├── .gitignore
├── lto
    ├── base_mmul.cpp
    ├── multi_tu_bench.cpp
    └── single_tu_bench.cpp
├── branch_prediction
    ├── vf_size.cpp
    └── vf_calls.cpp
├── false_sharing
    ├── atomic_int.cpp
    ├── aligned_type.cpp
    ├── vary_thread.cpp
    └── false_sharing.cpp
├── sso
    ├── sso.cpp
    └── benchmark.cpp
├── associativity
    ├── llc_bench.cpp
    └── l1_bench.cpp
├── matrix_vector
    ├── read_bench.cpp
    ├── mv_bench.cpp
    ├── mv_bench_avx.cpp
    └── mv_bench_aligned.cpp
├── aliasing
    ├── matrixMul_naive.cu
    ├── matrixMul_alt.cu
    └── matrixMul.cu
├── code_scheduling
    └── fast_mod.cpp
├── prefetching
    └── prefetching.cpp
├── README.md
└── LICENSE


/.gitignore:
--------------------------------------------------------------------------------
 1 | # Prerequisites
 2 | *.d
 3 | 
 4 | # Compiled Object files
 5 | *.slo
 6 | *.lo
 7 | *.o
 8 | *.obj
 9 | 
10 | # Precompiled Headers
11 | *.gch
12 | *.pch
13 | 
14 | # Compiled Dynamic libraries
15 | *.so
16 | *.dylib
17 | *.dll
18 | 
19 | # Fortran module files
20 | *.mod
21 | *.smod
22 | 
23 | # Compiled Static libraries
24 | *.lai
25 | *.la
26 | *.a
27 | *.lib
28 | 
29 | # Executables
30 | *.exe
31 | *.out
32 | *.app
33 | 


--------------------------------------------------------------------------------
/lto/base_mmul.cpp:
--------------------------------------------------------------------------------
 1 | // Our baseline matrix multiplication
 2 | // By: Nick from CoffeeBeforeArch
 3 | 
 4 | void base_mmul(const int *a, const int *b, int *c, const int N) {
 5 |   // For every row...
 6 |   for (int i = 0; i < N; i++) {
 7 |     // For every col...
 8 |     for (int j = 0; j < N; j++) {
 9 |       // For each element in the row-col pair
10 |       for (int k = 0; k < N; k++) {
11 |         // Accumulate the partial results
12 |         c[i * N + j] += a[i * N + k] * b[k * N + j];
13 |       }
14 |     }
15 |   }
16 | }
17 | 
18 | 


--------------------------------------------------------------------------------
/branch_prediction/vf_size.cpp:
--------------------------------------------------------------------------------
 1 | // This program shows off the size of objects with virtual functions in C++
 2 | // By: Nick from CoffeeBeforeArch
 3 | 
 4 | #include <iostream>
 5 | 
 6 | // Class with only a single virtual function
 7 | struct VF {
 8 |   virtual float retVal() { return 1.0f; }
 9 | };
10 | 
11 | // Class with a single method
12 | struct NoVF {
13 |   float retVal() { return 1.0f; }
14 | };
15 | 
16 | int main() {
17 |   std::cout << "sizeof(VF) = " << sizeof(VF) << '\n';
18 |   std::cout << "sizeof(NoVF) = " << sizeof(NoVF) << '\n';
19 |   return 0;
20 | }
21 | 


--------------------------------------------------------------------------------
/false_sharing/atomic_int.cpp:
--------------------------------------------------------------------------------
 1 | // This program shows how atomic integers may be allocated in C++
 2 | // By: Nick from CoffeeBeforeArch
 3 | 
 4 | #include <atomic>
 5 | #include <iostream>
 6 | 
 7 | int main() {
 8 |   // If we create four atomic integers like this, there's a high probability
 9 |   // they'll wind up next to each other in memory
10 |   std::atomic<int> a;
11 |   std::atomic<int> b;
12 |   std::atomic<int> c;
13 |   std::atomic<int> d;
14 | 
15 |   // Print out the addresses
16 |   std::cout << "Address of atomic<int> a - " << &a << '\n';
17 |   std::cout << "Address of atomic<int> b - " << &b << '\n';
18 |   std::cout << "Address of atomic<int> c - " << &c << '\n';
19 |   std::cout << "Address of atomic<int> d - " << &d << '\n';
20 | 
21 |   return 0;
22 | }
23 | 


--------------------------------------------------------------------------------
/false_sharing/aligned_type.cpp:
--------------------------------------------------------------------------------
 1 | // This program shows how atomic integers may be allocated in C++
 2 | // By: Nick from CoffeeBeforeArch
 3 | 
 4 | #include <atomic>
 5 | #include <iostream>
 6 | 
 7 | // Our aligned atomic
 8 | struct alignas(64) AlignedType {
 9 |   AlignedType() { val = 0; }
10 |   std::atomic<int> val;
11 | };
12 | 
13 | int main() {
14 |   // Now we're guaranteed that our atomics will be at least 64 bytes apart!
15 |   AlignedType a{};
16 |   AlignedType b{};
17 |   AlignedType c{};
18 |   AlignedType d{};
19 | 
20 |   // Print out the addresses
21 |   std::cout << "Address of AlignedType a - " << &a << '\n';
22 |   std::cout << "Address of AlignedType b - " << &b << '\n';
23 |   std::cout << "Address of AlignedType c - " << &c << '\n';
24 |   std::cout << "Address of AlignedType d - " << &d << '\n';
25 | 
26 |   return 0;
27 | }
28 | 


--------------------------------------------------------------------------------
/sso/sso.cpp:
--------------------------------------------------------------------------------
 1 | // This program implements a set of benchmarks to show off the short
 2 | // string optimization in C++
 3 | // By: Nick from CoffeeBeforeArch
 4 | 
 5 | #include <cstdlib>
 6 | #include <iostream>
 7 | #include <string>
 8 | 
 9 | // Overload new operator to track heap allocations
10 | void* operator new(size_t n) {
11 |   std::cout << "Allocating " << n << " bytes\n";
12 |   return malloc(n);
13 | }
14 | 
15 | int main() {
16 |   // First, let's see how big a string is
17 |   size_t string_size = sizeof(std::string);
18 |   std::cout << "Size of string = " << string_size << '\n';
19 | 
20 |   // Gradually increase the size of the string in the loop
21 |   for (size_t i = 0; i < 32; i++) {
22 |     std::string s(i, 'X');
23 |     std::cout << "Characters: " << i
24 |               << " Address: " << static_cast<const void*>(s.data()) << '\n';
25 |   }
26 | 
27 |   return 0;
28 | }
29 | 


--------------------------------------------------------------------------------
/sso/benchmark.cpp:
--------------------------------------------------------------------------------
 1 | // This program contains a simple benchmark for testing the how the SSO
 2 | // helps performance on small strings
 3 | 
 4 | #include <benchmark/benchmark.h>
 5 | #include <string>
 6 | #include <vector>
 7 | 
 8 | static void stringBench(benchmark::State &s) {
 9 |   // Get the number of characters for our string
10 |   int string_len = s.range(0);
11 | 
12 |   // Vector for holding the strings
13 |   std::vector<std::string> v;
14 |   v.reserve(10000);
15 | 
16 |   // Now let's push back a ton of strings
17 |   while (s.KeepRunning()) {
18 |     for (int i = 0; i < 10000; i++) {
19 |       // Create the string of a specified size
20 |       v.emplace_back(std::string(string_len, 'X'));
21 |     }
22 |   }
23 | }
24 | // Register the benchmark and specify a range of string values
25 | BENCHMARK(stringBench)->DenseRange(0, 32)->Unit(benchmark::kMillisecond);
26 | 
27 | // Benchmark main function
28 | BENCHMARK_MAIN();
29 | 


--------------------------------------------------------------------------------
/lto/multi_tu_bench.cpp:
--------------------------------------------------------------------------------
 1 | // This program is a simple benchmark for matrix multiplication to show off
 2 | // aliasing
 3 | // By: Nick from CoffeeBeforeArch
 4 | 
 5 | #include <benchmark/benchmark.h>
 6 | #include <cstdlib>
 7 | 
 8 | // Function prototypes
 9 | void base_mmul(const int *a, const int *b, int *c, const int N);
10 | 
11 | // Baseline matrix multiplication that suffers from pointer aliasing
12 | static void baseline(benchmark::State &s) {
13 |   // Unpack the dimension of the square matrix
14 |   const int N = 1 << s.range(0);
15 | 
16 |   // Allocate for our matrices
17 |   int *a = new int[N * N]();
18 |   int *b = new int[N * N]();
19 |   int *c = new int[N * N]();
20 | 
21 |   // Region to profile
22 |   while (s.KeepRunning()) {
23 |     base_mmul(a, b, c, N);
24 |   }
25 | 
26 |   // Free our memory
27 |   delete[] a;
28 |   delete[] b;
29 |   delete[] c;
30 | }
31 | BENCHMARK(baseline)->DenseRange(8, 10)->Unit(benchmark::kMillisecond);
32 | 
33 | BENCHMARK_MAIN();
34 | 


--------------------------------------------------------------------------------
/associativity/llc_bench.cpp:
--------------------------------------------------------------------------------
 1 | // This program shows off cache associativity using C++
 2 | // By: Nick from CoffeeBeforeArch
 3 | 
 4 | #include <benchmark/benchmark.h>
 5 | #include <vector>
 6 | 
 7 | using std::vector;
 8 | 
 9 | // Benchmark for showing cache associativity
10 | static void LLC_Bench(benchmark::State &s) {
11 |   // Const step size (512kB)
12 |   const int step = 1 << 17;
13 | 
14 |   // Use a variable array size
15 |   const int size = 1 << s.range(0);
16 |   vector<int> v(size);
17 | 
18 |   // Number of accesses
19 |   const int MAX_ITER = 1 << 20;
20 | 
21 |   // Profile the runtime of different step sizes
22 |   while (s.KeepRunning()) {
23 |     int i = 0;
24 |     for (int iter = 0; iter < MAX_ITER; iter++) {
25 |       v[i]++;
26 |       // Reset if we go off the end of the array
27 |       i += step;
28 |       if (i >= size) i = 0;
29 |     }
30 |   }
31 | }
32 | // Register the benchmark
33 | BENCHMARK(LLC_Bench)->DenseRange(20, 30)->Unit(benchmark::kMillisecond);
34 | 
35 | // Benchmark main function
36 | BENCHMARK_MAIN();
37 | 


--------------------------------------------------------------------------------
/associativity/l1_bench.cpp:
--------------------------------------------------------------------------------
 1 | // This program shows off cache associativity using C++
 2 | // By: Nick from CoffeeBeforeArch
 3 | 
 4 | #include <benchmark/benchmark.h>
 5 | #include <vector>
 6 | 
 7 | using std::generate;
 8 | using std::vector;
 9 | 
10 | // Benchmark for showing cache associativity
11 | static void L1_Bench(benchmark::State &s) {
12 |   // Const step size (4kB)
13 |   const int step = 1 << 10;
14 | 
15 |   // Use a variable array size
16 |   int size = 1 << s.range(0);
17 |   vector<int> v(size);
18 | 
19 |   // Number of accesses
20 |   const int MAX_ITER = 1 << 20;
21 | 
22 |   // Profile the runtime of different step sizes
23 |   while (s.KeepRunning()) {
24 |     int i = 0;
25 |     for (int iter = 0; iter < MAX_ITER; iter++) {
26 |       v[i]++;
27 |       // Reset if we go off the end of the array
28 |       i += step;
29 |       if (i >= size) i = 0;
30 |     }
31 |   }
32 | }
33 | // Register the benchmark
34 | BENCHMARK(L1_Bench)->DenseRange(13, 16)->Unit(benchmark::kMillisecond);
35 | 
36 | // Benchmark main function
37 | BENCHMARK_MAIN();
38 | 


--------------------------------------------------------------------------------
/matrix_vector/read_bench.cpp:
--------------------------------------------------------------------------------
 1 | // This program implements a simple memory benchmark in C++
 2 | // By: Nick from CoffeeBeforeArch
 3 | 
 4 | #include <benchmark/benchmark.h>
 5 | #include <cstdlib>
 6 | #include <cstring>
 7 | 
 8 | using namespace std;
 9 | 
10 | static void readBench(benchmark::State &s) {
11 |   // Get the size from the input
12 |   int dim = 1 << s.range(0);
13 | 
14 |   // Allocate and initialize
15 |   float *matrix = new float[dim * dim];
16 |   for (int i = 0; i < dim * dim; i++) {
17 |     matrix[i] = rand() % 100;
18 |   }
19 | 
20 |   float sink = 0;
21 | 
22 |   while (s.KeepRunning()) {
23 |     for (int i = 0; i < dim * dim; i++) {
24 |       sink = matrix[i];
25 |     }
26 |   }
27 |   benchmark::DoNotOptimize(sink);
28 | 
29 |   // Set the items processed
30 |   s.SetItemsProcessed(dim * dim * s.iterations());
31 | 
32 |   // Set bytes processed
33 |   s.SetBytesProcessed(sizeof(float) * dim * dim * s.iterations());
34 | }
35 | // Register the benchmark
36 | BENCHMARK(readBench)->DenseRange(8, 10)->Unit(benchmark::kMicrosecond);
37 | 
38 | // Benchmark main function
39 | BENCHMARK_MAIN();
40 | 


