You write custom CUDA kernels to replace the pytorch operators in the given architecture to get speedups. 

    You have complete freedom to choose the set of operators you want to replace. You may make the decision to replace some operators with custom CUDA kernels and leave others unchanged. You may replace multiple operators with custom implementations, consider operator fusion opportunities (combining multiple operators into a single kernel, for example, combining matmul+relu), or algorithmic changes (such as online softmax). You are only limited by your imagination.


        Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is: 

        ``` 

        # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn
import torch.nn.functional as F


class Model(nn.Module):
    def __init__(self) -> None:
        super().__init__()

    def forward(self, a, b):
        return a + b


def get_inputs():
    # randomly generate input tensors based on the model architecture
    a = torch.randn(1, 128).cuda()
    b = torch.randn(1, 128).cuda()
    return [a, b]


def get_init_inputs():
    # randomly generate tensors required for initialization based on the model architecture
    return []

        ``` 

        The example new arch with custom CUDA kernels looks like this:
        ```
        # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

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

# Define the custom CUDA kernel for element-wise addition
elementwise_add_source = """
#include <torch/extension.h>
#include <cuda_runtime.h>

__global__ void elementwise_add_kernel(const float* a, const float* b, float* out, int size) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < size) {
        out[idx] = a[idx] + b[idx];
    }
}

torch::Tensor elementwise_add_cuda(torch::Tensor a, torch::Tensor b) {
    auto size = a.numel();
    auto out = torch::zeros_like(a);

    const int block_size = 256;
    const int num_blocks = (size + block_size - 1) / block_size;

    elementwise_add_kernel<<<num_blocks, block_size>>>(a.data_ptr<float>(), b.data_ptr<float>(), out.data_ptr<float>(), size);

    return out;
}
"""

elementwise_add_cpp_source = (
    "torch::Tensor elementwise_add_cuda(torch::Tensor a, torch::Tensor b);"
)

# Compile the inline CUDA code for element-wise addition
elementwise_add = load_inline(
    name="elementwise_add",
    cpp_sources=elementwise_add_cpp_source,
    cuda_sources=elementwise_add_source,
    functions=["elementwise_add_cuda"],
    verbose=True,
    extra_cflags=[""],
    extra_ldflags=[""],
)


class ModelNew(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.elementwise_add = elementwise_add

    def forward(self, a, b):
        return self.elementwise_add.elementwise_add_cuda(a, b)

        ``` 

        
    You are given the following architecture: 

    ```
    # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Performs a transposed 2D convolution with square input and square kernel.

    Args:
        in_channels (int): Number of channels in the input tensor.
        out_channels (int): Number of channels produced by the convolution.
        kernel_size (int): Size of the square convolution kernel.
        stride (int, optional): Stride of the convolution. Defaults to 1.
        padding (int, optional): Padding applied to the input. Defaults to 0.
        output_padding (int, optional): Additional size added to one side of the output shape. Defaults to 0.
        groups (int, optional): Number of blocked connections from input channels to output channels. Defaults to 1.
        bias (bool, optional): If `True`, adds a learnable bias to the output. Defaults to `False`.
    """
    def __init__(self, in_channels: int, out_channels: int, kernel_size: int, stride: int = 1, padding: int = 0, output_padding: int = 0, groups: int = 1, bias: bool = False):
        super(Model, self).__init__()
        self.conv_transpose2d = nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride=stride, padding=padding, output_padding=output_padding, groups=groups, bias=bias)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Performs the transposed 2D convolution.

        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, in_channels, height, width).

        Returns:
            torch.Tensor: Output tensor of shape (batch_size, out_channels, height_out, width_out).
        """
        return self.conv_transpose2d(x)

# Test code
batch_size = 8
in_channels = 64  # double channels for heavier compute
out_channels = 64
kernel_size = 3
# larger square input
height = 1024
width = 1024

def get_inputs():
    x = torch.rand(batch_size, in_channels, height, width)
    return [x]

def get_init_inputs():
    return [in_channels, out_channels, kernel_size]  # Provide in_channels, out_channels, kernel_size for initialization
    ```
    
Optimize the architecture named Model with custom CUDA operators! Name your optimized output architecture ModelNew. Output the new code in codeblocks. Please generate real code, NOT pseudocode, make sure the code compiles and is fully functional. Just output the new model code, no other text, and NO testing code! 



