Added cuda ops

op/__init__.py

@@ -0,0 +1,2 @@
+from .fused_act import FusedLeakyReLU, fused_leaky_relu
+from .upfirdn2d import upfirdn2d

op/conv2d_gradfix.py

@@ -0,0 +1,227 @@
+import contextlib
+import warnings
+
+import torch
+from torch import autograd
+from torch.nn import functional as F
+
+enabled = True
+weight_gradients_disabled = False
+
+
+@contextlib.contextmanager
+def no_weight_gradients():
+    global weight_gradients_disabled
+
+    old = weight_gradients_disabled
+    weight_gradients_disabled = True
+    yield
+    weight_gradients_disabled = old
+
+
+def conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1):
+    if could_use_op(input):
+        return conv2d_gradfix(
+            transpose=False,
+            weight_shape=weight.shape,
+            stride=stride,
+            padding=padding,
+            output_padding=0,
+            dilation=dilation,
+            groups=groups,
+        ).apply(input, weight, bias)
+
+    return F.conv2d(
+        input=input,
+        weight=weight,
+        bias=bias,
+        stride=stride,
+        padding=padding,
+        dilation=dilation,
+        groups=groups,
+    )
+
+
+def conv_transpose2d(
+    input,
+    weight,
+    bias=None,
+    stride=1,
+    padding=0,
+    output_padding=0,
+    groups=1,
+    dilation=1,
+):
+    if could_use_op(input):
+        return conv2d_gradfix(
+            transpose=True,
+            weight_shape=weight.shape,
+            stride=stride,
+            padding=padding,
+            output_padding=output_padding,
+            groups=groups,
+            dilation=dilation,
+        ).apply(input, weight, bias)
+
+    return F.conv_transpose2d(
+        input=input,
+        weight=weight,
+        bias=bias,
+        stride=stride,
+        padding=padding,
+        output_padding=output_padding,
+        dilation=dilation,
+        groups=groups,
+    )
+
+
+def could_use_op(input):
+    if (not enabled) or (not torch.backends.cudnn.enabled):
+        return False
+
+    if input.device.type != "cuda":
+        return False
+
+    if any(torch.__version__.startswith(x) for x in ["1.7.", "1.8."]):
+        return True
+
+    warnings.warn(
+        f"conv2d_gradfix not supported on PyTorch {torch.__version__}. Falling back to torch.nn.functional.conv2d()."
+    )
+
+    return False
+
+
+def ensure_tuple(xs, ndim):
+    xs = tuple(xs) if isinstance(xs, (tuple, list)) else (xs,) * ndim
+
+    return xs
+
+
+conv2d_gradfix_cache = dict()
+
+
+def conv2d_gradfix(
+    transpose, weight_shape, stride, padding, output_padding, dilation, groups
+):
+    ndim = 2
+    weight_shape = tuple(weight_shape)
+    stride = ensure_tuple(stride, ndim)
+    padding = ensure_tuple(padding, ndim)
+    output_padding = ensure_tuple(output_padding, ndim)
+    dilation = ensure_tuple(dilation, ndim)
+
+    key = (transpose, weight_shape, stride, padding, output_padding, dilation, groups)
+    if key in conv2d_gradfix_cache:
+        return conv2d_gradfix_cache[key]
+
+    common_kwargs = dict(
+        stride=stride, padding=padding, dilation=dilation, groups=groups
+    )
+
+    def calc_output_padding(input_shape, output_shape):
+        if transpose:
+            return [0, 0]
+
+        return [
+            input_shape[i + 2]
+            - (output_shape[i + 2] - 1) * stride[i]
+            - (1 - 2 * padding[i])
+            - dilation[i] * (weight_shape[i + 2] - 1)
+            for i in range(ndim)
+        ]
+
+    class Conv2d(autograd.Function):
+        @staticmethod
+        def forward(ctx, input, weight, bias):
+            if not transpose:
+                out = F.conv2d(input=input, weight=weight, bias=bias, **common_kwargs)
+
+            else:
+                out = F.conv_transpose2d(
+                    input=input,
+                    weight=weight,
+                    bias=bias,
+                    output_padding=output_padding,
+                    **common_kwargs,
+                )
+
+            ctx.save_for_backward(input, weight)
+
+            return out
+
+        @staticmethod
+        def backward(ctx, grad_output):
+            input, weight = ctx.saved_tensors
+            grad_input, grad_weight, grad_bias = None, None, None
+
+            if ctx.needs_input_grad[0]:
+                p = calc_output_padding(
+                    input_shape=input.shape, output_shape=grad_output.shape
+                )
+                grad_input = conv2d_gradfix(
+                    transpose=(not transpose),
+                    weight_shape=weight_shape,
+                    output_padding=p,
+                    **common_kwargs,
+                ).apply(grad_output, weight, None)
+
+            if ctx.needs_input_grad[1] and not weight_gradients_disabled:
+                grad_weight = Conv2dGradWeight.apply(grad_output, input)
+
+            if ctx.needs_input_grad[2]:
+                grad_bias = grad_output.sum((0, 2, 3))
+
+            return grad_input, grad_weight, grad_bias
+
+    class Conv2dGradWeight(autograd.Function):
+        @staticmethod
+        def forward(ctx, grad_output, input):
+            op = torch._C._jit_get_operation(
+                "aten::cudnn_convolution_backward_weight"
+                if not transpose
+                else "aten::cudnn_convolution_transpose_backward_weight"
+            )
+            flags = [
+                torch.backends.cudnn.benchmark,
+                torch.backends.cudnn.deterministic,
+                torch.backends.cudnn.allow_tf32,
+            ]
+            grad_weight = op(
+                weight_shape,
+                grad_output,
+                input,
+                padding,
+                stride,
+                dilation,
+                groups,
+                *flags,
+            )
+            ctx.save_for_backward(grad_output, input)
+
+            return grad_weight
+
+        @staticmethod
+        def backward(ctx, grad_grad_weight):
+            grad_output, input = ctx.saved_tensors
+            grad_grad_output, grad_grad_input = None, None
+
+            if ctx.needs_input_grad[0]:
+                grad_grad_output = Conv2d.apply(input, grad_grad_weight, None)
+
+            if ctx.needs_input_grad[1]:
+                p = calc_output_padding(
+                    input_shape=input.shape, output_shape=grad_output.shape
+                )
+                grad_grad_input = conv2d_gradfix(
+                    transpose=(not transpose),
+                    weight_shape=weight_shape,
+                    output_padding=p,
+                    **common_kwargs,
+                ).apply(grad_output, grad_grad_weight, None)
+
+            return grad_grad_output, grad_grad_input
+
+    conv2d_gradfix_cache[key] = Conv2d
+
+    return Conv2d