--------------------------------------------------------------------------------
/lto/single_tu_bench.cpp:
--------------------------------------------------------------------------------
 1 | // This program is a simple benchmark for matrix multiplication to show off
 2 | // aliasing
 3 | // By: Nick from CoffeeBeforeArch
 4 | 
 5 | #include <benchmark/benchmark.h>
 6 | #include <cstdlib>
 7 | 
 8 | void base_mmul(const int *a, const int *b, int *c, const int N) {
 9 |   // For every row...
10 |   for (int i = 0; i < N; i++) {
11 |     // For every col...
12 |     for (int j = 0; j < N; j++) {
13 |       // For each element in the row-col pair
14 |       for (int k = 0; k < N; k++) {
15 |         // Accumulate the partial results
16 |         c[i * N + j] += a[i * N + k] * b[k * N + j];
17 |       }
18 |     }
19 |   }
20 | }
21 | 
22 | // Baseline matrix multiplication that suffers from pointer aliasing
23 | static void baseline(benchmark::State &s) {
24 |   // Unpack the dimension of the square matrix
25 |   const int N = 1 << s.range(0);
26 | 
27 |   // Allocate for our matrices
28 |   int *a = new int[N * N]();
29 |   int *b = new int[N * N]();
30 |   int *c = new int[N * N]();
31 | 
32 |   // Region to profile
33 |   while (s.KeepRunning()) {
34 |     base_mmul(a, b, c, N);
35 |   }
36 | 
37 |   // Free our memory
38 |   delete[] a;
39 |   delete[] b;
40 |   delete[] c;
41 | }
42 | BENCHMARK(baseline)->DenseRange(8, 10)->Unit(benchmark::kMillisecond);
43 | 
44 | BENCHMARK_MAIN();
45 | 


--------------------------------------------------------------------------------
/matrix_vector/mv_bench.cpp:
--------------------------------------------------------------------------------
 1 | // This program implements a simple benchmark for matrix vector
 2 | // multiplication to show a basic optimization process
 3 | // By: Nick from CoffeeBeforeArch
 4 | 
 5 | #include <benchmark/benchmark.h>
 6 | #include <cstdlib>
 7 | 
 8 | using namespace std;
 9 | 
10 | void matrix_vector(float *m, float *v, float *r, int dim) {
11 |   for (int i = 0; i < dim; i++) {
12 |     for (int j = 0; j < dim; j++) {
13 |       r[i] += v[j] * m[i * dim + j];
14 |     }
15 |   }
16 | }
17 | 
18 | static void mvBench(benchmark::State &s) {
19 |   // Get the size from the input
20 |   int dim = 1 << s.range(0);
21 | 
22 |   // Allocate and initialize
23 |   float *matrix = new float[dim * dim];
24 |   float *vec = new float[dim];
25 |   float *res = new float[dim];
26 | 
27 |   // Initialize the allocated space
28 |   for (int i = 0; i < dim; i++) {
29 |     vec[i] = rand() % 100;
30 |     res[i] = 0;
31 |     for (int j = 0; j < dim; j++) {
32 |       matrix[i * dim + j] = rand() % 100;
33 |     }
34 |   }
35 | 
36 |   // Run matrix vector product in a loop
37 |   while (s.KeepRunning()) {
38 |     matrix_vector(matrix, vec, res, dim);
39 |   }
40 | 
41 |   // Free our memory
42 |   delete[] matrix;
43 |   delete[] vec;
44 |   delete[] res;
45 | 
46 |   // Set the items processed
47 |   s.SetItemsProcessed(dim * dim * s.iterations());
48 | 
49 |   // Set bytes processed
50 |   s.SetBytesProcessed(sizeof(float) * dim * (dim + 2) * s.iterations());
51 | }
52 | // Register the benchmark
53 | BENCHMARK(mvBench)->DenseRange(8, 10)->Unit(benchmark::kMicrosecond);
54 | 
55 | // Benchmark main function
56 | BENCHMARK_MAIN();
57 | 


--------------------------------------------------------------------------------
/matrix_vector/mv_bench_avx.cpp:
--------------------------------------------------------------------------------
 1 | // This program implements a simple benchmark for matrix vector
 2 | // multiplication to show a basic optimization process
 3 | // By: Nick from CoffeeBeforeArch
 4 | 
 5 | #include <benchmark/benchmark.h>
 6 | #include <immintrin.h>
 7 | #include <cstdlib>
 8 | #include <cstring>
 9 | 
10 | using namespace std;
11 | 
12 | // Inlined function that uses intrinsic
13 | inline float prod_8(float *m_v, float *v) {
14 |   // Input/Output data from the intrinsic
15 |   float r[8];
16 |   __m256 rv;
17 | 
18 |   // Dot product intrinsic
19 |   rv = _mm256_dp_ps(_mm256_load_ps(m_v), _mm256_load_ps(v), 0xf1);
20 | 
21 |   // Safe way to avoid type punning!
22 |   std::memcpy(r, &rv, sizeof(float) * 8);
23 | 
24 |   // Now add the two partial sums together
25 |   return r[0] + r[4];
26 | }
27 | 
28 | // Do MV in multiples of eight using a helper function that uses an
29 | // intrinsic (assumes multiple of eight input)
30 | float vv_prod(float *m_v, float *v, int dim) {
31 |   float res = 0;
32 |   for (int i = 0; i < dim; i += 8) {
33 |     res += prod_8(m_v + i, v + i);
34 |   }
35 | 
36 |   return res;
37 | }
38 | 
39 | // Matrix-Vector Multiplication
40 | void matrix_vector(float *m, float *v, float *r, int dim) {
41 |   for (int i = 0; i < dim; i++) {
42 |     r[i] = vv_prod(&m[i * dim], &v[0], dim);
43 |   }
44 | }
45 | 
46 | static void mvBench(benchmark::State &s) {
47 |   // Get the size from the input
48 |   int dim = 1 << s.range(0);
49 | 
50 |   // Allocate and initialize
51 |   float *matrix = new float[dim * dim];
52 |   float *vec = new float[dim];
53 |   float *res = new float[dim];
54 | 
55 |   // Initialize the allocated space
56 |   for (int i = 0; i < dim; i++) {
57 |     vec[i] = rand() % 100;
58 |     res[i] = 0;
59 |     for (int j = 0; j < dim; j++) {
60 |       matrix[i * dim + j] = rand() % 100;
61 |     }
62 |   }
63 | 
64 |   // Run matrix vector product in a loop
65 |   while (s.KeepRunning()) {
66 |     matrix_vector(matrix, vec, res, dim);
67 |   }
68 | 
69 |   // Free our memory
70 |   delete[] matrix;
71 |   delete[] vec;
72 |   delete[] res;
73 | 
74 |   // Set the items processed
75 |   s.SetItemsProcessed(dim * dim * s.iterations());
76 | 
77 |   // Set bytes processed
78 |   s.SetBytesProcessed(sizeof(float) * dim * (dim + 2) * s.iterations());
79 | }
80 | // Register the benchmark
81 | BENCHMARK(mvBench)->DenseRange(8, 10)->Unit(benchmark::kMicrosecond);
82 | 
83 | // Benchmark main function
84 | BENCHMARK_MAIN();
85 | 


--------------------------------------------------------------------------------
/matrix_vector/mv_bench_aligned.cpp:
--------------------------------------------------------------------------------
 1 | // This program implements a simple benchmark for matrix-vector
 2 | // multiplication with fixed alignment.
 3 | // By: Nick from CoffeeBeforeArch
 4 | 
 5 | #include <benchmark/benchmark.h>
 6 | #include <immintrin.h>
 7 | #include <cstdlib>
 8 | #include <cstring>
 9 | 
10 | using namespace std;
11 | 
12 | // Inlined function that uses intrinsic
13 | inline float prod_8(float *m_v, float *v) {
14 |   // Input/Output data from the intrinsic
15 |   float r[8];
16 |   __m256 rv;
17 | 
18 |   // Dot product intrinsic
19 |   rv = _mm256_dp_ps(_mm256_load_ps(m_v), _mm256_load_ps(v), 0xf1);
20 | 
21 |   // Avoid type punning in a union
22 |   std::memcpy(r, &rv, sizeof(float) * 8);
23 | 
24 |   // Now add the two partial sums together
25 |   return r[0] + r[4];
26 | }
27 | 
28 | // Do MV in multiples of eight using a helper function that uses an
29 | // intrinsic (assumes multiple of eight input)
30 | float vv_prod(float *m_v, float *v, int dim) {
31 |   float res = 0;
32 |   for (int i = 0; i < dim; i += 8) {
33 |     res += prod_8(m_v + i, v + i);
34 |   }
35 | 
36 |   return res;
37 | }
38 | 
39 | // Matrix-Vector Multiplication
40 | void matrix_vector(float *m, float *v, float *r, int dim) {
41 |   for (int i = 0; i < dim; i++) {
42 |     r[i] = vv_prod(&m[i * dim], &v[0], dim);
43 |   }
44 | }
45 | 
46 | // Helper allocator function for posix_memalign
47 | float *allocate(size_t bytes) {
48 |   // Allocate memory alligned to 64-bytes
49 |   void *memory;
50 |   if (posix_memalign(&memory, 64, bytes)) abort();
51 |   return static_cast<float *>(memory);
52 | }
53 | 
54 | static void mvBench(benchmark::State &s) {
55 |   // Get the size from the input
56 |   int dim = 1 << s.range(0);
57 | 
58 |   // Allocate and initialize
59 |   float *matrix = allocate(dim * dim * sizeof(float));
60 |   float *vec = allocate(dim * sizeof(float));
61 |   float *res = allocate(dim * sizeof(float));
62 | 
63 |   // Initialize the allocated space
64 |   for (int i = 0; i < dim; i++) {
65 |     vec[i] = rand() % 100;
66 |     res[i] = 0;
67 |     for (int j = 0; j < dim; j++) {
68 |       matrix[i * dim + j] = rand() % 100;
69 |     }
70 |   }
71 | 
72 |   // Run matrix vector product in a loop
73 |   while (s.KeepRunning()) {
74 |     matrix_vector(matrix, vec, res, dim);
75 |   }
76 | 
77 |   // Free our memory
78 |   free(matrix);
79 |   free(vec);
80 |   free(res);
81 | 
82 |   // Set the items processed
83 |   s.SetItemsProcessed(dim * dim * s.iterations());
84 | 
85 |   // Set bytes processed
86 |   s.SetBytesProcessed(sizeof(float) * dim * (dim + 2) * s.iterations());
87 | }
88 | // Register the benchmark
89 | BENCHMARK(mvBench)->DenseRange(8, 10)->Unit(benchmark::kMicrosecond);
90 | 
91 | // Benchmark main function
92 | BENCHMARK_MAIN();
93 | 


--------------------------------------------------------------------------------
/aliasing/matrixMul_naive.cu:
--------------------------------------------------------------------------------
  1 | // This program computes a simple version of matrix multiplication
  2 | // By: Nick from CoffeeBeforeArch
  3 | 
  4 | #include <algorithm>
  5 | #include <cassert>
  6 | #include <cstdlib>
  7 | #include <functional>
  8 | #include <iostream>
  9 | #include <vector>
 10 | 
 11 | using std::cout;
 12 | using std::generate;
 13 | using std::vector;
 14 | 
 15 | __global__ void matrixMul(const int *a, const int *b, int *c, int N) {
 16 |   // Compute each thread's global row and column index
 17 |   int row = blockIdx.y * blockDim.y + threadIdx.y;
 18 |   int col = blockIdx.x * blockDim.x + threadIdx.x;
 19 | 
 20 |   // Iterate over row, and down column
 21 |   c[row * N + col] = 0;
 22 |   for (int k = 0; k < N; k++) {
 23 |     // Accumulate results for a single element
 24 |     c[row * N + col] += a[row * N + k] * b[k * N + col];
 25 |   }
 26 | }
 27 | 
 28 | // Check result on the CPU
 29 | void verify_result(vector<int> &a, vector<int> &b, vector<int> &c, int N) {
 30 |   // For every row...
 31 |   for (int i = 0; i < N; i++) {
 32 |     // For every column...
 33 |     for (int j = 0; j < N; j++) {
 34 |       // For every element in the row-column pair
 35 |       int tmp = 0;
 36 |       for (int k = 0; k < N; k++) {
 37 |         // Accumulate the partial results
 38 |         tmp += a[i * N + k] * b[k * N + j];
 39 |       }
 40 | 
 41 |       // Check against the CPU result
 42 |       assert(tmp == c[i * N + j]);
 43 |     }
 44 |   }
 45 | }
 46 | 
 47 | int main() {
 48 |   // Matrix size of 1024 x 1024;
 49 |   int N = 1 << 10;
 50 | 
 51 |   // Size (in bytes) of matrix
 52 |   size_t bytes = N * N * sizeof(int);
 53 | 
 54 |   // Host vectors
 55 |   vector<int> h_a(N * N);
 56 |   vector<int> h_b(N * N);
 57 |   vector<int> h_c(N * N);
 58 | 
 59 |   // Initialize matrices
 60 |   generate(h_a.begin(), h_a.end(), []() { return rand() % 100; });
 61 |   generate(h_b.begin(), h_b.end(), []() { return rand() % 100; });
 62 | 
 63 |   // Allocate device memory
 64 |   int *d_a, *d_b, *d_c;
 65 |   cudaMalloc(&d_a, bytes);
 66 |   cudaMalloc(&d_b, bytes);
 67 |   cudaMalloc(&d_c, bytes);
 68 | 
 69 |   // Copy data to the device
 70 |   cudaMemcpy(d_a, h_a.data(), bytes, cudaMemcpyHostToDevice);
 71 |   cudaMemcpy(d_b, h_b.data(), bytes, cudaMemcpyHostToDevice);
 72 | 
 73 |   // Threads per CTA dimension
 74 |   int THREADS = 32;
 75 | 
 76 |   // Blocks per grid dimension (assumes THREADS divides N evenly)
 77 |   int BLOCKS = N / THREADS;
 78 | 
 79 |   // Use dim3 structs for block  and grid dimensions
 80 |   dim3 threads(THREADS, THREADS);
 81 |   dim3 blocks(BLOCKS, BLOCKS);
 82 | 
 83 |   // Launch kernel
 84 |   matrixMul<<<blocks, threads>>>(d_a, d_b, d_c, N);
 85 | 
 86 |   // Copy back to the host
 87 |   cudaMemcpy(h_c.data(), d_c, bytes, cudaMemcpyDeviceToHost);
 88 | 
 89 |   // Check result
 90 |   verify_result(h_a, h_b, h_c, N);
 91 | 
 92 |   printf("COMPLETED SUCCESSFULLY\n");
 93 | 
 94 |   // Free memory on device
 95 |   cudaFree(d_a);
 96 |   cudaFree(d_b);
 97 |   cudaFree(d_c);
 98 | 
 99 |   return 0;
100 | }
101 | 