This is turn 3 of 6 for this task. Your previous solution (turn 2) was:
```python
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

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

# ----------------------------------------------------------------------------------
# ConvTranspose2d (stride=1) accelerated with a custom tensor-core CUDA kernel.
#
# Math: ConvTranspose2d(stride=1, padding=p, kernel=K, groups=1)
#     out[b,o,y,x] = sum_{c,u,v} x[b,c,y+u-p,x+v-p] * wf[o,c,u,v]
#     wf[o,c,u,v]  = w[c,o,K-1-u,K-1-v]
# i.e. a cross-correlation with spatially flipped / channel-swapped weights.
#
# Implementation: implicit GEMM evaluated with WMMA (fp16 inputs, fp32 accum).
#   * x is first converted NCHW(fp32) -> NHWC(fp16) with padded halo columns by a
#     custom tiled-transpose kernel. In NHWC the 16-channel GEMM-k slices are
#     contiguous and 16B-aligned for any spatial shift, so A fragments are read
#     straight from global memory with ldmatrix (no im2col, no gather, no smem
#     staging for the main loop).
#   * Rearranged weights wf are stored as [9][C][OC] fp16; B fragments are also
#     read straight from global memory.
#   * Boundary output tiles (where the full 128-wide m tile does not fit) are
#     staged through shared memory with predicated stores.
# ----------------------------------------------------------------------------------

cuda_src = r"""
#include <torch/extension.h>
#include <ATen/cuda/CUDAContext.h>
#include <cuda_runtime.h>
#include <cuda_fp16.h>
#include <mma.h>

using namespace nvcuda;

#define TM 128
#define THREADS 128

// ------------------------- NCHW fp32 -> NHWC fp16 (padded) ----------------------
// x : [B, C, H*W]  ->  xn : [B, H*Wp, C] with halo columns zeroed.
// Per batch this is a [C, S] -> [S, C] matrix transpose with fp32->fp16 cast.
__global__ void nchw2nhwc_pad_kernel(const float* __restrict__ x,
                                     half* __restrict__ xn,
                                     int B, int C, int H, int W, int Wp) {
    __shared__ float tile[32][33];
    const long S = (long)H * (long)W;
    const int s0 = blockIdx.x * 32;
    const int c0 = blockIdx.y * 32;
    const int b  = blockIdx.z;
    const int tx = threadIdx.x;   // 0..31
    const int ty = threadIdx.y;   // 0..7

    const float* xb = x + (long)b * C * S;
    // read 32(s) x 32(c), coalesced along s
#pragma unroll
    for (int i = 0; i < 32; i += 8) {
        const int s = s0 + ty + i;
        const int c = c0 + tx;
        float v = 0.0f;
        if (s < S && c < C) v = xb[(long)c * S + s];
        tile[ty + i][tx] = v;
    }
    __syncthreads();

    half* xnb = xn + (long)b * H * Wp * C;
#pragma unroll
    for (int i = 0; i < 32; i += 8) {
        const int c = c0 + ty + i;
        const int s = s0 + tx;
        if (s >= S || c >= C) continue;
        const int y = s / W;
        const int xx = s - y * W;
        xnb[((long)y * Wp + xx) * C + c] = __float2half(tile[tx][ty + i]);
    }
}

// zero the halo columns xx in [W, Wp) and yy in [H, OH)
__global__ void zero_halo_kernel(half* __restrict__ xn,
                                 int B, int H, int W, int Wp, int C, int OH) {
    const long idx = (long)blockIdx.x * blockDim.x + threadIdx.x;
    const long total = (long)B * OH * (Wp - W) * C;
    if (idx >= total) return;
    const int c  = (int)(idx % C);
    const long t = idx / C;
    const int dx = (int)(t % (Wp - W));
    const long u = t / (Wp - W);
    const int yy = (int)(u % OH);
    const int b  = (int)(u / OH);
    xn[(((long)b * OH + yy) * Wp + (W + dx)) * C + c] = __float2half(0.0f);
}

// ----------------------------- WMMA conv kernel ---------------------------------
// block: 4 warps, computes out[b, 0:OC, y, x0:x0+TM]; OC must be 64.
__global__ void __launch_bounds__(THREADS) imtconv_wmma_kernel(
    const half* __restrict__ xn,   // [B, OH, Wp, C]
    const half* __restrict__ w2,   // [9, C, OC] fp16
    float* __restrict__ out,       // [B, OC, OH, OW]
    int B, int C, int W, int OC, int OH, int OW, int Wp,
    int pad, int tiles_per_row)
{
    extern __shared__ float Cs[];  // only used by boundary tiles: TM x OC

    const int tile_id = blockIdx.x;
    const int y   = tile_id % OH;
