Add large matrix benchmarks, refactor tests

Refeactos the codebase to split the tests and benchmarks into separate
files. Adds some large matrix multiplication benchmarks for the serial
and parallel implementations.

implementation/matmul-c/Makefile

@@ -27,6 +27,8 @@ matmul-c: \
 	src/operations.c \
 	src/parse_header.c \
 	src/threadpool.c \
+	src/utils.c \
+	src/benchmarks.c \
 	src/tests.c
 	$(CC) $(CFLAGS) -o$@ $^ $(CLIBS)
 

implementation/matmul-c/include/benchmarks.h

@@ -0,0 +1,24 @@
+/* vim: noet:ts=2:sts=2:sw=2 */ 
+
+/* SPDX-License-Identifier: MIT */
+/* Copyright © 2024 David Llewellyn-Jones */
+
+#include <stdbool.h>
+#include <stdio.h>
+
+#ifndef __MATRIX_BENCHMARKS_H
+#define __MATRIX_BENCHMARKS_H (1)
+
+typedef struct _Benchmark Benchmark;
+
+Benchmark * new_benchmark();
+Benchmark * delete_benchmark(Benchmark *benchmark);
+
+void benchmarks_start(Benchmark *benchmark, uint32_t operations);
+void benchmarks_end(Benchmark *benchmark);
+void benchmark_set_quiet(Benchmark *benchmark, bool quiet);
+void benchmarks_multiply_big(ThreadPool *pool);
+void benchmarks_multiply_small(Matrices *a, Matrices *b, Matrices *d);
+
+#endif /* __MATRIX_BENCHMARKS_H */
+

implementation/matmul-c/include/load.h

@@ -23,4 +23,4 @@ void matrix_npz_load(char *filename, Matrices *matrices);
 Matrices *new_matrices(uint32_t count);
 Matrices *delete_matrices(Matrices *matrices);
 
-#endif /* __MATRIX_OPERATIONS_H */
+#endif /* __MATRIX_LOAD_H */

implementation/matmul-c/include/tests.h

@@ -6,6 +6,10 @@
 #include <stdbool.h>
 #include <stdio.h>
 
+#include "matrix.h"
+#include "load.h"
+#include "threadpool.h"
+
 #include "load.h"
 
 #ifndef __MATRIX_TESTS_H
@@ -14,5 +18,7 @@
 uint32_t tests_load_matrices(Matrices *a, Matrices *b, Matrices *c);
 bool tests_allocate_results(Matrices *c, Matrices *d);
 
+void tests_compare(Matrices *a, Matrices *b, Matrices *c, Matrices *d, ThreadPool *pool);
+
 #endif /* __MATRIX_TESTS_H */
 

implementation/matmul-c/include/utils.h

@@ -0,0 +1,19 @@
+/* vim: noet:ts=2:sts=2:sw=2 */ 
+
+/* SPDX-License-Identifier: MIT */
+/* Copyright © 2024 David Llewellyn-Jones */
+
+#ifndef __MATRIX_UTILS_H
+#define __MATRIX_UTILS_H (1)
+
+#include <stdint.h>
+
+typedef struct _Rand Rand;
+
+Rand * new_rand();
+Rand * delete_rand(Rand *rand);
+void rand_seed(Rand *rand, uint32_t seed);
+double rand_next(Rand *rand);
+double rand_digit(Rand *rand);
+
+#endif /* __MATRIX_UTILS_H */

implementation/matmul-c/main.c

@@ -10,12 +10,7 @@
 #include "load.h"
 #include "tests.h"
 #include "threadpool.h"
-
-#define BENCHMARK_REPEAT_SMALL (32768)
-#define BENCHMARK_REPEAT_LARGE (1)
-
-#define HEIGHT (2048)
-#define WIDTH (2048)
+#include "benchmarks.h"
 
 int main(int argc, char *argv[]) {
 	Matrix *A;
@@ -58,93 +53,13 @@ int main(int argc, char *argv[]) {
 	result = tests_allocate_results(c, d);
 
