Thanks for using Compiler Explorer
Sponsors
Jakt
C++
Ada
Algol68
Analysis
Android Java
Android Kotlin
Assembly
C
C3
Carbon
C with Coccinelle
C++ with Coccinelle
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#
GLSL
Go
Haskell
HLSL
Hook
Hylo
IL
ispc
Java
Julia
Kotlin
LLVM IR
LLVM MIR
Modula-2
Mojo
Nim
Numba
Nix
Objective-C
Objective-C++
OCaml
Odin
OpenCL C
Pascal
Pony
PTX
Python
Racket
Raku
Ruby
Rust
Sail
Snowball
Scala
Slang
Solidity
Spice
SPIR-V
Swift
LLVM TableGen
Toit
Triton
TypeScript Native
V
Vala
Visual Basic
Vyper
WASM
Yul (Solidity IR)
Zig
Javascript
GIMPLE
Ygen
sway
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
16.0.0 sm_90 CUDA-11.8
17.0.1(libc++) sm_90 CUDA-12.1
18.1.0(libc++) sm_90 CUDA-12.3.1
19.1.0 sm_90 CUDA-12.5.1
20.1.0 sm_90 CUDA-12.5.1
20.1.0 sm_90 CUDA-12.6.1
20.1.0 sm_90 CUDA-12.6.2
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 12.6.1
NVCC 12.6.2
NVCC 12.8.1
NVCC 12.9.0
NVCC 12.9.1
NVCC 13.0.0
NVCC 13.0.1
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
NVRTC 12.2.1
NVRTC 12.3.1
NVRTC 12.4.1
NVRTC 12.5.1
NVRTC 12.6.1
NVRTC 12.6.2
NVRTC 12.8.1
NVRTC 12.9.0
NVRTC 12.9.1
NVRTC 13.0.0
NVRTC 13.0.1
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 rocm-6.2.4
clang rocm-6.3.3
clang rocm-6.4.0
clang rocm-7.0.1
clang staging rocm-6.1.2
clang staging rocm-6.2.4
clang staging rocm-6.3.3
clang staging rocm-6.4.0
clang staging rocm-7.0.1
clang trunk rocm-6.1.2
clang trunk rocm-6.2.4
clang trunk rocm-6.3.3
clang trunk rocm-6.4.0
clang trunk rocm-7.0.1
trunk sm_100a CUDA-12.8.1
Options
Source code
#define GGML_USE_HIP #define WARP_SIZE 32 #if defined(GGML_USE_HIP) template <int dpp_ctrl, typename T, int row_mask = 0xf, int bank_mask = 0xf, bool bound_ctrl = true> static __device__ T hip_move_dpp(T old, T v) { return __builtin_bit_cast( T, __builtin_amdgcn_update_dpp( __builtin_bit_cast(int, old), __builtin_bit_cast(int, v), dpp_ctrl, row_mask, bank_mask, bound_ctrl ) ); } template <int mask, typename T> static __device__ T hip_ds_swizzle(T v) { return __builtin_bit_cast(T, __builtin_amdgcn_ds_swizzle(__builtin_bit_cast(int, v), mask)); } #endif // GGML_USE_HIP template<int width = WARP_SIZE, typename T> static __device__ __forceinline__ T ggml_cuda_shfl_xor_sync(T x, int offset) { #if defined(GGML_USE_HIP) static T old; // for some reason clang (v20) will not unroll loops with just the plain `int offset` // inside the switch, but it will optimize the switch case out when it does. switch (~offset) { // width != warp size shouldn't affect normal butterfly shuffle pattern case ~1: return hip_move_dpp<0x160 + 1>(old, x); // row_xor_mask: offset case ~2: return hip_move_dpp<0x160 + 2>(old, x); case ~4: return hip_move_dpp<0x160 + 4>(old, x); case ~8: return hip_move_dpp<0x160 + 8>(old, x); case ~16: return hip_ds_swizzle<0x401f>(x); // swap neighboring groups of 16 default: return __shfl_xor(x, offset, width); } #else return __shfl_xor_sync(0xffffffff, x, offset, width); #endif // GGML_USE_HIP } template<int width = WARP_SIZE> static __device__ __forceinline__ int warp_reduce_sum(int x) { #pragma unroll for (int offset = width/2; offset > 0; offset >>= 1) { x += ggml_cuda_shfl_xor_sync<width>(x, offset); } return x; } __global__ void test_64(int x, int * o) { *o = warp_reduce_sum<64>(x); } __global__ void test_32(int x, int * o) { *o = warp_reduce_sum<32>(x); } __global__ void test_32_no_unroll(int x, int width, int * o) { #pragma unroll for (int offset = width/2; offset > 0; offset >>= 1) { x += ggml_cuda_shfl_xor_sync<32>(x, offset); } *o = x; } __global__ void test_32_old_shfl(int x, int * o) { #pragma unroll for (int offset = 32/2; offset > 0; offset >>= 1) { x += __shfl_xor(x, offset, 32); } *o = x; } __global__ void test_32__shfl_no_unroll(int x, int width, int * o) { #pragma unroll for (int offset = width/2; offset > 0; offset >>= 1) { x += __shfl_xor(x, offset, 32); } *o = x; } __global__ void test_16(int x, int * o) { *o = warp_reduce_sum<16>(x); } __global__ void test_8(int x, int * o) { *o = warp_reduce_sum<8>(x); } __global__ void test_noop(int x, int * o) { *o = x; }
Become a Patron
Sponsor on GitHub
Donate via PayPal
Compiler Explorer Shop
Source on GitHub
Mailing list
Installed libraries
Wiki
Report an issue
How it works
Contact the author
CE on Mastodon
CE on Bluesky
Statistics
Changelog
Version tree