import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Performs a transposed 1D convolution operation with asymmetric input and square kernel.
    Supports padding, striding, and dilation.

    Args:
        in_channels (int): Number of channels in the input tensor.
        out_channels (int): Number of channels produced by the convolution.
        kernel_size (int): Size of the square convolution kernel.
        stride (int, optional): Stride of the convolution. Defaults to 1.
        padding (int, optional): Padding applied to the input. Defaults to 0.
        dilation (int, optional): Spacing between kernel elements. Defaults to 1.
        bias (bool, optional): If `True`, adds a learnable bias to the output. Defaults to `False`.
    """
    def __init__(self, in_channels: int, out_channels: int, kernel_size: int, stride: int = 1, padding: int = 0, dilation: int = 1, bias: bool = False):
        super(Model, self).__init__()
        self.conv1d_transpose = nn.ConvTranspose1d(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias)
        
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Performs the transposed 1D convolution.

        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, in_channels, length).

        Returns:
            torch.Tensor: Output tensor of shape (batch_size, out_channels, length_out).
        """
        return self.conv1d_transpose(x)

# Test code
batch_size = 16
in_channels = 32
out_channels = 64
kernel_size = 3
# long sequence
length = 131072
stride = 2
padding = 1
dilation = 2

def get_inputs():
    x = torch.rand(batch_size, in_channels, length)
    return [x]

def get_init_inputs():
    return [in_channels, out_channels, kernel_size, stride, padding, dilation]

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline

_cpp_src = r'''
#include <torch/extension.h>

torch::Tensor conv_transpose1d_forward_cuda(torch::Tensor x, torch::Tensor weight, torch::Tensor bias,
                                            int64_t stride, int64_t padding, int64_t dilation);
'''

_cuda_src = r'''
#include <torch/extension.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cuda.h>
#include <cuda_runtime.h>

// Specialized fast path for the benchmark case:
// Bx32xL input, 32x64x3 weights, stride=2, padding=1, dilation=2, no bias.
// Only odd output locations are non-zero.  Even locations are zeroed by cudaMemset.
__global__ __launch_bounds__(256, 2)
void convt1d_s2_p1_d2_k3_32_64_kernel(const float* __restrict__ x,
                                       const float* __restrict__ w,
                                       float* __restrict__ y,
                                       int L, int Lout) {
    const int m  = blockIdx.x * 32 + threadIdx.x;       // input/output odd-position index
    const int oc = blockIdx.y * 8  + threadIdx.y;
    const int n  = blockIdx.z;

    if (m >= L || oc >= 64) return;

    const float* __restrict__ xn = x + ((size_t)n * 32) * (size_t)L;
    float acc = 0.0f;

    if (m > 0 && m + 1 < L) {
        #pragma unroll
        for (int ic = 0; ic < 32; ++ic) {
            const float* __restrict__ xb = xn + (size_t)ic * (size_t)L + m;
            const float* __restrict__ wb = w + ((ic * 64 + oc) * 3);
            acc = fmaf(xb[1],  wb[0], acc);  // k=0, input m+1
            acc = fmaf(xb[0],  wb[1], acc);  // k=1, input m
            acc = fmaf(xb[-1], wb[2], acc);  // k=2, input m-1
        }
    } else if (m == 0) {
        #pragma unroll
        for (int ic = 0; ic < 32; ++ic) {
            const float* __restrict__ xb = xn + (size_t)ic * (size_t)L;
            const float* __restrict__ wb = w + ((ic * 64 + oc) * 3);
            acc = fmaf(xb[1], wb[0], acc);
            acc = fmaf(xb[0], wb[1], acc);
        }
    } else { // m == L-1
        #pragma unroll
        for (int ic = 0; ic < 32; ++ic) {
            const float* __restrict__ xb = xn + (size_t)ic * (size_t)L + m;
            const float* __restrict__ wb = w + ((ic * 64 + oc) * 3);
            acc = fmaf(xb[0],  wb[1], acc);
            acc = fmaf(xb[-1], wb[2], acc);
        }
    }

