diff options
Diffstat (limited to 'test/monniaux/float_mat')
-rw-r--r-- | test/monniaux/float_mat/Makefile | 4 | ||||
-rw-r--r-- | test/monniaux/float_mat/float_mat.c | 228 | ||||
-rw-r--r-- | test/monniaux/float_mat/float_mat.h | 55 | ||||
-rw-r--r-- | test/monniaux/float_mat/float_mat_run.c | 133 |
4 files changed, 420 insertions, 0 deletions
diff --git a/test/monniaux/float_mat/Makefile b/test/monniaux/float_mat/Makefile new file mode 100644 index 00000000..69621159 --- /dev/null +++ b/test/monniaux/float_mat/Makefile @@ -0,0 +1,4 @@ +TARGET=float_mat +MEASURES="c1 c2 c3 c4 c5 c6 c7 c8" + +include ../rules.mk diff --git a/test/monniaux/float_mat/float_mat.c b/test/monniaux/float_mat/float_mat.c new file mode 100644 index 00000000..45612635 --- /dev/null +++ b/test/monniaux/float_mat/float_mat.c @@ -0,0 +1,228 @@ +#include "float_mat.h" +#include <stddef.h> + +#define ADD += +#define MUL * + +void REAL_mat_mul1(unsigned m, unsigned n, unsigned p, + REAL * restrict c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b) { + for(unsigned i=0; i<m; i++) { + for(unsigned k=0; k<p; k++) { + c[i*stride_c+k] = 0; + } + } + for(unsigned i=0; i<m; i++) { + for(unsigned k=0; k<p; k++) { + for(unsigned j=0; j<n; j++) { + c[i*stride_c+k] ADD (a[i*stride_a+j] MUL b[j*stride_b+k]); + } + } + } +} + +void REAL_mat_mul2(unsigned m, unsigned n, unsigned p, + REAL * restrict c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b) { + for(unsigned i=0; i<m; i++) { + for(unsigned k=0; k<p; k++) { + REAL total = 0; + for(unsigned j=0; j<n; j++) { + total ADD (a[i*stride_a + j] MUL b[j*stride_b + k]); + } + c[i*stride_c+k] = total; + } + } +} + +void REAL_mat_mul3(unsigned m, unsigned n, unsigned p, + REAL * restrict c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b) { + for(unsigned i=0; i<m; i++) { + for(unsigned k=0; k<p; k++) { + REAL total0 = 0, total1 = 0; + unsigned j; + for(j=0; j+1<n; j+=2) { + total0 ADD (a[i*stride_a + j] MUL b[j*stride_b + k]); + total1 ADD (a[i*stride_a + (j+1)] MUL b[(j+1)*stride_b + k]); + } + if (j < n) { + total0 ADD a[i*stride_a + j] MUL b[j*stride_b + k]; + } + total0 ADD total1; + c[i*stride_c+k] = total0; + } + } +} + +void REAL_mat_mul4(unsigned m, unsigned n, unsigned p, + REAL * c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b) { + const REAL *pa_i = a; + REAL * pc_i = c; + for(unsigned i=0; i<m; i++) { + for(unsigned k=0; k<p; k++) { + const REAL *pb_j_k = b+k, *pa_i_j = pa_i; + REAL total = 0; + for(unsigned j=0; j<n; j++) { + total ADD (*pa_i_j MUL *pb_j_k); + pa_i_j ++; + pb_j_k += stride_b; + } + pc_i[k] = total; + } + pa_i += stride_a; + pc_i += stride_c; + } +} + +void REAL_mat_mul5(unsigned m, unsigned n, unsigned p, + REAL * c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b) { + const REAL *pa_i = a; + REAL * pc_i = c; + for(unsigned i=0; i<m; i++) { + for(unsigned k=0; k<p; k++) { + const REAL *pb_j_k = b+k, *pa_i_j = pa_i; + REAL total = 0; + for(unsigned j2=0, n2=n/2; j2<n2; j2++) { + REAL p0 = *pa_i_j MUL *pb_j_k; + pa_i_j ++; + pb_j_k += stride_b; + REAL p1 = *pa_i_j MUL *pb_j_k; + pa_i_j ++; + pb_j_k += stride_b; + total ADD p0; + total ADD p1; + } + if (n%2) { + total ADD *pa_i_j MUL *pb_j_k; + } + pc_i[k] = total; + } + pa_i += stride_a; + pc_i += stride_c; + } +} + +#define CHUNK \ + total ADD (*pa_i_j MUL *pb_j_k); \ + pa_i_j ++; \ + pb_j_k += stride_b; + +void REAL_mat_mul6(unsigned m, unsigned n, unsigned p, + REAL * c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b) { + const