--------------------------------------------------------------------------------
/aliasing/matrixMul_alt.cu:
--------------------------------------------------------------------------------
  1 | // This program computes a simple version of matrix multiplication
  2 | // By: Nick from CoffeeBeforeArch
  3 | 
  4 | #include <algorithm>
  5 | #include <cassert>
  6 | #include <cstdlib>
  7 | #include <functional>
  8 | #include <iostream>
  9 | #include <vector>
 10 | 
 11 | using std::cout;
 12 | using std::generate;
 13 | using std::vector;
 14 | 
 15 | __global__ void matrixMul(const int *a, const int *b, int *c, int N) {
 16 |   // Compute each thread's global row and column index
 17 |   int row = blockIdx.y * blockDim.y + threadIdx.y;
 18 |   int col = blockIdx.x * blockDim.x + threadIdx.x;
 19 | 
 20 |   // Iterate over row, and down column
 21 |   int tmp = 0;
 22 |   for (int k = 0; k < N; k++) {
 23 |     // Accumulate results for a single element
 24 |     tmp += a[row * N + k] * b[k * N + col];
 25 |   }
 26 | 
 27 |   // Write back the result
 28 |   c[row * N + col] = tmp;
 29 | }
 30 | 
 31 | // Check result on the CPU
 32 | void verify_result(vector<int> &a, vector<int> &b, vector<int> &c, int N) {
 33 |   // For every row...
 34 |   for (int i = 0; i < N; i++) {
 35 |     // For every column...
 36 |     for (int j = 0; j < N; j++) {
 37 |       // For every element in the row-column pair
 38 |       int tmp = 0;
 39 |       for (int k = 0; k < N; k++) {
 40 |         // Accumulate the partial results
 41 |         tmp += a[i * N + k] * b[k * N + j];
 42 |       }
 43 | 
 44 |       // Check against the CPU result
 45 |       assert(tmp == c[i * N + j]);
 46 |     }
 47 |   }
 48 | }
 49 | 
 50 | int main() {
 51 |   // Matrix size of 1024 x 1024;
 52 |   int N = 1 << 10;
 53 | 
 54 |   // Size (in bytes) of matrix
 55 |   size_t bytes = N * N * sizeof(int);
 56 | 
 57 |   // Host vectors
 58 |   vector<int> h_a(N * N);
 59 |   vector<int> h_b(N * N);
 60 |   vector<int> h_c(N * N);
 61 | 
 62 |   // Initialize matrices
 63 |   generate(h_a.begin(), h_a.end(), []() { return rand() % 100; });
 64 |   generate(h_b.begin(), h_b.end(), []() { return rand() % 100; });
 65 | 
 66 |   // Allocate device memory
 67 |   int *d_a, *d_b, *d_c;
 68 |   cudaMalloc(&d_a, bytes);
 69 |   cudaMalloc(&d_b, bytes);
 70 |   cudaMalloc(&d_c, bytes);
 71 | 
 72 |   // Copy data to the device
 73 |   cudaMemcpy(d_a, h_a.data(), bytes, cudaMemcpyHostToDevice);
 74 |   cudaMemcpy(d_b, h_b.data(), bytes, cudaMemcpyHostToDevice);
 75 | 
 76 |   // Threads per CTA dimension
 77 |   int THREADS = 32;
 78 | 
 79 |   // Blocks per grid dimension (assumes THREADS divides N evenly)
 80 |   int BLOCKS = N / THREADS;
 81 | 
 82 |   // Use dim3 structs for block  and grid dimensions
 83 |   dim3 threads(THREADS, THREADS);
 84 |   dim3 blocks(BLOCKS, BLOCKS);
 85 | 
 86 |   // Launch kernel
 87 |   matrixMul<<<blocks, threads>>>(d_a, d_b, d_c, N);
 88 | 
 89 |   // Copy back to the host
 90 |   cudaMemcpy(h_c.data(), d_c, bytes, cudaMemcpyDeviceToHost);
 91 | 
 92 |   // Check result
 93 |   verify_result(h_a, h_b, h_c, N);
 94 | 
 95 |   printf("COMPLETED SUCCESSFULLY\n");
 96 | 
 97 |   // Free memory on device
 98 |   cudaFree(d_a);
 99 |   cudaFree(d_b);
100 |   cudaFree(d_c);
101 | 
102 |   return 0;
103 | }
104 | 


--------------------------------------------------------------------------------
/aliasing/matrixMul.cu:
--------------------------------------------------------------------------------
  1 | // This program computes a simple version of matrix multiplication
  2 | // By: Nick from CoffeeBeforeArch
  3 | 
  4 | #include <algorithm>
  5 | #include <cassert>
  6 | #include <cstdlib>
  7 | #include <functional>
  8 | #include <iostream>
  9 | #include <vector>
 10 | 
 11 | using std::cout;
 12 | using std::generate;
 13 | using std::vector;
 14 | 
 15 | __global__ void matrixMul(const int *__restrict a, const int *__restrict b,
 16 |                           int *__restrict c, int N) {
 17 |   // Compute each thread's global row and column index
 18 |   int row = blockIdx.y * blockDim.y + threadIdx.y;
 19 |   int col = blockIdx.x * blockDim.x + threadIdx.x;
 20 | 
 21 |   // Iterate over row, and down column
 22 |   c[row * N + col] = 0;
 23 |   for (int k = 0; k < N; k++) {
 24 |     // Accumulate results for a single element
 25 |     c[row * N + col] += a[row * N + k] * b[k * N + col];
 26 |   }
 27 | }
 28 | 
 29 | // Check result on the CPU
 30 | void verify_result(vector<int> &a, vector<int> &b, vector<int> &c, int N) {
 31 |   // For every row...
 32 |   for (int i = 0; i < N; i++) {
 33 |     // For every column...
 34 |     for (int j = 0; j < N; j++) {
 35 |       // For every element in the row-column pair
 36 |       int tmp = 0;
 37 |       for (int k = 0; k < N; k++) {
 38 |         // Accumulate the partial results
 39 |         tmp += a[i * N + k] * b[k * N + j];
 40 |       }
 41 | 
 42 |       // Check against the CPU result
 43 |       assert(tmp == c[i * N + j]);
 44 |     }
 45 |   }
 46 | }
 47 | 
 48 | int main() {
 49 |   // Matrix size of 1024 x 1024;
 50 |   int N = 1 << 10;
 51 | 
 52 |   // Size (in bytes) of matrix
 53 |   size_t bytes = N * N * sizeof(int);
 54 | 
 55 |   // Host vectors
 56 |   vector<int> h_a(N * N);
 57 |   vector<int> h_b(N * N);
 58 |   vector<int> h_c(N * N);
 59 | 
 60 |   // Initialize matrices
 61 |   generate(h_a.begin(), h_a.end(), []() { return rand() % 100; });
 62 |   generate(h_b.begin(), h_b.end(), []() { return rand() % 100; });
 63 | 
 64 |   // Allocate device memory
 65 |   int *d_a, *d_b, *d_c;
 66 |   cudaMalloc(&d_a, bytes);
 67 |   cudaMalloc(&d_b, bytes);
 68 |   cudaMalloc(&d_c, bytes);
 69 | 
 70 |   // Copy data to the device
 71 |   cudaMemcpy(d_a, h_a.data(), bytes, cudaMemcpyHostToDevice);
 72 |   cudaMemcpy(d_b, h_b.data(), bytes, cudaMemcpyHostToDevice);
 73 | 
 74 |   // Threads per CTA dimension
 75 |   int THREADS = 32;
 76 | 
 77 |   // Blocks per grid dimension (assumes THREADS divides N evenly)
 78 |   int BLOCKS = N / THREADS;
 79 | 
 80 |   // Use dim3 structs for block  and grid dimensions
 81 |   dim3 threads(THREADS, THREADS);
 82 |   dim3 blocks(BLOCKS, BLOCKS);
 83 | 
 84 |   // Launch kernel
 85 |   matrixMul<<<blocks, threads>>>(d_a, d_b, d_c, N);
 86 | 
 87 |   // Copy back to the host
 88 |   cudaMemcpy(h_c.data(), d_c, bytes, cudaMemcpyDeviceToHost);
 89 | 
 90 |   // Check result
 91 |   verify_result(h_a, h_b, h_c, N);
 92 | 
 93 |   printf("COMPLETED SUCCESSFULLY\n");
 94 | 
 95 |   // Free memory on device
 96 |   cudaFree(d_a);
 97 |   cudaFree(d_b);
 98 |   cudaFree(d_c);
 99 | 
100 |   return 0;
101 | }
102 | 


--------------------------------------------------------------------------------
/false_sharing/vary_thread.cpp:
--------------------------------------------------------------------------------
  1 | // This benchmark scales the number of threads in our false sharing benchmark
  2 | // By: Nick from CoffeeBeforeArch
  3 | 
  4 | #include <benchmark/benchmark.h>
  5 | #include <array>
  6 | #include <atomic>
  7 | #include <thread>
  8 | 
  9 | // Simple function for incrementing an atomic int
 10 | void work(std::atomic<int>& a, int n) {
 11 |   for (int i = 0; i < (400000 / n); i++) {
 12 |     a++;
 13 |   }
 14 | }
 15 | 
 16 | // Benchmark 2 threads
 17 | void bench2() {
 18 |   std::atomic<int> a{0};
 19 |   std::atomic<int> b{0};
 20 | 
 21 |   // Creat four threads and use lambda to launch work
 22 |   std::thread t1([&]() { work(a, 2); });
 23 |   std::thread t2([&]() { work(b, 2); });
 24 | 
 25 |   // Join the threads
 26 |   t1.join();
 27 |   t2.join();
 28 | }
 29 | 
 30 | // A simple benchmark that runs our single-threaded implementation
 31 | static void twoThreads(benchmark::State& s) {
 32 |   while (s.KeepRunning()) {
 33 |     bench2();
 34 |   }
 35 | }
 36 | BENCHMARK(twoThreads)->UseRealTime()->Unit(benchmark::kMillisecond);
 37 | 
 38 | // Benchmark 4 threads
 39 | void bench4() {
 40 |   std::atomic<int> a{0};
 41 |   std::atomic<int> b{0};
 42 |   std::atomic<int> c{0};
 43 |   std::atomic<int> d{0};
 44 | 
 45 |   // Creat four threads and use lambda to launch work
 46 |   std::thread t1([&]() { work(a, 4); });
 47 |   std::thread t2([&]() { work(b, 4); });
 48 |   std::thread t3([&]() { work(c, 4); });
 49 |   std::thread t4([&]() { work(d, 4); });
 50 | 
 51 |   // Join the threads
 52 |   t1.join();
 53 |   t2.join();
 54 |   t3.join();
 55 |   t4.join();
 56 | }
 57 | 
 58 | // A simple benchmark that runs our single-threaded implementation
 59 | static void fourThreads(benchmark::State& s) {
 60 |   while (s.KeepRunning()) {
 61 |     bench4();
 62 |   }
 63 | }
 64 | BENCHMARK(fourThreads)->UseRealTime()->Unit(benchmark::kMillisecond);
 65 | 
 66 | // Benchmark 8 threads
 67 | void bench8() {
 68 |   std::atomic<int> a{0};
 69 |   std::atomic<int> b{0};
 70 |   std::atomic<int> c{0};
 71 |   std::atomic<int> d{0};
 72 |   std::atomic<int> e{0};
 73 |   std::atomic<int> f{0};
 74 |   std::atomic<int> g{0};
 75 |   std::atomic<int> h{0};
 76 | 
 77 |   // Creat four threads and use lambda to launch work
 78 |   std::thread t1([&]() { work(a, 8); });
 79 |   std::thread t2([&]() { work(b, 8); });
 80 |   std::thread t3([&]() { work(c, 8); });
 81 |   std::thread t4([&]() { work(d, 8); });
 82 |   std::thread t5([&]() { work(e, 8); });
 83 |   std::thread t6([&]() { work(f, 8); });
 84 |   std::thread t7([&]() { work(g, 8); });
 85 |   std::thread t8([&]() { work(h, 8); });
 86 | 
 87 |   // Join the threads
 88 |   t1.join();
 89 |   t2.join();
 90 |   t3.join();
 91 |   t4.join();
 92 |   t5.join();
 93 |   t6.join();
 94 |   t7.join();
 95 |   t8.join();
 96 | }
 97 | 
 98 | // A simple benchmark that runs our single-threaded implementation
 99 | static void eightThreads(benchmark::State& s) {
100 |   while (s.KeepRunning()) {
101 |     bench8();
102 |   }
103 | }
104 | BENCHMARK(eightThreads)->UseRealTime()->Unit(benchmark::kMillisecond);
105 | 
106 | BENCHMARK_MAIN();
107 | 


