import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs a 3D convolution, applies minimum operation along a specific dimension, 
    and then applies softmax.
    """
    def __init__(self, in_channels, out_channels, kernel_size, dim):
        super(Model, self).__init__()
        self.conv = nn.Conv3d(in_channels, out_channels, kernel_size)
        self.dim = dim

    def forward(self, x):
        """
        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, in_channels, D, H, W)
        Returns:
            torch.Tensor: Output tensor of shape (batch_size, out_channels, H, W)
        """
        x = self.conv(x)
        x = torch.min(x, dim=self.dim)[0]  # Apply minimum along the specified dimension
        x = torch.softmax(x, dim=1)  # Apply softmax along the channel dimension
        return x

batch_size = 128
in_channels = 3
out_channels = 24  # Increased output channels
D, H, W = 24, 32, 32  # Increased depth
kernel_size = 3
dim = 2  # Dimension along which to apply minimum operation (e.g., depth)

def get_inputs():
    return [torch.rand(batch_size, in_channels, D, H, W)]

def get_init_inputs():
    return [in_channels, out_channels, kernel_size, dim]

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_min_softmax(torch::Tensor x, torch::Tensor weight, torch::Tensor bias);
'''

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

template<int WPB>
__global__ __launch_bounds__(256, 2) void conv3d_min_softmax_k3_kernel(
    const float* __restrict__ x,
    const float* __restrict__ weight,
    const float* __restrict__ bias,
    float* __restrict__ out,
    int N)
{
    constexpr int Cin = 3;
    constexpr int Cout = 24;
    constexpr int D = 24;
    constexpr int H = 32;
    constexpr int W = 32;
    constexpr int Do = 22;
    constexpr int Ho = 30;
    constexpr int Wo = 30;
    constexpr int HW = H*W;
    constexpr int CHW = D*HW;
    constexpr int K = 3;
    int total = N * Ho * Wo;

    int tid = threadIdx.x;
    int lane = tid & 31;
    int warp_in_block = tid >> 5;
    int pix = (int)blockIdx.x * WPB + warp_in_block;
    if (pix >= total) return;

    int ow = pix % Wo;
    int t = pix / Wo;
    int oh = t % Ho;
    int n = t / Ho;

    float minv = INFINITY;
    if (lane < Cout) {
        const int co = lane;
        const float* __restrict__ wco = weight + co * (Cin*K*K*K);
        const float* __restrict__ xbase = x + n * (Cin*CHW) + oh * W + ow;
        float b = __ldg(bias + co);
        float acc[Do];
        #pragma unroll
        for (int od = 0; od < Do; ++od) acc[od] = b;

        #pragma unroll
        for (int ci = 0; ci < Cin; ++ci) {
            #pragma unroll
            for (int kd = 0; kd < K; ++kd) {
                #pragma unroll
                for (int kh = 0; kh < K; ++kh) {
                    #pragma unroll
                    for (int kw = 0; kw < K; ++kw) {
                        float ww = __ldg(wco + ((ci*K + kd)*K + kh)*K + kw);
                        const float* __restrict__ xp = xbase + ci*CHW + kd*HW + kh*W + kw;
                        #pragma unroll
                        for (int od = 0; od < Do; ++od) {
                            acc[od] = fmaf(__ldg(xp + od*HW), ww, acc[od]);
                        }
                    }
                }
            }
        }
        #pragma unroll
        for (int od = 0; od < Do; ++od) minv = fminf(minv, acc[od]);
    }

    float v = (lane < Cout) ? minv : -INFINITY;
    float m = v;
    unsigned mask = 0xffffffffu;
    #pragma unroll
    for (int off = 16; off > 0; off >>= 1) m = fmaxf(m, __shfl_down_sync(mask, m, off));
    m = __shfl_sync(mask, m, 0);
    float e = (lane < Cout) ? __expf(v - m) : 0.0f;
    float s = e;
    #pragma unroll
    for (int off = 16; off > 0; off >>= 1) s += __shfl_down_sync(mask, s, off);
    s = __shfl_sync(mask, s, 0);
    if (lane < Cout) out[((n*Cout + lane)*Ho + oh)*Wo + ow] = e / s;
}

torch::Tensor conv_min_softmax(torch::Tensor x, torch::Tensor weight, torch::Tensor bias) {
    int N = (int)x.size(0);
    auto out = torch::empty({N, 24, 30, 30}, x.options());
    constexpr int WPB = 8;
    int total = N * 30 * 30;
    dim3 block(WPB * 32);
    dim3 grid((total + WPB - 1) / WPB);
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    conv3d_min_softmax_k3_kernel<WPB><<<grid, block, 0, stream>>>(
        x.data_ptr<float>(), weight.data_ptr<float>(), bias.data_ptr<float>(), out.data_ptr<float>(), N);
    return out;
}
'''

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

class ModelNew(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size, dim):
        super(ModelNew, self).__init__()
        self.conv = nn.Conv3d(in_channels, out_channels, kernel_size)
        self.dim = dim
    def forward(self, x):
        return _ext.conv_min_softmax(x, self.conv.weight, self.conv.bias)