48_Mean_reduction_over_a_dimension • hybrid_reduce_warp_shared_baseimport torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(x: torch.Tensor, dim: int) -> torch.Tensor:
"""
Reduces the input tensor along the specified dimension by taking the mean.
Args:
x (torch.Tensor): Input tensor of arbitrary shape.
dim (int): The dimension to reduce over.
Returns:
torch.Tensor: Output tensor with reduced dimension. The shape of the output is the same as the input except for the reduced dimension which is removed.
"""
return torch.mean(x, dim=dim)
class Model(nn.Module):
"""
Simple model that performs mean reduction over a specific dimension.
"""
def __init__(self, dim: int):
"""
Initializes the model with the dimension to reduce over.
Args:
dim (int): The dimension to reduce over.
"""
super(Model, self).__init__()
self.dim = dim
def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
"""
Reduces the input tensor along the specified dimension by taking the mean.
Args:
x (torch.Tensor): Input tensor of arbitrary shape.
Returns:
torch.Tensor: Output tensor with reduced dimension. The shape of the output is the same as the input except for the reduced dimension which is removed.
"""
return fn(x, self.dim)
batch_size = 16
dim1 = 256
dim2 = 256
def get_inputs():
x = torch.randn(batch_size, dim1, dim2)
return [x]
def get_init_inputs():
return [1]
import torch
import torch.nn as nn
class Model(nn.Module):
"""
Simple model that performs mean reduction over a specific dimension.
"""
def __init__(self, dim: int):
"""
Initializes the model with the dimension to reduce over.
Args:
dim (int): The dimension to reduce over.
"""
super(Model, self).__init__()
self.dim = dim
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Reduces the input tensor along the specified dimension by taking the mean.
Args:
x (torch.Tensor): Input tensor of arbitrary shape.
Returns:
torch.Tensor: Output tensor with reduced dimension. The shape of the output is the same as the input except for the reduced dimension which is removed.
"""
return torch.mean(x, dim=self.dim)
batch_size = 16
dim1 = 256
dim2 = 256
def get_inputs():
x = torch.randn(batch_size, dim1, dim2)
return [x]
def get_init_inputs():
return [1]#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
template <typename scalar_t>
__device__ __forceinline__ scalar_t warp_reduce(scalar_t val) {
#pragma unroll
for (int offset = 16; offset > 0; offset /= 2) {
val += __shfl_down_sync(0xffffffff, val, offset);
}
return val;
}
template <typename scalar_t>
__global__ void hybrid_reduce_mean_kernel(
const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
int64_t outer_size,
int64_t dim_size,
int64_t inner_size) {
constexpr int BLOCK_SIZE = 256;
constexpr int WARP_SIZE = 32;
__shared__ scalar_t shared_data[BLOCK_SIZE];
const int tid = threadIdx.x;
const int wid = tid / WARP_SIZE;
const int lane = tid % WARP_SIZE;
const int output_idx = blockIdx.x;
if (output_idx >= outer_size * inner_size) return;
const int outer_idx = output_idx / inner_size;
const int inner_idx = output_idx % inner_size;
const int input_offset = outer_idx * dim_size * inner_size + inner_idx;
// First level reduction: threads cooperatively load and sum values
scalar_t thread_sum = 0;
#pragma unroll 4
for (int i = tid; i < dim_size; i += BLOCK_SIZE) {
thread_sum += input[input_offset + i * inner_size];
}
// Store partial sum in shared memory
shared_data[tid] = thread_sum;
__syncthreads();
// Second level reduction: warp-level reduction of shared memory
if (tid < WARP_SIZE) {
scalar_t warp_sum = 0;
#pragma unroll
for (int i = tid; i < BLOCK_SIZE; i += WARP_SIZE) {
warp_sum += shared_data[i];
}
// Final warp reduction using shuffle
warp_sum = warp_reduce(warp_sum);
// First thread writes result
if (lane == 0) {
output[output_idx] = warp_sum / static_cast<scalar_t>(dim_size);
}
}
}
torch::Tensor mean_reduce_cuda(torch::Tensor input, int64_t dim) {
if (dim < 0) dim += input.dim();
auto sizes = input.sizes().vec();
int64_t dim_size = sizes[dim];
int64_t outer_size = 1;
for (int i = 0; i < dim; i++) {
outer_size *= sizes[i];
}
int64_t inner_size = 1;
for (int i = dim + 1; i < sizes.size(); i++) {
inner_size *= sizes[i];
}
sizes.erase(sizes.begin() + dim);
auto output = torch::empty(sizes, input.options());
const int BLOCK_SIZE = 256;
const int num_blocks = outer_size * inner_size;
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "hybrid_reduce_mean_cuda", ([&] {
hybrid_reduce_mean_kernel<scalar_t><<<num_blocks, BLOCK_SIZE>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
outer_size,
dim_size,
inner_size
);
}));
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &mean_reduce_cuda, "Mean reduction using hybrid approach (CUDA)");
}| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 1.522 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 1.218 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 38.460 | % | 0.054 | 5 |
| Issued Ipc Active | 1.538 | inst/cycle | 0.000 | 5 |
| SM Busy | 38.460 | % | 0.054 | 5 |
| Memory Throughput | 372491063610.610 | byte/second | 262171594067876288.000 | 5 |
| Mem Busy | 65.822 | % | 0.199 | 5 |
| Max Bandwidth | 30.070 | % | 0.016 | 5 |
| L1/TEX Hit Rate | 1.386 | % | 0.281 | 5 |
| L2 Hit Rate | 85.640 | % | 0.281 | 5 |
| Mem Pipes Busy | 14.446 | % | 0.001 | 5 |
| Warp Cycles Per Issued Instruction | 30.878 | cycle | 0.054 | 5 |
| Warp Cycles Per Executed Instruction | 31.204 | cycle | 0.057 | 5 |
| Avg. Active Threads Per Warp | 31.460 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 29.110 | 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 | 10.000 | block | 0.000 | 5 |
| Block Limit Shared Mem | 16.000 | 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 | 74.434 | % | 0.015 | 5 |
| Achieved Active Warps Per SM | 47.640 | warp | 0.006 | 5 |
| Rule | Description |
|---|---|
| INF HighPipeUtilization | ALU is the highest-utilized pipeline (22.9%) based on active cycles, taking into account the rates of its different instructions. It executes integer and logic operations. It is well-utilized, but should not be a bottleneck. |
| 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 (74.3%) 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 | 584957.09 | μs |
| Device Time | 453.05 | μs |
| Self CPU Time | 40.44 | 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 | 584916.65 | μs |
| Device Time | 453.05 | μs |
| Self CPU Time | 100.72 | 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 | 584111.32 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 79.95 | 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 | 583794.55 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 583794.55 | 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 |
| cudaLaunchKernel | ||
| CPU Time | 539926.18 | μs |
| Device Time | 21262.90 | μs |
| Self CPU Time | 539926.18 | μs |
| Self Device Time | 21262.90 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void hybrid_reduce_mean_kernel<float>(float const*, float*, long, long, long) | ||
| CPU Time | 0.00 | μs |
| Device Time | 84006.88 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 84006.88 | 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 | 18246.20 | μs |
| Device Time | 42289.51 | μs |
| Self CPU Time | 18246.20 | μs |
| Self Device Time | 42289.51 | 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 | 67729.63 | μs |
| Device Time | 632088.07 | μs |
| Self CPU Time | 14819.45 | 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 | 52911.23 | μs |
| Device Time | 632088.07 | μs |
| Self CPU Time | 15330.53 | μs |
| Self Device Time | 632088.07 | 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 | 632166.59 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 632166.59 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45286 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.