TRITON_REPO/python/tutorials/
├── 01-vector-add.py             # Triton 基础: @triton.jit, program_id, load/store
├── 02-fused-softmax.py          # 内核融合, reduction, tl.max/tl.sum/tl.exp
├── 03-matrix-multiplication.py  # Block matmul, L2 cache, @triton.autotune
├── 04-low-memory-dropout.py     # 并行 PRNG, tl.rand, seed-based dropout
├── 05-layer-norm.py             # Backward pass, atomic ops, tl.atomic_cas
├── 06-fused-attention.py        # Flash Attention v2, causal mask, FP8, warp spec
├── 07-extern-functions.py       # libdevice 外部函数调用
├── 08-grouped-gemm.py           # Group GEMM, TMA, tensor descriptors
├── 09-persistent-matmul.py      # 持久化内核, TMA, warp specialization, FP8
├── 10-block-scaled-matmul.py    # FP4/FP8, MXFP4, tl.dot_scaled
└── 11-programmatic-dependent-launch.py  # PDL, gdc_wait, gdc_launch_dependents

TRITON_REPO/python/tutorials/gluon/
├── 01-intro.py                  # Gluon vs Triton, tile-based SPMD, @gluon.jit
├── 02-layouts.py                # BlockedLayout, size_per_thread, warps_per_cta
├── 03-async-copy.py             # cp.async, 流水线, shared memory
├── 04-tma.py                    # Tensor Memory Accelerator, tensor desc, mbarrier
├── 05-wgmma.py                  # Warp-Group MMA, Hopper Tensor Core, async MMA
├── 06-tcgen05.py                # Blackwell Tensor Core, Tensor Memory, tcgen05_mma
├── 07-persistence.py            # 持久化内核, work assignment, 多级流水线
├── 08-warp-specialization.py    # Warp 特化, 任务重叠
├── 09-tma-gather-scatter.py     # Native TMA Gather/Scatter (Blackwell)
├── 10-tcgen05-copy.py           # tcgen05_copy, shared→tensor memory
├── 11-tcgen05-mma-scaled.py     # tcgen05_mma_scaled, nvfp4/mxfp4/mxfp8
└── 12-cluster-launch-control.py # CLC, 动态 work distribution

TRITON_REPO/python/examples/gluon/
└── 01-attention-forward.py      # Flash Attention forward (Blackwell)
                                 # 完整的 producer/consumer, TMA, tcgen05_mma

TRITON_REPO/python/triton_kernels/triton_kernels/
├── matmul.py                    # 矩阵乘法 API (融合激活/MoE/ragged)
├── matmul_details/
│   ├── _matmul.py               # Dense GEMM kernel (TMA, mxfp4/8)
│   ├── _p_matmul.py             # Persistent GEMM kernel (ragged TMA)
│   └── _common.py               # 偏移计算, XCD swizzle
├── reduce.py                    # Reduction kernel (mask/scale/mxfp/flexpoint)
├── topk.py                      # Top-K selection (forward/backward, bitmatrix)
├── swiglu.py                    # SwiGLU activation kernel
├── compaction.py                # Masked compaction kernel
├── numerics.py                  # FP8/MXFP 数值配置
├── numerics_details/
│   ├── mxfp.py                  # MXFP 量化/反量化
│   └── flexpoint.py             # Flexpoint 缩放
├── tensor.py                    # Tensor/Layout 抽象 (TMA descriptors)
├── tensor_details/
│   └── layout_details/          # Blackwell/Hopper/CDNA4 MX 布局
├── distributed.py               # 分布式 MoE, SymmetricMemory
├── testing.py                   # 测试工具 (assert_close, compute_sanitizer)
└── roofline.py                  # Roofline 性能分析

TRITON_REPO/python/triton/language/           # tl.* 操作的定义和语义
TRITON_REPO/python/triton/experimental/gluon/ # gluon.* 操作的定义
TRITON_REPO/python/triton/runtime/            # JIT 编译, 缓存, 解释器
TRITON_REPO/python/triton/compiler/           # 代码生成
TRITON_REPO/python/triton/tools/              # Tensor descriptor 工具

TRITON_REPO/include/triton/
├── Dialect/
│   ├── Triton/           # Triton IR dialect 定义 (.td, .h)
│   ├── TritonGPU/        # TritonGPU dialect (layouts, encodings)
│   ├── TritonNvidiaGPU/  # NVIDIA 特定 ops (wgmma, tma, tcgen05)
│   └── Gluon/            # Gluon dialect
├── Conversion/           # IR lowering passes (TritonGPU -> LLVM)
├── Analysis/             # Alias, Allocation, AxisInfo, Membar
└── Tools/                # 工具类

TRITON_REPO/lib/
├── Dialect/
│   ├── Triton/           # Triton ops 实现, canonicalize
│   ├── TritonGPU/        # GPU layout 优化 passes
│   ├── TritonNvidiaGPU/  # NVIDIA lowering
│   └── Gluon/            # Gluon ops 实现
├── Conversion/           # Lowering pass 实现 (TritonGPU -> LLVM IR)
├── Analysis/             # 分析 pass 实现
└── Target/               # 代码生成目标

undefined

undefined

undefined

undefined

undefined

undefined

undefined

undefined

import triton
import triton.language as tl

@triton.jit
def kernel(x_ptr, y_ptr, n_elements, BLOCK_SIZE: tl.constexpr):
    pid = tl.program_id(0)
    offsets = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
    mask = offsets < n_elements
    x = tl.load(x_ptr + offsets, mask=mask)
    y = x * 2  # 计算
    tl.store(y_ptr + offsets, y, mask=mask)

undefined

@triton.autotune(
    configs=[
        triton.Config({'BLOCK_M': 128, 'BLOCK_N': 256, 'BLOCK_K': 64}, num_stages=3, num_warps=8),
        triton.Config({'BLOCK_M': 64, 'BLOCK_N': 256, 'BLOCK_K': 32}, num_stages=4, num_warps=4),
    ],
    key=['M', 'N', 'K'],
)
@triton.jit
def matmul_kernel(a_ptr, b_ptr, c_ptr, M, N, K, ...):
    ...

from triton.experimental import gluon

@gluon.jit
def kernel(x: gluon.tensor[M, N, tl.float16]):
    # 直接使用 tile 操作
    y = x + 1.0
    return y

undefined

undefined

undefined

undefined