 	// Perform 512 multiplications and compare against the results from NumPy
-	printf("Performing unit tests...\n");
-	uint32_t passed = 0;
-	for (uint32_t index = 0; index < total; ++index) {
-		result = multiply(d->matrices[index].matrix, a->matrices[index].matrix, b->matrices[index].matrix);
-		result = result && equals(c->matrices[index].matrix, d->matrices[index].matrix);
-		if (result) {
-			passed += 1;
-		}
-		else {
-			printf("Incorrect result\n");
-			matrix_print(c->matrices[index].matrix);
-			matrix_print(d->matrices[index].matrix);
-		}
-	}
-	for (uint32_t index = 0; index < total; ++index) {
-		result = multiply_parallel(pool, d->matrices[index].matrix, a->matrices[index].matrix, b->matrices[index].matrix);
-		result = result && equals(c->matrices[index].matrix, d->matrices[index].matrix);
-		if (result) {
-			passed += 1;
-		}
-		else {
-			printf("Incorrect result\n");
-			matrix_print(c->matrices[index].matrix);
-			matrix_print(d->matrices[index].matrix);
-		}
-	}
-	printf("Multiplication tests passed: %u out of %u\n", passed, total * 2);
-
-	// Measure time taken to perform 16777216 multiplications
-	struct timespec start_time, end_time;
-	uint32_t operations;
-	double elapsed;
-	double ops_per_sec;
-
-	// Create a pair of big random matrices
-	A = new_matrix(HEIGHT, WIDTH);
-	B = new_matrix(WIDTH, HEIGHT);
-	D = new_matrix(HEIGHT, HEIGHT);
-	matrix_fill(A, 8);
-	matrix_fill(B, 16);
-	operations = BENCHMARK_REPEAT_LARGE;
-
-	printf("Benchmarking...\n");
-
-	clock_gettime(CLOCK_MONOTONIC, &start_time);
-	for (uint32_t count = 0; count < operations; ++count) {
-		multiply(D, A, B);
-	}
-	clock_gettime(CLOCK_MONOTONIC, &end_time);
-	elapsed = (end_time.tv_sec - start_time.tv_sec);
-	elapsed += (end_time.tv_nsec - start_time.tv_nsec) / 1000000000.0;
-
-	printf("Time taken to perform %u standard large multiply operations: %.02f seconds\n", operations, elapsed);
-	ops_per_sec = operations / elapsed;
-	printf("Equivalent to %.02f operations per second\n", ops_per_sec);
-
-	clock_gettime(CLOCK_MONOTONIC, &start_time);
-	for (uint32_t count = 0; count < operations; ++count) {
-		multiply_parallel(pool, D, A, B);
-	}
-	clock_gettime(CLOCK_MONOTONIC, &end_time);
-	elapsed = (end_time.tv_sec - start_time.tv_sec);
-	elapsed += (end_time.tv_nsec - start_time.tv_nsec) / 1000000000.0;
-
-	printf("Time taken to perform %u standard large parallel multiply operations: %.02f seconds\n", operations, elapsed);
-	ops_per_sec = operations / elapsed;
-	printf("Equivalent to %.02f operations per second\n", ops_per_sec);
+	tests_compare(a, b, c, d, pool);
 
-	A = delete_matrix(A);
-	B = delete_matrix(B);
-	D = delete_matrix(D);
+	// Benchmark large matrix multiplications
+	benchmarks_multiply_big(pool);
 