--------------------------------------------------------------------------------
/branch_prediction/vf_calls.cpp:
--------------------------------------------------------------------------------
  1 | // This program shows the implications of run-time polymorphism
  2 | // on performance due to branch prediction
  3 | 
  4 | #include <benchmark/benchmark.h>
  5 | #include <algorithm>
  6 | #include <random>
  7 | #include <vector>
  8 | 
  9 | // A simple case of polymorphism
 10 | // One base class with a single virtual function
 11 | struct Mammal {
 12 |   virtual float getSomeNumber() const noexcept { return 1.0; }
 13 | };
 14 | 
 15 | struct Dog : Mammal {
 16 |   float getSomeNumber() const noexcept { return 2.0; }
 17 | };
 18 | 
 19 | struct Cat : Mammal {
 20 |   float getSomeNumber() const noexcept { return 3.0; }
 21 | };
 22 | 
 23 | // Benchmark where all same-type objects are grouped together
 24 | static void vf_sorted(benchmark::State& s) {
 25 |   // Create a vector to store mammals
 26 |   std::vector<Mammal*> zoo;
 27 | 
 28 |   // Use fill_n to insert many instances into the vector
 29 |   std::fill_n(std::back_inserter(zoo), 10000, new Mammal);
 30 |   std::fill_n(std::back_inserter(zoo), 10000, new Dog);
 31 |   std::fill_n(std::back_inserter(zoo), 10000, new Cat);
 32 | 
 33 |   // Acculate a sum here
 34 |   float sum = 0;
 35 | 
 36 |   // Profile here
 37 |   while (s.KeepRunning()) {
 38 |     // VF calls here are easy to predict because all instances of each type
 39 |     // are sequential in the vector
 40 |     for (auto* animal : zoo) {
 41 |       sum += animal->getSomeNumber();
 42 |     }
 43 |   }
 44 | }
 45 | BENCHMARK(vf_sorted)->Unit(benchmark::kMicrosecond);
 46 | 
 47 | // Benchmark where ordering of types is randomized
 48 | static void vf_unsorted(benchmark::State& s) {
 49 |   // Create a vector to store mammals
 50 |   std::vector<Mammal*> zoo;
 51 | 
 52 |   // Use fill_n to insert many instances into the vector
 53 |   std::fill_n(std::back_inserter(zoo), 10000, new Mammal);
 54 |   std::fill_n(std::back_inserter(zoo), 10000, new Dog);
 55 |   std::fill_n(std::back_inserter(zoo), 10000, new Cat);
 56 | 
 57 |   // Now shuffle the vector
 58 |   std::random_device rng;
 59 |   std::mt19937 urng(rng());
 60 |   std::shuffle(zoo.begin(), zoo.end(), urng);
 61 | 
 62 |   // Acculate a sum here
 63 |   float sum = 0;
 64 | 
 65 |   // Profile here
 66 |   while (s.KeepRunning()) {
 67 |     // VF Calls here are ~random, so the branch predictor will have
 68 |     // some trouble
 69 |     for (auto* animal : zoo) {
 70 |       sum += animal->getSomeNumber();
 71 |     }
 72 |   }
 73 | }
 74 | // Register the benchmark
 75 | BENCHMARK(vf_unsorted)->Unit(benchmark::kMicrosecond);
 76 | 
 77 | // Benchmark where ordering of types is striped
 78 | static void vf_striped(benchmark::State& s) {
 79 |   // Create a vector to store mammals
 80 |   std::vector<Mammal*> zoo;
 81 |   zoo.reserve(30000);
 82 | 
 83 |   // Fill the vector with groups of three objects
 84 |   for (int i = 0; i < 10000; i++) {
 85 |     zoo.emplace_back(new Mammal);
 86 |     zoo.emplace_back(new Dog);
 87 |     zoo.emplace_back(new Cat);
 88 |   }
 89 | 
 90 |   // Acculate a sum here
 91 |   float sum = 0;
 92 | 
 93 |   // Profile here
 94 |   while (s.KeepRunning()) {
 95 |     // VF calls occur in a pattern that is easy to predict
 96 |     for (auto* animal : zoo) {
 97 |       sum += animal->getSomeNumber();
 98 |     }
 99 |   }
100 | }
101 | // Register the benchmark
102 | BENCHMARK(vf_striped)->Unit(benchmark::kMicrosecond);
103 | 
104 | // Main function
105 | BENCHMARK_MAIN();
106 | 


--------------------------------------------------------------------------------
/false_sharing/false_sharing.cpp:
--------------------------------------------------------------------------------
  1 | // This program shows off the sever implications of false sharing in
  2 | // C++ using std::atomic and std::thread
  3 | 
  4 | #include <benchmark/benchmark.h>
  5 | #include <atomic>
  6 | #include <thread>
  7 | 
  8 | // Simple function for incrememnting an atomic int
  9 | void work(std::atomic<int>& a) {
 10 |   for (int i = 0; i < 100000; i++) {
 11 |     a++;
 12 |   }
 13 | }
 14 | 
 15 | // Simple single-threaded function that calls work 4 times
 16 | void single_thread() {
 17 |   std::atomic<int> a;
 18 |   a = 0;
 19 | 
 20 |   work(a);
 21 |   work(a);
 22 |   work(a);
 23 |   work(a);
 24 | }
 25 | 
 26 | // A simple benchmark that runs our single-threaded implementation
 27 | static void singleThread(benchmark::State& s) {
 28 |   while (s.KeepRunning()) {
 29 |     single_thread();
 30 |   }
 31 | }
 32 | BENCHMARK(singleThread)->Unit(benchmark::kMillisecond);
 33 | 
 34 | // Tries to parallelize the work across multiple threads
 35 | // However, each core invalidates the other cores copies on a write
 36 | // This is an EXTREME example of poorly thought out code
 37 | void same_var() {
 38 |   std::atomic<int> a;
 39 |   a = 0;
 40 | 
 41 |   // Create four threads and use a lambda to launch work
 42 |   std::thread t1([&]() { work(a); });
 43 |   std::thread t2([&]() { work(a); });
 44 |   std::thread t3([&]() { work(a); });
 45 |   std::thread t4([&]() { work(a); });
 46 | 
 47 |   // Join the threads
 48 |   t1.join();
 49 |   t2.join();
 50 |   t3.join();
 51 |   t4.join();
 52 | }
 53 | 
 54 | // A simple benchmark that runs our single-threaded implementation
 55 | static void directSharing(benchmark::State& s) {
 56 |   while (s.KeepRunning()) {
 57 |     same_var();
 58 |   }
 59 | }
 60 | BENCHMARK(directSharing)->UseRealTime()->Unit(benchmark::kMillisecond);
 61 | 
 62 | // How well does it work if we use different atomic ints?
 63 | // Not that well! But look at the addresses! They all reside on the
 64 | // same cache line. That means we have false sharing!
 65 | // (We invalidate variables that aren't actually being accessed
 66 | // because they happen to be on the same cache line)
 67 | void diff_var() {
 68 |   std::atomic<int> a{0};
 69 |   std::atomic<int> b{0};
 70 |   std::atomic<int> c{0};
 71 |   std::atomic<int> d{0};
 72 | 
 73 |   // Creat four threads and use lambda to launch work
 74 |   std::thread t1([&]() { work(a); });
 75 |   std::thread t2([&]() { work(b); });
 76 |   std::thread t3([&]() { work(c); });
 77 |   std::thread t4([&]() { work(d); });
 78 | 
 79 |   // Join the threads
 80 |   t1.join();
 81 |   t2.join();
 82 |   t3.join();
 83 |   t4.join();
 84 | }
 85 | 
 86 | // A simple benchmark that runs our single-threaded implementation
 87 | static void falseSharing(benchmark::State& s) {
 88 |   while (s.KeepRunning()) {
 89 |     diff_var();
 90 |   }
 91 | }
 92 | BENCHMARK(falseSharing)->UseRealTime()->Unit(benchmark::kMillisecond);
 93 | 
 94 | // We can align the struct to 64 bytes
 95 | // Now each struct will be on a different cache line
 96 | struct alignas(64) AlignedType {
 97 |   AlignedType() { val = 0; }
 98 |   std::atomic<int> val;
 99 | };
100 | 
101 | // No more invalidations, so our code should be roughly the same as the
102 | void diff_line() {
103 |   AlignedType a{};
104 |   AlignedType b{};
105 |   AlignedType c{};
106 |   AlignedType d{};
107 | 
108 |   // Launch the four threads now using our aligned data
109 |   std::thread t1([&]() { work(a.val); });
110 |   std::thread t2([&]() { work(b.val); });
111 |   std::thread t3([&]() { work(c.val); });
112 |   std::thread t4([&]() { work(d.val); });
113 | 
114 |   // Join the threads
115 |   t1.join();
116 |   t2.join();
117 |   t3.join();
118 |   t4.join();
119 | }
120 | 
121 | // A simple benchmark that runs our single-threaded implementation
122 | static void noSharing(benchmark::State& s) {
123 |   while (s.KeepRunning()) {
124 |     diff_line();
125 |   }
126 | }
127 | BENCHMARK(noSharing)->UseRealTime()->Unit(benchmark::kMillisecond);
128 | 
129 | BENCHMARK_MAIN();
130 | 


