import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs a LogSoftmax activation.
    """
    def __init__(self, dim: int = 1):
        super(Model, self).__init__()
        self.dim = dim
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Applies LogSoftmax activation to the input tensor.

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

        Returns:
            torch.Tensor: Output tensor with LogSoftmax applied, same shape as input.
        """
        return torch.log_softmax(x, dim=self.dim)

batch_size = 4096
dim = 393216

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

def get_init_inputs():
    return []  # No special initialization inputs needed

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

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

// CUDA declaration
torch::Tensor logsoftmax_forward_cuda(torch::Tensor x, int64_t dim);

// Host wrapper exposed to Python
torch::Tensor logsoftmax_forward(torch::Tensor x, int64_t dim) {
    return logsoftmax_forward_cuda(x, dim);
}
"""

_cuda_src = r"""
#include <torch/extension.h>
#include <ATen/cuda/CUDAContext.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <float.h>

#ifndef WARP_SIZE
#define WARP_SIZE 32
#endif

template<int BLOCK>
__device__ __forceinline__ float block_reduce_max(float v) {
    __shared__ float smem[BLOCK / WARP_SIZE];
    int lane = threadIdx.x & (WARP_SIZE - 1);
    int wid  = threadIdx.x >> 5;
    #pragma unroll
    for (int off = 16; off > 0; off >>= 1) v = fmaxf(v, __shfl_down_sync(0xffffffff, v, off));
    if (lane == 0) smem[wid] = v;
    __syncthreads();
    v = (threadIdx.x < (BLOCK / WARP_SIZE)) ? smem[lane] : -FLT_MAX;
    if (wid == 0) {
        #pragma unroll
        for (int off = 16; off > 0; off >>= 1) v = fmaxf(v, __shfl_down_sync(0xffffffff, v, off));
    }
    return __shfl_sync(0xffffffff, v, 0);
}

template<int BLOCK>
__device__ __forceinline__ float block_reduce_sum(float v) {
    __shared__ float smem[BLOCK / WARP_SIZE];
    int lane = threadIdx.x & (WARP_SIZE - 1);
    int wid  = threadIdx.x >> 5;
    #pragma unroll
    for (int off = 16; off > 0; off >>= 1) v += __shfl_down_sync(0xffffffff, v, off);
    if (lane == 0) smem[wid] = v;
    __syncthreads();
    v = (threadIdx.x < (BLOCK / WARP_SIZE)) ? smem[lane] : 0.0f;
    if (wid == 0) {
        #pragma unroll
        for (int off = 16; off > 0; off >>= 1) v += __shfl_down_sync(0xffffffff, v, off);
    }
    return __shfl_sync(0xffffffff, v, 0);
}

template<int BLOCK>
__global__ void logsoftmax_dim1_kernel(const float* __restrict__ x, float* __restrict__ y,
                                       int rows, int cols) {
    int row = blockIdx.x;
    if (row >= rows) return;
    const float* __restrict__ xr = x + (long long)row * cols;
    float* __restrict__ yr = y + (long long)row * cols;

    float vmax = -FLT_MAX;
    // vector-friendly strided loop. cols is large; use all threads in one CTA per row.
    for (int c = threadIdx.x; c < cols; c += BLOCK) {
        vmax = fmaxf(vmax, xr[c]);
    }
    vmax = block_reduce_max<BLOCK>(vmax);

    float sum = 0.0f;
    for (int c = threadIdx.x; c < cols; c += BLOCK) {
        sum += expf(xr[c] - vmax);
    }
    sum = block_reduce_sum<BLOCK>(sum);
    float lsum = logf(sum) + vmax;

    for (int c = threadIdx.x; c < cols; c += BLOCK) {
        yr[c] = xr[c] - lsum;
    }
}

// Special case for dim=0, not expected for the benchmark but implemented for completeness.
template<int BLOCK>
__global__ void logsoftmax_dim0_kernel(const float* __restrict__ x, float* __restrict__ y,
                                       int rows, int cols) {
    int col = blockIdx.x;
    if (col >= cols) return;
    float vmax = -FLT_MAX;
    for (int r = threadIdx.x; r < rows; r += BLOCK) {
        vmax = fmaxf(vmax, x[(long long)r * cols + col]);
    }
    vmax = block_reduce_max<BLOCK>(vmax);
    float sum = 0.0f;
    for (int r = threadIdx.x; r < rows; r += BLOCK) {
        sum += expf(x[(long long)r * cols + col] - vmax);
    }
    sum = block_reduce_sum<BLOCK>(sum);
    float lsum = logf(sum) + vmax;
    for (int r = threadIdx.x; r < rows; r += BLOCK) {
        y[(long long)r * cols + col] = x[(long long)r * cols + col] - lsum;
    }
}

torch::Tensor logsoftmax_forward_cuda(torch::Tensor x, int64_t dim) {
    TORCH_CHECK(x.is_cuda(), "x must be CUDA");
    TORCH_CHECK(x.scalar_type() == at::kFloat, "only float32 is supported");
    TORCH_CHECK(x.dim() == 2, "only 2D tensors are supported");
    if (!x.is_contiguous()) x = x.contiguous();
    int64_t nd = x.dim();
    if (dim < 0) dim += nd;
    TORCH_CHECK(dim == 0 || dim == 1, "dim out of range");
    auto y = torch::empty_like(x);
    int rows = (int)x.size(0);
    int cols = (int)x.size(1);
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    constexpr int BLOCK = 256;
    if (dim == 1) {
        logsoftmax_dim1_kernel<BLOCK><<<rows, BLOCK, 0, stream>>>(x.data_ptr<float>(), y.data_ptr<float>(), rows, cols);
    } else {
        logsoftmax_dim0_kernel<BLOCK><<<cols, BLOCK, 0, stream>>>(x.data_ptr<float>(), y.data_ptr<float>(), rows, cols);
    }
    return y;
}
"""

_ext = load_inline(
    name="logsoftmax_custom_ext_v1",
    cpp_sources=_cpp_src,
    cuda_sources=_cuda_src,
    functions=["logsoftmax_forward"],
    extra_cflags=["-O3"],
    extra_cuda_cflags=["-O3", "--use_fast_math"],
    verbose=False,
)

class ModelNew(nn.Module):
    """
    Simple model that performs a LogSoftmax activation using a custom CUDA kernel.
    """
    def __init__(self, dim: int = 1):
        super(ModelNew, self).__init__()
        self.dim = dim

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return _ext.logsoftmax_forward(x, int(self.dim))