"""
FMHA D2: Multi-CTA grid with flat_divide + tma_partition inside kernel.

Proper Blackwell approach following CUTLASS reference pattern:
- Q/K/V/O tensors with embedded head dimensions: (s, d, ((h_r, h_k), batch))
- flat_divide + tma_partition inside warp blocks (runtime block_idx)
- Direct TMA bulk copy for O output (not epilogue_tma_store)
- Grid: (M_tiles, h_q, batch) — one CTA per (M-tile, head, batch) triple

DSV4 is MQA: h_r = num_query_heads, h_k = 1, K/V shared.
"""
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 reference_attention(q, k, v, scale):
    """FP32 reference attention. q: (batch, n_h, T, hd), k/v: (batch, s_k, hd)"""
    qf = q.float()
    kf = k.float()
    vf = v.float()
    batch, n_h, T, hd = q.shape
    s_k = k.shape[1]
    ref = torch.zeros_like(qf)
    for b in range(batch):
        for h in range(n_h):
            attn = qf[b, h] @ kf[b].T * scale
            attn_max = attn.max(dim=-1, keepdim=True)[0]
            attn_exp = torch.exp(attn - attn_max)
            attn_sum = attn_exp.sum(dim=-1, keepdim=True)
            ref[b, h] = (attn_exp / attn_sum) @ vf[b]
    return ref


def test_multicta(hd=64, n_h=1, batch=1, T=128, s_k=128):
    """Test multi-CTA grid FMHA with n_h query heads (MQA)."""
    torch.manual_seed(42)
    scale = 1.0 / math.sqrt(hd)

    q = torch.randn(batch, n_h, T, hd, dtype=torch.bfloat16, device='cuda')
    k = torch.randn(batch, s_k, hd, dtype=torch.bfloat16, device='cuda')
    v = torch.randn(batch, s_k, hd, dtype=torch.bfloat16, device='cuda')
    o = torch.zeros(batch, n_h, T, hd, dtype=torch.bfloat16, device='cuda')
    lse = torch.zeros(batch, n_h, T, dtype=torch.float32, device='cuda')

    # FP32 reference
    ref = reference_attention(q, k, v, scale)

    # Run kernel with multi-CTA grid
    kernel = FmhaKernel(
        head_dim=hd, s_k=s_k, num_query_heads=n_h, batch_size=batch,
        use_smem_p=(hd > 64), normalize=True,
    )
    stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
    kernel(q, k, v, o, stream, lse=lse)
    torch.cuda.synchronize()

    cos = torch.nn.functional.cosine_similarity(
        o.flatten().float().unsqueeze(0), ref.flatten().unsqueeze(0)
    ).item()

    status = "PASS" if cos >= 0.99 else "FAIL"
    print(f'  hd={hd}, n_h={n_h}, batch={batch}, T={T}, s_k={s_k}: cos {cos:.6f}  {status}')
    return cos >= 0.99


def test():
    print("=== D2: Multi-CTA Grid (flat_divide approach) ===\n")

    all_pass = True

    # n_h=1 regression (should match existing single-head behavior)
    all_pass &= test_multicta(64, 1, 1, 128, 128)

    # n_h=2, single batch
    all_pass &= test_multicta(64, 2, 1, 128, 128)

    # n_h=4, batch=2
    all_pass &= test_multicta(64, 4, 2, 128, 128)

    # n_h=8
    all_pass &= test_multicta(64, 8, 1, 128, 128)

    print(f'\nOverall: {"ALL PASS" if all_pass else "SOME FAILED"}')


if __name__ == '__main__':
    test()