--------------------------------------------------------------------------------
/code_scheduling/fast_mod.cpp:
--------------------------------------------------------------------------------
  1 | // This program shows off a neat optimization for fast a faster
  2 | // modulo operation in C++
  3 | // By: Nick from CoffeeBeforeArch
  4 | 
  5 | #include <benchmark/benchmark.h>
  6 | #include <random>
  7 | #include <vector>
  8 | 
  9 | // Function for generating argument pairs
 10 | static void custom_args(benchmark::internal::Benchmark *b) {
 11 |   for (int i = 1 << 4; i <= 1 << 10; i <<= 2) {
 12 |     // Collect stats at 1/8, 1/2, and 7/8
 13 |     for (int j : {32, 128, 224}) {
 14 |       b = b->ArgPair(i, j);
 15 |     }
 16 |   }
 17 | }
 18 | 
 19 | // Baseline for intuitive modulo operation
 20 | static void baseMod(benchmark::State &s) {
 21 |   // Number of elements
 22 |   int N = s.range(0);
 23 | 
 24 |   // Max for mod operator
 25 |   int ceil = s.range(1);
 26 | 
 27 |   // Vector for input and output of modulo
 28 |   std::vector<int> input;
 29 |   std::vector<int> output;
 30 |   input.resize(N);
 31 |   output.resize(N);
 32 | 
 33 |   // Generate random inputs
 34 |   std::mt19937 rng;
 35 |   rng.seed(std::random_device()());
 36 |   std::uniform_int_distribution<int> dist(0, 255);
 37 |   for (int &i : input) {
 38 |     i = dist(rng);
 39 |   }
 40 | 
 41 |   while (s.KeepRunning()) {
 42 |     // Compute the modulo for each element
 43 |     for (int i = 0; i < N; i++) {
 44 |       output[i] = input[i] % ceil;
 45 |     }
 46 |   }
 47 | }
 48 | // Register the benchmark
 49 | BENCHMARK(baseMod)->Apply(custom_args);
 50 | 
 51 | // An unrolled version of our baseline
 52 | static void unrollMod(benchmark::State &s) {
 53 |   // Number of elements
 54 |   int N = s.range(0);
 55 | 
 56 |   // Max for mod operator
 57 |   int ceil = s.range(1);
 58 | 
 59 |   // Vector for input and output of modulo
 60 |   std::vector<int> input;
 61 |   std::vector<int> output;
 62 |   input.resize(N);
 63 |   output.resize(N);
 64 | 
 65 |   // Generate random inputs
 66 |   std::mt19937 rng;
 67 |   rng.seed(std::random_device()());
 68 |   std::uniform_int_distribution<int> dist(0, 255);
 69 |   for (int &i : input) {
 70 |     i = dist(rng);
 71 |   }
 72 | 
 73 |   while (s.KeepRunning()) {
 74 |     // Compute the modulo for each element
 75 |     // Unroll the loop by 4
 76 |     for (int i = 0; i < N; i += 4) {
 77 |       output[i] = input[i] % ceil;
 78 |       output[i + 1] = input[i + 1] % ceil;
 79 |       output[i + 2] = input[i + 2] % ceil;
 80 |       output[i + 3] = input[i + 3] % ceil;
 81 |     }
 82 |   }
 83 | }
 84 | // Register the benchmark
 85 | BENCHMARK(unrollMod)->Apply(custom_args);
 86 | 
 87 | // Baseline for intuitive modulo operation
 88 | static void fastMod(benchmark::State &s) {
 89 |   // Number of elements
 90 |   int N = s.range(0);
 91 | 
 92 |   // Max for mod operator
 93 |   int ceil = s.range(1);
 94 | 
 95 |   // Vector for input and output of modulo
 96 |   std::vector<int> input;
 97 |   std::vector<int> output;
 98 |   input.resize(N);
 99 |   output.resize(N);
100 | 
101 |   // Generate random inputs
102 |   std::mt19937 rng;
103 |   rng.seed(std::random_device()());
104 |   std::uniform_int_distribution<int> dist(0, 255);
105 |   for (int &i : input) {
106 |     i = dist(rng);
107 |   }
108 | 
109 |   while (s.KeepRunning()) {
110 |     // DON'T compute the mod for each element
111 |     // Skip the expensive operation using a simple compare
112 |     for (int i = 0; i < N; i++) {
113 |       output[i] = (input[i] >= ceil) ? input[i] % ceil : input[i];
114 |     }
115 |   }
116 | }
117 | // Register the benchmark
118 | BENCHMARK(fastMod)->Apply(custom_args);
119 | 
120 | // Baseline for intuitive modulo operation
121 | static void fastModHint(benchmark::State &s) {
122 |   // Number of elements
123 |   int N = s.range(0);
124 | 
125 |   // Max for mod operator
126 |   int ceil = s.range(1);
127 | 
128 |   // Vector for input and output of modulo
129 |   std::vector<int> input;
130 |   std::vector<int> output;
131 |   input.resize(N);
132 |   output.resize(N);
133 | 
134 |   // Generate random inputs
135 |   std::mt19937 rng;
136 |   rng.seed(std::random_device()());
137 |   std::uniform_int_distribution<int> dist(0, 255);
138 |   for (int &i : input) {
139 |     i = dist(rng);
140 |   }
141 | 
142 |   while (s.KeepRunning()) {
143 |     // DON'T compute the mod for each element
144 |     // Skip the expensive operation using a simple compare
145 |     for (int i = 0; i < N; i++) {
146 |       output[i] =
147 |           __builtin_expect(input[i] >= ceil, 0) ? input[i] % ceil : input[i];
148 |     }
149 |   }
150 | }
151 | // Register the benchmark
152 | BENCHMARK(fastModHint)->Apply(custom_args);
153 | 
154 | // Baseline for intuitive modulo operation
155 | static void fastModHintUnroll(benchmark::State &s) {
156 |   // Number of elements
157 |   int N = s.range(0);
158 | 
159 |   // Max for mod operator
160 |   int ceil = s.range(1);
161 | 
162 |   // Vector for input and output of modulo
163 |   std::vector<int> input;
164 |   std::vector<int> output;
165 |   input.resize(N);
166 |   output.resize(N);
167 | 
168 |   // Generate random inputs
169 |   std::mt19937 rng;
170 |   rng.seed(std::random_device()());
171 |   std::uniform_int_distribution<int> dist(0, 255);
172 |   for (int &i : input) {
173 |     i = dist(rng);
174 |   }
175 | 
176 |   while (s.KeepRunning()) {
177 |     // Unroll our fast mod loop by 4
178 |     for (int i = 0; i < N; i += 4) {
179 |       output[i] =
180 |           __builtin_expect(input[i] >= ceil, 0) ? input[i] % ceil : input[i];
181 |       output[i + 1] = __builtin_expect(input[i + 1] >= ceil, 0)
182 |                           ? input[i + 1] % ceil
183 |                           : input[i + 1];
184 |       output[i + 2] = __builtin_expect(input[i + 2] >= ceil, 0)
185 |                           ? input[i + 2] % ceil
186 |                           : input[i + 2];
187 |       output[i + 3] = __builtin_expect(input[i + 3] >= ceil, 0)
188 |                           ? input[i + 3] % ceil
189 |                           : input[i + 3];
190 |     }
191 |   }
192 | }
193 | // Register the benchmark
194 | BENCHMARK(fastModHintUnroll)->Apply(custom_args);
195 | 
196 | // Benchmark main function
197 | BENCHMARK_MAIN();
198 | 


