Implement 6D parallelization with 1D and no tiling

4 files changed, 555 insertions, 0 deletions

diff --git a/src/fastpath.c b/src/fastpath.c
index 6abbebe..b914ff0 100644
--- a/src/fastpath.c
+++ b/src/fastpath.c
@@ -1085,6 +1085,163 @@ PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_5d_tile_2d_fastpath(
 	pthreadpool_fence_release();
 }
 
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_6d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_6d_t task = (pthreadpool_task_6d_t) pthreadpool_load_relaxed_void_p(&threadpool->task);
+	void *const argument = pthreadpool_load_relaxed_void_p(&threadpool->argument);
+
+	const size_t threads_count = threadpool->threads_count.value;
+	const size_t range_threshold = -threads_count;
+
+	/* Process thread's own range of items */
+	const size_t range_start = pthreadpool_load_relaxed_size_t(&thread->range_start);
+	const struct fxdiv_divisor_size_t range_lmn = threadpool->params.parallelize_6d.range_lmn;
+	const struct fxdiv_result_size_t index_ijk_lmn = fxdiv_divide_size_t(range_start, range_lmn);
+	const struct fxdiv_divisor_size_t range_k = threadpool->params.parallelize_6d.range_k;
+	const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(index_ijk_lmn.quotient, range_k);
+	const struct fxdiv_divisor_size_t range_n = threadpool->params.parallelize_6d.range_n;
+	const struct fxdiv_result_size_t index_lm_n = fxdiv_divide_size_t(index_ijk_lmn.remainder, range_n);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_6d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+	const struct fxdiv_divisor_size_t range_m = threadpool->params.parallelize_6d.range_m;
+	const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(index_lm_n.quotient, range_m);
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_ij_k.remainder;
+	size_t l = index_l_m.quotient;
+	size_t m = index_l_m.remainder;
+	size_t n = index_lm_n.remainder;
+
+	const size_t range_l = threadpool->params.parallelize_6d.range_l;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j, k, l, m, n);
+		if (++n == range_n.value) {
+			n = 0;
+			if (++m == range_m.value) {
+				m = 0;
+				if (++l == range_l) {
+					l = 0;
+					if (++k == range_k.value) {
+						k = 0;
+						if (++j == range_j.value) {
+							j = 0;
+							i += 1;
+						}
+					}
+				}
+			}
+		}
+	}
+
+
+	/* There still may be other threads with work */
+	const size_t thread_number = thread->thread_number;
+	for (size_t tid = modulo_decrement(thread_number, threads_count);
+		tid != thread_number;
+		tid = modulo_decrement(tid, threads_count))
+	{
+		struct thread_info* other_thread = &threadpool->threads[tid];
+		while (pthreadpool_decrement_fetch_relaxed_size_t(&other_thread->range_length) < range_threshold) {
+			const size_t linear_index = pthreadpool_decrement_fetch_relaxed_size_t(&other_thread->range_end);
+			const struct fxdiv_result_size_t index_ijk_lmn = fxdiv_divide_size_t(linear_index, range_lmn);
+			const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(index_ijk_lmn.quotient, range_k);
+			const struct fxdiv_result_size_t index_lm_n = fxdiv_divide_size_t(index_ijk_lmn.remainder, range_n);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+			const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(index_lm_n.quotient, range_m);
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_ij_k.remainder, index_l_m.quotient, index_l_m.remainder, index_lm_n.remainder);
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_6d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_6d_tile_1d_t task = (pthreadpool_task_6d_tile_1d_t) pthreadpool_load_relaxed_void_p(&threadpool->task);
+	void *const argument = pthreadpool_load_relaxed_void_p(&threadpool->argument);
+
+	const size_t threads_count = threadpool->threads_count.value;
+	const size_t range_threshold = -threads_count;
+
+	/* Process thread's own range of items */
+	const size_t range_start = pthreadpool_load_relaxed_size_t(&thread->range_start);
+	const struct fxdiv_divisor_size_t tile_range_lmn = threadpool->params.parallelize_6d_tile_1d.tile_range_lmn;
+	const struct fxdiv_result_size_t tile_index_ijk_lmn = fxdiv_divide_size_t(range_start, tile_range_lmn);
+	const struct fxdiv_divisor_size_t range_k = threadpool->params.parallelize_6d_tile_1d.range_k;
+	const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(tile_index_ijk_lmn.quotient, range_k);
+	const struct fxdiv_divisor_size_t tile_range_n = threadpool->params.parallelize_6d_tile_1d.tile_range_n;
+	const struct fxdiv_result_size_t tile_index_lm_n = fxdiv_divide_size_t(tile_index_ijk_lmn.remainder, tile_range_n);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_6d_tile_1d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+	const struct fxdiv_divisor_size_t range_m = threadpool->params.parallelize_6d_tile_1d.range_m;
