"""
D2: Multi-CTA grid test.

Tests the multi-CTA FMHA kernel with Q shape (n_h, T, hd, 1).
Each CTA (indexed by block_idx_y) handles one query head.
K/V are shared (MQA) — all CTAs load the same K/V.
"""
import torch, math
import cutlass.cute as cute
import cutlass.torch as ct
import cuda.bindings.driver as cuda
from dsv4.kernels.attention.fmha import FmhaKernel


def test_multi_cta(hd=64, n_h=2, s_k=128):
    T = 128
    torch.manual_seed(42)

    # Q: (n_h, T, hd, 1) — head dimension outermost
    q = torch.randn(n_h, T, hd, 1, dtype=torch.bfloat16, device='cuda')
    # K/V: (s_k, hd, 1) — shared KV (no head dim)
    k = torch.randn(s_k, hd, 1, dtype=torch.bfloat16, device='cuda')
    v = torch.randn(s_k, hd, dtype=torch.bfloat16, device='cuda')
    # O: (n_h, T, hd, 1)
    o = torch.zeros(n_h, T, hd, 1, dtype=torch.bfloat16, device='cuda')

    # FP32 reference (un-normalized)
    qf = q[:, :, :, 0].float()  # (n_h, T, hd)
    kf = k[:, :, 0].float()     # (s_k, hd)
    vf = v.float()               # (s_k, hd)
    scale = 1.0 / math.sqrt(hd)

    ref_unnorm = torch.zeros(n_h, T, hd, dtype=torch.float32, device='cuda')
    for h in range(n_h):
        attn = qf[h] @ kf.T * scale
        attn_max = attn.max(dim=-1, keepdim=True)[0]
        attn_exp = torch.exp(attn - attn_max)
        ref_unnorm[h] = attn_exp @ vf

    lse_tensor = torch.zeros(T, 1, 1, dtype=torch.float32, device='cuda')

    kernel = FmhaKernel(head_dim=hd, s_k=s_k, use_smem_p=False, normalize=False,
                        num_query_heads=n_h, batch_size=1)
    pv_n_tile = kernel.pv_n_tile
    stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)

    # Compile with Q having head dimension
    v_tile = v[:, 0:pv_n_tile].contiguous()
    v_kernel = v_tile.unsqueeze(-1)
    c_tile = torch.zeros(T, pv_n_tile, 1, dtype=torch.bfloat16, device='cuda')

    mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
    mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
    mV = ct.from_dlpack(v_kernel).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_kernel))
    mC = ct.from_dlpack(c_tile).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c_tile))
    mLSE = ct.from_dlpack(lse_tensor).mark_layout_dynamic(leading_dim=ct.get_leading_dim(lse_tensor))

    print(f'  Compiling (hd={hd}, n_h={n_h})...', flush=True)
    compiled = cute.compile(kernel, mQ, mK, mV, mC, stream, mLSE)
    compiled(mQ, mK, mV, mC, stream, mLSE)
    torch.cuda.synchronize()

    # Check output
    out = o[:, :, :, 0].float()  # (n_h, T, hd)
    cos = torch.nn.functional.cosine_similarity(
        out.flatten().unsqueeze(0), ref_unnorm.flatten().unsqueeze(0)
    ).item()
    print(f'  hd={hd}, n_h={n_h}: cos {cos:.6f}  {"PASS" if cos >= 0.99 else "FAIL"}')


def test():
    print("=== D2: Multi-CTA Grid ===\n")
    test_multi_cta(64, 2)


if __name__ == '__main__':
    test()