op/fused_act.py

@@ -0,0 +1,119 @@
+import os
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.autograd import Function
+from torch.utils.cpp_extension import load
+
+
+module_path = os.path.dirname(__file__)
+fused = load(
+    "fused",
+    sources=[
+        os.path.join(module_path, "fused_bias_act.cpp"),
+        os.path.join(module_path, "fused_bias_act_kernel.cu"),
+    ],
+)
+
+
+class FusedLeakyReLUFunctionBackward(Function):
+    @staticmethod
+    def forward(ctx, grad_output, out, bias, negative_slope, scale):
+        ctx.save_for_backward(out)
+        ctx.negative_slope = negative_slope
+        ctx.scale = scale
+
+        empty = grad_output.new_empty(0)
+
+        grad_input = fused.fused_bias_act(
+            grad_output.contiguous(), empty, out, 3, 1, negative_slope, scale
+        )
+
+        dim = [0]
+
+        if grad_input.ndim > 2:
+            dim += list(range(2, grad_input.ndim))
+
+        if bias:
+            grad_bias = grad_input.sum(dim).detach()
+
+        else:
+            grad_bias = empty
+
+        return grad_input, grad_bias
+
+    @staticmethod
+    def backward(ctx, gradgrad_input, gradgrad_bias):
+        out, = ctx.saved_tensors
+        gradgrad_out = fused.fused_bias_act(
+            gradgrad_input, gradgrad_bias, out, 3, 1, ctx.negative_slope, ctx.scale
+        )
+
+        return gradgrad_out, None, None, None, None
+
+
+class FusedLeakyReLUFunction(Function):
+    @staticmethod
+    def forward(ctx, input, bias, negative_slope, scale):
+        empty = input.new_empty(0)
+
+        ctx.bias = bias is not None
+
+        if bias is None:
+            bias = empty
+
+        out = fused.fused_bias_act(input, bias, empty, 3, 0, negative_slope, scale)
+        ctx.save_for_backward(out)
+        ctx.negative_slope = negative_slope
+        ctx.scale = scale
+
+        return out
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        out, = ctx.saved_tensors
+
+        grad_input, grad_bias = FusedLeakyReLUFunctionBackward.apply(
+            grad_output, out, ctx.bias, ctx.negative_slope, ctx.scale
+        )
+
+        if not ctx.bias:
+            grad_bias = None
+
+        return grad_input, grad_bias, None, None
+
+
+class FusedLeakyReLU(nn.Module):
+    def __init__(self, channel, bias=True, negative_slope=0.2, scale=2 ** 0.5):
+        super().__init__()
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(channel))
+
+        else:
+            self.bias = None
+
+        self.negative_slope = negative_slope
+        self.scale = scale
+
+    def forward(self, input):
+        return fused_leaky_relu(input, self.bias, self.negative_slope, self.scale)
+
+
+def fused_leaky_relu(input, bias=None, negative_slope=0.2, scale=2 ** 0.5):
+    if input.device.type == "cpu":
+        if bias is not None:
+            rest_dim = [1] * (input.ndim - bias.ndim - 1)
+            return (
+                F.leaky_relu(
+                    input + bias.view(1, bias.shape[0], *rest_dim), negative_slope=0.2
+                )
+                * scale
+            )
+
+        else:
+            return F.leaky_relu(input, negative_slope=0.2) * scale
+
+    else:
+        return FusedLeakyReLUFunction.apply(input, bias, negative_slope, scale)

