| 1 | /* VerifyThis 2016 - Challenge 1: Matrix Multiplication
|
|---|
| 2 | * Consider the following pseudocode algorithm, which is naive
|
|---|
| 3 | * implementation of matrix multiplication. For simplicity we assume that
|
|---|
| 4 | * the matrices are square.
|
|---|
| 5 |
|
|---|
| 6 | int[][] matrixMultiply(int[][] A, int[][] B) {
|
|---|
| 7 | int n = A.length;
|
|---|
| 8 |
|
|---|
| 9 | // initialise C
|
|---|
| 10 | int[][] C = new int[n][n];
|
|---|
| 11 |
|
|---|
| 12 | for (int i = 0; i < n; i++) {
|
|---|
| 13 | for (int k = 0; k < n; k++) {
|
|---|
| 14 | for (int j = 0; j < n; j++) {
|
|---|
| 15 | C[i][j] += A[i][k] * B[k][j];
|
|---|
| 16 | }
|
|---|
| 17 | }
|
|---|
| 18 | }
|
|---|
| 19 | return C;
|
|---|
| 20 | }
|
|---|
| 21 |
|
|---|
| 22 | * Tasks.
|
|---|
| 23 | * (1) Provide a specification to describe the behaviour of this algorithm,
|
|---|
| 24 | * and prove that it correctly implements its specification.
|
|---|
| 25 | * (2) Show that matrix multiplication is associative, i.e., the order in
|
|---|
| 26 | * which matrices are multiplied can be disregarded: A(BC) = (AB)C. To show
|
|---|
| 27 | * this, you should write a program that performs the two different
|
|---|
| 28 | * computations, and then prove that the result of the two computations is
|
|---|
| 29 | * always the same.
|
|---|
| 30 | * (3) In the literature, there exist many proposals for more efficient
|
|---|
| 31 | * matrix multiplication algorithms. Strassen’s algorithm was one of the
|
|---|
| 32 | * first. The key idea of the algorithm is to use a recurisive algorithm
|
|---|
| 33 | * that reduces the number of multiplications on submatrices (from 8 to 7),
|
|---|
| 34 | * see https://en.wikipedia.org/wiki/Strassen_algorithm for an explanation.
|
|---|
| 35 | * A relatively clean Java implementation (and Python and C++) can be found
|
|---|
| 36 | * here: https://martin-thoma.com/strassen-algorithm-in-python-java-cpp/.
|
|---|
| 37 | * Prove that the naive algorithm above has the same behaviour as Strassen’s
|
|---|
| 38 | * algorithm. Proving it for a restricted case, like a 2x2 matrix should be
|
|---|
| 39 | * straightforward, the challenge is to prove it for arbitrary matrices with
|
|---|
| 40 | * size 2^n.
|
|---|
| 41 | *
|
|---|
| 42 | * Author: Stephen Siegel
|
|---|
| 43 | */
|
|---|
| 44 |
|
|---|
| 45 | #include <civlc.cvh>
|
|---|
| 46 | #include <stdio.h>
|
|---|
| 47 | #include <pointer.cvh>
|
|---|
| 48 |
|
|---|
| 49 | // upper bound on N, the size of the matrices
|
|---|
| 50 | $input int BOUND = 4; // can go up to 16 if you have a few minutes
|
|---|
| 51 | $assume(BOUND >= 1);
|
|---|
| 52 | $input int N=4; // the size of the matrices
|
|---|
| 53 | $assume(1<=N && N<=BOUND);
|
|---|
| 54 | // some arbitrary input matrices...
|
|---|
| 55 | $input float A0[N][N];
|
|---|
| 56 | $input float B0[N][N];
|
|---|
| 57 | $input float C0[N][N];
|
|---|
| 58 |
|
|---|
| 59 | // the "leaf size" for Strassen...