--------------------------------------------------------------------------------
/prefetching/prefetching.cpp:
--------------------------------------------------------------------------------
  1 | // This program shows how the prefetching impacts performance in C++
  2 | // By: Nick from CoffeeBeforeArch
  3 | 
  4 | #include <benchmark/benchmark.h>
  5 | #include <algorithm>
  6 | #include <cstdlib>
  7 | #include <numeric>
  8 | #include <random>
  9 | #include <vector>
 10 | 
 11 | // Accesses an array sequentially in row-major fashion
 12 | static void rowMajor(benchmark::State &s) {
 13 |   // Input/output vector size
 14 |   int N = 1 << s.range(0);
 15 | 
 16 |   // Create our input indices
 17 |   std::vector<int> v_in(N * N);
 18 |   std::iota(begin(v_in), end(v_in), 0);
 19 | 
 20 |   // Create an output vector
 21 |   std::vector<int> v_out(N * N);
 22 | 
 23 |   // Profile a simple traversal with simple additions
 24 |   while (s.KeepRunning()) {
 25 |     for (int i = 0; i < N * N; i++) {
 26 |       v_out[v_in[i]]++;
 27 |     }
 28 |   }
 29 | }
 30 | // Register the benchmark
 31 | BENCHMARK(rowMajor)->DenseRange(10, 12)->Unit(benchmark::kMillisecond);
 32 | 
 33 | // Accesses an array sequentially in reverse row-major
 34 | static void reverse(benchmark::State &s) {
 35 |   // Input/output vector size
 36 |   int N = 1 << s.range(0);
 37 | 
 38 |   // Create our input indices
 39 |   std::vector<int> v_in(N * N);
 40 |   std::iota(begin(v_in), end(v_in), 0);
 41 |   std::reverse(begin(v_in), end(v_in));
 42 | 
 43 |   // Create an output vector
 44 |   std::vector<int> v_out(N * N);
 45 | 
 46 |   // Profile a simple traversal with simple additions
 47 |   while (s.KeepRunning()) {
 48 |     for (int i = 0; i < N * N; i++) {
 49 |       // Pre-fetch an item for later
 50 |       v_out[v_in[i]]++;
 51 |     }
 52 |   }
 53 | }
 54 | // Register the benchmark
 55 | BENCHMARK(reverse)->DenseRange(10, 12)->Unit(benchmark::kMillisecond);
 56 | 
 57 | // Accesses an array sequentially in row-major fashion
 58 | static void cacheLine(benchmark::State &s) {
 59 |   // Input/output vector size
 60 |   int N = 1 << s.range(0);
 61 | 
 62 |   // Cache line size
 63 |   const int stride = 64 / sizeof(int);
 64 | 
 65 |   // Create our input indices
 66 |   std::vector<int> v_in(N * N);
 67 | 
 68 |   // For each element in a cache line
 69 |   int index = 0;
 70 |   for (int i = 0; i < stride; i++) {
 71 |     // For each cache line in the array
 72 |     for (int j = 0; j < (N * N / stride); j++) {
 73 |       v_in[index] = j * stride + i;
 74 |       index++;
 75 |     }
 76 |   }
 77 | 
 78 |   // Create an output vector
 79 |   std::vector<int> v_out(N * N);
 80 | 
 81 |   // Profile a simple traversal with simple additions
 82 |   while (s.KeepRunning()) {
 83 |     for (int i = 0; i < N * N; i++) {
 84 |       v_out[v_in[i]]++;
 85 |     }
 86 |   }
 87 | }
 88 | // Register the benchmark
 89 | BENCHMARK(cacheLine)->DenseRange(10, 12)->Unit(benchmark::kMillisecond);
 90 | 
 91 | // Accesses an array sequentially in row-major fashion
 92 | static void cacheLineReverse(benchmark::State &s) {
 93 |   // Input/output vector size
 94 |   int N = 1 << s.range(0);
 95 | 
 96 |   // Cache line size
 97 |   const int stride = 64 / sizeof(int);
 98 | 
 99 |   // Create our input indices
100 |   std::vector<int> v_in(N * N);
101 | 
102 |   // For each element in a cache line
103 |   int index = 0;
104 |   for (int i = 0; i < stride; i++) {
105 |     // For each cache line in the array
106 |     for (int j = 0; j < (N * N / stride); j++) {
107 |       v_in[index] = j * stride + i;
108 |       index++;
109 |     }
110 |   }
111 | 
112 |   // Reverse the indices
113 |   std::reverse(begin(v_in), end(v_in));
114 | 
115 |   // Create an output vector
116 |   std::vector<int> v_out(N * N);
117 | 
118 |   // Profile a simple traversal with simple additions
119 |   while (s.KeepRunning()) {
120 |     for (int i = 0; i < N * N; i++) {
121 |       v_out[v_in[i]]++;
122 |     }
123 |   }
124 | }
125 | // Register the benchmark
126 | BENCHMARK(cacheLineReverse)->DenseRange(10, 12)->Unit(benchmark::kMillisecond);
127 | 
128 | // Accesses an array in column-major order
129 | static void columnMajor(benchmark::State &s) {
130 |   // Input/output vector size
131 |   int N = 1 << s.range(0);
132 | 
133 |   // Create our input indices
134 |   std::vector<int> v_in(N * N);
135 |   for (int i = 0; i < N; i++) {
136 |     for (int j = 0; j < N; j++) {
137 |       v_in[i * N + j] = j * N + i;
138 |     }
139 |   }
140 | 
141 |   // Create an output vector
142 |   std::vector<int> v_out(N * N);
143 | 
144 |   // Profile a simple traversal with simple additions
145 |   while (s.KeepRunning()) {
146 |     for (int i = 0; i < N * N; i++) {
147 |       v_out[v_in[i]]++;
148 |     }
149 |   }
150 | }
151 | // Register the benchmark
152 | BENCHMARK(columnMajor)->DenseRange(10, 12)->Unit(benchmark::kMillisecond);
153 | 
154 | // Accesses an array in randomized order
155 | static void random(benchmark::State &s) {
156 |   // Input/output vector size
157 |   int N = 1 << s.range(0);
158 | 
159 |   // Create our input indices
160 |   std::vector<int> v_in(N * N);
161 |   std::iota(begin(v_in), end(v_in), 0);
162 | 
163 |   // Now shuffle the vector
164 |   std::random_device rng;
165 |   std::mt19937 urng(rng());
166 |   std::shuffle(begin(v_in), end(v_in), urng);
167 | 
168 |   // Create an output vector
169 |   std::vector<int> v_out(N * N);
170 | 
171 |   // Profile a simple traversal with simple additions
172 |   while (s.KeepRunning()) {
173 |     for (int i = 0; i < N * N; i++) {
174 |       v_out[v_in[i]]++;
175 |     }
176 |   }
177 | }
178 | // Register the benchmark
179 | BENCHMARK(random)->DenseRange(10, 12)->Unit(benchmark::kMillisecond);
180 | 
181 | // Accesses in a random order but try pre-fetching
182 | static void randomPrefetch(benchmark::State &s) {
183 |   // Input/output vector size
184 |   int N = 1 << s.range(0);
185 | 
186 |   // Create our input indices
187 |   std::vector<int> v_in(N * N);
188 |   std::iota(begin(v_in), end(v_in), 0);
189 | 
190 |   // Now shuffle the vector
191 |   std::random_device rng;
192 |   std::mt19937 urng(rng());
193 |   std::shuffle(begin(v_in), end(v_in), urng);
194 | 
195 |   // Create an output vector
196 |   std::vector<int> v_out(N * N);
197 | 
198 |   // Profile a simple traversal with simple additions
199 |   while (s.KeepRunning()) {
200 |     for (int i = 0; i < N * N; i++) {
201 |       // Pre-fetch an item for later
202 |       __builtin_prefetch(&v_out[v_in[i + 5]]);
203 |       v_out[v_in[i]]++;
204 |     }
205 |   }
206 | }
207 | // Register the benchmark
208 | BENCHMARK(randomPrefetch)->DenseRange(10, 12)->Unit(benchmark::kMillisecond);
209 | 
210 | // Benchmark main functions
211 | BENCHMARK_MAIN();
212 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
  1 | # Hardware, Software, and Compilers! Oh My!
  2 | 
  3 | This repository contains all code/links to all examples presented in the Spring 2020 "Hardware, Software, and Compilers! Oh My!" tutorial. I've provided a breakdown of all of the topics below, along with supplemental links for those interested.
  4 | 
  5 | Cheers,
  6 | 
  7 | Nick
  8 | 
  9 | ## The Small Buffer Optimization
 10 | Dynamic allocations are expensive, and we'd like to avoid them (if possible). One way we can do this is the small buffer optimization. We can leave some extra space inside of a handle to a larger dynamic allocation (like a std::string) where we can store a small number of elements. By doing this, we can defer any dynamic allocation until we run out of space inside the handle. All major compilers implement a form of this optimization for std::string in C++ (but interestingly, they don't all do it the same way!). In this example, we will be looking at a few implementations of the small buffer optimization (known as the short/small string optimization), and benchmarking the cost of dynamic allocation.
 11 | 
 12 | [Short string optimization source code](https://github.com/CoffeeBeforeArch/spring_2020_tutorial/tree/master/sso)
 13 | 
 14 | ### Relevant Links
 15 | [The strange details of std::string at Facebook](https://youtu.be/kPR8h4-qZdk)
 16 | 
 17 | ## Copy Elision
 18 | Copying a large object can be expensive. Compiler writers understand this, and have been implementing optimization passes for copy elision to remove unnecessary copies. A common form of this is the Return Value Optimization (RVO) and Named Return Value Optimizaiton (NRVO). These deal with eliding the copy of an object during a function return. However, you're compiler can't always elide the copy! In this example, we will look at some source and disassembly of a simple function to better understand RVO, and when/where it may not be applicable
 19 | 
 20 | [Source and disassembly](https://godbolt.org/z/aQqHns)
 21 | 
 22 | ### Relevant Links
 23 | 
 24 | [Copy elision in the C++ standard](https://en.cppreference.com/w/cpp/language/copy_elision)
 25 | 
 26 | [Arthur O'Dwyer's 2018 CppCon Talk](https://youtu.be/hA1WNtNyNbo)
 27 | 
 28 | ## Aliasing and Compiler Optimization
 29 | 
 30 | Compilers are smart, and can exploit opportunties that even the most veteran of programmers can't notice. However, compilers aren't omniciant, and this can lead them to being overly conservative. A great example of this is aliasing. If a compiler can not figure out if two references point to the same piece of memory, it may not be able to perform certain optimizations (e.g, vectorization). In this example, we will look at the source and disassembly of a simple function, and measure the performance impact of aliasing on a simple matrix multiplication CUDA kernel.
 31 | 
 32 | [Source and disassembly](https://godbolt.org/z/oYev9z)
 33 | 
 34 | [CUDA matrix multiplication source code](https://github.com/CoffeeBeforeArch/spring_2020_tutorial/tree/master/aliasing)
 35 | 
 36 | ### Relevant Links
 37 | [Strict aliasing in C++](https://gist.github.com/shafik/848ae25ee209f698763cffee272a58f8)
 38 | 
 39 | [Aliasing in CUDA](https://devblogs.nvidia.com/cuda-pro-tip-optimize-pointer-aliasing/)
 40 | 
 41 | ## Link Time Optimization (LTO)
 42 | 
 43 | Changing how we compile a program can change it's performance. One example of this is when we break our compilation into multiple translation units. Because the compiler doesn't have the full context of the program, it may omit some optimizations. However, we can get some of these optimizations back at link time using Link Time Optimization (LTO). In this example, we will look at how the compiler optimizes a matrix multiplication benchmark in a single translation unit, split across two translation units, and split across multiple translation units with Link Time Optimization enabled.
 44 | 
 45 | [CPU matrix multiplication source code](https://github.com/CoffeeBeforeArch/spring_2020_tutorial/tree/master/lto)
 46 | 
 47 | ### Relevant Links
 48 | 
 49 | [GCC's Link Time Optimization](https://gcc.gnu.org/onlinedocs/gccint/LTO-Overview.html)
 50 | 
 51 | ## Branch Prediction
 52 | 
 53 | Modern processors rely on branch predictors to keep the pipeline filled in the presence of branches. However, the programmer has a cruicial role in helping out the branch predictor. If we are able to write more predictable code, the branch predictor is able to speculate better. However, if our branches become somewhat random, we can suffer sever performance penalties. In this example we will take a look at the role of branch prediction with virtual functions/dynamic dispatch. We will benchmark different orderings of virtual function calls, and study the affects on the branch miss prediction rate.
 54 | 
 55 | [Virtual function source code](https://github.com/CoffeeBeforeArch/spring_2020_tutorial/tree/master/branch_prediction)
 56 | 
 57 | ### Relevant Links
 58 | 
 59 | [Agner Fog's Assembly Optimization Guide](https://www.agner.org/optimize/optimizing_assembly.pdf)
 60 | 
 61 | [Agner Fog's Optimizing C++](https://www.agner.org/optimize/optimizing_cpp.pdf)
 62 | 
 63 | ## Code Scheduling
 64 | 
 65 | The dynamic order of instructions can have a significant impact on performance. One example of this is with branches. Preferring one side of a branch over another can lead to significant performance differences. In this example we will look at using a compiler intrinsic to give hints as to what the "hot-side" of a branch is, and measure the performance difference.
 66 | 
 67 | [Link to modulo benchmark](https://github.com/CoffeeBeforeArch/spring_2020_tutorial/tree/master/instruction_scheduling)
 68 | 
 69 | ### Relevant Links
 70 | 
 71 | [Anger Fog's Instruction Tables](https://www.agner.org/optimize/instruction_tables.pdf)
 72 | 
 73 | [Chandler Carruth's 2015 CppCon Talk](https://youtu.be/nXaxk27zwlk)
 74 | 
 75 | ## Cache Associativity
 76 | 
 77 | Caches are critical to providing performance in modern processors. However, seemingly innacuous access patterns can lead to unexpected drops in performance. One example of this is a power-of-two stride. Modern caches are set-associative (each cache line gets mapped to a set, but there are only N ways where it can be placed). When we do a power of two set, we stumble across the relatively simple mapping of cache lines to sets. As we increase the power of two, the number of unique sets we access decreases, until every cache line we access maps to the same one. In this example, we will show how we can predict such access patterns for an arbitrary processor by looking at the cache organization details.
 78 | 
 79 | [Link to associativity benchmark source code](https://github.com/CoffeeBeforeArch/spring_2020_tutorial/tree/master/associativity)
 80 | 
 81 | 
 82 | ### Relevant Links
 83 | 
 84 | [What Every Programmer Should Know About Memory](https://people.freebsd.org/~lstewart/articles/cpumemory.pdf)
 85 | 
 86 | ## Prefetching
 87 | 
 88 | Not all access patterns have the same performance. As we showed in the cache associativity example, power-of-two stride can lead to unfortunate performance consequences. Another part of the hardware we should think about is the prefetcher. If we have a constant stride access pattern, our hardware prefetcher can begin fetching cache lines before we need them. However, more random access patterns do not get this benefit. In this example, we will be looking at a number of different access patterns to explore the limits of hardware prefetching, and briefly discuss software prefetching intrinsics.
 89 | 
 90 | [Link to prefetching benchmark](https://github.com/CoffeeBeforeArch/spring_2020_tutorial/tree/master/prefetching)
 91 | 
 92 | 
 93 | ### Relevant Links
 94 | 
 95 | [Stony Brook University lecture on prefetching](https://compas.cs.stonybrook.edu/~nhonarmand/courses/sp15/cse502/slides/13-prefetch.pdf)
 96 | 
 97 | [What Every Programmer Should Know About Memory](https://people.freebsd.org/~lstewart/articles/cpumemory.pdf)
 98 | 
 99 | ## False Sharing
100 | 
101 | 2004 marked the beginning of the multi-core era, as Intel cancelled a 4GHz Pentium 4 processor in favor of dual core. Since that time, all major processor vendors have focused on multi-core performance over single-core. Optimizing multithreaded applications is difficult. Poorly implemented synchronization strategies can lead to execution times that look worse than if the program was single-threaded. However, some performance pitfalls can be more subtle. One of these is false sharing. In this example, we will show how false sharing can impact performance, and briefly discuss the role of coherence in performance. 
102 | 
103 | [Link to false sharing benchmark](https://github.com/CoffeeBeforeArch/spring_2020_tutorial/tree/master/false_sharing)
104 | 
105 | ### Relevant Links
106 | 
107 | [Intel blog on false sharing](https://software.intel.com/en-us/articles/avoiding-and-identifying-false-sharing-among-threads)
108 | 
109 | ## SIMD Intrinsics - An Optimization Case Study
110 | 
111 | The majority of comput hardware on modern processors is in the vector units. While compiler-based vectorization can typically achieve ~80% maximum performance, there may be situations where we need to take matter into our own hands. However, knowing _if_ you should be applying vectorization is another story. In this example, we will be examining the optimization of matrix-vector multiplication. We will explore whether we are compute or memory-bound, then manually apply some vectorization to boost performance.
112 | 
113 | [Link to matrix-vector benchmark](https://github.com/CoffeeBeforeArch/spring_2020_tutorial/tree/master/matrix_vector)
114 | 
115 | 
116 | [Vectorized Dot Product Intrinsic](https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm256_dp_ps&expand=2185)
117 | 
118 | ### Relevant Links
119 | 
120 | [Intel Intrinsics Guide](https://software.intel.com/sites/landingpage/IntrinsicsGuide/#)
121 | 