op/fused_bias_act.cpp

@@ -0,0 +1,32 @@
+
+#include <ATen/ATen.h>
+#include <torch/extension.h>
+
+torch::Tensor fused_bias_act_op(const torch::Tensor &input,
+                                const torch::Tensor &bias,
+                                const torch::Tensor &refer, int act, int grad,
+                                float alpha, float scale);
+
+#define CHECK_CUDA(x)                                                          \
+  TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x)                                                    \
+  TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_INPUT(x)                                                         \
+  CHECK_CUDA(x);                                                               \
+  CHECK_CONTIGUOUS(x)
+
+torch::Tensor fused_bias_act(const torch::Tensor &input,
+                             const torch::Tensor &bias,
+                             const torch::Tensor &refer, int act, int grad,
+                             float alpha, float scale) {
+  CHECK_INPUT(input);
+  CHECK_INPUT(bias);
+
+  at::DeviceGuard guard(input.device());
+
+  return fused_bias_act_op(input, bias, refer, act, grad, alpha, scale);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("fused_bias_act", &fused_bias_act, "fused bias act (CUDA)");
+}

op/fused_bias_act_kernel.cu

@@ -0,0 +1,105 @@
+// Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+//
+// This work is made available under the Nvidia Source Code License-NC.
+// To view a copy of this license, visit
+// https://nvlabs.github.io/stylegan2/license.html
+
+#include <torch/types.h>
+
+#include <ATen/ATen.h>
+#include <ATen/AccumulateType.h>
+#include <ATen/cuda/CUDAApplyUtils.cuh>
+#include <ATen/cuda/CUDAContext.h>
+
+
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+template <typename scalar_t>
+static __global__ void
+fused_bias_act_kernel(scalar_t *out, const scalar_t *p_x, const scalar_t *p_b,
+                      const scalar_t *p_ref, int act, int grad, scalar_t alpha,
+                      scalar_t scale, int loop_x, int size_x, int step_b,
+                      int size_b, int use_bias, int use_ref) {
+  int xi = blockIdx.x * loop_x * blockDim.x + threadIdx.x;
+
+  scalar_t zero = 0.0;
+
+  for (int loop_idx = 0; loop_idx < loop_x && xi < size_x;
+       loop_idx++, xi += blockDim.x) {
+    scalar_t x = p_x[xi];
+
+    if (use_bias) {
+      x += p_b[(xi / step_b) % size_b];
+    }
+
+    scalar_t ref = use_ref ? p_ref[xi] : zero;
+
+    scalar_t y;
+
+    switch (act * 10 + grad) {
+    default:
+    case 10:
+      y = x;
+      break;
+    case 11:
+      y = x;
+      break;
+    case 12:
+      y = 0.0;
+      break;
+
+    case 30:
+      y = (x > 0.0) ? x : x * alpha;
+      break;
+    case 31:
+      y = (ref > 0.0) ? x : x * alpha;
+      break;
+    case 32:
+      y = 0.0;
+      break;
+    }
+
+    out[xi] = y * scale;
+  }
+}
+
+torch::Tensor fused_bias_act_op(const torch::Tensor &input,
+                                const torch::Tensor &bias,
+                                const torch::Tensor &refer, int act, int grad,
+                                float alpha, float scale) {
+  int curDevice = -1;
+  cudaGetDevice(&curDevice);
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  auto x = input.contiguous();
+  auto b = bias.contiguous();
+  auto ref = refer.contiguous();
+
+  int use_bias = b.numel() ? 1 : 0;
+  int use_ref = ref.numel() ? 1 : 0;
+
+  int size_x = x.numel();
+  int size_b = b.numel();
+  int step_b = 1;
+
+  for (int i = 1 + 1; i < x.dim(); i++) {
+    step_b *= x.size(i);
+  }
+
+  int loop_x = 4;
+  int block_size = 4 * 32;
+  int grid_size = (size_x - 1) / (loop_x * block_size) + 1;
+
+  auto y = torch::empty_like(x);
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      x.scalar_type(), "fused_bias_act_kernel", [&] {
+        fused_bias_act_kernel<scalar_t><<<grid_size, block_size, 0, stream>>>(
+            y.data_ptr<scalar_t>(), x.data_ptr<scalar_t>(),
+            b.data_ptr<scalar_t>(), ref.data_ptr<scalar_t>(), act, grad, alpha,
+            scale, loop_x, size_x, step_b, size_b, use_bias, use_ref);
+      });
+
+  return y;
+}

