Triton

Purpose

Guide agents through writing GPU kernels in OpenAI Triton: the @triton.jit decorator, block-oriented tl.load/tl.store with masking, atomic operations, shared memory via tl.constexpr, benchmarking with triton.testing.Benchmark, PyTorch integration, and debugging with barriers.

When to Use

• Writing custom PyTorch ops faster than pure PyTorch but without raw CUDA
• Prototyping fused kernels (e.g., softmax + scale + bias)
• Comparing block sizes and warp counts with Triton's autotuner
• Porting NumPy-style elementwise ops to GPU
• Learning GPU programming with higher-level Python syntax
• Benchmarking kernel variants systematically

Workflow

1. Minimal Triton kernel

import torch
import triton
import triton.language as tl

@triton.jit
def add_kernel(x_ptr, y_ptr, out_ptr, n, BLOCK: tl.constexpr):
    pid = tl.program_id(0)
    offsets = pid * BLOCK + tl.arange(0, BLOCK)
    mask = offsets < n
    x = tl.load(x_ptr + offsets, mask=mask)
    y = tl.load(y_ptr + offsets, mask=mask)
    tl.store(out_ptr + offsets, x + y, mask=mask)

def add(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
    n = x.numel()
    out = torch.empty_like(x)
    grid = lambda meta: (triton.cdiv(n, meta["BLOCK"]),)
    add_kernel[grid](x, y, out, n, BLOCK=1024)
    return out

Key concepts:

• tl.program_id(0) — block index (like blockIdx.x)
• tl.arange(0, BLOCK) — vector of thread indices within block
• mask — predication for tail elements (no separate bounds kernel)
• BLOCK: tl.constexpr — compile-time constant, enables unrolling

2. Load/store and masking

@triton.jit
def masked_load_example(ptr, n, BLOCK: tl.constexpr):
    pid = tl.program_id(0)
    offs = pid * BLOCK + tl.arange(0, BLOCK)
    mask = offs < n
    # masked load returns 0 for masked-off lanes
    vals = tl.load(ptr + offs, mask=mask, other=0.0)
    return vals

Block pointers (Triton 2.x+) for structured 2D access:

@triton.jit
def matvec_kernel(a_ptr, x_ptr, y_ptr, M, N, BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr):
    pid_m = tl.program_id(0)
    offs_m = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
    acc = tl.zeros((BLOCK_M,), dtype=tl.float32)
    for start_n in range(0, N, BLOCK_N):
        offs_n = start_n + tl.arange(0, BLOCK_N)
        a = tl.load(a_ptr + offs_m[:, None] * N + offs_n[None, :])
        x = tl.load(x_ptr + offs_n)
        acc += tl.sum(a * x[None, :], axis=1)
    tl.store(y_ptr + offs_m, acc, mask=offs_m < M)

3. Atomic operations

@triton.jit
def atomic_histogram(data_ptr, hist_ptr, n, BLOCK: tl.constexpr):
    pid = tl.program_id(0)
    offs = pid * BLOCK + tl.arange(0, BLOCK)
    mask = offs < n
    data = tl.load(data_ptr + offs, mask=mask)
    bucket = (data % 256).to(tl.int32)
    tl.atomic_add(hist_ptr + bucket, 1, mask=mask)

Use atomics sparingly — they serialize memory updates. Prefer block-level reduction then single atomic per block.

4. Shared memory via constexpr

@triton.jit
def reduce_kernel(x_ptr, out_ptr, n, BLOCK: tl.constexpr):
    pid = tl.program_id(0)
    offs = pid * BLOCK + tl.arange(0, BLOCK)
    mask = offs < n
    x = tl.load(x_ptr + offs, mask=mask, other=0.0)
    # Block reduction
    x = tl.sum(x, axis=0)
    tl.atomic_add(out_ptr, x)

BLOCK as tl.constexpr lets the compiler allocate shared memory and unroll loops at compile time.

5. Autotuning

@triton.autotune(
    configs=[
        triton.Config({"BLOCK": 128}, num_warps=4),
        triton.Config({"BLOCK": 256}, num_warps=4),
        triton.Config({"BLOCK": 512}, num_warps=8),
    ],
    key=["n"],
)
@triton.jit
def tuned_kernel(x_ptr, y_ptr, out_ptr, n, BLOCK: tl.constexpr):
    # ... kernel body ...
    pass

Autotuner benchmarks configs on first run and caches the best for each key shape.

6. Benchmarking

from triton.testing import Benchmark

def benchmark_add():
    n = 1024 * 1024
    x = torch.randn(n, device="cuda")
    y = torch.randn(n, device="cuda")

    def triton_add():
        return add(x, y)

    def torch_add():
        return x + y

    bench = Benchmark(
        x_names=["n"],
        x_vals=[2**i for i in range(10, 24)],
        line_arg="provider",
        line_vals=["triton", "torch"],
        line_names=["Triton", "PyTorch"],
        plot_name="add-bench",
        args={},
    )
    bench.run(lambda n, provider: {
        "triton": lambda: add(x[:n], y[:n]),
        "torch": lambda: x[:n] + y[:n],
    }[provider](), quantiles=[0.5, 0.9])

# Quick timing in REPL
import triton.testing as tt
ms = tt.do_bench(lambda: add(x, y))
print(f"{ms:.3f} ms")

7. PyTorch integration

import torch
from torch.library import custom_op

@custom_op("mylib::triton_add", mutates_args=())
def triton_add_op(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
    return add(x, y)

@triton_add_op.register_fake
def _(x, y):
    return torch.empty_like(x)

# Use in model
class MyModule(torch.nn.Module):
    def forward(self, x, y):
        return triton_add_op(x, y)

For torch.compile compatibility, register fake/meta kernels and avoid Python side effects in the JIT function.

8. Debugging

@triton.jit
def debug_kernel(x_ptr, n, BLOCK: tl.constexpr):
    pid = tl.program_id(0)
    offs = pid * BLOCK + tl.arange(0, BLOCK)
    mask = offs < n
    x = tl.load(x_ptr + offs, mask=mask)
    # Synchronize threads within block for inspection
    tl.debug_barrier()
    tl.store(x_ptr + offs, x * 2.0, mask=mask)

# Dump generated PTX/LLVM IR
TRITON_PRINT_AUTOTUNING=1 python script.py
# Set cache dir to inspect compiled kernels
export TRITON_CACHE_DIR=/tmp/triton_cache

Compare against PyTorch reference on small inputs before scaling up.

Common Problems

Related Skills

• skills/gpu/cuda — underlying CUDA concepts when Triton limits are hit
• skills/gpu/cuda-profiling — Nsight profiling of Triton-compiled kernels
• skills/gpu/gpu-memory-model — coalescing and occupancy theory
• skills/gpu/hip-rocm — AMD GPU path (Triton supports ROCm)
• skills/hpc/openmp — CPU parallelism alongside GPU kernels