+	const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(tile_index_lm_n.quotient, range_m);
+	const size_t tile_n = threadpool->params.parallelize_6d_tile_1d.tile_n;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_ij_k.remainder;
+	size_t l = index_l_m.quotient;
+	size_t m = index_l_m.remainder;
+	size_t start_n = tile_index_lm_n.remainder * tile_n;
+
+	const size_t range_n = threadpool->params.parallelize_6d_tile_1d.range_n;
+	const size_t range_l = threadpool->params.parallelize_6d_tile_1d.range_l;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j, k, l, m, start_n, min(range_n - start_n, tile_n));
+		start_n += tile_n;
+		if (start_n >= range_n) {
+			start_n = 0;
+			if (++m == range_m.value) {
+				m = 0;
+				if (++l == range_l) {
+					l = 0;
+					if (++k == range_k.value) {
+						k = 0;
+						if (++j == range_j.value) {
+							j = 0;
+							i += 1;
+						}
+					}
+				}
+			}
+		}
+	}
+
+
+	/* There still may be other threads with work */
+	const size_t thread_number = thread->thread_number;
+	for (size_t tid = modulo_decrement(thread_number, threads_count);
+		tid != thread_number;
+		tid = modulo_decrement(tid, threads_count))
+	{
+		struct thread_info* other_thread = &threadpool->threads[tid];
+		while (pthreadpool_decrement_fetch_relaxed_size_t(&other_thread->range_length) < range_threshold) {
+			const size_t linear_index = pthreadpool_decrement_fetch_relaxed_size_t(&other_thread->range_end);
+			const struct fxdiv_result_size_t tile_index_ijk_lmn = fxdiv_divide_size_t(linear_index, tile_range_lmn);
+			const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(tile_index_ijk_lmn.quotient, range_k);
+			const struct fxdiv_result_size_t tile_index_lm_n = fxdiv_divide_size_t(tile_index_ijk_lmn.remainder, tile_range_n);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+			const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(tile_index_lm_n.quotient, range_m);
+			const size_t start_n = tile_index_lm_n.remainder * tile_n;
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_ij_k.remainder, index_l_m.quotient, index_l_m.remainder,
+				start_n, min(range_n - start_n, tile_n));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
 PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_6d_tile_2d_fastpath(
 	struct pthreadpool* threadpool,
 	struct thread_info* thread)
diff --git a/src/portable-api.c b/src/portable-api.c
index ef36578..42d0369 100644
--- a/src/portable-api.c
+++ b/src/portable-api.c
@@ -1002,6 +1002,153 @@ static void thread_parallelize_5d_tile_2d(struct pthreadpool* threadpool, struct
 	pthreadpool_fence_release();
 }
 
+static void thread_parallelize_6d(struct pthreadpool* threadpool, struct thread_info* thread) {
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_6d_t task = (pthreadpool_task_6d_t) pthreadpool_load_relaxed_void_p(&threadpool->task);
+	void *const argument = pthreadpool_load_relaxed_void_p(&threadpool->argument);
+
+	/* Process thread's own range of items */
+	const size_t range_start = pthreadpool_load_relaxed_size_t(&thread->range_start);
+	const struct fxdiv_divisor_size_t range_lmn = threadpool->params.parallelize_6d.range_lmn;
+	const struct fxdiv_result_size_t index_ijk_lmn = fxdiv_divide_size_t(range_start, range_lmn);
+	const struct fxdiv_divisor_size_t range_k = threadpool->params.parallelize_6d.range_k;
+	const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(index_ijk_lmn.quotient, range_k);
+	const struct fxdiv_divisor_size_t range_n = threadpool->params.parallelize_6d.range_n;
+	const struct fxdiv_result_size_t index_lm_n = fxdiv_divide_size_t(index_ijk_lmn.remainder, range_n);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_6d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+	const struct fxdiv_divisor_size_t range_m = threadpool->params.parallelize_6d.range_m;
+	const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(index_lm_n.quotient, range_m);
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_ij_k.remainder;
+	size_t l = index_l_m.quotient;
+	size_t m = index_l_m.remainder;
+	size_t n = index_lm_n.remainder;
+
+	const size_t range_l = threadpool->params.parallelize_6d.range_l;
+	while (pthreadpool_try_decrement_relaxed_size_t(&thread->range_length)) {
+		task(argument, i, j, k, l, m, n);
+		if (++n == range_n.value) {
+			n = 0;
+			if (++m == range_m.value) {
+				m = 0;
+				if (++l == range_l) {
+					l = 0;
+					if (++k == range_k.value) {
+						k = 0;
+						if (++j == range_j.value) {