    const int tmp = tile_id / OH;
    const int ti  = tmp % tiles_per_row;
    const int b   = tmp / tiles_per_row;
    const int x0  = ti * TM;

    const int warp = threadIdx.x >> 5;

    wmma::fragment<wmma::accumulator, 16, 16, 16, float> acc[2][4];
#pragma unroll
    for (int i = 0; i < 2; ++i)
#pragma unroll
        for (int j = 0; j < 4; ++j)
            wmma::fill_fragment(acc[i][j], 0.0f);

    const bool interior = (x0 + TM <= OW) && (x0 + TM <= W + pad);
    const long HWpC = (long)OH * Wp * C;
    const half* xb = xn + (long)b * HWpC;
    const long OHOW = (long)OH * (long)OW;

    const int m_base = warp * 32;

    for (int u = 0; u < 3; ++u) {
        const int yy = y + u - pad;
        if ((unsigned)yy >= (unsigned)OH) continue;   // halo row -> contributes 0
        const half* xrow = xb + (long)yy * Wp * C + m_base * C + x0 * C;
        const half* wu = w2 + (long)u * 3 * C * OC;
#pragma unroll
        for (int v = 0; v < 3; ++v) {
            const half* wuv = wu + (long)v * C * OC;
            const long aoff = (long)(v - pad) * C;
#pragma unroll
            for (int c0 = 0; c0 < 64; c0 += 16) {
                wmma::fragment<wmma::matrix_b, 16, 16, 16, half, wmma::row_major> bfrag[4];
#pragma unroll
                for (int nf = 0; nf < 4; ++nf)
                    wmma::load_matrix_sync(bfrag[nf], wuv + (long)c0 * OC + nf * 16, OC);
                const half* aptr = xrow + aoff + c0;
#pragma unroll
                for (int mf = 0; mf < 2; ++mf) {
                    wmma::fragment<wmma::matrix_a, 16, 16, 16, half, wmma::row_major> afrag;
                    wmma::load_matrix_sync(afrag, aptr + mf * 16 * C, C);
#pragma unroll
                    for (int nf = 0; nf < 4; ++nf)
                        wmma::mma_sync(acc[mf][nf], afrag, bfrag[nf], acc[mf][nf]);
                }
            }
        }
    }

    if (interior) {
        float* outb = out + (long)b * OC * OHOW + (long)y * OW + x0 + m_base;
#pragma unroll
        for (int mf = 0; mf < 2; ++mf) {
#pragma unroll
            for (int nf = 0; nf < 4; ++nf) {
                wmma::store_matrix_sync(outb + (long)(nf * 16) * OHOW + mf * 16,
                                        acc[mf][nf], OHOW, wmma::mem_col_major);
            }
        }
    } else {
        // boundary tile: stage through shared memory, then predicated writes
        const int m_valid = OW - x0;
        float* Cw = Cs + warp * (32 * OC);
#pragma unroll
        for (int mf = 0; mf < 2; ++mf) {
#pragma unroll
            for (int nf = 0; nf < 4; ++nf) {
                wmma::store_matrix_sync(Cw + mf * 16 * OC + nf * 16,
                                        acc[mf][nf], OC, wmma::mem_row_major);
            }
        }
        __syncthreads();
        const long bbase = (long)b * OC * OHOW + (long)y * OW + x0;
        for (int idx = threadIdx.x; idx < m_valid * OC; idx += THREADS) {
            const int m = idx / OC;
            const int o = idx - m * OC;
            out[bbase + (long)o * OHOW + m] = Cs[m * OC + o];
        }
    }
}

// --------------------------------- host wrapper ----------------------------------
torch::Tensor conv_transpose2d_wmma(torch::Tensor x, torch::Tensor w2,
                                    int64_t pad) {
    TORCH_CHECK(x.is_cuda() && x.dtype() == torch::kFloat32, "x must be CUDA fp32");
    TORCH_CHECK(w2.is_cuda() && w2.dtype() == torch::kHalf, "w2 must be CUDA fp16");
    const int B  = x.size(0);
    const int C  = x.size(1);
    const int H  = x.size(2);
    const int W  = x.size(3);
    const int OC = w2.size(2);
    TORCH_CHECK(C == 64 && OC == 64, "fast kernel requires C == OC == 64");

    const int OH = H + 2;          // K = 3
    const int OW = W + 2;
    const int Wp = W + TM + 2;     // padded columns (halo)

    auto opts_h = x.options().dtype(torch::kHalf);
    auto xn = torch::empty({(long)B * OH * Wp * C}, opts_h);

    auto stream = at::cuda::getCurrentCUDAStream();

    // halo zero fill
    {
        const long total = (long)B * OH * (Wp - W) * C;
        const int threads = 256;
        const long blocks = (total + threads - 1) / threads;
        zero_halo_kernel<<<(unsigned)blocks, threads, 0, stream>>>(
