dylib open/call/close io functions, initial ffi api

[enricozb]

Overview

Adds DL_OPEN, DL_CALL, and DL_CLOSE IO functions.

C Runtime

Example usage looks like this: A user first defines some C functions they want to invoke through HVM at runtime:

#include "src/hvm.h"
#include <stdio.h>

Port add_two_nums(Net* net, Book* book, Port argm) {
  Tup tup = readback_tup(net, book, argm, 2);
  u32 num1 = get_u24(get_val(tup.elem_buf[0]));
  u32 num2 = get_u24(get_val(tup.elem_buf[1]));

  printf("adding numbers, %u and %u\n", num1, num2);

  return new_port(NUM, new_u24(num1 + num2));
}

Port print_num(Net* net, Book* book, Port argm) {
  u32 num = get_u24(get_val(argm));

  printf("printing number %u\n", num);

  return new_port(ERA, 0);
}

Functions must have the signature Port (my_func)(Net*, Book*, Port).

This file must be compiled as a shared library (.so). For example, gcc -shared my-funcs.c -o my-funcs.so. The file can then be loaded and symbols can be accessed. In Bend this looks like:

def main():
  x = 123
  y = 456

  with IO_T:
    dl <- call_io("DL_OPEN", ("./my-funcs.so", 0))
    res <- call_io("DL_CALL", (dl, "add_two_nums", x, y))
    * <- call_io("DL_CALL", (dl, "print_num", res))
    * <- call_io("DL_CLOSE", dl)

    return 42

C Compiled Mode

When compiling a generated HVM C file, you must use the -rdynamic flag to enable the shared library to access symbols from the main binary. For example,

cargo run -r -- gen-c testing-ffi.hvm > testing-ffi.c
gcc -rdynamic -lm testing-ffi.c -o testing-ffi

CUDA Runtime

The FFI is a little different, the above C file would look like this instead:

#include "src/hvm.cuh"
#include <stdio.h>

Port add_two_nums(GNet* gnet, Port argm) {
  Tup tup = gnet_readback_tup(gnet, argm, 2);
  u32 num1 = get_u24(get_val(tup.elem_buf[0]));
  u32 num2 = get_u24(get_val(tup.elem_buf[1]));

  printf("adding numbers, %u and %u\n", num1, num2);

  return new_port(NUM, new_u24(num1 + num2));
}

Port print_num(GNet* net, Port argm) {
  u32 num = get_u24(get_val(argm));

  printf("printing number %u\n", num);

  return new_port(ERA, 0);
}

And functions must have the signature Port (my_func)(GNet*, Port).

CUDA Compiled Mode

When compiling a generated HVM C file, you must use the -rdynamic flag to the host compiler to enable the shared library to access symbols from the main binary. For example,

cargo run -r -- gen-cu testing-ffi.hvm > testing-ffi.cu
nvcc --compiler-options=-rdynamic testing-ffi.c -o testing-ffi

HVM FFI API

Not everything is exposed to users at the moment, we expose

• numeric readback and creation functions (since these are allocation-free)
• readback_str, readback_tup, readback_bytes, and inject_bytes

See hvm.h for users of the C runtime.
See hvm.cuh for users of the CUDA runtime.

[HigherOrderBot]

Perf run for 0fc5635:

compiled
========

file            runtime         main            (local)       
==============================================================
sort_bitonic    c                        5.53s           4.28s
                cuda                     0.24s           0.23s
--------------------------------------------------------------
sum_rec         c                        1.42s           1.42s
                cuda                     0.14s           0.14s
--------------------------------------------------------------
sum_tree        c                        0.12s           0.13s
                cuda                     0.11s           0.10s
--------------------------------------------------------------
tuples          c                        3.72s           4.16s
                cuda                   timeout         timeout
--------------------------------------------------------------

interpreted
===========

file            runtime         main            (local)       
==============================================================
sort_bitonic    c                        6.48s           4.42s
                cuda                     0.24s           0.24s
                rust                   timeout         timeout
--------------------------------------------------------------
sum_rec         c                        1.83s           2.03s
                cuda                     0.14s           0.13s
                rust                    13.69s          14.10s
--------------------------------------------------------------
sum_tree        c                        0.25s           0.17s
                cuda                     0.08s           0.08s
                rust                     0.83s           0.84s
--------------------------------------------------------------
tuples          c                        2.52s           2.51s
                cuda                   timeout         timeout
                rust                     3.76s           3.82s
--------------------------------------------------------------