import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(x: torch.Tensor, dim: int) -> torch.Tensor:
"""
Performs a cumulative sum operation.
Args:
x (torch.Tensor): Input tensor.
dim (int): The dimension along which to perform the cumulative sum.
Returns:
torch.Tensor: Output tensor.
"""
return torch.cumsum(x, dim=dim)
class Model(nn.Module):
"""
A simple model that performs a cumulative sum (prefix sum) operation along a specified dimension.
"""
def __init__(self, dim):
"""
Initialize the Scan model.
Args:
dim (int): The dimension along which to perform the cumulative sum.
"""
super(Model, self).__init__()
self.dim = dim
def forward(self, x, fn=module_fn):
"""
Forward pass for the Scan model, computing the cumulative sum along the specified dimension.
Args:
x (torch.Tensor): Input tensor of shape (batch_size, *input_shape)
fn (callable): Function to compute the output, defaults to module_fn
"""
return fn(x, self.dim)
# Define input dimensions and parameters
batch_size = 128
input_shape = (4000,) # Example shape (arbitrary)
dim = 1
def get_inputs():
"""
Generates random inputs for testing the Scan model.
Returns:
list: A list containing a single randomly generated tensor with shape
(batch_size, *input_shape).
"""
return [torch.randn(batch_size, *input_shape)]
def get_init_inputs():
"""
Returns the initialization parameters for the Scan model.
Returns:
list: A list containing the `dim` parameter for model initialization.
"""
return [dim]
import torch
import torch.nn as nn
class Model(nn.Module):
"""
A simple model that performs a cumulative sum (prefix sum) operation along a specified dimension.
Parameters:
dim (int): The dimension along which to perform the scan operation.
"""
def __init__(self, dim):
"""
Initialize the Scan model.
Args:
dim (int): The dimension along which to perform the cumulative sum.
"""
super(Model, self).__init__()
self.dim = dim
def forward(self, x):
"""
Forward pass for the Scan model, computing the cumulative sum along the specified dimension.
Args:
x (torch.Tensor): Input tensor of shape (batch_size, *input_shape), where `*input_shape`
can vary depending on the use case.
Returns:
torch.Tensor: Tensor of the same shape as `x` after applying cumulative sum along `dim`.
"""
return torch.cumsum(x, dim=self.dim)
# Define input dimensions and parameters
batch_size = 128
input_shape = (4000,) # Example shape (arbitrary)
dim = 1
def get_inputs():
"""
Generates random inputs for testing the Scan model.
Returns:
list: A list containing a single randomly generated tensor with shape
(batch_size, *input_shape).
"""
return [torch.randn(batch_size, *input_shape)]
def get_init_inputs():
"""
Returns the initialization parameters for the Scan model.
Returns:
list: A list containing the `dim` parameter for model initialization.
"""
return [dim]#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#define CHECK_CUDA(x) TORCH_CHECK(x.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)
__device__ __forceinline__ float warp_scan(float val, const unsigned mask = 0xffffffff) {
#pragma unroll
for (int offset = 1; offset < 32; offset *= 2) {
float n = __shfl_up_sync(mask, val, offset);
if (threadIdx.x % 32 >= offset) val += n;
}
return val;
}
__global__ void cumsum_kernel_optimized(const float* __restrict__ input,
float* __restrict__ output,
float* __restrict__ warp_sums,
const int inner_size,
const int stride) {
const int idx = blockIdx.x;
const int outer_idx = idx / inner_size;
const int inner_idx = idx % inner_size;
const int lane_id = threadIdx.x % 32;
const int warp_id = threadIdx.x / 32;
const int warps_per_block = blockDim.x / 32;
// Base index for this thread's work
const int base_idx = outer_idx * stride * inner_size + inner_idx;
// Process elements in chunks of 32 (warp size)
for (int warp_start = warp_id * 32; warp_start < stride; warp_start += warps_per_block * 32) {
float val = 0.0f;
const int pos = warp_start + lane_id;
if (pos < stride) {
val = input[base_idx + pos * inner_size];
}
// Perform warp-level scan
val = warp_scan(val);
// Last thread in warp stores sum for next warp
if (lane_id == 31 && pos < stride) {
warp_sums[outer_idx * ((stride + 31)/32) + warp_start/32] = val;
}
// Synchronize only when necessary
if (warp_start + 32 < stride) {
__syncthreads();
}
// Add previous warps' sums
if (pos < stride && warp_start > 0) {
float prev_sum = 0.0f;
const int warp_idx = warp_start/32;
#pragma unroll 4
for (int w = 0; w < warp_idx; w++) {
