import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(predictions: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
"""
Computes the Kullback-Leibler Divergence for comparing two distributions.
Args:
predictions (torch.Tensor): Predicted values.
targets (torch.Tensor): Target values.
Returns:
torch.Tensor: Kullback-Leibler Divergence.
"""
return F.kl_div(torch.log(predictions), targets, reduction="batchmean")
class Model(nn.Module):
"""
A model that computes Kullback-Leibler Divergence for comparing two distributions.
Parameters:
None
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, predictions, targets, fn=module_fn):
return fn(predictions, targets)
batch_size = 128
input_shape = (4096,)
dim = 1
def get_inputs():
return [
torch.randn(batch_size, *input_shape).softmax(dim=-1),
torch.randn(batch_size, *input_shape).softmax(dim=-1),
]
def get_init_inputs():
return []
import torch
import torch.nn as nn
class Model(nn.Module):
"""
A model that computes Kullback-Leibler Divergence for comparing two distributions.
Parameters:
None
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, predictions, targets):
return torch.nn.functional.kl_div(torch.log(predictions), targets, reduction='batchmean')
batch_size = 128
input_shape = (4096, )
dim = 1
def get_inputs():
return [torch.randn(batch_size, *input_shape).softmax(dim=-1), torch.randn(batch_size, *input_shape).softmax(dim=-1)]
def get_init_inputs():
return []
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
// Define constants
constexpr int WARP_SIZE = 32;
constexpr int ELEMENTS_PER_THREAD = 8;
// Optimized CUDA kernel for KL divergence using grid-stride loop, loop unrolling, and warp-level reduction
__global__ void fast_strided_kl_kernel(
const float* __restrict__ log_predictions,
const float* __restrict__ targets,
float* __restrict__ output,
const int n) {
const int total_threads = gridDim.x * blockDim.x;
const int tid = blockIdx.x * blockDim.x + threadIdx.x;
float sum = 0.f;
// Each thread processes multiple elements using a grid-stride loop and unrolling
for (int stride = 0;; stride++) {
// Compute base index for this iteration
int base = tid + stride * total_threads * ELEMENTS_PER_THREAD;
if (base >= n) break;
#pragma unroll
for (int i = 0; i < ELEMENTS_PER_THREAD; i++) {
int idx = base + i * total_threads;
if (idx < n) {
// Use __ldg for read-only cache load
float lp = __ldg(log_predictions + idx);
float t = __ldg(targets + idx);
sum += expf(lp) - t * lp;
}
}
}
// Intra-warp reduction using shuffle
for (int offset = WARP_SIZE / 2; offset > 0; offset /= 2) {
sum += __shfl_down_sync(0xffffffff, sum, offset);
}
// Allocate shared memory for each warp's result
extern __shared__ float warp_sums[];
int warp_id = threadIdx.x / WARP_SIZE;
int lane = threadIdx.x % WARP_SIZE;
// First thread in each warp writes its result
if (lane == 0) {
warp_sums[warp_id] = sum;
}
__syncthreads();
// Final reduction: let the first warp reduce the per-warp sums
int num_warps = (blockDim.x + WARP_SIZE - 1) / WARP_SIZE;
sum = (threadIdx.x < num_warps) ? warp_sums[threadIdx.x] : 0.f;
for (int offset = WARP_SIZE / 2; offset > 0; offset /= 2) {
sum += __shfl_down_sync(0xffffffff, sum, offset);
}
// The first thread of the block atomically adds the block's sum to the global output
if (threadIdx.x == 0) {
atomicAdd(output, sum);
}
}
// Host function to launch the optimized kernel
torch::Tensor fast_strided_kl_forward(
torch::Tensor log_predictions,
torch::Tensor targets) {
const int n = log_predictions.numel();
auto output = torch::zeros({1}, log_predictions.options());
// Define kernel launch parameters
const int threads = 256;
int blocks = (n + threads * ELEMENTS_PER_THREAD - 1) / (threads * ELEMENTS_PER_THREAD);
// Optionally limit the number of blocks to ensure sufficient work per block
const int max_blocks = 256;
blocks = (blocks < max_blocks) ? blocks : max_blocks;
// Calculate shared memory size: one float per warp
int shared_mem = ((threads + WARP_SIZE - 1) / WARP_SIZE) * sizeof(float);
fast_strided_kl_kernel<<<blocks, threads, shared_mem>>>(
log_predictions.data_ptr<float>(),
targets.data_ptr<float>(),
output.data_ptr<float>(),
n
);
return output / static_cast<float>(n);
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &fast_strided_kl_forward, "Optimized KL divergence (CUDA)");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 0.572 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.306 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 15.142 | % | 0.014 | 5 |
| Issued Ipc Active | 0.606 | inst/cycle | 0.000 | 5 |
| SM Busy | 15.142 | % | 0.014 | 5 |
| Memory Throughput | 844536043915.252 | byte/second | 65376945529851543552.000 | 5 |
| Mem Busy | 14.610 | % | 0.023 | 5 |
| Max Bandwidth | 25.308 | % | 0.073 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 18.480 | % | 0.001 | 5 |
| Mem Pipes Busy | 9.664 | % | 0.010 | 5 |
| Warp Cycles Per Issued Instruction | 24.718 | cycle | 0.003 | 5 |
| Warp Cycles Per Executed Instruction | 26.146 | cycle | 0.004 | 5 |
| Avg. Active Threads Per Warp | 31.900 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 29.100 | 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 | 28.000 | 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 | 23.508 | % | 0.002 | 5 |
| Achieved Active Warps Per SM | 15.046 | 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 (23.5%) 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::zeros | ||
| CPU Time | 5147720.82 | μs |
| Device Time | 227599.52 | μs |
| Self CPU Time | 135404.98 | 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::zero_ | ||
| CPU Time | 5468448.81 | μs |
| Device Time | 7528904.81 | μs |
| Self CPU Time | 268501.32 | 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::fill_ | ||
| CPU Time | 5199949.47 | μs |
| Device Time | 7528904.81 | μs |
| Self CPU Time | 381673.46 | μs |
| Self Device Time | 7528902.22 | 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 | 5544420.75 | μs |
| Device Time | 24648.81 | μs |
| Self CPU Time | 5544420.75 | μs |
| Self Device Time | 24648.81 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| fast_strided_kl_kernel(float const*, float const*, float*, int) | ||
| CPU Time | 0.00 | μs |
| Device Time | 452684.92 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 452684.92 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| aten::div | ||
| CPU Time | 951330.81 | μs |
| Device Time | 249758.90 | μs |
| Self CPU Time | 573920.58 | μs |
| Self Device Time | 249678.84 | 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 | 260645.12 | μs |
| Device Time | 680526.63 | μs |
| Self CPU Time | 260645.12 | μs |
| Self Device Time | 680526.63 | 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 | 7301305.28 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 7301305.28 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45284 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.