op/upfirdn2d.cpp

@@ -0,0 +1,31 @@
+#include <ATen/ATen.h>
+#include <torch/extension.h>
+
+torch::Tensor upfirdn2d_op(const torch::Tensor &input,
+                           const torch::Tensor &kernel, int up_x, int up_y,
+                           int down_x, int down_y, int pad_x0, int pad_x1,
+                           int pad_y0, int pad_y1);
+
+#define CHECK_CUDA(x)                                                          \
+  TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x)                                                    \
+  TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_INPUT(x)                                                         \
+  CHECK_CUDA(x);                                                               \
+  CHECK_CONTIGUOUS(x)
+
+torch::Tensor upfirdn2d(const torch::Tensor &input, const torch::Tensor &kernel,
+                        int up_x, int up_y, int down_x, int down_y, int pad_x0,
+                        int pad_x1, int pad_y0, int pad_y1) {
+  CHECK_INPUT(input);
+  CHECK_INPUT(kernel);
+
+  at::DeviceGuard guard(input.device());
+
+  return upfirdn2d_op(input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1,
+                      pad_y0, pad_y1);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("upfirdn2d", &upfirdn2d, "upfirdn2d (CUDA)");
+}

op/upfirdn2d.py

@@ -0,0 +1,209 @@
+from collections import abc
+import os
+
+import torch
+from torch.nn import functional as F
+from torch.autograd import Function
+from torch.utils.cpp_extension import load
+
+
+module_path = os.path.dirname(__file__)
+upfirdn2d_op = load(
+    "upfirdn2d",
+    sources=[
+        os.path.join(module_path, "upfirdn2d.cpp"),
+        os.path.join(module_path, "upfirdn2d_kernel.cu"),
+    ],
+)
+
+
+class UpFirDn2dBackward(Function):
+    @staticmethod
+    def forward(
+        ctx, grad_output, kernel, grad_kernel, up, down, pad, g_pad, in_size, out_size
+    ):
+
+        up_x, up_y = up
+        down_x, down_y = down
+        g_pad_x0, g_pad_x1, g_pad_y0, g_pad_y1 = g_pad
+
+        grad_output = grad_output.reshape(-1, out_size[0], out_size[1], 1)
+
+        grad_input = upfirdn2d_op.upfirdn2d(
+            grad_output,
+            grad_kernel,
+            down_x,
+            down_y,
+            up_x,
+            up_y,
+            g_pad_x0,
+            g_pad_x1,
+            g_pad_y0,
+            g_pad_y1,
+        )
+        grad_input = grad_input.view(in_size[0], in_size[1], in_size[2], in_size[3])
+
+        ctx.save_for_backward(kernel)
+
+        pad_x0, pad_x1, pad_y0, pad_y1 = pad
+
+        ctx.up_x = up_x
+        ctx.up_y = up_y
+        ctx.down_x = down_x
+        ctx.down_y = down_y
+        ctx.pad_x0 = pad_x0
+        ctx.pad_x1 = pad_x1
+        ctx.pad_y0 = pad_y0
+        ctx.pad_y1 = pad_y1
+        ctx.in_size = in_size
+        ctx.out_size = out_size
+
+        return grad_input
+
+    @staticmethod
+    def backward(ctx, gradgrad_input):
+        kernel, = ctx.saved_tensors
+
+        gradgrad_input = gradgrad_input.reshape(-1, ctx.in_size[2], ctx.in_size[3], 1)
+
+        gradgrad_out = upfirdn2d_op.upfirdn2d(
+            gradgrad_input,
+            kernel,
+            ctx.up_x,
+            ctx.up_y,
+            ctx.down_x,
+            ctx.down_y,
+            ctx.pad_x0,
+            ctx.pad_x1,
+            ctx.pad_y0,
+            ctx.pad_y1,
+        )
+        # gradgrad_out = gradgrad_out.view(ctx.in_size[0], ctx.out_size[0], ctx.out_size[1], ctx.in_size[3])
+        gradgrad_out = gradgrad_out.view(
+            ctx.in_size[0], ctx.in_size[1], ctx.out_size[0], ctx.out_size[1]
+        )
+
+        return gradgrad_out, None, None, None, None, None, None, None, None
+
+
+class UpFirDn2d(Function):
