├── .gitignore
├── CMakeLists.txt
├── README.md
└── particle-repel.cc


/.gitignore:
--------------------------------------------------------------------------------
1 | wasi-sdk-5.0*
2 | build
3 | /CMakeFiles
4 | /cmake_install.cmake
5 | /build-wasm
6 | 


--------------------------------------------------------------------------------
/CMakeLists.txt:
--------------------------------------------------------------------------------
 1 | cmake_minimum_required(VERSION 2.8.12)
 2 | 
 3 | project(particle-repl)
 4 | 
 5 | if (NOT CMAKE_BUILD_TYPE)
 6 |     message(STATUS "No build type selected, default to Release")
 7 |     set(CMAKE_BUILD_TYPE "Release")
 8 | endif()
 9 | 
10 | message(STATUS "Building in ${CMAKE_BUILD_TYPE} mode")
11 | message(STATUS "Building with ${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION} on ${CMAKE_SYSTEM}")
12 | 
13 | set(CMAKE_CXX_FLAGS "$ENV{CFLAGS} -std=c++11 -O2")
14 | 
15 | add_executable(particle-repel-simd
16 |     particle-repel.cc
17 | )
18 | set_target_properties(particle-repel-simd PROPERTIES COMPILE_FLAGS "-msimd128")
19 | 
20 | add_executable(particle-repel
21 |     particle-repel.cc
22 | )
23 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # C++ Wasm SIMD 128 example
 2 | 
 3 | This repo is made to showcase how to emit Wasm SIMD 128 instructions from C++, and use it with [Wasmer](https://github.com/wasmerio/wasmer).
 4 | 
 5 | ## Build
 6 | 
 7 | ```bash
 8 | ./build.sh
 9 | ```
10 | 
11 | It will download the [WASI SDK 5](https://github.com/CraneStation/wasi-sdk/releases/tag/wasi-sdk-5) (the macOS or linux version, depending on your system).
12 | 
13 | ## Run it!
14 | 
15 | You can run the SIMD version with [Wasmer](https://wasmer.io/):
16 | 
17 | ```bash
18 | wasmer run --backend=llvm --enable-simd build/particle-repel-simd.wasm
19 | ```
20 | 
21 | Or run the non-SIMD version
22 | 
23 | ```bash
24 | wasmer run --backend=llvm build/particle-repel.wasm
25 | ```
26 | 
27 | Or run the Native SIMD version
28 | 
29 | ```bash
30 | ./build/particle-repel
31 | ```
32 | 
33 | 
34 | After running it, the results should be something like:
35 | 
36 | ```
37 | Beginning simulation.
38 | 5.44089e-21
39 | 9.19434e-22
40 | 9.98077e-20
41 | 1.01414e-17
42 | 1.89481e-17
43 | 3.61239e-16
44 | 2.24645e-14
45 | 1.24237e-13
46 | 2.39859e-13
47 | 3.13435e-13
48 | 4.13311e-13
49 | 6.31935e-13
50 | 4.59395e-10
51 | 6.40114e-10
52 | 1.15545e-09
53 | 2.24414e-10
54 | 1.03695e-09
55 | 4.06071e-10
56 | 7.64414e-10
57 | 7.57652e-10
58 | ```
59 | 
60 | ## Benchmarks
61 | 
62 | ```
63 | # Native
64 | time ./build/particle-repel # 4.930 total
65 | 
66 | # Wasmer SIMD
67 | time wasmer run --backend=llvm --enable-simd build/particle-repel-simd.wasm # 4.980 total
68 | 
69 | # Wasmer Non-SIMD
70 | time wasmer run --backend=llvm build/particle-repel.wasm # 11.630 total
71 | ```
72 | 


--------------------------------------------------------------------------------
/particle-repel.cc:
--------------------------------------------------------------------------------
 1 | #include <cstdio>
 2 | #include <random>
 3 | 
 4 | typedef float float4 __attribute__((ext_vector_type(4)));
 5 | 
 6 | namespace {
 7 | constexpr int object_count = 10000;
 8 | 
 9 | float4 positions[object_count];
10 | float4 velocities[object_count];
11 | 
12 | void init_positions() {
13 |   std::mt19937_64 mt;
14 |   std::uniform_real_distribution<float> dist(-127.0, 127.0);
15 |   for (int i = 0; i < object_count; ++i) {
16 |     positions[i] = (float4){dist(mt), dist(mt), dist(mt), dist(mt)};
17 |   }
18 | }
19 | 
20 | void init_velocities() {
21 |   for (int i = 0; i < object_count; ++i) {
22 |     velocities[i] = (float4){0, 0, 0, 0};
23 |   }
24 | }
25 | 
26 | float horizontal_add(float4 x) {
27 |   auto partial_sum = x.xy + x.zw;
28 |   return partial_sum.x + partial_sum.y;
29 | }
30 | 
31 | float distance(float4 x, float4 y) {
32 |   float4 d = y - x;
33 |   float4 d2 = d * d;
34 |   return horizontal_add(d2);
35 | }
36 | 
37 | void update() {
38 |   for (int i = 0; i < object_count; ++i) {
39 |     for (int j = 0; j < object_count; ++j) {
40 |       float dist = distance(positions[i], positions[j]);
41 |       dist *= dist;
42 |       if (dist == 0)
43 |         continue;
44 |       velocities[i] += (positions[j] - positions[i]) / dist;
45 |     }
46 |   }
47 | 
48 |   for (int i = 0; i < object_count; ++i) {
49 |     positions[i] += velocities[i];
50 |   }
51 | }
52 | 
53 | float avg_motion() {
54 |   float4 avg_motion = velocities[0] / (double)object_count;
55 |   for (int i = 1; i < object_count; ++i) {
56 |     avg_motion += velocities[i] / (double)object_count;
57 |   }
58 |   return distance((float4){0, 0, 0, 0}, avg_motion);
59 | }
60 | }
61 | 
62 | int main(void) {
63 |   init_positions();
64 |   init_velocities();
65 |   std::printf("Beginning simulation.\n");
66 |   fflush(stdout);
67 |   for (int i = 0; i < 20; ++i) {
68 |     update();
69 |     std::printf("%g\n", avg_motion());
70 |     fflush(stdout);
71 |   }
72 | }
73 | 


--------------------------------------------------------------------------------