import torch
import torch.nn as nn
import torch.nn.functional as F
import math
def module_fn(x: torch.Tensor) -> torch.Tensor:
"""
Implementation of the Gaussian Error Linear Units (GELU) activation function currently in Google BERT repo (identical to OpenAI GPT).
Args:
x (torch.Tensor): Input tensor.
Returns:
torch.Tensor: Output tensor.
"""
return (
0.5
* x
* (
1.0
+ torch.tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * torch.pow(x, 3.0)))
)
)
class Model(nn.Module):
"""
Implementation of the GELU activation function currently in Google BERT repo (identical to OpenAI GPT).
Reference: Gaussian Error Linear Units (GELU) paper: https://arxiv.org/abs/1606.08415
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x, fn=module_fn):
return fn(x)
batch_size = 2000
dim = 2000
def get_inputs():
return [torch.randn(batch_size, dim)]
def get_init_inputs():
return []
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
# From https://github.com/karpathy/minGPT/blob/master/mingpt/model.py
class Model(nn.Module):
"""
Implementation of the GELU activation function currently in Google BERT repo (identical to OpenAI GPT).
Reference: Gaussian Error Linear Units (GELU) paper: https://arxiv.org/abs/1606.08415
"""
def __init__(self):
super(Model, self).__init__()
def forward(self, x):
return (
0.5
* x
* (
1.0
+ torch.tanh(
math.sqrt(2.0 / math.pi) * (x + 0.044715 * torch.pow(x, 3.0))
)
)
)
batch_size = 2000
dim = 2000
def get_inputs():
return [torch.randn(batch_size, dim)]
def get_init_inputs():
return []
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <math.h>
// Device function for GELU activation computation
__device__ float compute_gelu(float x) {
const float sqrt_2_over_pi = 0.7978845608f;
const float coeff = 0.044715f;
float x_cubed = x * x * x;
float inner = x + coeff * x_cubed;
inner *= sqrt_2_over_pi;
float tanh_val = tanhf(inner);
return 0.5f * x * (1.0f + tanh_val);
}
// Optimized kernel that applies the GELU activation using shared memory
__global__ void gelu_kernel_optimized(const float* __restrict__ x, float* __restrict__ y, int n) {
extern __shared__ float shared_x[];
const int tid = threadIdx.x;
const int gid = blockIdx.x * blockDim.x * 4 + tid;
// Load data into shared memory
#pragma unroll
for (int i = 0; i < 4; i++) {
const int idx = gid + i * blockDim.x;
if (idx < n) {
shared_x[tid + i * blockDim.x] = x[idx];
}
}
__syncthreads();
// Process elements using shared memory
#pragma unroll
for (int i = 0; i < 4; i++) {
const int idx = gid + i * blockDim.x;
if (idx < n) {
y[idx] = compute_gelu(shared_x[tid + i * blockDim.x]);
}
}
}
// Torch binding to launch optimized GELU kernel
torch::Tensor gelu_forward(torch::Tensor x) {
TORCH_CHECK(x.is_cuda(), "Input tensor must be on CUDA");
TORCH_CHECK(x.is_contiguous(), "Input tensor must be contiguous");
auto y = torch::empty_like(x);
int n = x.numel();
const int threads = 256;
int blocks = (n + threads * 4 - 1) / (threads * 4);
// Allocate shared memory for the block
size_t shared_mem_size = threads * 4 * sizeof(float);
gelu_kernel_optimized<<<blocks, threads, shared_mem_size>>>(
x.data_ptr<float>(),
y.data_ptr<float>(),
n
);
return y;
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.def("forward", &gelu_forward, "Optimized GELU forward CUDA implementation");
}| Metric | Value | Unit | Variance | Samples |
|---|---|---|---|---|
| Executed Ipc Active | 2.856 | inst/cycle | 0.000 | 5 |
| Executed Ipc Elapsed | 2.256 | inst/cycle | 0.000 | 5 |
| Issue Slots Busy | 71.524 | % | 0.027 | 5 |
| Issued Ipc Active | 2.862 | inst/cycle | 0.000 | 5 |
| SM Busy | 71.524 | % | 0.027 | 5 |
| Memory Throughput | 1361378074709.264 | byte/second | 131282521743893102592.000 | 5 |
| Mem Busy | 33.626 | % | 0.083 | 5 |
| Max Bandwidth | 40.700 | % | 0.134 | 5 |
| L1/TEX Hit Rate | 0.000 | % | 0.000 | 5 |
| L2 Hit Rate | 50.540 | % | 0.043 | 5 |
| Mem Pipes Busy | 32.252 | % | 0.037 | 5 |
| Warp Cycles Per Issued Instruction | 19.050 | cycle | 0.006 | 5 |
| Warp Cycles Per Executed Instruction | 19.090 | cycle | 0.006 | 5 |
| Avg. Active Threads Per Warp | 25.350 | 0.000 | 5 | |
| Avg. Not Predicated Off Threads Per Warp | 24.340 | 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 | 20.000 | 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 | 85.600 | % | 0.005 | 5 |
| Achieved Active Warps Per SM | 54.784 | warp | 0.002 | 5 |
| Rule | Description |
|---|---|
| INF HighPipeUtilization | FMA is the highest-utilized pipeline (31.6%) based on active cycles, taking into account the rates of its different instructions. It executes 32-bit floating point (FADD, FMUL, FMAD, ...) and integer (IMUL, IMAD) 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 (85.6%) 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 | 306800.69 | μs |
| Device Time | 1644.47 | μs |
| Self CPU Time | 39.81 | 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 | 306760.88 | μs |
| Device Time | 1644.47 | μs |
| Self CPU Time | 89.30 | 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 | 322656.93 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 17350.13 | 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 | 303589.51 | μs |
| Device Time | 0.00 | μs |
| Self CPU Time | 303589.51 | 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 | 522395.96 | μs |
| Device Time | 21118.46 | μs |
| Self CPU Time | 522395.96 | μs |
| Self Device Time | 21118.46 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
| gelu_kernel_optimized(float const*, float*, int) | ||
| CPU Time | 0.00 | μs |
| Device Time | 105522.71 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 105522.71 | 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 | 18684.16 | μs |
| Device Time | 40594.80 | μs |
| Self CPU Time | 18684.16 | μs |
| Self Device Time | 40594.80 | 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 | 80223.35 | μs |
| Device Time | 601821.36 | μs |
| Self CPU Time | 11994.76 | 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 | 68233.18 | μs |
| Device Time | 601821.36 | μs |
| Self CPU Time | 14891.09 | μs |
| Self Device Time | 601821.36 | 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 | 601821.36 | μs |
| Self CPU Time | 0.00 | μs |
| Self Device Time | 601821.36 | μs |
| CPU Memory Usage | 0 | B |
| Device Memory Usage | 0 | B |
| Self CPU Memory Usage | 0 | B |
| Self Device Memory Usage | 0 | B |
45281 warnings generated when compiling for host. Suppressed 45321 warnings (45274 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.