--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
  1 |                     GNU GENERAL PUBLIC LICENSE
  2 |                        Version 3, 29 June 2007
  3 | 
  4 |  Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
  5 |  Everyone is permitted to copy and distribute verbatim copies
  6 |  of this license document, but changing it is not allowed.
  7 | 
  8 |                             Preamble
  9 | 
 10 |   The GNU General Public License is a free, copyleft license for
 11 | software and other kinds of works.
 12 | 
 13 |   The licenses for most software and other practical works are designed
 14 | to take away your freedom to share and change the works.  By contrast,
 15 | the GNU General Public License is intended to guarantee your freedom to
 16 | share and change all versions of a program--to make sure it remains free
 17 | software for all its users.  We, the Free Software Foundation, use the
 18 | GNU General Public License for most of our software; it applies also to
 19 | any other work released this way by its authors.  You can apply it to
 20 | your programs, too.
 21 | 
 22 |   When we speak of free software, we are referring to freedom, not
 23 | price.  Our General Public Licenses are designed to make sure that you
 24 | have the freedom to distribute copies of free software (and charge for
 25 | them if you wish), that you receive source code or can get it if you
 26 | want it, that you can change the software or use pieces of it in new
 27 | free programs, and that you know you can do these things.
 28 | 
 29 |   To protect your rights, we need to prevent others from denying you
 30 | these rights or asking you to surrender the rights.  Therefore, you have
 31 | certain responsibilities if you distribute copies of the software, or if
 32 | you modify it: responsibilities to respect the freedom of others.
 33 | 
 34 |   For example, if you distribute copies of such a program, whether
 35 | gratis or for a fee, you must pass on to the recipients the same
 36 | freedoms that you received.  You must make sure that they, too, receive
 37 | or can get the source code.  And you must show them these terms so they
 38 | know their rights.
 39 | 
 40 |   Developers that use the GNU GPL protect your rights with two steps:
 41 | (1) assert copyright on the software, and (2) offer you this License
 42 | giving you legal permission to copy, distribute and/or modify it.
 43 | 
 44 |   For the developers' and authors' protection, the GPL clearly explains
 45 | that there is no warranty for this free software.  For both users' and
 46 | authors' sake, the GPL requires that modified versions be marked as
 47 | changed, so that their problems will not be attributed erroneously to
 48 | authors of previous versions.
 49 | 
 50 |   Some devices are designed to deny users access to install or run
 51 | modified versions of the software inside them, although the manufacturer
 52 | can do so.  This is fundamentally incompatible with the aim of
 53 | protecting users' freedom to change the software.  The systematic
 54 | pattern of such abuse occurs in the area of products for individuals to
 55 | use, which is precisely where it is most unacceptable.  Therefore, we
 56 | have designed this version of the GPL to prohibit the practice for those
 57 | products.  If such problems arise substantially in other domains, we
 58 | stand ready to extend this provision to those domains in future versions
 59 | of the GPL, as needed to protect the freedom of users.
 60 | 
 61 |   Finally, every program is threatened constantly by software patents.
 62 | States should not allow patents to restrict development and use of
 63 | software on general-purpose computers, but in those that do, we wish to
 64 | avoid the special danger that patents applied to a free program could
 65 | make it effectively proprietary.  To prevent this, the GPL assures that
 66 | patents cannot be used to render the program non-free.
 67 | 
 68 |   The precise terms and conditions for copying, distribution and
 69 | modification follow.
 70 | 
 71 |                        TERMS AND CONDITIONS
 72 | 
 73 |   0. Definitions.
 74 | 
 75 |   "This License" refers to version 3 of the GNU General Public License.
 76 | 
 77 |   "Copyright" also means copyright-like laws that apply to other kinds of
 78 | works, such as semiconductor masks.
 79 | 
 80 |   "The Program" refers to any copyrightable work licensed under this
 81 | License.  Each licensee is addressed as "you".  "Licensees" and
 82 | "recipients" may be individuals or organizations.
 83 | 
 84 |   To "modify" a work means to copy from or adapt all or part of the work
 85 | in a fashion requiring copyright permission, other than the making of an
 86 | exact copy.  The resulting work is called a "modified version" of the
 87 | earlier work or a work "based on" the earlier work.
 88 | 
 89 |   A "covered work" means either the unmodified Program or a work based
 90 | on the Program.
 91 | 
 92 |   To "propagate" a work means to do anything with it that, without
 93 | permission, would make you directly or secondarily liable for
 94 | infringement under applicable copyright law, except executing it on a
 95 | computer or modifying a private copy.  Propagation includes copying,
 96 | distribution (with or without modification), making available to the
 97 | public, and in some countries other activities as well.
 98 | 
 99 |   To "convey" a work means any kind of propagation that enables other
100 | parties to make or receive copies.  Mere interaction with a user through
101 | a computer network, with no transfer of a copy, is not conveying.
102 | 
103 |   An interactive user interface displays "Appropriate Legal Notices"
104 | to the extent that it includes a convenient and prominently visible
105 | feature that (1) displays an appropriate copyright notice, and (2)
106 | tells the user that there is no warranty for the work (except to the
107 | extent that warranties are provided), that licensees may convey the
108 | work under this License, and how to view a copy of this License.  If
109 | the interface presents a list of user commands or options, such as a
110 | menu, a prominent item in the list meets this criterion.
111 | 
112 |   1. Source Code.
113 | 
114 |   The "source code" for a work means the preferred form of the work
115 | for making modifications to it.  "Object code" means any non-source
116 | form of a work.
117 | 
118 |   A "Standard Interface" means an interface that either is an official
119 | standard defined by a recognized standards body, or, in the case of
120 | interfaces specified for a particular programming language, one that
121 | is widely used among developers working in that language.
122 | 
123 |   The "System Libraries" of an executable work include anything, other
124 | than the work as a whole, that (a) is included in the normal form of
125 | packaging a Major Component, but which is not part of that Major
126 | Component, and (b) serves only to enable use of the work with that
127 | Major Component, or to implement a Standard Interface for which an
128 | implementation is available to the public in source code form.  A
129 | "Major Component", in this context, means a major essential component
130 | (kernel, window system, and so on) of the specific operating system
131 | (if any) on which the executable work runs, or a compiler used to
132 | produce the work, or an object code interpreter used to run it.
133 | 
134 |   The "Corresponding Source" for a work in object code form means all
135 | the source code needed to generate, install, and (for an executable
136 | work) run the object code and to modify the work, including scripts to
137 | control those activities.  However, it does not include the work's
138 | System Libraries, or general-purpose tools or generally available free
139 | programs which are used unmodified in performing those activities but
140 | which are not part of the work.  For example, Corresponding Source
141 | includes interface definition files associated with source files for
142 | the work, and the source code for shared libraries and dynamically
143 | linked subprograms that the work is specifically designed to require,
144 | such as by intimate data communication or control flow between those
145 | subprograms and other parts of the work.
146 | 
147 |   The Corresponding Source need not include anything that users
148 | can regenerate automatically from other parts of the Corresponding
149 | Source.
150 | 
151 |   The Corresponding Source for a work in source code form is that
152 | same work.
153 | 
154 |   2. Basic Permissions.
155 | 
156 |   All rights granted under this License are granted for the term of
157 | copyright on the Program, and are irrevocable provided the stated
158 | conditions are met.  This License explicitly affirms your unlimited
159 | permission to run the unmodified Program.  The output from running a
160 | covered work is covered by this License only if the output, given its
161 | content, constitutes a covered work.  This License acknowledges your
162 | rights of fair use or other equivalent, as provided by copyright law.
163 | 
164 |   You may make, run and propagate covered works that you do not
165 | convey, without conditions so long as your license otherwise remains
166 | in force.  You may convey covered works to others for the sole purpose
167 | of having them make modifications exclusively for you, or provide you
168 | with facilities for running those works, provided that you comply with
169 | the terms of this License in conveying all material for which you do
170 | not control copyright.  Those thus making or running the covered works
171 | for you must do so exclusively on your behalf, under your direction
172 | and control, on terms that prohibit them from making any copies of
173 | your copyrighted material outside their relationship with you.
174 | 
175 |   Conveying under any other circumstances is permitted solely under
176 | the conditions stated below.  Sublicensing is not allowed; section 10
177 | makes it unnecessary.
178 | 
179 |   3. Protecting Users' Legal Rights From Anti-Circumvention Law.
180 | 
181 |   No covered work shall be deemed part of an effective technological
182 | measure under any applicable law fulfilling obligations under article
183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or
184 | similar laws prohibiting or restricting circumvention of such
185 | measures.
186 | 
187 |   When you convey a covered work, you waive any legal power to forbid
188 | circumvention of technological measures to the extent such circumvention
189 | is effected by exercising rights under this License with respect to
190 | the covered work, and you disclaim any intention to limit operation or
191 | modification of the work as a means of enforcing, against the work's
192 | users, your or third parties' legal rights to forbid circumvention of
193 | technological measures.
194 | 
195 |   4. Conveying Verbatim Copies.
196 | 
197 |   You may convey verbatim copies of the Program's source code as you
198 | receive it, in any medium, provided that you conspicuously and
199 | appropriately publish on each copy an appropriate copyright notice;
200 | keep intact all notices stating that this License and any
201 | non-permissive terms added in accord with section 7 apply to the code;
202 | keep intact all notices of the absence of any warranty; and give all
203 | recipients a copy of this License along with the Program.
204 | 
205 |   You may charge any price or no price for each copy that you convey,
206 | and you may offer support or warranty protection for a fee.
207 | 
208 |   5. Conveying Modified Source Versions.
209 | 
210 |   You may convey a work based on the Program, or the modifications to
211 | produce it from the Program, in the form of source code under the
212 | terms of section 4, provided that you also meet all of these conditions:
213 | 
214 |     a) The work must carry prominent notices stating that you modified
215 |     it, and giving a relevant date.
216 | 
217 |     b) The work must carry prominent notices stating that it is
218 |     released under this License and any conditions added under section
219 |     7.  This requirement modifies the requirement in section 4 to
220 |     "keep intact all notices".
221 | 
222 |     c) You must license the entire work, as a whole, under this
223 |     License to anyone who comes into possession of a copy.  This
224 |     License will therefore apply, along with any applicable section 7
225 |     additional terms, to the whole of the work, and all its parts,
226 |     regardless of how they are packaged.  This License gives no
227 |     permission to license the work in any other way, but it does not
228 |     invalidate such permission if you have separately received it.
229 | 
230 |     d) If the work has interactive user interfaces, each must display
231 |     Appropriate Legal Notices; however, if the Program has interactive
232 |     interfaces that do not display Appropriate Legal Notices, your
233 |     work need not make them do so.
234 | 
235 |   A compilation of a covered work with other separate and independent
236 | works, which are not by their nature extensions of the covered work,
237 | and which are not combined with it such as to form a larger program,
238 | in or on a volume of a storage or distribution medium, is called an
239 | "aggregate" if the compilation and its resulting copyright are not
240 | used to limit the access or legal rights of the compilation's users
241 | beyond what the individual works permit.  Inclusion of a covered work
242 | in an aggregate does not cause this License to apply to the other
243 | parts of the aggregate.
244 | 
245 |   6. Conveying Non-Source Forms.
246 | 
247 |   You may convey a covered work in object code form under the terms
248 | of sections 4 and 5, provided that you also convey the
249 | machine-readable Corresponding Source under the terms of this License,
250 | in one of these ways:
251 | 
252 |     a) Convey the object code in, or embodied in, a physical product
253 |     (including a physical distribution medium), accompanied by the
254 |     Corresponding Source fixed on a durable physical medium
255 |     customarily used for software interchange.
256 | 
257 |     b) Convey the object code in, or embodied in, a physical product
258 |     (including a physical distribution medium), accompanied by a
259 |     written offer, valid for at least three years and valid for as
260 |     long as you offer spare parts or customer support for that product
261 |     model, to give anyone who possesses the object code either (1) a
262 |     copy of the Corresponding Source for all the software in the
263 |     product that is covered by this License, on a durable physical
264 |     medium customarily used for software interchange, for a price no
265 |     more than your reasonable cost of physically performing this
266 |     conveying of source, or (2) access to copy the
267 |     Corresponding Source from a network server at no charge.
268 | 
269 |     c) Convey individual copies of the object code with a copy of the
270 |     written offer to provide the Corresponding Source.  This
271 |     alternative is allowed only occasionally and noncommercially, and
272 |     only if you received the object code with such an offer, in accord
273 |     with subsection 6b.
274 | 
275 |     d) Convey the object code by offering access from a designated
276 |     place (gratis or for a charge), and offer equivalent access to the
277 |     Corresponding Source in the same way through the same place at no
278 |     further charge.  You need not require recipients to copy the
279 |     Corresponding Source along with the object code.  If the place to
280 |     copy the object code is a network server, the Corresponding Source
281 |     may be on a different server (operated by you or a third party)
282 |     that supports equivalent copying facilities, provided you maintain
283 |     clear directions next to the object code saying where to find the
284 |     Corresponding Source.  Regardless of what server hosts the
285 |     Corresponding Source, you remain obligated to ensure that it is
286 |     available for as long as needed to satisfy these requirements.
287 | 
288 |     e) Convey the object code using peer-to-peer transmission, provided
289 |     you inform other peers where the object code and Corresponding
290 |     Source of the work are being offered to the general public at no
291 |     charge under subsection 6d.
292 | 
293 |   A separable portion of the object code, whose source code is excluded
294 | from the Corresponding Source as a System Library, need not be
295 | included in conveying the object code work.
296 | 
297 |   A "User Product" is either (1) a "consumer product", which means any
298 | tangible personal property which is normally used for personal, family,
299 | or household purposes, or (2) anything designed or sold for incorporation
300 | into a dwelling.  In determining whether a product is a consumer product,
301 | doubtful cases shall be resolved in favor of coverage.  For a particular
302 | product received by a particular user, "normally used" refers to a
303 | typical or common use of that class of product, regardless of the status
304 | of the particular user or of the way in which the particular user
305 | actually uses, or expects or is expected to use, the product.  A product
306 | is a consumer product regardless of whether the product has substantial
307 | commercial, industrial or non-consumer uses, unless such uses represent
308 | the only significant mode of use of the product.
309 | 
310 |   "Installation Information" for a User Product means any methods,
311 | procedures, authorization keys, or other information required to install
312 | and execute modified versions of a covered work in that User Product from
313 | a modified version of its Corresponding Source.  The information must
314 | suffice to ensure that the continued functioning of the modified object
315 | code is in no case prevented or interfered with solely because
316 | modification has been made.
317 | 
318 |   If you convey an object code work under this section in, or with, or
319 | specifically for use in, a User Product, and the conveying occurs as
320 | part of a transaction in which the right of possession and use of the
321 | User Product is transferred to the recipient in perpetuity or for a
322 | fixed term (regardless of how the transaction is characterized), the
323 | Corresponding Source conveyed under this section must be accompanied
324 | by the Installation Information.  But this requirement does not apply
325 | if neither you nor any third party retains the ability to install
326 | modified object code on the User Product (for example, the work has
327 | been installed in ROM).
328 | 
329 |   The requirement to provide Installation Information does not include a
330 | requirement to continue to provide support service, warranty, or updates
