triton/third_party/tlx/tutorials/blackwell_gemm_2cta.py at main · facebookexperimental/triton · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
import torch

import triton
import triton.language as tl
import triton.language.extra.tlx as tlx

from typing import Optional

DEVICE = triton.runtime.driver.active.get_active_torch_device()


def alloc_fn(size: int, align: int, stream: Optional[int]):
    assert align == 128
    assert stream == 0
    return torch.empty(size, dtype=torch.int8, device="cuda")


@triton.jit
def tcgen5_dot_kernel2cta_tma(
    a_ptr,
    stride_am,
    stride_ak,
    b_ptr,
    stride_bk,
    stride_bn,
    c_ptr,
    stride_cm,
    stride_cn,
    BLOCK_M: tl.constexpr,
    BLOCK_N: tl.constexpr,
    BLOCK_K: tl.constexpr,
    OUT_DTYPE: tl.constexpr,
    M: tl.constexpr,
    N: tl.constexpr,
    K: tl.constexpr,
):
    pid_m = tl.program_id(axis=0)
    pid_n = tl.program_id(axis=1)

    # assuming CTA pairs along M dim
    cluster_cta_rank = tlx.cluster_cta_rank()  # 2cta specific
    pred_leader_cta = cluster_cta_rank % 2 == 0

    offs_am = pid_m * BLOCK_M
    offs_bn = pid_n * BLOCK_N + (cluster_cta_rank % 2) * (BLOCK_N // 2)  # 2cta specific

    offs_cm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
    offs_cn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)

    desc_a = tl.make_tensor_descriptor(
        a_ptr,
        shape=[M, K],
        strides=[stride_am, stride_ak],
        block_shape=[BLOCK_M, BLOCK_K],
    )

    desc_b = tl.make_tensor_descriptor(
        b_ptr,
        shape=[K, N],
        strides=[stride_bk, stride_bn],
        block_shape=[BLOCK_K, BLOCK_N // 2],
    )

    # async load a and b into SMEM
    buf_alloc_a = tlx.local_alloc((BLOCK_M, BLOCK_K), tlx.dtype_of(a_ptr), tl.constexpr(1))
    buf_alloc_b = tlx.local_alloc((BLOCK_K, BLOCK_N // 2), tlx.dtype_of(b_ptr), tl.constexpr(1))  # 2cta specific
    a_smem = tlx.local_view(buf_alloc_a, 0)
    b_smem = tlx.local_view(buf_alloc_b, 0)

    bars = tlx.alloc_barriers(tl.constexpr(3))
    bar_a = tlx.local_view(bars, 0)
    bar_b = tlx.local_view(bars, 1)

    # 2cta specific
    bar_cta = tlx.alloc_barriers(1, arrive_count=2)  # CTA0 waits for CTA1's data before mma
    bar_leader_cta = tlx.local_view(bar_cta, 0)

    buffers = tlx.local_alloc((BLOCK_M, BLOCK_N), tl.float32, tl.constexpr(1), tlx.storage_kind.tmem)
    acc_tmem = tlx.local_view(buffers, 0)

    acc_init = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
    tlx.local_store(acc_tmem, acc_init)

    dot_bars = tlx.alloc_barriers(num_barriers=1, arrive_count=1)

    phase = 0
    num_iter = tl.cdiv(K, BLOCK_K)
    for k in range(0, num_iter):
        offs_k = k * BLOCK_K

        tlx.barrier_expect_bytes(bar_a, BLOCK_M * BLOCK_K * 2)
        tlx.barrier_expect_bytes(bar_b, BLOCK_K * (BLOCK_N // 2) * 2)  # 2cta specific

        tlx.async_descriptor_load(desc_a, a_smem, [offs_am, offs_k], bar_a)
        tlx.async_descriptor_load(desc_b, b_smem, [offs_k, offs_bn], bar_b)

        tlx.barrier_wait(bar_a, phase)
        tlx.barrier_wait(bar_b, phase)

        # CTA0 needs to know CTA1 is done loading data before issuing MMA
        tlx.barrier_arrive(bar_leader_cta, 1, remote_cta_rank=cluster_cta_rank & ~1)
        tlx.barrier_wait(bar_leader_cta, phase=k % 2, pred=pred_leader_cta)

        # 2cta specific
        tlx.async_dot(a_smem, b_smem, acc_tmem, use_acc=True, mBarriers=[dot_bars[0]], two_ctas=True,
                      out_dtype=OUT_DTYPE)

        tlx.barrier_wait(dot_bars[0], phase)
        phase = phase ^ 1

    result = tlx.local_load(acc_tmem)

    c = result.to(tlx.dtype_of(c_ptr))
    c_ptrs = c_ptr + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
    tl.store(c_ptrs, c)


def matmul(a, b, config=None):
    """Matrix multiplication using TLX GEMM kernel."""
    # Check constraints.
    assert a.shape[1] == b.shape[0], "Incompatible dimensions"
    assert a.is_contiguous(), "Matrix A must be contiguous"

    M, K = a.shape
    K, N = b.shape
    # Allocates output.
    c = torch.empty((M, N), device=a.device, dtype=a.dtype)

    triton.set_allocator(alloc_fn)

    BLOCK_M, BLOCK_N, BLOCK_K = (128, 128, 128)

    kern_kwargs = {
        "BLOCK_M": BLOCK_M,
        "BLOCK_K": BLOCK_K,
        "BLOCK_N": BLOCK_N,
        "OUT_DTYPE": tl.float32,
        "M": M,
        "N": N,
        "K": K,
        "num_stages": 1,
        "ctas_per_cga": (4, 2, 1),
    }
    _ = tcgen5_dot_kernel2cta_tma[(M // BLOCK_M, N // BLOCK_N)](a, a.stride(0), a.stride(1), b, b.stride(0),
                                                                b.stride(1), c, c.stride(0), c.stride(1), **kern_kwargs)

    return c