+    @staticmethod
+    def forward(ctx, input, kernel, up, down, pad):
+        up_x, up_y = up
+        down_x, down_y = down
+        pad_x0, pad_x1, pad_y0, pad_y1 = pad
+
+        kernel_h, kernel_w = kernel.shape
+        batch, channel, in_h, in_w = input.shape
+        ctx.in_size = input.shape
+
+        input = input.reshape(-1, in_h, in_w, 1)
+
+        ctx.save_for_backward(kernel, torch.flip(kernel, [0, 1]))
+
+        out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h + down_y) // down_y
+        out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w + down_x) // down_x
+        ctx.out_size = (out_h, out_w)
+
+        ctx.up = (up_x, up_y)
+        ctx.down = (down_x, down_y)
+        ctx.pad = (pad_x0, pad_x1, pad_y0, pad_y1)
+
+        g_pad_x0 = kernel_w - pad_x0 - 1
+        g_pad_y0 = kernel_h - pad_y0 - 1
+        g_pad_x1 = in_w * up_x - out_w * down_x + pad_x0 - up_x + 1
+        g_pad_y1 = in_h * up_y - out_h * down_y + pad_y0 - up_y + 1
+
+        ctx.g_pad = (g_pad_x0, g_pad_x1, g_pad_y0, g_pad_y1)
+
+        out = upfirdn2d_op.upfirdn2d(
+            input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1
+        )
+        # out = out.view(major, out_h, out_w, minor)
+        out = out.view(-1, channel, out_h, out_w)
+
+        return out
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        kernel, grad_kernel = ctx.saved_tensors
+
+        grad_input = None
+
+        if ctx.needs_input_grad[0]:
+            grad_input = UpFirDn2dBackward.apply(
+                grad_output,
+                kernel,
+                grad_kernel,
+                ctx.up,
+                ctx.down,
+                ctx.pad,
+                ctx.g_pad,
+                ctx.in_size,
+                ctx.out_size,
+            )
+
+        return grad_input, None, None, None, None
+
+
+def upfirdn2d(input, kernel, up=1, down=1, pad=(0, 0)):
+    if not isinstance(up, abc.Iterable):
+        up = (up, up)
+
+    if not isinstance(down, abc.Iterable):
+        down = (down, down)
+
+    if len(pad) == 2:
+        pad = (pad[0], pad[1], pad[0], pad[1])
+
+    if input.device.type == "cpu":
+        out = upfirdn2d_native(input, kernel, *up, *down, *pad)
+
+    else:
+        out = UpFirDn2d.apply(input, kernel, up, down, pad)
+
+    return out
+
+
+def upfirdn2d_native(
+    input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1
+):
+    _, channel, in_h, in_w = input.shape
+    input = input.reshape(-1, in_h, in_w, 1)
+
+    _, in_h, in_w, minor = input.shape
+    kernel_h, kernel_w = kernel.shape
+
+    out = input.view(-1, in_h, 1, in_w, 1, minor)
+    out = F.pad(out, [0, 0, 0, up_x - 1, 0, 0, 0, up_y - 1])
+    out = out.view(-1, in_h * up_y, in_w * up_x, minor)
+
+    out = F.pad(
+        out, [0, 0, max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)]
+    )
+    out = out[
+        :,
+        max(-pad_y0, 0) : out.shape[1] - max(-pad_y1, 0),
+        max(-pad_x0, 0) : out.shape[2] - max(-pad_x1, 0),
+        :,
+    ]
+
+    out = out.permute(0, 3, 1, 2)
+    out = out.reshape(
+        [-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1]
+    )
+    w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
+    out = F.conv2d(out, w)
+    out = out.reshape(
+        -1,
+        minor,
+        in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1,
+        in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1,
+    )
+    out = out.permute(0, 2, 3, 1)
+    out = out[:, ::down_y, ::down_x, :]
+
+    out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h + down_y) // down_y
+    out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w + down_x) // down_x
+
+    return out.view(-1, channel, out_h, out_w)