|
|---|
| 60 | $input int LEAF_SIZE;
|
|---|
| 61 | $assume (2 <= LEAF_SIZE && LEAF_SIZE <= N);
|
|---|
| 62 |
|
|---|
| 63 | /* Part 1 */
|
|---|
| 64 |
|
|---|
| 65 | // impl: C is "out" variable
|
|---|
| 66 | void matrixMultiply(int n, float C[][], float A[][], float B[][]) {
|
|---|
| 67 | for (int i=0; i<n; i++)
|
|---|
| 68 | for (int j=0; j<n; j++)
|
|---|
| 69 | C[i][j] = 0.0;
|
|---|
| 70 |
|
|---|
| 71 | for (int i = 0; i < n; i++) {
|
|---|
| 72 | for (int k = 0; k < n; k++) {
|
|---|
| 73 | for (int j = 0; j < n; j++) {
|
|---|
| 74 | C[i][j] += A[i][k] * B[k][j];
|
|---|
| 75 | }
|
|---|
| 76 | }
|
|---|
| 77 | }
|
|---|
| 78 | }
|
|---|
| 79 |
|
|---|
| 80 | // Unfortunately no easy way to specify the sum...
|
|---|
| 81 |
|
|---|
| 82 | // Some "helpers" for verification...
|
|---|
| 83 |
|
|---|
| 84 | // Computes dot product of two vectors of length n...
|
|---|
| 85 | float dot(int n, float u[], float v[]) {
|
|---|
| 86 | float sum = 0;
|
|---|
| 87 |
|
|---|
| 88 | for (int i=0; i<n; i++)
|
|---|
| 89 | sum += u[i]*v[i];
|
|---|
| 90 | return sum;
|
|---|
| 91 | }
|
|---|
| 92 |
|
|---|
| 93 | // gets the index-th column of matrix mat, putting it in
|
|---|
| 94 | // contiguous memory starting from result. Returns
|
|---|
| 95 | // pointer to element 0 of result.
|
|---|
| 96 | float * column(int n, float result[], float mat[][], int index) {
|
|---|
| 97 | for (int i=0; i<n; i++)
|
|---|
| 98 | result[i] = mat[i][index];
|
|---|
| 99 | return &result[0];
|
|---|
| 100 | }
|
|---|
| 101 |
|
|---|
| 102 |
|
|---|
| 103 | // checks multiplication is correct: entry i,j should
|
|---|
| 104 | // be the dot product of i-th row of A and j-th column of B...
|
|---|
| 105 | void testMult(int n) {
|
|---|
| 106 | float actual[n][n];
|
|---|
| 107 | float tmp[n];
|
|---|
| 108 |
|
|---|
| 109 | matrixMultiply(n, actual, A0, B0);
|
|---|
| 110 | for (int i=0; i<n; i++)
|
|---|
| 111 | for (int j=0; j<n; j++)
|
|---|
| 112 | $assert(dot(n, A0[i], column(n, tmp, B0, j)) == actual[i][j]);
|
|---|
| 113 | }
|
|---|
| 114 |
|
|---|
| 115 |
|
|---|
| 116 |
|
|---|
| 117 | /* Part 2 */
|
|---|
| 118 |
|
|---|
| 119 | // tests associativity: (A0*B0)*C0 = A0*(B0*C0)
|
|---|
| 120 | void assoc(int n) {
|
|---|
| 121 | float T1[n][n], T2[n][n], R1[n][n], R2[n][n];
|
|---|
| 122 |
|
|---|
| 123 | matrixMultiply(n, T1, A0, B0);
|
|---|
| 124 | matrixMultiply(n, R1, T1, C0);
|
|---|
| 125 | matrixMultiply(n, T2, B0, C0);
|
|---|
| 126 | matrixMultiply(n, R2, A0, T2);
|
|---|
| 127 | $assert($equals(&R1, &R2));
|
|---|
| 128 | }
|
|---|
| 129 |
|
|---|
| 130 |
|
|---|
| 131 | /* Part 3 : Strassen */
|
|---|
| 132 |
|
|---|
| 133 | // adds two nxn matrices. C is "out" variable.
|
|---|
| 134 | void add(int n, float C[][], float A[][], float B[][]) {
|
|---|
| 135 | for (int i = 0; i < n; i++)
|
|---|
| 136 | for (int j = 0; j < n; j++)
|
|---|
| 137 | C[i][j] = A[i][j] + B[i][j];
|
|---|
| 138 | }
|
|---|
| 139 |
|
|---|
| 140 | // subtracts two nxn matrices. C is "out" variable.
|
|---|
| 141 | void subtract(int n, float C[][], float A[][], float B[][]) {
|
|---|
| 142 | for (int i = 0; i < n; i++)
|
|---|
| 143 | for (int j = 0; j < n; j++)
|
|---|
| 144 | C[i][j] = A[i][j] - B[i][j];
|
|---|
| 145 | }
|
|---|
| 146 |
|
|---|
| 147 |
|
|---|
| 148 | // Strassen algorithm from
|
|---|
| 149 | // https://martin-thoma.com/strassen-algorithm-in-python-java-cpp/
|
|---|
| 150 | // I'm just going to assume n is a power of 2.
