53_Min_reduction_over_a_dimension • min_reduce_dynamic_block_base_baseimport torch
import torch.nn as nn
import torch.functional as F
def module_fn(x: torch.Tensor, dim: int) -> torch.Tensor:
"""
Applies min 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 min reduction over the specified dimension
"""
return torch.min(x, dim)[0]
class Model(nn.Module):
"""
Simple model that performs min 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 min 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 min 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 min 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 min reduction over the specified dimension to the input tensor.
Args:
x (torch.Tensor): Input tensor.
Returns:
torch.Tensor: Output tensor after min reduction over the specified dimension.
"""
return torch.min(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>
#include <vector>
#include <limits>
#include <c10/cuda/CUDAStream.h>
// Optimized CUDA kernel using dynamic block sizes for efficient min reduction
// along a specified dimension using warp-level reductions.
template <typename scalar_t>
__global__ void min_reduce_dynamic_block_kernel(
const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
const int outer,
const int r,
const int inner) {
const int warpSize = 32;
int idx = blockIdx.x * blockDim.x + threadIdx.x;
int warp_id = idx / warpSize;
if (warp_id >= outer * inner) return;
int lane = threadIdx.x % warpSize;
int outer_idx = warp_id / inner;
int inner_idx = warp_id % inner;
int base = outer_idx * (r * inner) + inner_idx;
// Initialize local_min to the maximum possible value
scalar_t local_min = std::numeric_limits<scalar_t>::max();
// Each thread in a warp processes different elements across 'r' dimension
#pragma unroll
for (int j = lane; j < r; j += warpSize) {
scalar_t val = input[base + j * inner];
local_min = (val < local_min) ? val : local_min;
}
// Reduce within the warp
for (int offset = warpSize / 2; offset > 0; offset /= 2) {
scalar_t other = __shfl_down_sync(0xffffffff, local_min, offset);
local_min = (other < local_min) ? other : local_min;
}
// First lane of each warp writes the result
if (lane == 0) {
output[warp_id] = local_min;
}
}
// Forward function that translates shapes, dimensions and launches the kernel
torch::Tensor forward(torch::Tensor input, int64_t dim) {
TORCH_CHECK(input.is_cuda(), "input must be a CUDA tensor");
if (!input.is_contiguous()) {
input = input.contiguous();
}
int ndim = input.dim();
TORCH_CHECK(dim >= 0 && dim < ndim, "dim out of range");
int outer = 1;
for (int i = 0; i < dim; i++) {
outer *= input.size(i);
}
int r = input.size(dim);
int inner = 1;
for (int i = dim + 1; i < ndim; i++) {
inner *= input.size(i);
}
std::vector<int64_t> output_shape;
for (int i = 0; i < ndim; i++) {
if (i != dim) {
output_shape.push_back(input.size(i));
}
}
auto output = torch::empty(output_shape, input.options());
int total_warps = outer * inner;
const int block_sizes[] = {32, 64, 128, 256, 512};
const int threads_per_block = block_sizes[(total_warps * 32 / 512) % 5];
int num_blocks = (total_warps * 32 + threads_per_block - 1) / threads_per_block;
AT_DISPATCH_ALL_TYPES(input.scalar_type(), "min_reduce_dynamic_block_cuda", ([&] {
min_reduce_dynamic_block_kernel<scalar_t><<<num_blocks, threads_per_block, 0,
c10::cuda::getCurrentCUDAStream().stream()>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
outer,
r,
inner);
}));
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &forward, "Optimized min reduction using dynamic block sizes and warp-level primitives (CUDA)");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 0.360 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.266 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 9.356 | % | 0.003 | 5 |
| Issued Ipc Active | 0.374 | inst/cycle | 0.000 | 5 |
| SM Busy | 9.356 | % | 0.003 | 5 |
| Memory Throughput | 461133518415.172 | byte/second | 51234458738480144384.000 | 5 |
| Mem Busy | 56.890 | % | 0.823 | 5 |
| Max Bandwidth | 20.234 | % | 0.437 | 5 |
| L1/TEX Hit Rate | 48.108 | % | 0.058 | 5 |
| L2 Hit Rate | 75.122 | % | 0.564 | 5 |
| Mem Pipes Busy | 3.244 | % | 0.002 | 5 |
| Warp Cycles Per Issued Instruction | 68.932 | cycle | 0.172 | 5 |
| Warp Cycles Per Executed Instruction | 71.424 | cycle | 0.189 | 5 |
| Avg. Active Threads Per Warp | 30.480 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 28.110 | 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 | 42.000 | block | 0.000 | 5 |
| Block Limit Shared Mem | 32.000 | block | 0.000 | 5 |
| Block Limit Warps | 32.000 | block | 0.000 | 5 |
| Theoretical Active Warps per SM | 64.000 | warp | 0.000 | 5 |
| Theoretical Occupancy | 100.000 | % | 0.000 | 5 |
| Achieved Occupancy | 40.412 | % | 0.007 | 5 |
| Achieved Active Warps Per SM | 25.864 | warp | 0.003 | 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 (40.3%) 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 | 371638.76 | μs |
| Device Time | 352.96 | μs |
| Self CPU Time | 34.96 | 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 | 371603.80 | μs |
| Device Time | 352.96 | μs |
| Self CPU Time | 90.98 | 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 | 370908.97 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 72.46 | 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 | 370627.52 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 370627.52 | 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 | 508392.17 | μs |
| Device Time | 705.41 | μs |
| Self CPU Time | 508392.17 | μs |
| Self Device Time | 705.41 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| void min_reduce_dynamic_block_kernel<float>(float const*, float*, int, int, int) | ||
| CPU Time | 0.00 | μs |
| Device Time | 61073.41 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 61073.41 | 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 | 66504.23 | μs |
| Device Time | 621372.15 | μs |
| Self CPU Time | 14228.54 | 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 | 52279.94 | μs |
| Device Time | 621372.15 | μs |
| Self CPU Time | 16549.45 | μs |
| Self Device Time | 621372.15 | 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 | 621372.15 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 621372.15 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45296 warnings generated when compiling for host. Suppressed 45327 warnings (45280 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.