import torch
import torch.nn as nn
import torch.nn.functional as F
def module_fn(x: torch.Tensor, dim: int) -> torch.Tensor:
"""
Applies LogSoftmax activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of shape (batch_size, dim)
dim (int): Dimension along which to apply LogSoftmax
Returns:
torch.Tensor: Output tensor with LogSoftmax applied, same shape as input
"""
return F.log_softmax(x, dim=dim)
class Model(nn.Module):
"""
Simple model that performs a LogSoftmax activation.
"""
def __init__(self, dim):
super(Model, self).__init__()
self.dim = dim
def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
return fn(x, self.dim)
batch_size = 16
dim = 16384
sm_dim = 1
def get_inputs():
x = torch.randn(batch_size, dim)
return [x]
def get_init_inputs():
return [sm_dim]
import torch
import torch.nn as nn
class Model(nn.Module):
"""
Simple model that performs a LogSoftmax activation.
"""
def __init__(self, dim: int = 1):
super(Model, self).__init__()
self.dim = dim
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Applies LogSoftmax activation to the input tensor.
Args:
x (torch.Tensor): Input tensor of shape (batch_size, dim).
Returns:
torch.Tensor: Output tensor with LogSoftmax applied, same shape as input.
"""
return torch.log_softmax(x, dim=self.dim)
batch_size = 16
dim = 16384
sm_dim = 1
def get_inputs():
x = torch.randn(batch_size, dim)
return [x]
def get_init_inputs():
return [sm_dim]
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <vector>
#include <limits>
#include <cmath>
// Hybrid kernel combining unrolling and warp-level primitives for efficiency
template <typename scalar_t, int BLOCK_SIZE>
__global__ void hybrid_logsoftmax_kernel(
const scalar_t* __restrict__ input,
scalar_t* __restrict__ output,
int dim_size) {
int batch_idx = blockIdx.x;
const scalar_t* input_row = input + batch_idx * dim_size;
scalar_t* output_row = output + batch_idx * dim_size;
__shared__ scalar_t sdata[BLOCK_SIZE];
// Phase 1: Compute the maximum value in the row using warp-level reduction
scalar_t local_max = -std::numeric_limits<scalar_t>::infinity();
for (int idx = threadIdx.x; idx < dim_size; idx += BLOCK_SIZE) {
scalar_t val = input_row[idx];
local_max = max(local_max, val);
}
unsigned int mask = 0xffffffff;
for (int offset = warpSize/2; offset > 0; offset /= 2) {
scalar_t other = __shfl_down_sync(mask, local_max, offset);
local_max = max(local_max, other);
}
int warp_id = threadIdx.x / warpSize;
int lane = threadIdx.x % warpSize;
if (lane == 0) {
sdata[warp_id] = local_max;
}
__syncthreads();
if (threadIdx.x == 0) {
scalar_t block_max = -std::numeric_limits<scalar_t>::infinity();
for (int i = 0; i < BLOCK_SIZE / warpSize; i++) {
block_max = max(block_max, sdata[i]);
}
sdata[0] = block_max;
}
__syncthreads();
scalar_t max_val = sdata[0];
// Phase 2: Compute the sum of exp(x - max_val) using warp-level reduction
scalar_t local_sum = 0;
for (int idx = threadIdx.x; idx < dim_size; idx += BLOCK_SIZE) {
scalar_t exp_val = exp(input_row[idx] - max_val);
local_sum += exp_val;
}
for (int offset = warpSize/2; offset > 0; offset /= 2) {
local_sum += __shfl_down_sync(mask, local_sum, offset);
}
if (lane == 0) {
atomicAdd(&sdata[0], local_sum);
}
__syncthreads();
scalar_t sum = sdata[0];
scalar_t log_sum = log(sum);
// Phase 3: Write back the final LogSoftmax values
for (int idx = threadIdx.x; idx < dim_size; idx += BLOCK_SIZE) {
output_row[idx] = (input_row[idx] - max_val) - log_sum;
}
}
// Host function
torch::Tensor hybrid_logsoftmax_cuda_forward(torch::Tensor input, int64_t dim) {
TORCH_CHECK(input.is_cuda(), "input must be a CUDA tensor");
TORCH_CHECK(
input.scalar_type() == torch::kFloat32 || input.scalar_type() == torch::kFloat64,
"input must be float32 or float64");
int64_t ndim = input.dim();
TORCH_CHECK(dim >= -ndim && dim < ndim, "dim out of range");
dim = dim >= 0 ? dim : dim + ndim;
std::vector<int64_t> permute_dims;
for (int64_t i = 0; i < ndim; ++i) {
if (i != dim) {
permute_dims.push_back(i);
}
}
permute_dims.push_back(dim);
input = input.permute(permute_dims).contiguous();
int64_t batch_size = input.numel() / input.size(-1);
int64_t dim_size = input.size(-1);
auto output = torch::empty_like(input);
int optimal_block_size = 256;
if (dim_size <= 32) {
optimal_block_size = 32;
} else if (dim_size <= 64) {
optimal_block_size = 64;
} else if (dim_size <= 128) {
optimal_block_size = 128;
} else if (dim_size <= 256) {