op/upfirdn2d_kernel.cu

@@ -0,0 +1,369 @@
+// Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+//
+// This work is made available under the Nvidia Source Code License-NC.
+// To view a copy of this license, visit
+// https://nvlabs.github.io/stylegan2/license.html
+
+#include <torch/types.h>
+
+#include <ATen/ATen.h>
+#include <ATen/AccumulateType.h>
+#include <ATen/cuda/CUDAApplyUtils.cuh>
+#include <ATen/cuda/CUDAContext.h>
+
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+static __host__ __device__ __forceinline__ int floor_div(int a, int b) {
+  int c = a / b;
+
+  if (c * b > a) {
+    c--;
+  }
+
+  return c;
+}
+
+struct UpFirDn2DKernelParams {
+  int up_x;
+  int up_y;
+  int down_x;
+  int down_y;
+  int pad_x0;
+  int pad_x1;
+  int pad_y0;
+  int pad_y1;
+
+  int major_dim;
+  int in_h;
+  int in_w;
+  int minor_dim;
+  int kernel_h;
+  int kernel_w;
+  int out_h;
+  int out_w;
+  int loop_major;
+  int loop_x;
+};
+
+template <typename scalar_t>
+__global__ void upfirdn2d_kernel_large(scalar_t *out, const scalar_t *input,
+                                       const scalar_t *kernel,
+                                       const UpFirDn2DKernelParams p) {
+  int minor_idx = blockIdx.x * blockDim.x + threadIdx.x;
+  int out_y = minor_idx / p.minor_dim;
+  minor_idx -= out_y * p.minor_dim;
+  int out_x_base = blockIdx.y * p.loop_x * blockDim.y + threadIdx.y;
+  int major_idx_base = blockIdx.z * p.loop_major;
+
+  if (out_x_base >= p.out_w || out_y >= p.out_h ||
+      major_idx_base >= p.major_dim) {
+    return;
+  }
+
+  int mid_y = out_y * p.down_y + p.up_y - 1 - p.pad_y0;
+  int in_y = min(max(floor_div(mid_y, p.up_y), 0), p.in_h);
+  int h = min(max(floor_div(mid_y + p.kernel_h, p.up_y), 0), p.in_h) - in_y;
+  int kernel_y = mid_y + p.kernel_h - (in_y + 1) * p.up_y;
+
+  for (int loop_major = 0, major_idx = major_idx_base;
+       loop_major < p.loop_major && major_idx < p.major_dim;
+       loop_major++, major_idx++) {
+    for (int loop_x = 0, out_x = out_x_base;
+         loop_x < p.loop_x && out_x < p.out_w; loop_x++, out_x += blockDim.y) {
+      int mid_x = out_x * p.down_x + p.up_x - 1 - p.pad_x0;
+      int in_x = min(max(floor_div(mid_x, p.up_x), 0), p.in_w);
+      int w = min(max(floor_div(mid_x + p.kernel_w, p.up_x), 0), p.in_w) - in_x;
+      int kernel_x = mid_x + p.kernel_w - (in_x + 1) * p.up_x;
+
+      const scalar_t *x_p =
+          &input[((major_idx * p.in_h + in_y) * p.in_w + in_x) * p.minor_dim +
+                 minor_idx];
+      const scalar_t *k_p = &kernel[kernel_y * p.kernel_w + kernel_x];
+      int x_px = p.minor_dim;
+      int k_px = -p.up_x;
+      int x_py = p.in_w * p.minor_dim;
+      int k_py = -p.up_y * p.kernel_w;
+
+      scalar_t v = 0.0f;
+
+      for (int y = 0; y < h; y++) {
+        for (int x = 0; x < w; x++) {
+          v += static_cast<scalar_t>(*x_p) * static_cast<scalar_t>(*k_p);
+          x_p += x_px;
+          k_p += k_px;
+        }
+
+        x_p += x_py - w * x_px;
+        k_p += k_py - w * k_px;
+      }
+
+      out[((major_idx * p.out_h + out_y) * p.out_w + out_x) * p.minor_dim +
+          minor_idx] = v;
+    }
+  }
+}
+
+template <typename scalar_t, int up_x, int up_y, int down_x, int down_y,
+          int kernel_h, int kernel_w, int tile_out_h, int tile_out_w>
+__global__ void upfirdn2d_kernel(scalar_t *out, const scalar_t *input,
+                                 const scalar_t *kernel,
+                                 const UpFirDn2DKernelParams p) {
+  const int tile_in_h = ((tile_out_h - 1) * down_y + kernel_h - 1) / up_y + 1;
+  const int tile_in_w = ((tile_out_w - 1) * down_x + kernel_w - 1) / up_x + 1;
+
+  __shared__ volatile float sk[kernel_h][kernel_w];
+  __shared__ volatile float sx[tile_in_h][tile_in_w];
+
+  int minor_idx = blockIdx.x;
+  int tile_out_y = minor_idx / p.minor_dim;
+  minor_idx -= tile_out_y * p.minor_dim;
+  tile_out_y *= tile_out_h;
+  int tile_out_x_base = blockIdx.y * p.loop_x * tile_out_w;
+  int major_idx_base = blockIdx.z * p.loop_major;
