toyc

A toy expression compiler with CPU and GPU (Vulkan compute) backends, written in Python. Source text is lexed, parsed to an AST, and executed either by a tree-walking CPU evaluator or by a compute shader acting as a register machine — then both results can be compared.

Install

pip install toyc

You also need a Vulkan driver plus one GLSL→SPIR-V compiler (only used to build the bundled shaders, automatically at first run):

sudo apt install glslang-tools   # glslangValidator (preferred)
# or
sudo apt install glslc           # shaderc (fallback)

Requires Python 3.10+ and the vulkan + pyopengl packages (installed as dependencies).

The language

Arithmetic expressions with correct precedence: + - * /, parentheses, int/float literals, and variables.

1 + 2 * 3        → 7
(a * b) + (c * d) / (a - b)

Quickstart

from toyc import Cpu, GpuVulkan

Cpu.run("program.toy")          # 15        (CPU tree-walk / bytecode VM)
GpuVulkan.run("program.toy")    # 15.0      (Vulkan compute shader)

How it works (pipeline)

source text ──► Lexer ──► Parser ──► AST ──┬──► Evaluator / Cpu ──► CPU result
                                            │
                                            └──► Flattener ──► flat bytecode
                                                  [op, dest, src1, src2] ──► GPU ──► GPU result

CPU backend

from toyc import Cpu, Compiler, Lexer, Parser

Cpu.run("program.toy")                          # lex+parse+compile+run, prints 15
Cpu.run("program.toy", env={"a": 1})            # variable bindings
Cpu.run("program.toy", silent=True)             # capture without printing
Cpu.run("out.toyc")                             # run a compiled file
Compiler.compile("program.toy")                 # write program.toyc to disk

tokens = Lexer.tokenize("a + b * 2")
ast    = Parser.parse(tokens)

Cpu.run accepts a .toy path, a .toyc path, or an in-memory list[Instr].

GPU backend (single value)

from toyc import GpuVulkan

GpuVulkan.run("program.toy")                    # 15.0
GpuVulkan.run("program.toy", env={"a": 10.0, "b": 5.0})
GpuVulkan.run("program.toy", silent=True)       # capture without printing
GpuVulkan.run("program.toyc")                   # cached 4-byte result, no dispatch
GpuVulkan.compile("program.toy")                # evaluate + cache program.toyc

How a single run works: the AST is flattened to [n_instrs, n_consts, n_vars, op, dest, src1, src2, …] plus a constant pool, your env values, and an output + error slot. That buffer is uploaded to one storage buffer, one compute workgroup runs the comp shader (a 256-register interpreter loop), and the output float is mapped back. The .toyc file written next to the source holds the cached 4-byte result, so run("program.toyc") decodes it without touching the GPU.

Session lifetime

Creating the Vulkan instance + device + pipelines costs ~15–40 ms; a dispatch costs ~1 ms. So the backend keeps one shared session per process instead of rebuilding it per call:

GpuVulkan.startup()    # build the session now (optional; first call does it lazily)
GpuVulkan.run(...)     # reuses the live session — never tears it down
GpuVulkan.shutdown()   # release it (also happens automatically at process exit)

The session is guarded by a lock, so concurrent run() calls from threads share it safely. Recompiling the shaders transparently rebuilds the pipelines (the cache key includes both .spv mtimes).

GPU backend (batch — where the GPU wins)

run_batch evaluates one expression for many variable sets with one dispatch per chunk — one shader invocation per instance (local_size_x = 64, comp_batch shader):

envs = [{"a": 1.0, "b": 2.0}, {"a": 3.0, "b": 4.0}, ...]  # N sets
results = GpuVulkan.run_batch("prog.toy", envs)          # [3.0, 7.0, ...]

• The AST is flattened once; the variable pool holds all N sets (instance-major) and the shader writes N outputs. Bigger envs lists are split into chunks of the recommended batch size (gpu_detect.batch, e.g. 1,638 on an 8 GB card) with a VRAM-budget safety cap — never silently wrong.
• Returns one float per instance, in input order.
• Same error contract as everywhere: NameError for a variable missing from any set, ZeroDivisionError if any instance divides by zero (signalled through a shared error slot, since a shader cannot throw).

Rule of thumb (RX 580): single expressions → CPU (~0.1 ms vs ~1 ms GPU dispatch + overhead); bulk evaluation → run_batch, where the fixed per-call overhead amortises and throughput wins (~10× at N=500 and growing).

Timing, debug output, and benchmarking

Every entry point takes timed=False (print a report + return timing data as a second value) and the GPU calls also take debug=False (print device/VRAM diagnostics):

cpu_val, cpu_t = Cpu.run("program.toy", timed=True)
# [CPU timing] total=0.105 ms (resolve=0.099 ms, execute=0.006 ms)

gpu_raw, gpu_t = GpuVulkan.run("program.toy", timed=True)
# [GPU timing] total=0.843 ms (flatten=0.069 ms, select=0.010 ms,
#              init=0.027 ms, exec=0.666 ms, teardown=0.000 ms, decode=0.000 ms)

Timing dicts hold seconds (perf_counter); stages that didn't run are 0.0. Untimed calls return exactly what they always did, so existing code is unaffected.

toyc.bench — table-ready profiling for pandas

Dependency-free (stdlib only). Every call returns plain row-dicts — one per run — that drop straight into a DataFrame:

from toyc.bench import profile, compare, compare_batch, to_csv

a = 10
b = 20
rows = compare(["1 + 2 * 7", "a + b * 2"], repeats=5)
# env auto-collected from your variables — no envs= needed

import pandas as pd
df = pd.DataFrame(rows)
df.groupby("backend")["total"].mean().plot.bar()   # CPU vs GPU

• env_sets accepts a list of env dicts or a column dict of equal-length value lists: {"a": [1, 3], "b": [2, 4]}.
• max_err is the max abs GPU-vs-CPU difference in that repeat (0.0 on CPU rows) — the table proves the GPU results, so vectors aren't stored per row.
• Single-value result is the decoded float; batch vectors collapse to max_err (see above).
• The GPU session is warmed up before repeats, per-run prints are suppressed unless verbose=True, and temp .toy files are cleaned up automatically.

Plotting later (matplotlib — yours to add)

The library ships no plotters on purpose. When you're ready, the hook is the row table — seconds in total/per_eval, one row per repeat:

import matplotlib.pyplot as plt
import pandas as pd
from toyc.bench import compare_batch

rows = compare_batch("a + b * 2",
                     {"a": list(range(2000)), "b": list(range(2000))},
                     repeats=3)
df = pd.DataFrame(rows)
df["ms"] = df["total"] * 1e3
df.groupby("backend")["ms"].mean().plot.bar()   # cpu vs gpu-batch
plt.ylabel("ms")
plt.show()

Useful columns to plot: total vs n (scaling curve), per_eval (cost per instance), exec (pure dispatch time), max_err (validity). to_csv(rows, "timings.csv") saves anything for later sessions.

Semantics (CPU ≡ GPU, guaranteed)

License

See LICENSE.