Thanks for using Compiler Explorer
Sponsors
Jakt
C++
Ada
Analysis
Android Java
Android Kotlin
Assembly
C
C3
Carbon
C++ (Circle)
CIRCT
Clean
CMake
CMakeScript
COBOL
C++ for OpenCL
MLIR
Cppx
Cppx-Blue
Cppx-Gold
Cpp2-cppfront
Crystal
C#
CUDA C++
D
Dart
Elixir
Erlang
Fortran
F#
Go
Haskell
HLSL
Hook
Hylo
ispc
Java
Julia
Kotlin
LLVM IR
LLVM MIR
Modula-2
Nim
Objective-C
Objective-C++
OCaml
OpenCL C
Pascal
Pony
Python
Racket
Ruby
Rust
Snowball
Scala
Solidity
Spice
Swift
LLVM TableGen
Toit
TypeScript Native
V
Vala
Visual Basic
WASM
Zig
Javascript
GIMPLE
cuda source #1
Output
Compile to binary object
Link to binary
Execute the code
Intel asm syntax
Demangle identifiers
Verbose demangling
Filters
Unused labels
Library functions
Directives
Comments
Horizontal whitespace
Debug intrinsics
Compiler
10.0.0 sm_75 CUDA-10.2
10.0.1 sm_75 CUDA-10.2
11.0.0 sm_75 CUDA-10.2
NVCC 10.0.130
NVCC 10.1.105
NVCC 10.1.168
NVCC 10.1.243
NVCC 10.2.89
NVCC 11.0.2
NVCC 11.0.3
NVCC 11.1.0
NVCC 11.1.1
NVCC 11.2.0
NVCC 11.2.1
NVCC 11.2.2
NVCC 11.3.0
NVCC 11.3.1
NVCC 11.4.0
NVCC 11.4.1
NVCC 11.4.2
NVCC 11.4.3
NVCC 11.4.4
NVCC 11.5.0
NVCC 11.5.1
NVCC 11.5.2
NVCC 11.6.0
NVCC 11.6.1
NVCC 11.6.2
NVCC 11.7.0
NVCC 11.7.1
NVCC 11.8.0
NVCC 12.0.0
NVCC 12.0.1
NVCC 12.1.0
NVCC 12.2.1
NVCC 12.3.1
NVCC 12.4.1
NVCC 12.5.1
NVCC 9.1.85
NVCC 9.2.88
NVRTC 11.0.2
NVRTC 11.0.3
NVRTC 11.1.0
NVRTC 11.1.1
NVRTC 11.2.0
NVRTC 11.2.1
NVRTC 11.2.2
NVRTC 11.3.0
NVRTC 11.3.1
NVRTC 11.4.0
NVRTC 11.4.1
NVRTC 11.5.0
NVRTC 11.5.1
NVRTC 11.5.2
NVRTC 11.6.0
NVRTC 11.6.1
NVRTC 11.6.2
NVRTC 11.7.0
NVRTC 11.7.1
NVRTC 11.8.0
NVRTC 12.0.0
NVRTC 12.0.1
NVRTC 12.1.0
clang 7.0.0 sm_70 CUDA-9.1
clang 8.0.0 sm_75 CUDA-10.0
clang 9.0.0 sm_75 CUDA-10.1
clang rocm-4.5.2
clang rocm-5.0.2
clang rocm-5.1.3
clang rocm-5.2.3
clang rocm-5.3.2
clang rocm-5.7.0
clang rocm-6.0.2
clang rocm-6.1.2
clang staging rocm-6.1.2
clang trunk rocm-6.1.2
trunk sm_86 CUDA-11.3
Options
Source code
#include <cuda_fp16.h> template<class ElementType> struct alignas(16) Packed128 { __device__ __forceinline__ Packed128() = default; __device__ __forceinline__ explicit Packed128(int4 bits) { static_assert(sizeof(bits) == sizeof(payload), "Size mismatch."); memcpy(&payload, &bits, sizeof(bits)); } __device__ __forceinline__ ElementType& operator[](int index) { return payload[index]; } __device__ __forceinline__ const ElementType& operator[](int index) const { return payload[index]; } __device__ __forceinline__ float fp32(int index) { return static_cast<float>(payload[index]); } __device__ __forceinline__ int4 get_bits() const { int4 bits; static_assert(sizeof(bits) == sizeof(payload), "Size mismatch."); memcpy(&bits, &payload, sizeof(bits)); return bits; } static constexpr const size_t size = sizeof(int4) / sizeof(ElementType); ElementType payload[size]; }; // load a Packet128 from an aligned memory address template<class ElementType> __device__ __forceinline__ Packed128<ElementType> load_aligned(const ElementType* address) { return Packed128<ElementType>{*reinterpret_cast<const int4*>(address)}; } // load a Packet128 from an aligned memory address with streaming cache hint template<class ElementType> __device__ __forceinline__ Packed128<ElementType> load_aligned_cs(const ElementType* address) { return Packed128<ElementType>{__ldcs(reinterpret_cast<const int4*>(address))}; } // store a Packet128 to an aligned memory address template<class ElementType> __device__ __forceinline__ void store_aligned(ElementType* target, Packed128<ElementType> value) { *reinterpret_cast<int4*>(target) = value.get_bits(); } // store a Packet128 to an aligned memory address with streaming cache hint template<class ElementType> __device__ __forceinline__ void store_aligned_cs(ElementType* target, Packed128<ElementType> value) { __stcs(reinterpret_cast<int4*>(target), value.get_bits()); } // ---------------------------------------------------------------------------- // CPU code reference #define GELU_SCALING_FACTOR sqrtf(2.0f / M_PI) // elementwise ops are nice and ez __global__ void gelu_kernel2(half* out, const half* inp, int N) { using packet_t = Packed128<half>; int i = (blockIdx.x * blockDim.x + threadIdx.x) * packet_t::size; if (i < N) { packet_t packet_out; packet_t packet_in = load_aligned_cs(inp + i); for(int k = 0; k < packet_in.size; ++k) { float xi = packet_in[k]; float cube = 0.044715f * xi * xi * xi; packet_out[k] = 0.5f * xi * (1.0f + tanhf(GELU_SCALING_FACTOR * (xi + cube))); } store_aligned(out + i, packet_out); } } // ---------------------------------------------------------------------------- // kernel launcher
Become a Patron
Sponsor on GitHub
Donate via PayPal
Source on GitHub
Mailing list
Installed libraries
Wiki
Report an issue
How it works
Contact the author
CE on Mastodon
About the author
Statistics
Changelog
Version tree