|
|---|
| 151 | // There is no problem dealing with the general case but need more
|
|---|
| 152 | // time!
|
|---|
| 153 |
|
|---|
| 154 | // multiplies two nxn matrices, storing result in C
|
|---|
| 155 | void strassenR(int n, float C[][], float A[][], float B[][]) {
|
|---|
| 156 | if (n <= LEAF_SIZE) {
|
|---|
| 157 | matrixMultiply(n, C, A, B);
|
|---|
| 158 | } else {
|
|---|
| 159 | // initializing the new sub-matrices
|
|---|
| 160 | int newSize = n / 2;
|
|---|
| 161 | float a11[newSize][newSize];
|
|---|
| 162 | float a12[newSize][newSize];
|
|---|
| 163 | float a21[newSize][newSize];
|
|---|
| 164 | float a22[newSize][newSize];
|
|---|
| 165 |
|
|---|
| 166 | float b11[newSize][newSize];
|
|---|
| 167 | float b12[newSize][newSize];
|
|---|
| 168 | float b21[newSize][newSize];
|
|---|
| 169 | float b22[newSize][newSize];
|
|---|
| 170 |
|
|---|
| 171 | float aResult[newSize][newSize];
|
|---|
| 172 | float bResult[newSize][newSize];
|
|---|
| 173 |
|
|---|
| 174 | // dividing the matrices in 4 sub-matrices:
|
|---|
| 175 | for (int i = 0; i < newSize; i++) {
|
|---|
| 176 | for (int j = 0; j < newSize; j++) {
|
|---|
| 177 | a11[i][j] = A[i][j]; // top left
|
|---|
| 178 | a12[i][j] = A[i][j + newSize]; // top right
|
|---|
| 179 | a21[i][j] = A[i + newSize][j]; // bottom left
|
|---|
| 180 | a22[i][j] = A[i + newSize][j + newSize]; // bottom right
|
|---|
| 181 |
|
|---|
| 182 | b11[i][j] = B[i][j]; // top left
|
|---|
| 183 | b12[i][j] = B[i][j + newSize]; // top right
|
|---|
| 184 | b21[i][j] = B[i + newSize][j]; // bottom left
|
|---|
| 185 | b22[i][j] = B[i + newSize][j + newSize]; // bottom right
|
|---|
| 186 | }
|
|---|
| 187 | }
|
|---|
| 188 | // Calculating p1 to p7:
|
|---|
| 189 | add(newSize, aResult, a11, a22);
|
|---|
| 190 | add(newSize, bResult, b11, b22);
|
|---|
| 191 | float p1[newSize][newSize];
|
|---|
| 192 | strassenR(newSize, p1, aResult, bResult);
|
|---|
| 193 | // p1 = (a11+a22) * (b11+b22)
|
|---|
| 194 |
|
|---|
| 195 | add(newSize, aResult, a21, a22); // a21 + a22
|
|---|
| 196 | float p2[newSize][newSize];
|
|---|
| 197 | strassenR(newSize, p2, aResult, b11); // p2 = (a21+a22) * (b11)
|
|---|
| 198 |
|
|---|
| 199 | subtract(newSize, bResult, b12, b22); // b12 - b22
|
|---|
| 200 | float p3[newSize][newSize];
|
|---|
| 201 | strassenR(newSize, p3, a11, bResult);
|
|---|
| 202 | // p3 = (a11) * (b12 - b22)
|
|---|
| 203 |
|
|---|
| 204 | subtract(newSize, bResult, b21, b11); // b21 - b11
|
|---|
| 205 | float p4[newSize][newSize];
|
|---|
| 206 | strassenR(newSize, p4, a22, bResult);
|
|---|
| 207 | // p4 = (a22) * (b21 - b11)
|
|---|
| 208 |
|
|---|
| 209 | add(newSize, aResult, a11, a12); // a11 + a12
|
|---|
| 210 | float p5[newSize][newSize];
|
|---|
| 211 | strassenR(newSize, p5, aResult, b22);
|
|---|
| 212 | // p5 = (a11+a12) * (b22)
|
|---|
| 213 |
