79_conv_transposed_1D_asymmetric_input_square_kernel___padded____strided____dilated__ • fast_conv_transpose1d_baseimport torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(
x: torch.Tensor,
weight: torch.Tensor,
bias: torch.Tensor,
stride: int,
padding: int,
dilation: int,
) -> torch.Tensor:
"""
Performs a transposed 1D convolution operation with asymmetric input and square kernel. Supports padding, striding, and dilation.
Args:
x (torch.Tensor): Input tensor
weight (torch.Tensor): Convolution weights
bias (torch.Tensor): Bias tensor (optional)
stride (int): Stride of the convolution
padding (int): Padding applied to the input
dilation (int): Spacing between kernel elements
Returns:
torch.Tensor: Output tensor
"""
return F.conv_transpose1d(
x, weight, bias=bias, stride=stride, padding=padding, dilation=dilation
)
class Model(nn.Module):
"""
Performs a transposed 1D convolution operation with asymmetric input and square kernel.
Supports padding, striding, and dilation.
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): Stride of the convolution.
padding (int): Padding applied to the input.
dilation (int): Spacing between kernel elements.
bias (bool): If `True`, adds a learnable bias to the output.
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int,
stride: int,
padding: int,
dilation: int,
bias: bool,
):
super(Model, self).__init__()
self.conv_transpose1d = nn.ConvTranspose1d(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
bias=bias,
)
# Copy the initialized parameters
self.weight = nn.Parameter(self.conv_transpose1d.weight.clone())
self.bias = nn.Parameter(self.conv_transpose1d.bias.clone()) if bias else None
self.stride = stride
self.padding = padding
self.dilation = dilation
def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
return fn(
x,
self.weight,
self.bias,
self.stride,
self.padding,
self.dilation,
)
# Constants
batch_size = 16
in_channels = 32
out_channels = 64
kernel_size = 3
length = 128
stride = 2
padding = 1
dilation = 2
bias = False
def get_inputs():
x = torch.randn(batch_size, in_channels, length)
return [x]
def get_init_inputs():
return [in_channels, out_channels, kernel_size, stride, padding, dilation, bias]
import torch
import torch.nn as nn
class Model(nn.Module):
"""
Performs a transposed 1D convolution operation with asymmetric input and square kernel.
Supports padding, striding, and dilation.
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.
dilation (int, optional): Spacing between kernel elements. 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,
dilation: int = 1,
bias: bool = False,
):
super(Model, self).__init__()
self.conv1d_transpose = nn.ConvTranspose1d(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
bias=bias,
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Performs the transposed 1D convolution.
Args:
x (torch.Tensor): Input tensor of shape (batch_size, in_channels, length).
Returns:
torch.Tensor: Output tensor of shape (batch_size, out_channels, length_out).
"""
return self.conv1d_transpose(x)
# Constants
batch_size = 16
in_channels = 32
out_channels = 64
kernel_size = 3
length = 128
stride = 2
padding = 1
dilation = 2
bias = False
def get_inputs():
x = torch.randn(batch_size, in_channels, length)
return [x]
def get_init_inputs():
return [in_channels, out_channels, kernel_size, stride, padding, dilation, bias]
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
// Computes the length of the output for ConvTranspose1D.
inline int compute_output_length(int input_length, int stride, int padding, int dilation, int kernel_size) {
return (input_length - 1) * stride - 2 * padding + dilation * (kernel_size - 1) + 1;
}
// This kernel assigns one thread per output element, fusing the improvements from the two kernels.
// It uses __restrict__ qualifiers, precomputed index arithmetic and optional bias handling to
// improve memory coalescing and reduce redundant computations.
