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 fused_conv_min_softmax_forward(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>

// Specialized for the problem dimensions:
// x:      [N, 3, 24, 32, 32]
// weight: [24, 3, 3, 3, 3]
// output: [N, 24, 30, 30]
__global__ __launch_bounds__(32, 8)
void fused_conv_min_softmax_kernel(const float* __restrict__ x,
                                   const float* __restrict__ w,
                                   const float* __restrict__ b,
                                   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 KD = 3;
    constexpr int KH = 3;
    constexpr int KW = 3;
    constexpr int DO = D - KD + 1;   // 22
    constexpr int HO = H - KH + 1;   // 30
    constexpr int WO = W - KW + 1;   // 30
    constexpr int HW = H * W;
    constexpr int CHW = D * H * W;
    constexpr int OHW = HO * WO;

    int pix = blockIdx.x;
    int lane = threadIdx.x;

    int n = pix / OHW;
    int r = pix - n * OHW;
    int oh = r / WO;
    int ow = r - oh * WO;

    float v = -CUDART_INF_F;

    if (lane < COUT) {
        int c = lane;
        const float* wc = w + c * (CIN * KD * KH * KW);
        float mn = CUDART_INF_F;
        float bias = __ldg(b + c);

        // one warp computes all output channels for a single (n, oh, ow)
        #pragma unroll 1
        for (int od = 0; od < DO; ++od) {
            float acc = bias;
            int x_base = n * (CIN * CHW) + od * HW + oh * W + ow;

            #pragma unroll
            for (int ci = 0; ci < CIN; ++ci) {
                int xb_ci = x_base + ci * CHW;
                int wb_ci = ci * (KD * KH * KW);
                #pragma unroll
                for (int kd = 0; kd < KD; ++kd) {
                    int xb_kd = xb_ci + kd * HW;
                    int wb_kd = wb_ci + kd * (KH * KW);
                    #pragma unroll
                    for (int kh = 0; kh < KH; ++kh) {
                        int xb_kh = xb_kd + kh * W;
                        int wb_kh = wb_kd + kh * KW;
                        acc = fmaf(__ldg(x + xb_kh + 0), __ldg(wc + wb_kh + 0), acc);
                        acc = fmaf(__ldg(x + xb_kh + 1), __ldg(wc + wb_kh + 1), acc);
                        acc = fmaf(__ldg(x + xb_kh + 2), __ldg(wc + wb_kh + 2), acc);
                    }
                }
            }
            mn = fminf(mn, acc);
        }
        v = mn;
    }

    unsigned mask = 0xffffffffu;
    float m = v;
    m = fmaxf(m, __shfl_down_sync(mask, m, 16));
    m = fmaxf(m, __shfl_down_sync(mask, m, 8));
    m = fmaxf(m, __shfl_down_sync(mask, m, 4));
    m = fmaxf(m, __shfl_down_sync(mask, m, 2));
    m = fmaxf(m, __shfl_down_sync(mask, m, 1));
    m = __shfl_sync(mask, m, 0);

    float e = (lane < COUT) ? expf(v - m) : 0.0f;
    float s = e;
    s += __shfl_down_sync(mask, s, 16);
    s += __shfl_down_sync(mask, s, 8);
    s += __shfl_down_sync(mask, s, 4);
    s += __shfl_down_sync(mask, s, 2);
    s += __shfl_down_sync(mask, s, 1);
    s = __shfl_sync(mask, s, 0);

    if (lane < COUT) {
        out[((n * COUT + lane) * HO + oh) * WO + ow] = e / s;
    }
}

torch::Tensor fused_conv_min_softmax_forward(torch::Tensor x, torch::Tensor weight, torch::Tensor bias) {
    const int64_t N64 = x.size(0);
    auto out = torch::empty({N64, 24, 30, 30}, x.options());
    int N = static_cast<int>(N64);
    int blocks = N * 30 * 30;
    dim3 grid(blocks);
    dim3 block(32);
    fused_conv_min_softmax_kernel<<<grid, block, 0, at::cuda::getCurrentCUDAStream()>>>(
        x.data_ptr<float>(), weight.data_ptr<float>(), bias.data_ptr<float>(), out.data_ptr<float>(), N);
    return out;
}
'''

_ext = load_inline(
    name='fused_conv_min_softmax_ext_v1',
    cpp_sources=_cpp_src,
    cuda_sources=_cuda_src,
    functions=['fused_conv_min_softmax_forward'],
    extra_cflags=['-O3'],
    extra_cuda_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.fused_conv_min_softmax_forward(x, self.conv.weight, self.conv.bias)