+
+  if (tile_out_x_base >= p.out_w | tile_out_y >= p.out_h |
+      major_idx_base >= p.major_dim) {
+    return;
+  }
+
+  for (int tap_idx = threadIdx.x; tap_idx < kernel_h * kernel_w;
+       tap_idx += blockDim.x) {
+    int ky = tap_idx / kernel_w;
+    int kx = tap_idx - ky * kernel_w;
+    scalar_t v = 0.0;
+
+    if (kx < p.kernel_w & ky < p.kernel_h) {
+      v = kernel[(p.kernel_h - 1 - ky) * p.kernel_w + (p.kernel_w - 1 - kx)];
+    }
+
+    sk[ky][kx] = v;
+  }
+
+  for (int loop_major = 0, major_idx = major_idx_base;
+       loop_major < p.loop_major & major_idx < p.major_dim;
+       loop_major++, major_idx++) {
+    for (int loop_x = 0, tile_out_x = tile_out_x_base;
+         loop_x < p.loop_x & tile_out_x < p.out_w;
+         loop_x++, tile_out_x += tile_out_w) {
+      int tile_mid_x = tile_out_x * down_x + up_x - 1 - p.pad_x0;
+      int tile_mid_y = tile_out_y * down_y + up_y - 1 - p.pad_y0;
+      int tile_in_x = floor_div(tile_mid_x, up_x);
+      int tile_in_y = floor_div(tile_mid_y, up_y);
+
+      __syncthreads();
+
+      for (int in_idx = threadIdx.x; in_idx < tile_in_h * tile_in_w;
+           in_idx += blockDim.x) {
+        int rel_in_y = in_idx / tile_in_w;
+        int rel_in_x = in_idx - rel_in_y * tile_in_w;
+        int in_x = rel_in_x + tile_in_x;
+        int in_y = rel_in_y + tile_in_y;
+
+        scalar_t v = 0.0;
+
+        if (in_x >= 0 & in_y >= 0 & in_x < p.in_w & in_y < p.in_h) {
+          v = input[((major_idx * p.in_h + in_y) * p.in_w + in_x) *
+                        p.minor_dim +
+                    minor_idx];
+        }
+
+        sx[rel_in_y][rel_in_x] = v;
+      }
+
+      __syncthreads();
+      for (int out_idx = threadIdx.x; out_idx < tile_out_h * tile_out_w;
+           out_idx += blockDim.x) {
+        int rel_out_y = out_idx / tile_out_w;
+        int rel_out_x = out_idx - rel_out_y * tile_out_w;
+        int out_x = rel_out_x + tile_out_x;
+        int out_y = rel_out_y + tile_out_y;
+
+        int mid_x = tile_mid_x + rel_out_x * down_x;
+        int mid_y = tile_mid_y + rel_out_y * down_y;
+        int in_x = floor_div(mid_x, up_x);
+        int in_y = floor_div(mid_y, up_y);
+        int rel_in_x = in_x - tile_in_x;
+        int rel_in_y = in_y - tile_in_y;
+        int kernel_x = (in_x + 1) * up_x - mid_x - 1;
+        int kernel_y = (in_y + 1) * up_y - mid_y - 1;
+
+        scalar_t v = 0.0;
+
+#pragma unroll
+        for (int y = 0; y < kernel_h / up_y; y++)
+#pragma unroll
+          for (int x = 0; x < kernel_w / up_x; x++)
+            v += sx[rel_in_y + y][rel_in_x + x] *
+                 sk[kernel_y + y * up_y][kernel_x + x * up_x];
+
+        if (out_x < p.out_w & out_y < p.out_h) {
+          out[((major_idx * p.out_h + out_y) * p.out_w + out_x) * p.minor_dim +
+              minor_idx] = v;
+        }
+      }
+    }
+  }
+}
+
+torch::Tensor upfirdn2d_op(const torch::Tensor &input,
+                           const torch::Tensor &kernel, int up_x, int up_y,
+                           int down_x, int down_y, int pad_x0, int pad_x1,
+                           int pad_y0, int pad_y1) {
+  int curDevice = -1;
+  cudaGetDevice(&curDevice);
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  UpFirDn2DKernelParams p;
+
+  auto x = input.contiguous();
+  auto k = kernel.contiguous();
+
+  p.major_dim = x.size(0);
+  p.in_h = x.size(1);
+  p.in_w = x.size(2);
+  p.minor_dim = x.size(3);
+  p.kernel_h = k.size(0);
+  p.kernel_w = k.size(1);
+  p.up_x = up_x;
+  p.up_y = up_y;
+  p.down_x = down_x;
+  p.down_y = down_y;
+  p.pad_x0 = pad_x0;
+  p.pad_x1 = pad_x1;
+  p.pad_y0 = pad_y0;
+  p.pad_y1 = pad_y1;
+
+  p.out_h = (p.in_h * p.up_y + p.pad_y0 + p.pad_y1 - p.kernel_h + p.down_y) /
+            p.down_y;
+  p.out_w = (p.in_w * p.up_x + p.pad_x0 + p.pad_x1 - p.kernel_w + p.down_x) /
+            p.down_x;
+
+  auto out =
+      at::empty({p.major_dim, p.out_h, p.out_w, p.minor_dim}, x.options());
+
+  int mode = -1;
+
+  int tile_out_h = -1;
+  int tile_out_w = -1;
+