__global__ void fast_conv_transpose1d_kernel(
const float* __restrict__ x,
const float* __restrict__ weight,
const float* __restrict__ bias, // can be nullptr if not provided
float* __restrict__ output,
int batch_size,
int in_channels,
int out_channels,
int input_length,
int output_length,
int kernel_size,
int stride,
int padding,
int dilation
) {
int idx = blockIdx.x * blockDim.x + threadIdx.x;
int total = batch_size * out_channels * output_length;
if (idx >= total) return;
// Decompose linear index to get batch, output channel, and spatial index
int o = idx % output_length;
int temp = idx / output_length;
int oc = temp % out_channels;
int b = temp / out_channels;
// Compute offsets to access the output memory
int out_offset = b * (out_channels * output_length) + oc * output_length;
// Each sample in x has 'in_channels * input_length' elements
int x_batch_stride = in_channels * input_length;
// Initialize the accumulated sum; add bias if available
float sum = (bias != nullptr) ? bias[oc] : 0.0f;
// Precompute padded output index
int o_pad = o + padding;
// Loop over the kernel positions
#pragma unroll
for (int k = 0; k < kernel_size; ++k) {
int i_pos = o_pad - k * dilation;
// Check if the position aligns with the stride
if (i_pos % stride != 0) continue;
int i = i_pos / stride;
if (i < 0 || i >= input_length) continue;
// Loop over the input channels
int x_base = b * x_batch_stride;
for (int ic = 0; ic < in_channels; ++ic) {
int x_idx = x_base + ic * input_length + i;
int weight_idx = ic * (out_channels * kernel_size) + oc * kernel_size + k;
sum += x[x_idx] * weight[weight_idx];
}
}
output[out_offset + o] = sum;
}
// Host function which sets up the tensors and launches the CUDA kernel
torch::Tensor forward_cuda(
torch::Tensor x,
torch::Tensor weight,
torch::optional<torch::Tensor> bias,
int stride,
int padding,
int dilation
) {
TORCH_CHECK(x.is_cuda(), "x must be a CUDA tensor");
TORCH_CHECK(weight.is_cuda(), "weight must be a CUDA tensor");
TORCH_CHECK(x.dim() == 3, "x must be 3D (batch, in_channels, input_length)");
TORCH_CHECK(weight.dim() == 3, "weight must be 3D (in_channels, out_channels, kernel_size)");
x = x.contiguous();
weight = weight.contiguous();
torch::Tensor bias_contig;
const float* bias_ptr = nullptr;
if (bias.has_value()) {
bias_contig = bias->contiguous();
TORCH_CHECK(bias_contig.is_cuda(), "bias must be a CUDA tensor");
TORCH_CHECK(bias_contig.dim() == 1, "bias must be 1D");
bias_ptr = bias_contig.data_ptr<float>();
}
int batch_size = x.size(0);
int in_channels = x.size(1);
int input_length = x.size(2);
int out_channels = weight.size(1);
int kernel_size = weight.size(2);
TORCH_CHECK(weight.size(0) == in_channels, "Weight in_channels mismatch.");
if (bias.has_value()) {
TORCH_CHECK(bias_contig.size(0) == out_channels, "Bias size must match out_channels");
}
int output_length = compute_output_length(input_length, stride, padding, dilation, kernel_size);
auto output = torch::zeros({batch_size, out_channels, output_length}, x.options());
int total_elements = batch_size * out_channels * output_length;
int threads_per_block = 256;
int num_blocks = (total_elements + threads_per_block - 1) / threads_per_block;
fast_conv_transpose1d_kernel<<<num_blocks, threads_per_block>>>(
x.data_ptr<float>(),
weight.data_ptr<float>(),
bias_ptr,
output.data_ptr<float>(),
batch_size,
in_channels,
out_channels,
input_length,
output_length,
kernel_size,
stride,
padding,
dilation
);
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &forward_cuda, "Fast ConvTranspose1D forward (CUDA)",
py::arg("x"), py::arg("weight"), py::arg("bias") = py::none(),
py::arg("stride"), py::arg("padding"), py::arg("dilation"));
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 2.184 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 1.808 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 54.774 | % | 0.010 | 5 |
| Issued Ipc Active | 2.194 | inst/cycle | 0.000 | 5 |
| SM Busy | 59.944 | % | 0.011 | 5 |
| Memory Throughput | 14983342487.770 | byte/second | 7080249737496516.000 | 5 |
| Mem Busy | 40.142 | % | 0.042 | 5 |
| Max Bandwidth | 39.618 | % | 0.036 | 5 |
| L1/TEX Hit Rate | 79.984 | % | 0.003 | 5 |
| L2 Hit Rate | 96.366 | % | 0.196 | 5 |
| Mem Pipes Busy | 39.550 | % | 0.038 | 5 |
| Warp Cycles Per Issued Instruction | 21.276 | cycle | 0.006 | 5 |
| Warp Cycles Per Executed Instruction | 21.354 | cycle | 0.006 | 5 |
| Avg. Active Threads Per Warp | 19.450 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 18.390 | 0.000 | 5 | |
