涉及到转置的算子有些时候出现算子精度不对

[Truth-Ke]

@triton.jit
def _conv2d_forward(input_ptr, weight_ptr, output_ptr, B, C, H, W, F, K,
    stride, padding, dilation, H_out, W_out, input_batch_stride,
    input_channel_stride, input_height_stride, input_width_stride,
    weight_out_channel_stride, weight_kernel_stride, output_batch_stride,
    output_channel_stride, output_height_stride, output_width_stride, M,
    K_total, BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr,
    BLOCK_SIZE_K: tl.constexpr):
    pid_m = tl.program_id(0)
    pid_n = tl.program_id(1)
    rm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
    rn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
    rk = tl.arange(0, BLOCK_SIZE_K)
    m_mask = rm < M
    batch_idx = rm // (H_out * W_out)
    spatial_idx = rm % (H_out * W_out)
    h_out = spatial_idx // W_out
    w_out = spatial_idx % W_out
    acc = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
    for k in range(0, tl.cdiv(K_total, BLOCK_SIZE_K)):
        rk = k * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K)
        k_mask = rk < K_total
        kernel_idx = rk
        c = kernel_idx // (K * K)
        kh = kernel_idx % (K * K) // K
        kw = kernel_idx % K
        h_in = h_out[:, None] * stride + kh[None, :] * dilation - padding
        w_in = w_out[:, None] * stride + kw[None, :] * dilation - padding
        in_bounds = (h_in >= 0) & (h_in < H) & (w_in >= 0) & (w_in < W)
        full_mask = m_mask[:, None] & k_mask[None, :] & in_bounds
        input_offsets = (batch_idx[:, None] * input_batch_stride + c[None,
            :] * input_channel_stride + h_in * input_height_stride + w_in *
            input_width_stride)
        input_block = tl.load(input_ptr + input_offsets, mask=full_mask,
            other=0.0)
        weight_block = tl.load(weight_ptr + rn[None, :] *
            weight_out_channel_stride + rk[:, None] * weight_kernel_stride,
            mask=k_mask[:, None] & (rn[None, :] < F), other=0.0)
        input_block = input_block.to(tl.float16)
        weight_block = weight_block.to(tl.float16)
        acc += tl.dot(input_block, weight_block)
    output_offsets = batch_idx[:, None] * output_batch_stride + rn[None, :
        ] * output_channel_stride + h_out[:, None
        ] * output_height_stride + w_out[:, None] * output_width_stride
    tl.store(output_ptr + output_offsets, acc, mask=m_mask[:, None] & (rn[
        None, :] < F))

Describe the bug
A clear and concise description of what the bug is.

To Reproduce
Steps to reproduce the behavior:

1. Go to '...'
2. Click on '....'
3. Scroll down to '....'
4. See error

Expected behavior
A clear and concise description of what you expected to happen.

Screenshots
If applicable, add screenshots to help explain your problem.

Desktop (please complete the following information):

• OS: [e.g. iOS]
• Browser [e.g. chrome, safari]
• Version [e.g. 22]

Smartphone (please complete the following information):

• Device: [e.g. iPhone6]
• OS: [e.g. iOS8.1]
• Browser [e.g. stock browser, safari]
• Version [e.g. 22]

Additional context
Add any other context about the problem here.