+  if (p.up_x == 1 && p.up_y == 1 && p.down_x == 1 && p.down_y == 1 &&
+      p.kernel_h <= 4 && p.kernel_w <= 4) {
+    mode = 1;
+    tile_out_h = 16;
+    tile_out_w = 64;
+  }
+
+  if (p.up_x == 1 && p.up_y == 1 && p.down_x == 1 && p.down_y == 1 &&
+      p.kernel_h <= 3 && p.kernel_w <= 3) {
+    mode = 2;
+    tile_out_h = 16;
+    tile_out_w = 64;
+  }
+
+  if (p.up_x == 2 && p.up_y == 2 && p.down_x == 1 && p.down_y == 1 &&
+      p.kernel_h <= 4 && p.kernel_w <= 4) {
+    mode = 3;
+    tile_out_h = 16;
+    tile_out_w = 64;
+  }
+
+  if (p.up_x == 2 && p.up_y == 2 && p.down_x == 1 && p.down_y == 1 &&
+      p.kernel_h <= 2 && p.kernel_w <= 2) {
+    mode = 4;
+    tile_out_h = 16;
+    tile_out_w = 64;
+  }
+
+  if (p.up_x == 1 && p.up_y == 1 && p.down_x == 2 && p.down_y == 2 &&
+      p.kernel_h <= 4 && p.kernel_w <= 4) {
+    mode = 5;
+    tile_out_h = 8;
+    tile_out_w = 32;
+  }
+
+  if (p.up_x == 1 && p.up_y == 1 && p.down_x == 2 && p.down_y == 2 &&
+      p.kernel_h <= 2 && p.kernel_w <= 2) {
+    mode = 6;
+    tile_out_h = 8;
+    tile_out_w = 32;
+  }
+
+  dim3 block_size;
+  dim3 grid_size;
+
+  if (tile_out_h > 0 && tile_out_w > 0) {
+    p.loop_major = (p.major_dim - 1) / 16384 + 1;
+    p.loop_x = 1;
+    block_size = dim3(32 * 8, 1, 1);
+    grid_size = dim3(((p.out_h - 1) / tile_out_h + 1) * p.minor_dim,
+                     (p.out_w - 1) / (p.loop_x * tile_out_w) + 1,
+                     (p.major_dim - 1) / p.loop_major + 1);
+  } else {
+    p.loop_major = (p.major_dim - 1) / 16384 + 1;
+    p.loop_x = 4;
+    block_size = dim3(4, 32, 1);
+    grid_size = dim3((p.out_h * p.minor_dim - 1) / block_size.x + 1,
+                     (p.out_w - 1) / (p.loop_x * block_size.y) + 1,
+                     (p.major_dim - 1) / p.loop_major + 1);
+  }
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "upfirdn2d_cuda", [&] {
+    switch (mode) {
+    case 1:
+      upfirdn2d_kernel<scalar_t, 1, 1, 1, 1, 4, 4, 16, 64>
+          <<<grid_size, block_size, 0, stream>>>(out.data_ptr<scalar_t>(),
+                                                 x.data_ptr<scalar_t>(),
+                                                 k.data_ptr<scalar_t>(), p);
+
+      break;
+
+    case 2:
+      upfirdn2d_kernel<scalar_t, 1, 1, 1, 1, 3, 3, 16, 64>
+          <<<grid_size, block_size, 0, stream>>>(out.data_ptr<scalar_t>(),
+                                                 x.data_ptr<scalar_t>(),
+                                                 k.data_ptr<scalar_t>(), p);
+
+      break;
+
+    case 3:
+      upfirdn2d_kernel<scalar_t, 2, 2, 1, 1, 4, 4, 16, 64>
+          <<<grid_size, block_size, 0, stream>>>(out.data_ptr<scalar_t>(),
+                                                 x.data_ptr<scalar_t>(),
+                                                 k.data_ptr<scalar_t>(), p);
+
+      break;
+
+    case 4:
+      upfirdn2d_kernel<scalar_t, 2, 2, 1, 1, 2, 2, 16, 64>
+          <<<grid_size, block_size, 0, stream>>>(out.data_ptr<scalar_t>(),
+                                                 x.data_ptr<scalar_t>(),
+                                                 k.data_ptr<scalar_t>(), p);
+
+      break;
+
+    case 5:
+      upfirdn2d_kernel<scalar_t, 1, 1, 2, 2, 4, 4, 8, 32>
+          <<<grid_size, block_size, 0, stream>>>(out.data_ptr<scalar_t>(),
+                                                 x.data_ptr<scalar_t>(),
+                                                 k.data_ptr<scalar_t>(), p);
+
+      break;
+
+    case 6:
+      upfirdn2d_kernel<scalar_t, 1, 1, 2, 2, 4, 4, 8, 32>
+          <<<grid_size, block_size, 0, stream>>>(out.data_ptr<scalar_t>(),
+                                                 x.data_ptr<scalar_t>(),
+                                                 k.data_ptr<scalar_t>(), p);
+
+      break;
+
+    default:
+      upfirdn2d_kernel_large<scalar_t><<<grid_size, block_size, 0, stream>>>(
+          out.data_ptr<scalar_t>(), x.data_ptr<scalar_t>(),
+          k.data_ptr<scalar_t>(), p);
+    }
+  });
+
+  return out;
+}