datatypes->datatype) * Remove GetStorageSize * Format custom datatypes like any other datatype Specifically, we now print the bits and lanes after the `custom[...]` string. * Correctly implement datatype lowering in Python * Remove unneeded include * Make function naming consistent * Use CHECK instead of internal_assert * Rename macro * Formatting * Rename functions * Implement Cast lowering `_datatype_register_op` is now able to lower both binary ops and Casts. * Formatting * Formatting * Clang format, google style * Fix std::string/extern "C" warnings * Formatting * Formatting * Lower Allocates and Loads during datatype lowering This should ensure that there are no custom datatypes remaining once datatype lowering is done. This will allow us to remove the code in the LLVM codegen which deals with custom datatypes. * Revert additions to codegen_llvm.cc which are now unneeded * Pass cpplint on lower_datatypes.cc * Add clarifying comment * Remove datatype lowering registration funcs from C++ * Add CHECKs * Remove TODO * Remove all references to storage size * Move and rename function * Rename function * Remove done TODOs and other handled comments * Remove irrelevant Load code and comments * Comment out the IR node types I'm not sure about yet * Add bfloat16 datatype unittest * Fix MakeConstScalar MakeConstScalar for a custom datatype will now call out to a function which can be registered on a per-datatype basis. The function will take a double and return the equivalent value in the custom datatype format. Note that these code paths are not actually used or tested at the moment. I have not yet written an example which uses const scalars of a custom datatype. * Formatting * Change pass name * Allow users to register whatever lowering function they want Tianqi pointed out that users should be able to register whatever lowering function they want, and should not be constrained to registering lowering functions which just call out to external libraries. I still provide a function for making lowering functions which call out to external libraries, for convenience. * Add clarifying comment * Remove unneeded comment * Remove unneeded function * Rename file * Undo unnecessary change * Undo unnecessary change * Make naming consistent Rename "datatypes" to "custom datatypes" in most contexts. * Revert an artifact of old code * Fix build warnings, add TODO * Lint * Remove unnecessary use of extern C by separating decl and impl * Error checking * Remove TODO * Missed a name change * Lint * Python lint * Correctly format datatype * Move bfloat16 to 3rdparty * "custom_datatypes" --> "datatype" in most places I left the pass as "LowerCustomDatatypes" to indicate that we're not lowering anything other than custom datatypes. Otherwise, everything else has been changed. * Upgrade datatype unittest I used a float calculator to generate some real testcases for the unittest. * Separate public includes and private implementation Specifically, create cleaner decoupling between datatypes stuff in packed_func and the datatype registry implementation. * Formatting * Limit custom datatype codes to >128 * Add TODOs * Fix comment * Formatting * Clean up datatype unittest * Remove un-exported functions in public headers; UIntImm->FloatImm More places where I accidentally was using implementation-only functions in public headers. Additionally, store custom datatype immediates as FloatImms. A later change will add new lowering logic to lower these FloatImms to UIntImms. Plus formatting change. * Lint * Use FloatImm (not UIntImm) to hold immediates of custom datatypes This change switches from using UIntImm to FloatImm for storing immediates of custom datatypes. The value of the number is stored in a double, which should be enough precision for now, for most custom types we will explore in the immediate future. In line with this change, we change the datatype lowering so that FloatImms are lowered to UInts of the appropriate size. Originally, this was going to be done by allowing the user to register a double->uint_<storage size>_t conversion which would be called at compile time to convert the value from the FloatImm to a UInt and store it in a UIntImm. After discussions with Tianqi, we decided to take the simpler route, and lower FloatImms just as we lower all other ops: by replacing them with Call nodes. In this case, presumably the user will Call out to a conversion function in their datatype library. The justification for this decision is due to the functionality added in apache#1486. This pull request adds the ability to load LLVM bytecode in at compile time. This applies in our case as follows: 1. The user writes their custom datatype programs and registers their lowering functions in the same way we've been doing it so far. All operations over custom datatypes are lowered to Calls to the datatype library. 2. The user compiles their datatype library to LLVM bytecode. 3. At TVM compile time, the user loads the LLVM bytecode. Depending on how the datatype library is written, Clang should be able to perform constant folding over the custom datatype immediates, even if their conversions are done with calls to the library. Additionally adds test to test the FloatImm codepath. * Re-add a change I removed accidentally during rebase * Cleanup * Remove unnecessary TVM_DLLs * Add custom datatype utilities source file to Go runtime pack * Revert "Remove unnecessary TVM_DLLs" This reverts commit 4b742b9. * Mark bfloat code as TVM_DLL * Moves custom datatype runtime utilities to c_runtime_api.cc * Revert "Add custom datatype utilities source file to Go runtime pack" This reverts commit aecbcde. * Move datatype parsing to its own function * Change comments * Remove unneeded function * Formatting * Formatting * Documentation * Add kCustomBegin, use it for checking for custom types * Documentation * Formatting * Move static definition to implementation * Remove comment * Decide toBeLowered before lowering arguments of Expr In the past, e.g. when lowering custom datatypes for an Add, we would lower a and b first, and then decide whether the resulting new Add needed to be lowered based on the (new) types of a and b. Now, instead, we need to check the types of a and b first (to see if they're custom types), and then lower them (so they'll become non-custom types), and then lower the new Add. * Revert "Move datatype parsing to its own function" This reverts commit d554a58. This broke parsing. Will figure this out later. There isn't a really clean way to separate this out given how the rest of the function is written. * Replace comment * Documentation * Remove comment and TVM_DLL * Better error messages * Remove artifact of rebase * Separate datatypes parsing to its own function * Add \returns * Comment changes; add TODO * Refactor tests