REAL *pa_i = a; + REAL * pc_i = c; + for(unsigned i=0; i<m; i++) { + for(unsigned k=0; k<p; k++) { + const REAL *pb_j_k = b+k, *pa_i_j = pa_i; + REAL total = 0; + unsigned j2=0, n2=n/2; + if (n2 > 0) { + do { + CHUNK + CHUNK + j2++; + } while (j2 < n2); + } + if (n%2) { + total ADD (*pa_i_j MUL *pb_j_k); + } + pc_i[k] = total; + } + pa_i += stride_a; + pc_i += stride_c; + } +} + +#define UNROLL 4 +void REAL_mat_mul7(unsigned m, unsigned n, unsigned p, + REAL * c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b) { + const REAL *pa_i = a; + REAL * pc_i = c; + for(unsigned i=0; i<m; i++) { + for(unsigned k=0; k<p; k++) { + const REAL *pb_j_k = b+k, *pa_i_j = pa_i; + REAL total = 0; + { + unsigned j4=0, n4=n/UNROLL; + if (n4 > 0) { + do { + CHUNK + CHUNK + CHUNK + CHUNK + j4++; + } while (j4 < n4); + } + } + { + unsigned j4=0, n4=n%UNROLL; + if (n4 > 0) { + do { + CHUNK + j4++; + } while (j4 < n4); + } + } + pc_i[k] = total; + } + pa_i += stride_a; + pc_i += stride_c; + } +} + +#undef CHUNK +#define CHUNK \ + total ADD (*pa_i_j MUL *pb_j_k); \ + pa_i_j ++; \ + pb_j_k = (REAL*) ((char*) pb_j_k + stride_b_scaled); + +void REAL_mat_mul8(unsigned m, unsigned n, unsigned p, + REAL * c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b) { + const REAL *pa_i = a; + REAL * pc_i = c; + size_t stride_b_scaled = sizeof(REAL) * stride_b; + for(unsigned i=0; i<m; i++) { + for(unsigned k=0; k<p; k++) { + const REAL *pb_j_k = b+k, *pa_i_j = pa_i; + REAL total = 0; + { + unsigned j4=0, n4=n/UNROLL; + if (n4 > 0) { + do { + CHUNK + CHUNK + CHUNK + CHUNK + j4++; + } while (j4 < n4); + } + } + { + unsigned j4=0, n4=n%UNROLL; + if (n4 > 0) { + do { + CHUNK + j4++; + } while (j4 < n4); + } + } + pc_i[k] = total; + } + pa_i += stride_a; + pc_i += stride_c; + } +} diff --git a/test/monniaux/float_mat/float_mat.h b/test/monniaux/float_mat/float_mat.h new file mode 100644 index 00000000..3b787b1f --- /dev/null +++ b/test/monniaux/float_mat/float_mat.h @@ -0,0 +1,55 @@ +#include <stdint.h> +#include <stdbool.h> + +typedef double REAL; + +void REAL_mat_mul1(unsigned m, unsigned n, unsigned p, + REAL * restrict c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b); + +void REAL_mat_mul2(unsigned m, unsigned n, unsigned p, + REAL * restrict c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b); + +void REAL_mat_mul3(unsigned m, unsigned n, unsigned p, + REAL * restrict c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b); + +void REAL_mat_mul4(unsigned m, unsigned n, unsigned p, + REAL * restrict c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b); + +void REAL_mat_mul5(unsigned m, unsigned n, unsigned p, + REAL * restrict c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b); + +void REAL_mat_mul6(unsigned m, unsigned n, unsigned p, + REAL * restrict c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b); + +void REAL_mat_mul7(unsigned m, unsigned n, unsigned p, + REAL * restrict c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b); + +void REAL_mat_mul8(unsigned m, unsigned n, unsigned p, + REAL * restrict c, unsigned stride_c, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b); + +REAL REAL_random(void); + +void REAL_mat_random(unsigned m, + unsigned n, + REAL *a, unsigned stride_a); + +bool REAL_mat_equal(unsigned m, + unsigned n, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b); diff --git a/test/monniaux/float_mat/float_mat_run.c b/test/monniaux/float_mat/float_mat_run.c