-	clock_gettime(CLOCK_MONOTONIC, &start_time);
-	for (uint32_t count = 0; count < BENCHMARK_REPEAT_SMALL; ++count) {
-		for (uint32_t index = 0; index < total; ++index) {
-			multiply(d->matrices[index].matrix, a->matrices[index].matrix, b->matrices[index].matrix);
-		}
-	}
-	clock_gettime(CLOCK_MONOTONIC, &end_time);
-
-	elapsed = (end_time.tv_sec - start_time.tv_sec);
-	elapsed += (end_time.tv_nsec - start_time.tv_nsec) / 1000000000.0;
-
-	operations = total * BENCHMARK_REPEAT_SMALL;
-	printf("Time taken to perform %u multiply operations: %.02f seconds\n", operations, elapsed);
-	ops_per_sec = operations / elapsed;
-	printf("Equivalent to %.02f operations per second\n", ops_per_sec);
+	// Measure time taken to perform 16777216 multiplications
+	benchmarks_multiply_small(a, b, d);
 
 	a = delete_matrices(a);
 	b = delete_matrices(b);

implementation/matmul-c/src/benchmarks.c

@@ -0,0 +1,162 @@
+/* vim: noet:ts=2:sts=2:sw=2 */ 
+
+/* SPDX-License-Identifier: MIT */
+/* Copyright © 2024 David Llewellyn-Jones */
+
+#include <stdlib.h>
+
+#include "matrix.h"
+#include "load.h"
+#include "operations.h"
+#include "threadpool.h"
+
+#include "benchmarks.h"
+
+#define BENCHMARK_REPEAT_SMALL (32768)
+#define BENCHMARK_REPEAT_LARGE (1)
+
+#define HEIGHT (2048)
+#define WIDTH (2048)
+
+struct _Benchmark {
+	struct timespec start_time;
+	struct timespec end_time;
+	uint32_t operations;
+	double elapsed;
+	double ops_per_sec;
+	bool quiet;
+};
+
+Benchmark * new_benchmark() {
+	Benchmark *benchmark = calloc(sizeof(Benchmark), sizeof(char));
+
+	return benchmark;
+}
+
+Benchmark * delete_benchmark(Benchmark *benchmark) {
+	if (benchmark) {
+		free(benchmark);
+	}
+	return NULL;
+}
+
+void benchmark_set_quiet(Benchmark *benchmark, bool quiet) {
+	benchmark->quiet = quiet;
+}
+
+void benchmarks_start(Benchmark *benchmark, uint32_t operations) {
+	if (benchmark) {
+		if (!benchmark->quiet) {
+			printf("Benchmarking...\n");
+		}
+		benchmark->operations = operations;
+		clock_gettime(CLOCK_MONOTONIC, &benchmark->start_time);
+	}
+}
+
+void benchmarks_end(Benchmark *benchmark) {
+	if (benchmark) {
+		clock_gettime(CLOCK_MONOTONIC, &benchmark->end_time);
+		benchmark->elapsed = (benchmark->end_time.tv_sec - benchmark->start_time.tv_sec);
+		benchmark->elapsed += (benchmark->end_time.tv_nsec - benchmark->start_time.tv_nsec) / 1000000000.0;
+		benchmark->ops_per_sec = benchmark->operations / benchmark->elapsed;
+
+		if (!benchmark->quiet) {
+			printf("Time taken to perform %u operations: %.02f seconds\n", benchmark->operations, benchmark->elapsed);
+			printf("Equivalent to %.02f operations per second\n", benchmark->ops_per_sec);
+		}
+	}
+}
+
+void benchmarks_multiply_big(ThreadPool *pool) {
+	Benchmark *benchmark;
+	Matrix *A;
+	Matrix *B;
+	Matrix *D;
+
+	printf("\n");
+	printf("## Large matrix multiplication\n");
+
+	benchmark = new_benchmark();
+	benchmark_set_quiet(benchmark, true);
+
+	for (uint32_t width = 128; width <= 2048; width += 128) {
+		uint32_t diag = width;
+		uint32_t height = width;
+		uint32_t const repeat = BENCHMARK_REPEAT_LARGE * (width < 512 ? 1024 : width < 1024 ? 16 : 1);
