Skip to content

Latest commit

 

History

History
107 lines (85 loc) · 4.75 KB

RTLIB.md

File metadata and controls

107 lines (85 loc) · 4.75 KB
Raw

JULEP RTLIB

  • Title: A runtime library for Julia
  • Authors: Valentin Churavy <v.churavy@gmail.com>
  • Created: November 11, 2016
  • Status: work in progress

Introduction

Currently there are two implementation of intrinsics supported in Julia. One of the implementations is defined in runtime_intrinsics.c and the second one is defined in intrinsics.cpp on top of LLVM intrinsics and LLVM instructions.

The first implementation specifies the semantics and behavior of the intrinsics and is used as a fallback. The LLVM based implementation is best understood as a pure performance optimization.

When a LLVM intrinsic is used and the compiler can't generate hardware instructions for it, a library call to the runtime library (compiler-rt or libgcc) is emitted. As a result Julia code using LLVM as a codegen backend needs to link against a runtime library. In the exploratory work (see JuliaLang/julia#18734) the LLVM compiler-rt library was choosen, due to its MIT license.

A drawback of compiler-rt is that it is not fully portable. As an example Float128 support is missing on 32bit platforms.

Motivating problem

How do we support Float16 and Float128 in a portable and performant manner? The current implementation of Float16 support in Julia is eagerly resolving to promotion to Float32 in order to implement most operations. This precludes optimization on platforms that natively support Float16 (most prominently GPUs). The second problem is how are we going to support Float128 across all platforms in a portable and still performant way. On 32bit systems we cannot rely on platform implementations.

Goals

Implement a runtime-library in a mix of C and Julia that contains Julia intrinsics and compiler-rt. This would allow us to use optimized implementations from compiler-rt, while having the flexibility of Julia to implement non-essential intrinsics.

When Julia uses LLVM as a compiler backend it should use a lazy libcall scheme. When the compiler can't emit optimized code paths (e.g. LLVM instructions and intrinsics), it will emit calls to calls to the intrinsics. Currently this done already for better error reporting. The interpreter can eagerly resolve to the implementations in the runtime library, which will require ccallfor the C based part of the runtime library.

The runtime library will consist of two stages. Stage-1 is implemented in C and contains compiler-rt, while Stage-2 is implemented in Julia. Stage-1 is required for bootstrapping a minimal Julia, on which Stage-2 can be implemented. Stage- 1 will also contain optimized (and platform dependent) implementations of the intrinsics, while Stage-2 will contain portable and general implementations.

Another goal is to define what is the minimal set of intrinsics that Julia requires (Stage-1) and what is the extended set (Stage-2). A well defined set of intrinsics would also be beneficial for alternative compilers.

Stage 1

C-based implementation of the essential Julia intrinsics + compiler-rt.

  • jl_reinterpret
  • jl_pointerset
  • jl_pointerref
  • Operations on Integers
    • Arithmetic
    • Comparisons
    • Conversion between integers
  • Operations on Floating Point (hardware based)
    • Arithmetic
    • Comparisons
    • Conversion to and from integers
    • Float32 and Float64 support

Stage 2

Julia based implementation for non-essential Julia intrinsics and implementations to supplement compiler-rt. This implementation will based on a reduced base library (early stages of the sysimage) that will is only allowed to use Stage-1 funcionality. The proper sysimage will be based upon Stage-1 and Stage-2. The basic idea is JuliaLang/julia#18927, which contains the initial port of compiler-rt to Julia.

  • Operations on Floating Point (software based)
    • Arithmetic
    • Comparisons
    • Conversion to and from integers
    • Necessary for Float16 and Float128 support
  • Scalar implementations for vectorized instructions

Building Stage-2

  1. Build Stage-1 and create a shared object file rtlib-stage1.so
  2. Build inference.ji
  3. Build Stage-2 with --rtlib=rtlib-stage1.so and --sysimage inference.ji
  4. Take the object files from Stage-1 and Stage-2 and create rtlib.so containing both Stage-1 and Stage-2.
  5. Build sys.so with --rtlib=rtlib.so and --sysimage inference.ji

Testing and Benchmarking

All implementations, but especially the runtime versions should be thoroughly tested for correctness and performance. As part of this Julep the testsuite needs to be extended to cover the current and future runtime intrinsics.

Non-Goals

  • No support for atomics at this time. Julia will continue to use libatomic.