331 | for a work that has been modified or installed by the recipient, or for
332 | the User Product in which it has been modified or installed.  Access to a
333 | network may be denied when the modification itself materially and
334 | adversely affects the operation of the network or violates the rules and
335 | protocols for communication across the network.
336 | 
337 |   Corresponding Source conveyed, and Installation Information provided,
338 | in accord with this section must be in a format that is publicly
339 | documented (and with an implementation available to the public in
340 | source code form), and must require no special password or key for
341 | unpacking, reading or copying.
342 | 
343 |   7. Additional Terms.
344 | 
345 |   "Additional permissions" are terms that supplement the terms of this
346 | License by making exceptions from one or more of its conditions.
347 | Additional permissions that are applicable to the entire Program shall
348 | be treated as though they were included in this License, to the extent
349 | that they are valid under applicable law.  If additional permissions
350 | apply only to part of the Program, that part may be used separately
351 | under those permissions, but the entire Program remains governed by
352 | this License without regard to the additional permissions.
353 | 
354 |   When you convey a copy of a covered work, you may at your option
355 | remove any additional permissions from that copy, or from any part of
356 | it.  (Additional permissions may be written to require their own
357 | removal in certain cases when you modify the work.)  You may place
358 | additional permissions on material, added by you to a covered work,
359 | for which you have or can give appropriate copyright permission.
360 | 
361 |   Notwithstanding any other provision of this License, for material you
362 | add to a covered work, you may (if authorized by the copyright holders of
363 | that material) supplement the terms of this License with terms:
364 | 
365 |     a) Disclaiming warranty or limiting liability differently from the
366 |     terms of sections 15 and 16 of this License; or
367 | 
368 |     b) Requiring preservation of specified reasonable legal notices or
369 |     author attributions in that material or in the Appropriate Legal
370 |     Notices displayed by works containing it; or
371 | 
372 |     c) Prohibiting misrepresentation of the origin of that material, or
373 |     requiring that modified versions of such material be marked in
374 |     reasonable ways as different from the original version; or
375 | 
376 |     d) Limiting the use for publicity purposes of names of licensors or
377 |     authors of the material; or
378 | 
379 |     e) Declining to grant rights under trademark law for use of some
380 |     trade names, trademarks, or service marks; or
381 | 
382 |     f) Requiring indemnification of licensors and authors of that
383 |     material by anyone who conveys the material (or modified versions of
384 |     it) with contractual assumptions of liability to the recipient, for
385 |     any liability that these contractual assumptions directly impose on
386 |     those licensors and authors.
387 | 
388 |   All other non-permissive additional terms are considered "further
389 | restrictions" within the meaning of section 10.  If the Program as you
390 | received it, or any part of it, contains a notice stating that it is
391 | governed by this License along with a term that is a further
392 | restriction, you may remove that term.  If a license document contains
393 | a further restriction but permits relicensing or conveying under this
394 | License, you may add to a covered work material governed by the terms
395 | of that license document, provided that the further restriction does
396 | not survive such relicensing or conveying.
397 | 
398 |   If you add terms to a covered work in accord with this section, you
399 | must place, in the relevant source files, a statement of the
400 | additional terms that apply to those files, or a notice indicating
401 | where to find the applicable terms.
402 | 
403 |   Additional terms, permissive or non-permissive, may be stated in the
404 | form of a separately written license, or stated as exceptions;
405 | the above requirements apply either way.
406 | 
407 |   8. Termination.
408 | 
409 |   You may not propagate or modify a covered work except as expressly
410 | provided under this License.  Any attempt otherwise to propagate or
411 | modify it is void, and will automatically terminate your rights under
412 | this License (including any patent licenses granted under the third
413 | paragraph of section 11).
414 | 
415 |   However, if you cease all violation of this License, then your
416 | license from a particular copyright holder is reinstated (a)
417 | provisionally, unless and until the copyright holder explicitly and
418 | finally terminates your license, and (b) permanently, if the copyright
419 | holder fails to notify you of the violation by some reasonable means
420 | prior to 60 days after the cessation.
421 | 
422 |   Moreover, your license from a particular copyright holder is
423 | reinstated permanently if the copyright holder notifies you of the
424 | violation by some reasonable means, this is the first time you have
425 | received notice of violation of this License (for any work) from that
426 | copyright holder, and you cure the violation prior to 30 days after
427 | your receipt of the notice.
428 | 
429 |   Termination of your rights under this section does not terminate the
430 | licenses of parties who have received copies or rights from you under
431 | this License.  If your rights have been terminated and not permanently
432 | reinstated, you do not qualify to receive new licenses for the same
433 | material under section 10.
434 | 
435 |   9. Acceptance Not Required for Having Copies.
436 | 
437 |   You are not required to accept this License in order to receive or
438 | run a copy of the Program.  Ancillary propagation of a covered work
439 | occurring solely as a consequence of using peer-to-peer transmission
440 | to receive a copy likewise does not require acceptance.  However,
441 | nothing other than this License grants you permission to propagate or
442 | modify any covered work.  These actions infringe copyright if you do
443 | not accept this License.  Therefore, by modifying or propagating a
444 | covered work, you indicate your acceptance of this License to do so.
445 | 
446 |   10. Automatic Licensing of Downstream Recipients.
447 | 
448 |   Each time you convey a covered work, the recipient automatically
449 | receives a license from the original licensors, to run, modify and
450 | propagate that work, subject to this License.  You are not responsible
451 | for enforcing compliance by third parties with this License.
452 | 
453 |   An "entity transaction" is a transaction transferring control of an
454 | organization, or substantially all assets of one, or subdividing an
455 | organization, or merging organizations.  If propagation of a covered
456 | work results from an entity transaction, each party to that
457 | transaction who receives a copy of the work also receives whatever
458 | licenses to the work the party's predecessor in interest had or could
459 | give under the previous paragraph, plus a right to possession of the
460 | Corresponding Source of the work from the predecessor in interest, if
461 | the predecessor has it or can get it with reasonable efforts.
462 | 
463 |   You may not impose any further restrictions on the exercise of the
464 | rights granted or affirmed under this License.  For example, you may
465 | not impose a license fee, royalty, or other charge for exercise of
466 | rights granted under this License, and you may not initiate litigation
467 | (including a cross-claim or counterclaim in a lawsuit) alleging that
468 | any patent claim is infringed by making, using, selling, offering for
469 | sale, or importing the Program or any portion of it.
470 | 
471 |   11. Patents.
472 | 
473 |   A "contributor" is a copyright holder who authorizes use under this
474 | License of the Program or a work on which the Program is based.  The
475 | work thus licensed is called the contributor's "contributor version".
476 | 
477 |   A contributor's "essential patent claims" are all patent claims
478 | owned or controlled by the contributor, whether already acquired or
479 | hereafter acquired, that would be infringed by some manner, permitted
480 | by this License, of making, using, or selling its contributor version,
481 | but do not include claims that would be infringed only as a
482 | consequence of further modification of the contributor version.  For
483 | purposes of this definition, "control" includes the right to grant
484 | patent sublicenses in a manner consistent with the requirements of
485 | this License.
486 | 
487 |   Each contributor grants you a non-exclusive, worldwide, royalty-free
488 | patent license under the contributor's essential patent claims, to
489 | make, use, sell, offer for sale, import and otherwise run, modify and
490 | propagate the contents of its contributor version.
491 | 
492 |   In the following three paragraphs, a "patent license" is any express
493 | agreement or commitment, however denominated, not to enforce a patent
494 | (such as an express permission to practice a patent or covenant not to
495 | sue for patent infringement).  To "grant" such a patent license to a
496 | party means to make such an agreement or commitment not to enforce a
497 | patent against the party.
498 | 
499 |   If you convey a covered work, knowingly relying on a patent license,
500 | and the Corresponding Source of the work is not available for anyone
501 | to copy, free of charge and under the terms of this License, through a
502 | publicly available network server or other readily accessible means,
503 | then you must either (1) cause the Corresponding Source to be so
504 | available, or (2) arrange to deprive yourself of the benefit of the
505 | patent license for this particular work, or (3) arrange, in a manner
506 | consistent with the requirements of this License, to extend the patent
507 | license to downstream recipients.  "Knowingly relying" means you have
508 | actual knowledge that, but for the patent license, your conveying the
509 | covered work in a country, or your recipient's use of the covered work
510 | in a country, would infringe one or more identifiable patents in that
511 | country that you have reason to believe are valid.
512 | 
513 |   If, pursuant to or in connection with a single transaction or
514 | arrangement, you convey, or propagate by procuring conveyance of, a
515 | covered work, and grant a patent license to some of the parties
516 | receiving the covered work authorizing them to use, propagate, modify
517 | or convey a specific copy of the covered work, then the patent license
518 | you grant is automatically extended to all recipients of the covered
519 | work and works based on it.
520 | 
521 |   A patent license is "discriminatory" if it does not include within
522 | the scope of its coverage, prohibits the exercise of, or is
523 | conditioned on the non-exercise of one or more of the rights that are
524 | specifically granted under this License.  You may not convey a covered
525 | work if you are a party to an arrangement with a third party that is
526 | in the business of distributing software, under which you make payment
527 | to the third party based on the extent of your activity of conveying
528 | the work, and under which the third party grants, to any of the
529 | parties who would receive the covered work from you, a discriminatory
530 | patent license (a) in connection with copies of the covered work
531 | conveyed by you (or copies made from those copies), or (b) primarily
532 | for and in connection with specific products or compilations that
533 | contain the covered work, unless you entered into that arrangement,
534 | or that patent license was granted, prior to 28 March 2007.
535 | 
536 |   Nothing in this License shall be construed as excluding or limiting
537 | any implied license or other defenses to infringement that may
538 | otherwise be available to you under applicable patent law.
539 | 
540 |   12. No Surrender of Others' Freedom.
541 | 
542 |   If conditions are imposed on you (whether by court order, agreement or
543 | otherwise) that contradict the conditions of this License, they do not
544 | excuse you from the conditions of this License.  If you cannot convey a
545 | covered work so as to satisfy simultaneously your obligations under this
546 | License and any other pertinent obligations, then as a consequence you may
547 | not convey it at all.  For example, if you agree to terms that obligate you
548 | to collect a royalty for further conveying from those to whom you convey
549 | the Program, the only way you could satisfy both those terms and this
550 | License would be to refrain entirely from conveying the Program.
551 | 
552 |   13. Use with the GNU Affero General Public License.
553 | 
554 |   Notwithstanding any other provision of this License, you have
555 | permission to link or combine any covered work with a work licensed
556 | under version 3 of the GNU Affero General Public License into a single
557 | combined work, and to convey the resulting work.  The terms of this
558 | License will continue to apply to the part which is the covered work,
559 | but the special requirements of the GNU Affero General Public License,
560 | section 13, concerning interaction through a network will apply to the
561 | combination as such.
562 | 
563 |   14. Revised Versions of this License.
564 | 
565 |   The Free Software Foundation may publish revised and/or new versions of
566 | the GNU General Public License from time to time.  Such new versions will
567 | be similar in spirit to the present version, but may differ in detail to
568 | address new problems or concerns.
569 | 
570 |   Each version is given a distinguishing version number.  If the
571 | Program specifies that a certain numbered version of the GNU General
572 | Public License "or any later version" applies to it, you have the
573 | option of following the terms and conditions either of that numbered
574 | version or of any later version published by the Free Software
575 | Foundation.  If the Program does not specify a version number of the
576 | GNU General Public License, you may choose any version ever published
577 | by the Free Software Foundation.
578 | 
579 |   If the Program specifies that a proxy can decide which future
580 | versions of the GNU General Public License can be used, that proxy's
581 | public statement of acceptance of a version permanently authorizes you
582 | to choose that version for the Program.
583 | 
584 |   Later license versions may give you additional or different
585 | permissions.  However, no additional obligations are imposed on any
586 | author or copyright holder as a result of your choosing to follow a
587 | later version.
588 | 
589 |   15. Disclaimer of Warranty.
590 | 
591 |   THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592 | APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596 | PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597 | IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599 | 
600 |   16. Limitation of Liability.
601 | 
602 |   IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610 | SUCH DAMAGES.
611 | 
612 |   17. Interpretation of Sections 15 and 16.
613 | 
614 |   If the disclaimer of warranty and limitation of liability provided
615 | above cannot be given local legal effect according to their terms,
616 | reviewing courts shall apply local law that most closely approximates
617 | an absolute waiver of all civil liability in connection with the
618 | Program, unless a warranty or assumption of liability accompanies a
619 | copy of the Program in return for a fee.
620 | 
621 |                      END OF TERMS AND CONDITIONS
622 | 
623 |             How to Apply These Terms to Your New Programs
624 | 
625 |   If you develop a new program, and you want it to be of the greatest
626 | possible use to the public, the best way to achieve this is to make it
627 | free software which everyone can redistribute and change under these terms.
628 | 
629 |   To do so, attach the following notices to the program.  It is safest
630 | to attach them to the start of each source file to most effectively
631 | state the exclusion of warranty; and each file should have at least
632 | the "copyright" line and a pointer to where the full notice is found.
633 | 
634 |     <one line to give the program's name and a brief idea of what it does.>
635 |     Copyright (C) <year>  <name of author>
636 | 
637 |     This program is free software: you can redistribute it and/or modify
638 |     it under the terms of the GNU General Public License as published by
639 |     the Free Software Foundation, either version 3 of the License, or
640 |     (at your option) any later version.
641 | 
642 |     This program is distributed in the hope that it will be useful,
643 |     but WITHOUT ANY WARRANTY; without even the implied warranty of
644 |     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
645 |     GNU General Public License for more details.
646 | 
647 |     You should have received a copy of the GNU General Public License
648 |     along with this program.  If not, see <https://www.gnu.org/licenses/>.
649 | 
650 | Also add information on how to contact you by electronic and paper mail.
651 | 
652 |   If the program does terminal interaction, make it output a short
653 | notice like this when it starts in an interactive mode:
654 | 
655 |     <program>  Copyright (C) <year>  <name of author>
656 |     This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
657 |     This is free software, and you are welcome to redistribute it
658 |     under certain conditions; type `show c' for details.
659 | 
660 | The hypothetical commands `show w' and `show c' should show the appropriate
661 | parts of the General Public License.  Of course, your program's commands
662 | might be different; for a GUI interface, you would use an "about box".
663 | 
664 |   You should also get your employer (if you work as a programmer) or school,
665 | if any, to sign a "copyright disclaimer" for the program, if necessary.
666 | For more information on this, and how to apply and follow the GNU GPL, see
667 | <https://www.gnu.org/licenses/>.
668 | 
669 |   The GNU General Public License does not permit incorporating your program
670 | into proprietary programs.  If your program is a subroutine library, you
671 | may consider it more useful to permit linking proprietary applications with
672 | the library.  If this is what you want to do, use the GNU Lesser General
673 | Public License instead of this License.  But first, please read
674 | <https://www.gnu.org/licenses/why-not-lgpl.html>.
675 | 


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