import torch
import torch.nn as nn

class Model(nn.Module):
    """
    A model that performs matrix multiplication, applies dropout, and then applies softmax.
    """
    def __init__(self, in_features, out_features, dropout_p):
        super(Model, self).__init__()
        self.matmul = nn.Linear(in_features, out_features)
        self.dropout = nn.Dropout(dropout_p)

    def forward(self, x):
        """
        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, in_features).

        Returns:
            torch.Tensor: Output tensor of shape (batch_size, out_features).
        """
        x = self.matmul(x)
        x = self.dropout(x)
        x = torch.softmax(x, dim=1)  # Softmax over features
        return x

batch_size = 128
in_features = 16384
out_features = 16384
dropout_p = 0.2

def get_inputs():
    return [torch.rand(batch_size, in_features)]

def get_init_inputs():
    return [in_features, out_features, dropout_p]

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

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

torch::Tensor softmax_forward(torch::Tensor input);
'''

_cuda_src = r'''
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <cfloat>

__inline__ __device__ float warp_reduce_max(float v) {
    for (int offset = 16; offset > 0; offset >>= 1)
        v = fmaxf(v, __shfl_down_sync(0xffffffff, v, offset));
    return v;
}

__inline__ __device__ float warp_reduce_sum(float v) {
    for (int offset = 16; offset > 0; offset >>= 1)
        v += __shfl_down_sync(0xffffffff, v, offset);
    return v;
}

template<int BLOCK>
__inline__ __device__ float block_reduce_max(float v) {
    __shared__ float shared[32];
    int lane = threadIdx.x & 31;
    int wid = threadIdx.x >> 5;
    v = warp_reduce_max(v);
    if (lane == 0) shared[wid] = v;
    __syncthreads();
    v = (threadIdx.x < (BLOCK >> 5)) ? shared[lane] : -FLT_MAX;
    if (wid == 0) v = warp_reduce_max(v);
    return v;
}

template<int BLOCK>
__inline__ __device__ float block_reduce_sum(float v) {
    __shared__ float shared[32];
    int lane = threadIdx.x & 31;
    int wid = threadIdx.x >> 5;
    v = warp_reduce_sum(v);
    if (lane == 0) shared[wid] = v;
    __syncthreads();
    v = (threadIdx.x < (BLOCK >> 5)) ? shared[lane] : 0.0f;
    if (wid == 0) v = warp_reduce_sum(v);
    return v;
}

template<int BLOCK>
__global__ void softmax_rows_kernel(const float* __restrict__ x,
                                    float* __restrict__ y,
                                    int rows, int cols) {
    int row = blockIdx.x;
    if (row >= rows) return;
    const float* row_x = x + (size_t)row * cols;
    float* row_y = y + (size_t)row * cols;

    float tmax = -FLT_MAX;
    for (int col = threadIdx.x; col < cols; col += BLOCK) {
        tmax = fmaxf(tmax, row_x[col]);
    }
    float m = block_reduce_max<BLOCK>(tmax);
    m = __shfl_sync(0xffffffff, m, 0);

    float tsum = 0.0f;
    for (int col = threadIdx.x; col < cols; col += BLOCK) {
        float e = __expf(row_x[col] - m);
        row_y[col] = e;
        tsum += e;
    }
    float s = block_reduce_sum<BLOCK>(tsum);
    s = __shfl_sync(0xffffffff, s, 0);
    float inv = 1.0f / s;
    for (int col = threadIdx.x; col < cols; col += BLOCK) {
        row_y[col] *= inv;
    }
}

torch::Tensor softmax_forward(torch::Tensor input) {
    TORCH_CHECK(input.is_cuda(), "input must be CUDA");
    TORCH_CHECK(input.scalar_type() == at::kFloat, "only float32 is supported");
    TORCH_CHECK(input.dim() == 2, "input must be 2D");
    auto x = input.contiguous();
    auto out = torch::empty_like(x);
    int rows = (int)x.size(0);
    int cols = (int)x.size(1);
    constexpr int BLOCK = 256;
    softmax_rows_kernel<BLOCK><<<rows, BLOCK, 0, at::cuda::getCurrentCUDAStream()>>>(
        x.data_ptr<float>(), out.data_ptr<float>(), rows, cols);
    return out;
}
'''

_ext = load_inline(
    name='modelnew_softmax_ext',
    cpp_sources=[_cpp_src],
    cuda_sources=[_cuda_src],
    functions=['softmax_forward'],
    extra_cuda_cflags=['-O3', '--use_fast_math'],
    extra_cflags=['-O3'],
    with_cuda=True,
    verbose=False,
)

class ModelNew(nn.Module):
    def __init__(self, in_features, out_features, dropout_p):
        super(ModelNew, self).__init__()
        self.matmul = nn.Linear(in_features, out_features)
        self.dropout = nn.Dropout(dropout_p)
        self.dropout_p = dropout_p

    def forward(self, x):
        y = F.linear(x, self.matmul.weight, self.matmul.bias)
        if self.training and self.dropout_p != 0.0:
            y = self.dropout(y)
        return _ext.softmax_forward(y)