| Max Active Clusters | 0.000 | cluster | 0.000 | 5 |
| Max Cluster Size | 8.000 | block | 0.000 | 5 |
| Overall GPU Occupancy | 0.000 | % | 0.000 | 5 |
| Cluster Occupancy | 0.000 | % | 0.000 | 5 |
| Block Limit SM | 32.000 | block | 0.000 | 5 |
| Block Limit Registers | 8.000 | block | 0.000 | 5 |
| Block Limit Shared Mem | 32.000 | block | 0.000 | 5 |
| Block Limit Warps | 8.000 | block | 0.000 | 5 |
| Theoretical Active Warps per SM | 64.000 | warp | 0.000 | 5 |
| Theoretical Occupancy | 100.000 | % | 0.000 | 5 |
| Achieved Occupancy | 72.540 | % | 0.142 | 5 |
| Achieved Active Warps Per SM | 46.428 | warp | 0.058 | 5 |
| Rule | Description |
|---|---|
| INF HighPipeUtilization | FMA is the highest-utilized pipeline (35.2%) based on active cycles, taking into account the rates of its different instructions. It executes 32-bit floating point (FADD, FMUL, FMAD, ...) and integer (IMUL, IMAD) operations. It is well-utilized, but should not be a bottleneck. |
| WRN ThreadDivergence | Instructions are executed in warps, which are groups of 32 threads. Optimal instruction throughput is achieved if all 32 threads of a warp execute the same instruction. The chosen launch configuration, early thread completion, and divergent flow control can significantly lower the number of active threads in a warp per cycle. This kernel achieves an average of 19.4 threads being active per cycle. This is further reduced to 18.4 threads per warp due to predication. The compiler may use predication to avoid an actual branch. Instead, all instructions are scheduled, but a per-thread condition code or predicate controls which threads execute the instructions. Try to avoid different execution paths within a warp when possible. In addition, ensure your kernel makes use of Independent Thread Scheduling, which allows a warp to reconverge after a data-dependent conditional block by explicitly calling __syncwarp(). |
| WRN Occupancy | This kernel's theoretical occupancy is not impacted by any block limit. The difference between calculated theoretical (100.0%) and measured achieved occupancy (73.2%) can be the result of warp scheduling overheads or workload imbalances during the kernel execution. Load imbalances can occur between warps within a block as well as across blocks of the same kernel. See the CUDA Best Practices Guide (https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#occupancy) for more details on optimizing occupancy. |
| Operation / Metric | Value | Unit |
|---|---|---|
| aten::to | ||
| CPU Time | 466420.17 | μs |
| Device Time | 16.48 | μs |
| Self CPU Time | 48.32 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::zeros | ||
| CPU Time | 6180958.75 | μs |
| Device Time | 231369.12 | μs |
| Self CPU Time | 135587.14 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::zero_ | ||
| CPU Time | 6783341.62 | μs |
| Device Time | 7132090.06 | μs |
| Self CPU Time | 299471.06 | μs |
| Self Device Time | 0.00 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::fill_ | ||
| CPU Time | 6483873.78 | μs |
| Device Time | 7132090.06 | μs |
| Self CPU Time | 375596.81 | μs |
| Self Device Time | 7132090.06 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| cudaLaunchKernel | ||
| CPU Time | 6417006.77 | μs |
| Device Time | 2702.42 | μs |
| Self CPU Time | 6417006.77 | μs |
| Self Device Time | 2702.42 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| fast_conv_transpose1d_kernel(float const*, float const*, float const*, float*, int, int, int, int, int, int, int, int, int) | ||
| CPU Time | 0.00 | μs |
| Device Time | 1451731.49 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 1451731.49 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| cudaEventRecord | ||
| CPU Time | 193670.01 | μs |
| Device Time | 1144180.99 | μs |
| Self CPU Time | 193670.01 | μs |
| Self Device Time | 1144180.99 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, at::detail::Array<char*, 1> >(int, at::native::FillFunctor<int>, at::detail::Array<char*, 1>) | ||
| CPU Time | 0.00 | μs |
| Device Time | 6901583.84 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 6901583.84 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45289 warnings generated when compiling for host. Suppressed 45326 warnings (45279 in non-user code, 47 NOLINT). Use -header-filter=.* to display errors from all non-system headers. Use -system-headers to display errors from system headers as well.