optimal_block_size = 256;
} else if (dim_size <= 512) {
optimal_block_size = 512;
} else {
optimal_block_size = 512;
}
const int blocks = batch_size;
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "hybrid_logsoftmax_cuda_forward", ([&] {
if (optimal_block_size == 32) {
hybrid_logsoftmax_kernel<scalar_t, 32><<<blocks, 32>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
dim_size);
} else if (optimal_block_size == 64) {
hybrid_logsoftmax_kernel<scalar_t, 64><<<blocks, 64>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
dim_size);
} else if (optimal_block_size == 128) {
hybrid_logsoftmax_kernel<scalar_t, 128><<<blocks, 128>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
dim_size);
} else if (optimal_block_size == 256) {
hybrid_logsoftmax_kernel<scalar_t, 256><<<blocks, 256>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
dim_size);
} else if (optimal_block_size == 512) {
hybrid_logsoftmax_kernel<scalar_t, 512><<<blocks, 512>>>(
input.data_ptr<scalar_t>(),
output.data_ptr<scalar_t>(),
dim_size);
}
}));
std::vector<int64_t> inverse_permute_dims(ndim);
for (size_t i = 0; i < permute_dims.size(); ++i) {
inverse_permute_dims[permute_dims[i]] = i;
}
output = output.permute(inverse_permute_dims);
return output;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &hybrid_logsoftmax_cuda_forward, "Hybrid LogSoftmax forward (CUDA)");
}
| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 1.116 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 0.100 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 28.066 | % | 0.021 | 5 |
| Issued Ipc Active | 1.124 | inst/cycle | 0.000 | 5 |
| SM Busy | 28.066 | % | 0.021 | 5 |
| Memory Throughput | 132012340371.506 | byte/second | 2046927348180559872.000 | 5 |
| Mem Busy | 6.372 | % | 0.007 | 5 |
| Max Bandwidth | 5.816 | % | 0.003 | 5 |
| L1/TEX Hit Rate | 50.000 | % | 0.000 | 5 |
| L2 Hit Rate | 67.580 | % | 0.257 | 5 |
| Mem Pipes Busy | 2.272 | % | 0.001 | 5 |
| Warp Cycles Per Issued Instruction | 13.720 | cycle | 0.015 | 5 |
| Warp Cycles Per Executed Instruction | 13.776 | cycle | 0.013 | 5 |
| Avg. Active Threads Per Warp | 30.778 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 29.558 | 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 | 4.000 | block | 0.000 | 5 |
| Block Limit Shared Mem | 10.000 | block | 0.000 | 5 |
| Block Limit Warps | 4.000 | block | 0.000 | 5 |
| Theoretical Active Warps per SM | 64.000 | warp | 0.000 | 5 |
| Theoretical Occupancy | 100.000 | % | 0.000 | 5 |
| Achieved Occupancy | 24.336 | % | 0.000 | 5 |
| Achieved Active Warps Per SM | 15.578 | 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. |
| 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 (24.3%) 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 | 393469.85 | μs |
| Device Time | 39.97 | μs |
| Self CPU Time | 38.64 | 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 | 393431.20 | μs |
| Device Time | 39.97 | μs |
| Self CPU Time | 87.94 | 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 | 413164.80 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 20156.04 | 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 | 391419.96 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 391419.96 | 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 |
| cudaLaunchKernel | ||
| CPU Time | 466024.94 | μs |
| Device Time | 21543.24 | μs |
| Self CPU Time | 466024.94 | μs |
| Self Device Time | 21543.24 | 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_logsoftmax_kernel<float, 512>(float const*, float*, int) | ||
| CPU Time | 0.00 | μs |
| Device Time | 60971.11 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 60971.11 | 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 | 18676.53 | μs |
| Device Time | 42460.12 | μs |
| Self CPU Time | 18676.53 | μs |
| Self Device Time | 42460.12 | 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 | 62858.53 | μs |
| Device Time | 620555.85 | μs |
| Self CPU Time | 11366.40 | 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 | 51493.67 | μs |
| Device Time | 620555.85 | μs |
| Self CPU Time | 17290.74 | μs |
| Self Device Time | 620555.85 | 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 | 620555.85 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 620555.85 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45287 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.