|
|---|
| 214 | subtract(newSize, aResult, a21, a11); // a21 - a11
|
|---|
| 215 | add(newSize, bResult, b11, b12); // b11 + b12
|
|---|
| 216 | float p6[newSize][newSize];
|
|---|
| 217 | strassenR(newSize, p6, aResult, bResult);
|
|---|
| 218 | // p6 = (a21-a11) * (b11+b12)
|
|---|
| 219 |
|
|---|
| 220 | subtract(newSize, aResult, a12, a22); // a12 - a22
|
|---|
| 221 | add(newSize, bResult, b21, b22); // b21 + b22
|
|---|
| 222 | float p7[newSize][newSize];
|
|---|
| 223 | strassenR(newSize, p7, aResult, bResult);
|
|---|
| 224 | // p7 = (a12-a22) * (b21+b22)
|
|---|
| 225 |
|
|---|
| 226 | // calculating c21, c21, c11 e c22:
|
|---|
| 227 | float c12[newSize][newSize];
|
|---|
| 228 | add(newSize, c12, p3, p5); // c12 = p3 + p5
|
|---|
| 229 | float c21[newSize][newSize];
|
|---|
| 230 | add(newSize, c21, p2, p4); // c21 = p2 + p4
|
|---|
| 231 |
|
|---|
| 232 | add(newSize, aResult, p1, p4); // p1 + p4
|
|---|
| 233 | add(newSize, bResult, aResult, p7); // p1 + p4 + p7
|
|---|
| 234 | float c11[newSize][newSize];
|
|---|
| 235 | subtract(newSize, c11, bResult, p5);
|
|---|
| 236 | // c11 = p1 + p4 - p5 + p7
|
|---|
| 237 |
|
|---|
| 238 | add(newSize, aResult, p1, p3); // p1 + p3
|
|---|
| 239 | add(newSize, bResult, aResult, p6); // p1 + p3 + p6
|
|---|
| 240 | float c22[newSize][newSize];
|
|---|
| 241 | subtract(newSize, c22, bResult, p2);
|
|---|
| 242 | // c22 = p1 + p3 - p2 + p6
|
|---|
| 243 |
|
|---|
| 244 | // Grouping the results obtained in a single matrix:
|
|---|
| 245 | for (int i = 0; i < newSize; i++) {
|
|---|
| 246 | for (int j = 0; j < newSize; j++) {
|
|---|
| 247 | C[i][j] = c11[i][j];
|
|---|
| 248 | C[i][j + newSize] = c12[i][j];
|
|---|
| 249 | C[i + newSize][j] = c21[i][j];
|
|---|
| 250 | C[i + newSize][j + newSize] = c22[i][j];
|
|---|
| 251 | }
|
|---|
| 252 | }
|
|---|
| 253 | }
|
|---|
| 254 | }
|
|---|
| 255 |
|
|---|
| 256 | // test Strassen multiplication agrees with the regular one...
|
|---|
| 257 | void testStrassen(int n) {
|
|---|
| 258 | float R1[n][n], R2[n][n];
|
|---|
| 259 |
|
|---|
| 260 | matrixMultiply(n, R1, A0, B0);
|
|---|
| 261 | strassenR(n, R2, A0, B0);
|
|---|
| 262 | $assert($equals(&R1, &R2));
|
|---|
| 263 | }
|
|---|
| 264 |
|
|---|
| 265 | // determines whether n is a power of 2
|
|---|
| 266 | _Bool isPowerOf2(int n) {
|
|---|
| 267 | while (n>1) {
|
|---|
| 268 | if (n%2 != 0)
|
|---|
| 269 | return $false;
|
|---|
| 270 | n = n/2;
|
|---|
| 271 | }
|
|---|
| 272 | return $true;
|
|---|
| 273 | }
|
|---|
| 274 |
|
|---|
| 275 | /* main: runs the three tests */
|
|---|
| 276 | int main() {
|
|---|
| 277 | //$elaborate(N); // hint to verifier
|
|---|
| 278 | printf("N=%d\n", N);
|
|---|
| 279 | //testMult(N);
|
|---|
| 280 | //assoc(N);
|
|---|
| 281 | //$assume(isPowerOf2(N));
|
|---|
| 282 | testStrassen(N);
|
|---|
| 283 | }
|
|---|