commit dd6bb53f1bf252880638f6d1c6155839f45d3d6a parent c25e9f916f001ef84681eda43f356628d55f29aa Author: NunoSempere <nuno.sempere@protonmail.com> Date: Fri, 12 Jan 2024 23:53:58 +0100 add more comments, undo loop unrolling. Diffstat:
17 files changed, 6 insertions(+), 32 deletions(-)
diff --git a/examples/more/00_example_template/example b/examples/more/00_example_template/example Binary files differ. diff --git a/examples/more/01_sample_from_cdf/example b/examples/more/01_sample_from_cdf/example Binary files differ. diff --git a/examples/more/02_sample_from_cdf_beta/example b/examples/more/02_sample_from_cdf_beta/example Binary files differ. diff --git a/examples/more/03_ci_beta/example b/examples/more/03_ci_beta/example Binary files differ. diff --git a/examples/more/04_nuclear_war/example b/examples/more/04_nuclear_war/example Binary files differ. diff --git a/examples/more/05_burn_10kg_fat/example b/examples/more/05_burn_10kg_fat/example Binary files differ. diff --git a/examples/more/06_nuclear_recovery/example b/examples/more/06_nuclear_recovery/example Binary files differ. diff --git a/examples/more/07_algebra/example b/examples/more/07_algebra/example Binary files differ. diff --git a/examples/more/08_algebra_and_conversion/example b/examples/more/08_algebra_and_conversion/example Binary files differ. diff --git a/examples/more/09_ergonomic_algebra/example b/examples/more/09_ergonomic_algebra/example Binary files differ. diff --git a/examples/more/10_twitter_thread_example/example b/examples/more/10_twitter_thread_example/example Binary files differ. diff --git a/examples/more/11_billion_lognormals_paralell/example b/examples/more/11_billion_lognormals_paralell/example Binary files differ. diff --git a/examples/more/12_time_to_botec_parallel/example b/examples/more/12_time_to_botec_parallel/example Binary files differ. diff --git a/examples/more/13_parallelize_min/example b/examples/more/13_parallelize_min/example Binary files differ. diff --git a/examples/more/14_check_confidence_interval/example b/examples/more/14_check_confidence_interval/example Binary files differ. diff --git a/makefile b/makefile @@ -24,7 +24,7 @@ lint: clang-tidy squiggle_more.c -- -lm profile: - OMP_NUM_THREADS=16 sudo perf record ./examples/more/12_time_to_botec_parallel/example + sudo perf record ./examples/more/12_time_to_botec_parallel/example sudo perf report rm perf.data diff --git a/squiggle_more.c b/squiggle_more.c @@ -64,42 +64,16 @@ void sampler_parallel(double (*sampler)(uint64_t* seed), double* results, int n_ { #pragma omp for for (i = 0; i < n_threads; i++) { - // Simple version - /* int quotient = n_samples / n_threads; int lower_bound_inclusive = i * quotient; int upper_bound_not_inclusive = ((i + 1) * quotient); // note the < in the for loop below, for (int j = lower_bound_inclusive; j < upper_bound_not_inclusive; j++) { results[j] = sampler(&(cache_box[i].seed)); - // Could also result in inefficient cache stuff, but hopefully not too often - } - */ - - // Version with loop unrolling - int quotient = n_samples / n_threads; - int lower_bound_inclusive = i * quotient; - int upper_bound_not_inclusive = ((i + 1) * quotient); // note the < in the for loop below, - int delta = quotient; - int eighth_of_deltas = delta/8; // why 8? a double in 8 bytes, 8 doubles is the size of a cache line - int k; - // to do: simplify these variables. Maybe divide by n_threads * 8 directly - for(int j=0; j<eighth_of_deltas; j++){ - k = lower_bound_inclusive + j*8; - results[k+0] = sampler(&(cache_box[i].seed)); - results[k+1] = sampler(&(cache_box[i].seed)); - results[k+2] = sampler(&(cache_box[i].seed)); - results[k+3] = sampler(&(cache_box[i].seed)); - results[k+4] = sampler(&(cache_box[i].seed)); - results[k+5] = sampler(&(cache_box[i].seed)); - results[k+6] = sampler(&(cache_box[i].seed)); - results[k+7] = sampler(&(cache_box[i].seed)); - // these all fit one single cache line - // name of the techique: loop unrolling. - } - for(int k=eighth_of_deltas*8; k<upper_bound_not_inclusive; k++){ - results[k] = sampler(&(cache_box[i].seed)); - - } + // In principle, these results[j] could also result in two threads competing for the same cache line. + // In practice, though, + // a) this would happen infrequently + // b) + } } }