new file mode 100644 index 00000000..2f590f98 --- /dev/null +++ b/test/monniaux/float_mat/float_mat_run.c @@ -0,0 +1,133 @@ +#include <stdbool.h> +#include <stdlib.h> +#include <stdio.h> +#include <inttypes.h> +#include "float_mat.h" +#include "../cycles.h" + +/* FIXME DMonniaux should be in the other but branches and float_of_int not implemented */ +bool REAL_mat_equal(unsigned m, + unsigned n, + const REAL *a, unsigned stride_a, + const REAL *b, unsigned stride_b) { + for(unsigned i=0; i<m; i++) { + for(unsigned j=0; j<n; j++) { + if (a[i*stride_a + j] != b[i*stride_b + j]) { + printf("at %u,%u: %g vs %g\n", i, j, + a[i*stride_a + j], b[i*stride_b + j]); + return false; + } + } + } + return true; +} + +REAL REAL_random(void) { + static uint64_t next = 1325997111; + next = next * 1103515249 + 12345; + return next % 1000; +} + +void REAL_mat_random(unsigned m, + unsigned n, + REAL *a, unsigned stride_a) { + for(unsigned i=0; i<m; i++) { + for(unsigned j=0; j<n; j++) { + a[i*stride_a + j] = REAL_random(); + } + } +} + +int main() { + const unsigned m = 60, n = 31, p = 50; + cycle_count_config(); + REAL *a = malloc(sizeof(REAL) * m * n); + REAL_mat_random(m, n, a, n); + REAL *b = malloc(sizeof(REAL) * n * p); + REAL_mat_random(n, p, b, p); + + REAL *c1 = malloc(sizeof(REAL) * m * p); + cycle_t c1_time = get_cycle(); + REAL_mat_mul1(m, n, p, c1, p, a, n, b, p); + c1_time = get_cycle()-c1_time; + + REAL *c2 = malloc(sizeof(REAL) * m * p); + cycle_t c2_time = get_cycle(); + REAL_mat_mul2(m, n, p, c2, p, a, n, b, p); + c2_time = get_cycle()-c2_time; + + REAL *c3 = malloc(sizeof(REAL) * m * p); + cycle_t c3_time = get_cycle(); + REAL_mat_mul3(m, n, p, c3, p, a, n, b, p); + c3_time = get_cycle()-c3_time; + + REAL *c4 = malloc(sizeof(REAL) * m * p); + cycle_t c4_time = get_cycle(); + REAL_mat_mul4(m, n, p, c4, p, a, n, b, p); + c4_time = get_cycle()-c4_time; + + REAL *c5 = malloc(sizeof(REAL) * m * p); + cycle_t c5_time = get_cycle(); + REAL_mat_mul5(m, n, p, c5, p, a, n, b, p); + c5_time = get_cycle()-c5_time; + + REAL *c6 = malloc(sizeof(REAL) * m * p); + cycle_t c6_time = get_cycle(); + REAL_mat_mul6(m, n, p, c6, p, a, n, b, p); + c6_time = get_cycle()-c6_time; + + REAL *c7 = malloc(sizeof(REAL) * m * p); + cycle_t c7_time = get_cycle(); + REAL_mat_mul7(m, n, p, c7, p, a, n, b, p); + c7_time = get_cycle()-c7_time; + + REAL *c8 = malloc(sizeof(REAL) * m * p); + cycle_t c8_time = get_cycle(); + REAL_mat_mul8(m, n, p, c8, p, a, n, b, p); + c8_time = get_cycle()-c8_time; + + printf("c1==c2: %s\n" + "c1==c3: %s\n" + "c1==c4: %s\n" + "c1==c5: %s\n" + "c1==c6: %s\n" + "c1==c7: %s\n" + "c1==c8: %s\n" + "c1 cycles: %" PRIu64 "\n" + "c2 cycles: %" PRIu64 "\n" + "c3 cycles: %" PRIu64 "\n" + "c4 cycles: %" PRIu64 "\n" + "c5 cycles: %" PRIu64 "\n" + "c6 cycles: %" PRIu64 "\n" + "c7 cycles: %" PRIu64 "\n" + "c8 cycles: %" PRIu64 "\n", + + REAL_mat_equal(m, n, c1, p, c2, p)?"true":"false", + REAL_mat_equal(m, n, c1, p, c3, p)?"true":"false", + REAL_mat_equal(m, n, c1, p, c4, p)?"true":"false", + REAL_mat_equal(m, n, c1, p, c5, p)?"true":"false", + REAL_mat_equal(m, n, c1, p, c6, p)?"true":"false", + REAL_mat_equal(m, n, c1, p, c7, p)?"true":"false", + REAL_mat_equal(m, n, c1, p, c8, p)?"true":"false", + + c1_time, + c2_time, + c3_time, + c4_time, + c5_time, + c6_time, + c7_time, + c8_time); + + free(a); + free(b); + free(c1); + free(c2); + free(c3); + free(c4); + free(c5); + free(c6); + free(c7); + free(c8); + return 0; +} |