Why triton convert to float8e5 will cause local memory read/write

[MARD1NO]

I just write a kernel and it contains a x.to(tl.float8e5) , in ncu I found it cause local memory read/store

[ThomasRaoux]

can you provide a simple kernel example?

[MARD1NO]

can you provide a simple kernel example?

Sure，I copy the tutorial gemm kernel as an example, just convert input matrix to tl.float8e5:

import torch

import triton
import triton.language as tl



@triton.jit
def matmul_kernel(
        # Pointers to matrices
        a_ptr, b_ptr, c_ptr,
        # Matrix dimensions
        M, N, K,
        # The stride variables represent how much to increase the ptr by when moving by 1
        # element in a particular dimension. E.g. `stride_am` is how much to increase `a_ptr`
        # by to get the element one row down (A has M rows).
        stride_am, stride_ak,  #
        stride_bk, stride_bn,  #
        stride_cm, stride_cn,
        # Meta-parameters
        BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr, BLOCK_SIZE_K: tl.constexpr,  #
        GROUP_SIZE_M: tl.constexpr,  #
):
    """Kernel for computing the matmul C = A x B.
    A has shape (M, K), B has shape (K, N) and C has shape (M, N)
    """
    # -----------------------------------------------------------
    # Map program ids `pid` to the block of C it should compute.
    # This is done in a grouped ordering to promote L2 data reuse.
    # See above `L2 Cache Optimizations` section for details.
    pid = tl.program_id(axis=0)
    num_pid_m = tl.cdiv(M, BLOCK_SIZE_M)
    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
    num_pid_in_group = GROUP_SIZE_M * num_pid_n
    group_id = pid // num_pid_in_group
    first_pid_m = group_id * GROUP_SIZE_M
    group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
    pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
    pid_n = (pid % num_pid_in_group) // group_size_m

    # ----------------------------------------------------------
    # Create pointers for the first blocks of A and B.
    # We will advance this pointer as we move in the K direction
    # and accumulate
    # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
    # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
    # See above `Pointer Arithmetic` section for details
    offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
    offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
    offs_k = tl.arange(0, BLOCK_SIZE_K)
    a_ptrs = a_ptr + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak)
    b_ptrs = b_ptr + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn)

    # -----------------------------------------------------------
    # Iterate to compute a block of the C matrix.
    # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
    # of fp32 values for higher accuracy.
    # `accumulator` will be converted back to fp16 after the loop.
    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
        # Load the next block of A and B, generate a mask by checking the K dimension.
        # If it is out of bounds, set it to 0.
        a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0)
        b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0)
        
        a = a.to(tl.float8e5)
        b = b.to(tl.float8e5)

        # We accumulate along the K dimension.
        accumulator = tl.dot(a, b, accumulator)
        # Advance the ptrs to the next K block.
        a_ptrs += BLOCK_SIZE_K * stride_ak
        b_ptrs += BLOCK_SIZE_K * stride_bk
    # You can fuse arbitrary activation functions here
    # while the accumulator is still in FP32!
    c = accumulator.to(tl.float16)

    # -----------------------------------------------------------
    # Write back the block of the output matrix C with masks.
    offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
    c_ptrs = c_ptr + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
    c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
    tl.store(c_ptrs, c, mask=c_mask)



def matmul(a, b):
    # 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=torch.float16)
    # 1D launch kernel where each block gets its own program.

    BLOCK_SIZE_M = 64
    BLOCK_SIZE_N = 128
    BLOCK_SIZE_K = 64  #
    GROUP_SIZE_M = 8

    grid = lambda META: (triton.cdiv(M, BLOCK_SIZE_M) * triton.cdiv(N, BLOCK_SIZE_N), )
    matmul_kernel[grid](
        a, b, c,  #
        M, N, K,  #
        a.stride(0), a.stride(1),  #
        b.stride(0), b.stride(1),  #
        c.stride(0), c.stride(1),  #
        BLOCK_SIZE_M, 
        BLOCK_SIZE_N, 
        BLOCK_SIZE_K, 
        GROUP_SIZE_M
    )
    return c

M = 16
K = 256
N = 256

a = torch.randn((M, K), device='cuda', dtype=torch.float16)
b = torch.randn((K, N), device='cuda', dtype=torch.float16)
b = b.permute(0, 1).contiguous()
b = b.permute(0, 1)


matmul(a, b)

I profile it in Nsight compute tool, it seems the datatype conversion related to the local memory:

I test it in h20

[Jokeren]

Likely a register spilling problem

[Jokeren]

Check your register usage of this kernel

[MARD1NO]

Check your register usage of this kernel

I don't think the problem is register spilling, nsight compute shows the register use only 168:

maybe the conversion use some non-constant index in array?...

[ThomasRaoux]

this PR might fix it. Can you try:
#4776

[MARD1NO]

this PR might fix it. Can you try: #4776

Thanks Thomas, I will try it :D

[MARD1NO]

this PR might fix it. Can you try: #4776

this commit still has local memory write

[Jokeren]

It has to do with cvt.rn.satfinite.e5m2x2.f16x2. Taking a look now

[MARD1NO]

It has to do with cvt.rn.satfinite.e5m2x2.f16x2. Taking a look now

yes, related SASS and ptx shows it use cvt.rn.satfinite.e5m2x2.f16x2

it seems the store local happens after convert to e5m2 from PTX

[Jokeren]

IMO it's probably caused by nvptx doesn't handle 8-bit vector type well. Let me trigger a discussion and get back to you