import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs min reduction over a specific dimension.
    """
    def __init__(self, dim: int):
        """
        Initializes the model with the dimension to reduce over.

        Args:
            dim (int): The dimension to reduce over.
        """
        super(Model, self).__init__()
        self.dim = dim

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Applies min reduction over the specified dimension to the input tensor.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            torch.Tensor: Output tensor after min reduction over the specified dimension.
        """
        return torch.min(x, dim=self.dim)[0]

batch_size = 128
dim1 = 4096
dim2 = 4095

def get_inputs():
    x = torch.rand(batch_size, dim1, dim2)
    return [x]

def get_init_inputs():
    return [1] # Example, change to desired dimension

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

cpp_src = r'''
#include <torch/extension.h>
torch::Tensor min_reduce_cuda(torch::Tensor x, int64_t dim);
'''

cuda_src = r'''
#include <torch/extension.h>
#include <ATen/cuda/CUDAContext.h>
#include <cuda_runtime.h>
#include <cfloat>

__global__ void min_dim1_kernel(const float* __restrict__ x, float* __restrict__ out,
                                int B, int D1, int D2) {
    int b = blockIdx.y;
    int col = blockIdx.x * blockDim.x + threadIdx.x;
    if (col >= D2) return;
    const float* p = x + ((long long)b * D1 * D2 + col);
    float mn = FLT_MAX;
    #pragma unroll 4
    for (int k = 0; k < D1; ++k) {
        float v = *p;
        mn = fminf(mn, v);
        p += D2;
    }
    out[(long long)b * D2 + col] = mn;
}

__global__ void min_dim2_kernel(const float* __restrict__ x, float* __restrict__ out,
                                int total_rows, int D2) {
    extern __shared__ float smem[];
    int row = blockIdx.x;
    int tid = threadIdx.x;
    const float* base = x + (long long)row * D2;
    float mn = FLT_MAX;
    for (int i = tid; i < D2; i += blockDim.x) {
        mn = fminf(mn, base[i]);
    }
    smem[tid] = mn;
    __syncthreads();
    for (int s = blockDim.x >> 1; s > 0; s >>= 1) {
        if (tid < s) smem[tid] = fminf(smem[tid], smem[tid+s]);
        __syncthreads();
    }
    if (tid == 0) out[row] = smem[0];
}

__global__ void min_dim0_kernel(const float* __restrict__ x, float* __restrict__ out,
                                int B, int D1, int D2) {
    long long idx = (long long)blockIdx.x * blockDim.x + threadIdx.x;
    long long total = (long long)D1 * D2;
    if (idx >= total) return;
    int k = idx / D2;
    int col = idx - (long long)k * D2;
    float mn = FLT_MAX;
    const float* p = x + (long long)k * D2 + col;
    long long stride = (long long)D1 * D2;
    for (int b = 0; b < B; ++b) {
        mn = fminf(mn, *p);
        p += stride;
    }
    out[idx] = mn;
}

torch::Tensor min_reduce_cuda(torch::Tensor x, int64_t dim) {
    TORCH_CHECK(x.is_cuda(), "x must be CUDA");
    TORCH_CHECK(x.scalar_type() == torch::kFloat32, "x must be float32");
    TORCH_CHECK(x.dim() == 3, "x must be 3D");
    if (dim < 0) dim += 3;
    TORCH_CHECK(dim >= 0 && dim < 3, "bad dim");
    auto xc = x.contiguous();
    int B = (int)xc.size(0);
    int D1 = (int)xc.size(1);
    int D2 = (int)xc.size(2);
    torch::Tensor out;
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    const float* xp = xc.data_ptr<float>();
    if (dim == 1) {
        out = torch::empty({B, D2}, xc.options());
        int threads = 512;
        dim3 grid((D2 + threads - 1) / threads, B);
        min_dim1_kernel<<<grid, threads, 0, stream>>>(xp, out.data_ptr<float>(), B, D1, D2);
    } else if (dim == 2) {
        out = torch::empty({B, D1}, xc.options());
        int threads = 256;
        min_dim2_kernel<<<B * D1, threads, threads * sizeof(float), stream>>>(xp, out.data_ptr<float>(), B * D1, D2);
    } else {
        out = torch::empty({D1, D2}, xc.options());
        int threads = 256;
        long long total = (long long)D1 * D2;
        min_dim0_kernel<<<(total + threads - 1) / threads, threads, 0, stream>>>(xp, out.data_ptr<float>(), B, D1, D2);
    }
    return out;
}
'''

_ext = load_inline(name='min_reduce_ext_final', cpp_sources=cpp_src, cuda_sources=cuda_src,
                   functions=['min_reduce_cuda'], extra_cuda_cflags=['-O3'], verbose=False)

class ModelNew(nn.Module):
    def __init__(self, dim: int):
        super(ModelNew, self).__init__()
        self.dim = dim
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return _ext.min_reduce_cuda(x, self.dim)