prev_sum += warp_sums[outer_idx * ((stride + 31)/32) + w];
}
val += prev_sum;
}
// Store result
if (pos < stride) {
output[base_idx + pos * inner_size] = val;
}
}
}
torch::Tensor forward(torch::Tensor x, int dim) {
CHECK_INPUT(x);
auto output = torch::empty_like(x);
int ndim = x.dim();
dim = (dim + ndim) % ndim;
int outer_size = 1;
for (int i = 0; i < dim; i++) {
outer_size *= x.size(i);
}
int inner_size = 1;
for (int i = dim + 1; i < ndim; i++) {
inner_size *= x.size(i);
}
int stride = x.size(dim);
// Allocate temporary storage for warp sums
auto warp_sums = torch::empty({outer_size * ((stride + 31)/32)}, x.options());
const int total_blocks = outer_size * inner_size;
const int threads_per_block = 256; // Use 8 warps per block
cumsum_kernel_optimized<<<total_blocks, threads_per_block>>>(
x.data_ptr<float>(),
output.data_ptr<float>(),
warp_sums.data_ptr<float>(),
inner_size,
stride
);
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &forward, "Optimized CUDA cumulative sum with selective synchronization");
}| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 0.692 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.534 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 17.336 | % | 0.005 | 5 |
| Issued Ipc Active | 0.692 | inst/cycle | 0.000 | 5 |
| SM Busy | 17.336 | % | 0.005 | 5 |
| Memory Throughput | 50922493530.508 | byte/second | 3686907656487795712.000 | 5 |
| Mem Busy | 12.910 | % | 0.014 | 5 |
| Max Bandwidth | 12.852 | % | 0.014 | 5 |
| L1/TEX Hit Rate | 88.136 | % | 0.000 | 5 |
| L2 Hit Rate | 64.180 | % | 1.628 | 5 |
| Mem Pipes Busy | 12.852 | % | 0.014 | 5 |
| Warp Cycles Per Issued Instruction | 10.974 | cycle | 0.002 | 5 |
| Warp Cycles Per Executed Instruction | 10.984 | cycle | 0.002 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 30.710 | 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 | 11.898 | % | 0.000 | 5 |
| Achieved Active Warps Per SM | 7.614 | warp | 0.000 | 5 |
| Rule | Description |
|---|---|
| WRN HighPipeUtilization | All compute pipelines are under-utilized. Either this kernel is very small or it doesn't issue enough warps per scheduler. Check the Launch Statistics and Scheduler Statistics sections for further details. |
| INF CPIStall | Check the Warp Stall Sampling (All Cycles) table for the top stall locations in your source based on sampling data. The Kernel Profiling Guide (https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-reference) provides more details on each stall reason. |
| 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 (11.9%) 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 | 617701.87 | μs |
| Device Time | 144.51 | μs |
| Self CPU Time | 38.00 | 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::_to_copy | ||
| CPU Time | 617663.87 | μs |
| Device Time | 144.51 | μs |
| Self CPU Time | 86.55 | 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::empty_strided | ||
| CPU Time | 631870.02 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 14744.70 | 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 |
| cudaDeviceGetStreamPriorityRange | ||
| CPU Time | 601296.38 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 601296.38 | 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 |
| cudaLaunchKernel | ||
| CPU Time | 545299.17 | μs |
| Device Time | 14337.01 | μs |
| Self CPU Time | 545299.17 | μs |
| Self Device Time | 14337.01 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| cumsum_kernel_optimized(float const*, float*, float*, int, int) | ||
| CPU Time | 0.00 | μs |
| Device Time | 222427.15 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 222427.15 | 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 | 13550.56 | μs |
| Device Time | 28064.41 | μs |
| Self CPU Time | 13550.56 | μs |
| Self Device Time | 28064.41 | 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 | 189023.92 | μs |
| Device Time | 489395.12 | μs |
| Self CPU Time | 10238.63 | 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 | 178788.73 | μs |
| Device Time | 489395.12 | μs |
| Self CPU Time | 12453.87 | μs |
| Self Device Time | 489395.12 | 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 | 489395.12 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 489395.12 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45288 warnings generated when compiling for host. Suppressed 45322 warnings (45275 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.