+
+		A = new_matrix(width, diag);
+		B = new_matrix(diag, height);
+		D = new_matrix(width, height);
+		matrix_fill(A, 8);
+		matrix_fill(B, 16);
+
+		benchmarks_start(benchmark, repeat);
+		for (uint32_t count = 0; count < repeat; ++count) {
+			multiply(D, A, B);
+		}
+		benchmarks_end(benchmark);
+		printf("Size: (%d, %d, %d), time per operation: %.02f seconds\n", width, diag, height, benchmark->elapsed / benchmark->operations);
+		uint64_t order = (uint64_t)width * (uint64_t)diag * (uint64_t)height;
+		double speed = (double)order / (benchmark->elapsed / benchmark->operations);
+		printf("Order: %llu, elements per second: %.02f\n", order, speed);
+
+		A = delete_matrix(A);
+		B = delete_matrix(B);
+		D = delete_matrix(D);
+	}
+
+	printf("\n");
+	printf("## Large parallel matrix multiplication\n");
+
+	for (uint32_t width = 128; width <= 2048; width += 128) {
+		uint32_t diag = width;
+		uint32_t height = width;
+		uint32_t const repeat = BENCHMARK_REPEAT_LARGE * (width < 512 ? 1024 : width < 1024 ? 16 : 1);
+
+		A = new_matrix(width, diag);
+		B = new_matrix(diag, height);
+		D = new_matrix(width, height);
+		matrix_fill(A, 8);
+		matrix_fill(B, 16);
+
+		benchmarks_start(benchmark, repeat);
+		for (uint32_t count = 0; count < repeat; ++count) {
+			multiply_parallel(pool, D, A, B);
+		}
+		benchmarks_end(benchmark);
+		printf("Size: (%d, %d, %d), time per operation: %.02f seconds\n", width, diag, height, benchmark->elapsed / benchmark->operations);
+		uint64_t order = (uint64_t)width * (uint64_t)diag * (uint64_t)height;
+		double speed = (double)order / (benchmark->elapsed / benchmark->operations);
+		printf("Order: %llu, elements per second: %.02f\n", order, speed);
+
+		A = delete_matrix(A);
+		B = delete_matrix(B);
+		D = delete_matrix(D);
+	}
+
+	benchmark = delete_benchmark(benchmark);
+}
+
+void benchmarks_multiply_small(Matrices *a, Matrices *b, Matrices *d) {
+	Benchmark *benchmark;
+	uint32_t total;
+
+	benchmark = new_benchmark();
+	total = d->count;
+
+	printf("\n");
+	printf("## Small matrix multiplication\n");
+	benchmarks_start(benchmark, total * BENCHMARK_REPEAT_SMALL);
+
+	for (uint32_t count = 0; count < BENCHMARK_REPEAT_SMALL; ++count) {
+		for (uint32_t index = 0; index < total; ++index) {
+			multiply(d->matrices[index].matrix, a->matrices[index].matrix, b->matrices[index].matrix);
+		}
+	}
+
+	benchmarks_end(benchmark);
+
+	benchmark = delete_benchmark(benchmark);
+}
+

implementation/matmul-c/src/matrix.c

@@ -6,6 +6,8 @@
 #include <stdio.h>
 #include <stdlib.h>
 
+#include "utils.h"
+
 #include "matrix.h"
 
 Matrix * new_matrix(uint16_t height, uint16_t width) {
@@ -57,12 +59,16 @@ void matrix_print(Matrix *A) {
 }
 
 void matrix_fill(Matrix *A, uint32_t seed) {
-	srand(seed);
+	Rand * rand = new_rand();
+	rand_seed(rand, seed);
+
 	if (A) {
 		uint32_t size = A->height * A->width;
 		for (uint32_t index = 0; index < size; ++index) {
-			A->elements[index] = ((uint32_t)(rand() / (RAND_MAX / 1000))) / 10;
+			A->elements[index] = rand_digit(rand);
 		}
 	}
+
+	rand = delete_rand(rand);
 }