"""FP8 e4m3 GEMM via Triton FP8 tensor-core tl.dot + per-channel scale."""
import torch
import torch.nn as nn
import triton
import triton.language as tl

E4M3_MAX = 448.0


@triton.jit
def _fp8_gemm_kernel(
    A,
    B,
    C,
    Scales,
    M,
    N,
    K,
    stride_am,
    stride_ak,
    stride_bk,
    stride_bn,
    stride_cm,
    stride_cn,
    BLOCK_M: tl.constexpr,
    BLOCK_N: tl.constexpr,
    BLOCK_K: tl.constexpr,
    GROUP_M: tl.constexpr,
    NUM_STAGES: tl.constexpr,
):
    pid = tl.program_id(0)
    num_pid_m = tl.cdiv(M, BLOCK_M)
    num_pid_n = tl.cdiv(N, BLOCK_N)
    num_pid_in_group = GROUP_M * num_pid_n
    group_id = pid // num_pid_in_group
    first_pid_m = group_id * GROUP_M
    group_size_m = min(num_pid_m - first_pid_m, GROUP_M)
    pid_m = first_pid_m + (pid % group_size_m)
    pid_n = (pid % num_pid_in_group) // group_size_m

    offs_m = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
    offs_n = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
    offs_k = tl.arange(0, BLOCK_K)

    a_ptrs = A + offs_m[:, None] * stride_am + offs_k[None, :] * stride_ak
    b_ptrs = B + offs_k[:, None] * stride_bk + offs_n[None, :] * stride_bn

    acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
    for k in tl.range(0, K, BLOCK_K, num_stages=NUM_STAGES):
        a = tl.load(a_ptrs, mask=offs_m[:, None] < M, other=0.0)
        b = tl.load(b_ptrs, mask=offs_n[None, :] < N, other=0.0)
        acc = tl.dot(a, b, acc)
        a_ptrs += BLOCK_K * stride_ak
        b_ptrs += BLOCK_K * stride_bk

    scales = tl.load(Scales + offs_n, mask=offs_n < N, other=1.0)
    acc = acc * scales[None, :]

    c_ptrs = C + offs_m[:, None] * stride_cm + offs_n[None, :] * stride_cn
    tl.store(c_ptrs, acc.to(tl.bfloat16), mask=(offs_m[:, None] < M) & (offs_n[None, :] < N))


@triton.jit
def _fp8_gemm_skinny_kernel(
    A,
    B,
    C,
    Scales,
    M,
    N,
    K,
    stride_am,
    stride_ak,
    stride_bk,
    stride_bn,
    stride_cm,
    stride_cn,
    BLOCK_M: tl.constexpr,
    BLOCK_N: tl.constexpr,
    BLOCK_K: tl.constexpr,
    NUM_STAGES: tl.constexpr,
):
    pid_n = tl.program_id(0)
    offs_m = tl.arange(0, BLOCK_M)
    offs_n = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
    offs_k = tl.arange(0, BLOCK_K)

    a_ptrs = A + offs_m[:, None] * stride_am + offs_k[None, :] * stride_ak
    b_ptrs = B + offs_k[:, None] * stride_bk + offs_n[None, :] * stride_bn

    acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
    for k in tl.range(0, K, BLOCK_K, num_stages=NUM_STAGES):
        a = tl.load(a_ptrs, mask=offs_m[:, None] < M, other=0.0)
        b = tl.load(b_ptrs, mask=offs_n[None, :] < N, other=0.0)
        acc = tl.dot(a, b, acc)
        a_ptrs += BLOCK_K * stride_ak
        b_ptrs += BLOCK_K * stride_bk

    scales = tl.load(Scales + offs_n, mask=offs_n < N, other=1.0)
    acc = acc * scales[None, :]

    c_ptrs = C + offs_m[:, None] * stride_cm + offs_n[None, :] * stride_cn
    tl.store(c_ptrs, acc.to(tl.bfloat16), mask=(offs_m[:, None] < M) & (offs_n[None, :] < N))


def _pick_config(M: int, N: int) -> tuple[int, int, int, int, int, int]:
    if M <= 64:
        return 32, 128, 256, 4, 3, 1
    if N >= 12000:
        return 128, 256, 128, 8, 3, 4
    return 128, 256, 128, 8, 3, 4


def _fp8_gemm(
    x: torch.Tensor,
    weight: torch.Tensor,
    weight_scale: torch.Tensor,
    M: int,
) -> torch.Tensor:
    K = x.shape[1]
    N, K_w = weight.shape
    assert K == K_w

    y = torch.empty((M, N), device=x.device, dtype=torch.bfloat16)
    block_m, block_n, block_k, num_warps, num_stages, group_m = _pick_config(M, N)

    if M <= 64:
        grid = (triton.cdiv(N, block_n),)
        _fp8_gemm_skinny_kernel[grid](
            x,
            weight,
            y,
            weight_scale,
            M,
            N,
            K,
            x.stride(0),
            x.stride(1),
            weight.stride(1),
            weight.stride(0),
            y.stride(0),
            y.stride(1),
            BLOCK_M=block_m,
            BLOCK_N=block_n,
            BLOCK_K=block_k,
            NUM_STAGES=num_stages,
            num_warps=num_warps,
        )
    else:
        grid = (triton.cdiv(M, block_m) * triton.cdiv(N, block_n),)
        _fp8_gemm_kernel[grid](
            x,
            weight,
            y,
            weight_scale,
            M,
            N,
            K,
            x.stride(0),
            x.stride(1),
            weight.stride(1),
            weight.stride(0),
            y.stride(0),
            y.stride(1),
            BLOCK_M=block_m,
            BLOCK_N=block_n,
            BLOCK_K=block_k,
            GROUP_M=group_m,
            NUM_STAGES=num_stages,
            num_warps=num_warps,
        )
    return y


class Model(nn.Module):
    def __init__(self, M: int, N: int, K: int):
        super().__init__()
        self.M, self.N, self.K = M, N, K
        w = torch.empty(N, K, dtype=torch.bfloat16)
        nn.init.normal_(w, std=0.02)
        s = (w.float().abs().amax(dim=1, keepdim=True) / E4M3_MAX).clamp(min=1e-12)
        w_fp8 = (w.float() / s).to(torch.float8_e4m3fn)
        self.register_buffer("weight", w_fp8)
        self.register_buffer("weight_scale", s.squeeze(1).to(torch.float32))
        self._weight_padded: torch.Tensor | None = None
        self._weight_version: int = -1

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        K = x.shape[1]
        pad_k = (128 - (K % 128)) % 128
        if pad_k:
            if self._weight_padded is None or self._weight_version != self.weight._version:
                self._weight_padded = torch.nn.functional.pad(self.weight, (0, pad_k))
                self._weight_version = self.weight._version
            x = torch.nn.functional.pad(x, (0, pad_k))
            weight = self._weight_padded
        else:
            weight = self.weight
        return _fp8_gemm(x, weight, self.weight_scale, self.M)


M = 4096
N = 4096
K = 4096


def get_inputs():
    x = (torch.rand(M, K) * 8 - 4).to(torch.float8_e4m3fn)
    return [x]


def get_init_inputs():
    return [M, N, K]