49_Max_reduction_over_a_dimension • modular_max_reduce_edit_1import torch
import torch.nn as nn
import torch.functional as F
def module_fn(x: torch.Tensor, dim: int) -> torch.Tensor:
"""
Applies Max reduction over the specified dimension to the input tensor.
Args:
x (torch.Tensor): Input tensor
dim (int): The dimension to reduce over
Returns:
torch.Tensor: Output tensor after Max reduction over the specified dimension
"""
return torch.max(x, dim=dim)[0]
class Model(nn.Module):
"""
Simple model that performs Max 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:
"""
Applies Max reduction over the specified dimension to the input tensor.
Args:
x (torch.Tensor): Input tensor
fn: Function to apply (defaults to module_fn)
Returns:
torch.Tensor: Output tensor after Max reduction over the specified dimension
"""
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] # Example, change to desired dimension
import torch
import torch.nn as nn
class Model(nn.Module):
"""
Simple model that performs Max 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:
"""
Applies Max reduction over the specified dimension to the input tensor.
Args:
x (torch.Tensor): Input tensor.
Returns:
torch.Tensor: Output tensor after Max reduction over the specified dimension.
"""
return torch.max(x, dim=self.dim)[0]
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] # Example, change to desired dimension#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
// Modular device function to perform max reduction on a single slice
// This function can be reused for different types of reduction operations
template <typename scalar_t>
__device__ scalar_t max_reduce_slice(const scalar_t* input, int64_t dim_size, int64_t inner_size, int64_t start_idx) {
const scalar_t* ptr = input + start_idx;
scalar_t max_val = *ptr;
for (int i = 1; i < dim_size; i++) {
ptr += inner_size;
max_val = max(max_val, *ptr);
}
return max_val;
}
// Kernel that utilizes the modular device function
template <typename scalar_t>
__global__ void modular_max_reduce_kernel(
const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
const int64_t outer_size,
const int64_t dim_size,
const int64_t inner_size
) {
int inner_idx = blockIdx.y * blockDim.x + threadIdx.x;
if (inner_idx >= inner_size) return;
int outer_idx = blockIdx.x;
const int64_t input_offset = outer_idx * dim_size * inner_size + inner_idx;
// Use the device function for max reduction
scalar_t max_val = max_reduce_slice(input, dim_size, inner_size, input_offset);
output[outer_idx * inner_size + inner_idx] = max_val;
}
torch::Tensor max_reduce_cuda_forward(torch::Tensor input, int64_t dim) {
if (dim < 0) dim += input.dim();
int64_t outer_size = 1;
for (int i = 0; i < dim; i++) {
outer_size *= input.size(i);
}
int64_t inner_size = 1;
for (int i = dim + 1; i < input.dim(); i++) {
inner_size *= input.size(i);
}
const int64_t dim_size = input.size(dim);
auto output_sizes = input.sizes().vec();
output_sizes.erase(output_sizes.begin() + dim);
auto output = torch::empty(output_sizes, input.options());
const int threads = 256;
const int blocks_y = (inner_size + threads - 1) / threads;
dim3 grid(outer_size, blocks_y);
AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "max_reduce_forward", ([&] {
modular_max_reduce_kernel<scalar_t><<<grid, threads>>>(
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", &max_reduce_cuda_forward, "Modular Max Reduction Forward (CUDA)");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 0.370 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.040 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 9.360 | % | 0.003 | 5 |
| Issued Ipc Active | 0.376 | inst/cycle | 0.000 | 5 |
| SM Busy | 9.752 | % | 0.004 | 5 |
| Memory Throughput | 211391218333.466 | byte/second | 2855431735867275776.000 | 5 |
| Mem Busy | 3.650 | % | 0.001 | 5 |
| Max Bandwidth | 6.314 | % | 0.003 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 18.780 | % | 0.008 | 5 |
| Mem Pipes Busy | 0.794 | % | 0.000 | 5 |
| Warp Cycles Per Issued Instruction | 21.068 | cycle | 0.006 | 5 |
| Warp Cycles Per Executed Instruction | 21.150 | cycle | 0.006 | 5 |
| Avg. Active Threads Per Warp | 32.000 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 31.730 | 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 | 12.274 | % | 0.002 | 5 |
| Achieved Active Warps Per SM | 7.858 | warp | 0.001 | 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 (12.3%) 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 | 498676.20 | μs |
| Device Time | 413.44 | μs |
| Self CPU Time | 34.68 | 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 | 498641.51 | μs |
| Device Time | 413.44 | μs |
| Self CPU Time | 107.35 | 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 | 497856.44 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 87.47 | 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 | 496381.36 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 496381.36 | 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 | 605848.75 | μs |
| Device Time | 21395.12 | μs |
| Self CPU Time | 605848.75 | μs |
| Self Device Time | 21395.12 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void modular_max_reduce_kernel<float>(float const*, float*, long, long, long) | ||
| CPU Time | 0.00 | μs |
| Device Time | 159364.69 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 159364.69 | 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 | 23879.68 | μs |
| Device Time | 42373.37 | μs |
| Self CPU Time | 23879.68 | μs |
| Self Device Time | 42373.37 | 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 | 132484.21 | μs |
| Device Time | 634192.80 | μs |
| Self CPU Time | 16062.67 | 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 | 116425.17 | μs |
| Device Time | 634192.80 | μs |
| Self CPU Time | 17540.38 | μs |
| Self Device Time | 634192.80 | 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 | 634271.33 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 634271.33 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45285 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.