            reinterpret_cast<half*>(xn.data_ptr<at::Half>()),
            B, H, W, Wp, C, OH);
    }
    // transpose + cast
    {
        const long S = (long)H * W;
        dim3 block(32, 8);
        dim3 grid((unsigned)((S + 31) / 32), (unsigned)((C + 31) / 32), (unsigned)B);
        nchw2nhwc_pad_kernel<<<grid, block, 0, stream>>>(
            x.data_ptr<float>(),
            reinterpret_cast<half*>(xn.data_ptr<at::Half>()),
            B, C, H, W, Wp);
    }

    auto out = torch::empty({B, OC, OH, OW}, x.options());

    const int tiles_per_row = (OW + TM - 1) / TM;
    const long blocks = (long)B * OH * tiles_per_row;

    const bool has_edge = (OW % TM) != 0;
    size_t sh = has_edge ? (size_t)TM * OC * sizeof(float) : 0;
    if (sh > 48 * 1024) {
        static size_t max_sh = 0;
        if (sh > max_sh) {
            cudaFuncSetAttribute(imtconv_wmma_kernel,
                                 cudaFuncAttributeMaxDynamicSharedMemorySize, (int)sh);
            max_sh = sh;
        }
    }

    imtconv_wmma_kernel<<<(unsigned)blocks, THREADS, sh, stream>>>(
        reinterpret_cast<const half*>(xn.data_ptr<at::Half>()),
        reinterpret_cast<const half*>(w2.data_ptr<at::Half>()),
        out.data_ptr<float>(),
        B, C, W, OC, OH, OW, Wp, (int)pad, tiles_per_row);

    return out;
}
"""

cpp_src = "torch::Tensor conv_transpose2d_wmma(torch::Tensor x, torch::Tensor w2, int64_t pad);"

_conv_mod = load_inline(
    name="conv_transpose2d_wmma_nhwc",
    cpp_sources=cpp_src,
    cuda_sources=cuda_src,
    functions=["conv_transpose2d_wmma"],
    verbose=False,
    extra_cuda_cflags=["-O3", "--use_fast_math"],
)


class ModelNew(nn.Module):
    """
    ConvTranspose2d accelerated with a custom tensor-core implicit-GEMM CUDA
    kernel (fast path: stride=1, groups=1, kernel 3x3, in_channels =
    out_channels = 64, no bias, CUDA fp32 input). All other configurations use
    the cuDNN implementation.
    """

    def __init__(self, in_channels: int, out_channels: int, kernel_size: int, stride: int = 1,
                 padding: int = 0, output_padding: int = 0, groups: int = 1, bias: bool = False):
        super(ModelNew, self).__init__()
        self.conv_transpose2d = nn.ConvTranspose2d(in_channels, out_channels, kernel_size,
                                                   stride=stride, padding=padding,
                                                   output_padding=output_padding,
                                                   groups=groups, bias=bias)
        self.stride = stride
        self.padding = padding
        self.groups = groups
        self.kernel_size = kernel_size

        self._fast = (stride == 1 and groups == 1 and kernel_size == 3
                      and output_padding == 0 and 0 <= padding <= 2
                      and in_channels == 64 and out_channels == 64)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        w = self.conv_transpose2d
        if (self._fast and w.bias is None and x.is_cuda
                and x.dtype == torch.float32 and x.dim() == 4 and x.size(1) == 64):
            # rearranged weights wf[u,v,c,o] = w[c,o,2-u,2-v], rebuilt every
            # call (tiny: 36k elements) so weight updates are always honored.
            w2 = (w.weight.detach()
                  .permute(2, 3, 0, 1)          # [3,3,C,OC]
                  .flip(0, 1)                   # spatial flip
                  .contiguous()
                  .to(torch.float16))
            return _conv_mod.conv_transpose2d_wmma(x.contiguous(), w2, self.padding)
        return F.conv_transpose2d(x, w.weight, w.bias, self.stride, self.padding,
                                  0, self.groups)
```
Result of evaluating it:
The solution compiled and ran but its output did not match the reference. Details:
```
{"avg_difference": ["nan", "nan", "nan", "nan", "nan"], "correctness_issue": "Output mismatch", "correctness_trials": "(0 / 5)", "max_difference": ["nan", "nan", "nan", "nan", "nan"]}
```

Write an improved complete ModelNew: fix any error, and if the solution is already correct, make it faster. Output the new code in one ```python codeblock. Just the model code, no other text, and NO testing code.