    y[((size_t)n * 64 + oc) * (size_t)Lout + (2 * (size_t)m + 1)] = acc;
}

__global__ void convt1d_generic_kernel(const float* __restrict__ x,
                                       const float* __restrict__ w,
                                       const float* __restrict__ b,
                                       float* __restrict__ y,
                                       int B, int Cin, int Cout, int L, int K,
                                       int stride, int padding, int dilation,
                                       int Lout, bool has_bias) {
    size_t idx = (size_t)blockIdx.x * blockDim.x + threadIdx.x;
    const size_t total = (size_t)B * Cout * Lout;
    if (idx >= total) return;

    int p = (int)(idx % Lout);
    size_t t0 = idx / Lout;
    int oc = (int)(t0 % Cout);
    int n = (int)(t0 / Cout);

    float acc = has_bias ? b[oc] : 0.0f;
    const float* __restrict__ xn = x + ((size_t)n * Cin) * (size_t)L;
    for (int ic = 0; ic < Cin; ++ic) {
        const float* __restrict__ xb = xn + (size_t)ic * (size_t)L;
        const float* __restrict__ wb = w + ((size_t)ic * Cout + oc) * (size_t)K;
        for (int k = 0; k < K; ++k) {
            int num = p + padding - k * dilation;
            if (num >= 0 && (num % stride) == 0) {
                int l = num / stride;
                if (l >= 0 && l < L) acc = fmaf(xb[l], wb[k], acc);
            }
        }
    }
    y[idx] = acc;
}

torch::Tensor conv_transpose1d_forward_cuda(torch::Tensor x, torch::Tensor weight, torch::Tensor bias,
                                            int64_t stride, int64_t padding, int64_t dilation) {
    const c10::cuda::CUDAGuard device_guard(x.device());

    const int B = (int)x.size(0);
    const int Cin = (int)x.size(1);
    const int L = (int)x.size(2);
    const int Cout = (int)weight.size(1);
    const int K = (int)weight.size(2);
    const int Lout = (int)((int64_t)(L - 1) * stride - 2 * padding + dilation * (K - 1) + 1);

    auto y = torch::empty({B, Cout, Lout}, x.options());
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    const bool has_bias = (bias.numel() != 0);

    if (!has_bias && x.scalar_type() == at::kFloat && weight.scalar_type() == at::kFloat &&
        Cin == 32 && Cout == 64 && K == 3 && stride == 2 && padding == 1 && dilation == 2) {
        AT_CUDA_CHECK(cudaMemsetAsync(y.data_ptr<float>(), 0, y.numel() * sizeof(float), stream));
        dim3 block(32, 8, 1);
        dim3 grid((L + 31) / 32, 8, B);
        convt1d_s2_p1_d2_k3_32_64_kernel<<<grid, block, 0, stream>>>(
            x.data_ptr<float>(), weight.data_ptr<float>(), y.data_ptr<float>(), L, Lout);
        C10_CUDA_KERNEL_LAUNCH_CHECK();
    } else {
        const int threads = 256;
        const size_t total = (size_t)B * Cout * Lout;
        const int blocks = (int)((total + threads - 1) / threads);
        convt1d_generic_kernel<<<blocks, threads, 0, stream>>>(
            x.data_ptr<float>(), weight.data_ptr<float>(), has_bias ? bias.data_ptr<float>() : nullptr,
            y.data_ptr<float>(), B, Cin, Cout, L, K, (int)stride, (int)padding, (int)dilation, Lout, has_bias);
        C10_CUDA_KERNEL_LAUNCH_CHECK();
    }
    return y;
}
'''

_ext = load_inline(
    name='convt1d_transpose_s2p1d2_ext',
    cpp_sources=_cpp_src,
    cuda_sources=_cuda_src,
    functions=['conv_transpose1d_forward_cuda'],
    extra_cflags=['-O3'],
    extra_cuda_cflags=['-O3', '--use_fast_math'],
    verbose=False,
)

class ModelNew(nn.Module):
    def __init__(self, in_channels: int, out_channels: int, kernel_size: int, stride: int = 1, padding: int = 0, dilation: int = 1, bias: bool = False):
        super(ModelNew, self).__init__()
        self.conv1d_transpose = nn.ConvTranspose1d(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias)
        self.stride = stride
        self.padding = padding
        self.dilation = dilation

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        b = self.conv1d_transpose.bias
        if b is None:
            b = torch.empty(0, device=x.device, dtype=x.dtype)
        return _ext.conv_transpose1d_forward_cuda(x, self.conv1d_transpose.weight, b, self.stride, self.padding, self.dilation)