+							j = 0;
+							i += 1;
+						}
+					}
+				}
+			}
+		}
+	}
+
+
+	/* There still may be other threads with work */
+	const size_t thread_number = thread->thread_number;
+	const size_t threads_count = threadpool->threads_count.value;
+	for (size_t tid = modulo_decrement(thread_number, threads_count);
+		tid != thread_number;
+		tid = modulo_decrement(tid, threads_count))
+	{
+		struct thread_info* other_thread = &threadpool->threads[tid];
+		while (pthreadpool_try_decrement_relaxed_size_t(&other_thread->range_length)) {
+			const size_t linear_index = pthreadpool_decrement_fetch_relaxed_size_t(&other_thread->range_end);
+			const struct fxdiv_result_size_t index_ijk_lmn = fxdiv_divide_size_t(linear_index, range_lmn);
+			const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(index_ijk_lmn.quotient, range_k);
+			const struct fxdiv_result_size_t index_lm_n = fxdiv_divide_size_t(index_ijk_lmn.remainder, range_n);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+			const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(index_lm_n.quotient, range_m);
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_ij_k.remainder, index_l_m.quotient, index_l_m.remainder, index_lm_n.remainder);
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+static void thread_parallelize_6d_tile_1d(struct pthreadpool* threadpool, struct thread_info* thread) {
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_6d_tile_1d_t task = (pthreadpool_task_6d_tile_1d_t) pthreadpool_load_relaxed_void_p(&threadpool->task);
+	void *const argument = pthreadpool_load_relaxed_void_p(&threadpool->argument);
+
+	/* Process thread's own range of items */
+	const size_t range_start = pthreadpool_load_relaxed_size_t(&thread->range_start);
+	const struct fxdiv_divisor_size_t tile_range_lmn = threadpool->params.parallelize_6d_tile_1d.tile_range_lmn;
+	const struct fxdiv_result_size_t tile_index_ijk_lmn = fxdiv_divide_size_t(range_start, tile_range_lmn);
+	const struct fxdiv_divisor_size_t range_k = threadpool->params.parallelize_6d_tile_1d.range_k;
+	const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(tile_index_ijk_lmn.quotient, range_k);
+	const struct fxdiv_divisor_size_t tile_range_n = threadpool->params.parallelize_6d_tile_1d.tile_range_n;
+	const struct fxdiv_result_size_t tile_index_lm_n = fxdiv_divide_size_t(tile_index_ijk_lmn.remainder, tile_range_n);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_6d_tile_1d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+	const struct fxdiv_divisor_size_t range_m = threadpool->params.parallelize_6d_tile_1d.range_m;
+	const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(tile_index_lm_n.quotient, range_m);
+	const size_t tile_n = threadpool->params.parallelize_6d_tile_1d.tile_n;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_ij_k.remainder;
+	size_t l = index_l_m.quotient;
+	size_t m = index_l_m.remainder;
+	size_t start_n = tile_index_lm_n.remainder * tile_n;
+
+	const size_t range_n = threadpool->params.parallelize_6d_tile_1d.range_n;
+	const size_t range_l = threadpool->params.parallelize_6d_tile_1d.range_l;
+	while (pthreadpool_try_decrement_relaxed_size_t(&thread->range_length)) {
+		task(argument, i, j, k, l, m, start_n, min(range_n - start_n, tile_n));
+		start_n += tile_n;
+		if (start_n >= range_n) {
+			start_n = 0;
+			if (++m == range_m.value) {
+				m = 0;
+				if (++l == range_l) {
+					l = 0;
+					if (++k == range_k.value) {
+						k = 0;
+						if (++j == range_j.value) {
+							j = 0;
+							i += 1;
+						}
+					}
+				}
+			}
+		}
+	}
+
+
+	/* There still may be other threads with work */
+	const size_t thread_number = thread->thread_number;
+	const size_t threads_count = threadpool->threads_count.value;
+	for (size_t tid = modulo_decrement(thread_number, threads_count);
+		tid != thread_number;
+		tid = modulo_decrement(tid, threads_count))
+	{
+		struct thread_info* other_thread = &threadpool->threads[tid];
+		while (pthreadpool_try_decrement_relaxed_size_t(&other_thread->range_length)) {
+			const size_t linear_index = pthreadpool_decrement_fetch_relaxed_size_t(&other_thread->range_end);
+			const struct fxdiv_result_size_t tile_index_ijk_lmn = fxdiv_divide_size_t(linear_index, tile_range_lmn);
+			const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(tile_index_ijk_lmn.quotient, range_k);
+			const struct fxdiv_result_size_t tile_index_lm_n = fxdiv_divide_size_t(tile_index_ijk_lmn.remainder, tile_range_n);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+			const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(tile_index_lm_n.quotient, range_m);
+			const size_t start_n = tile_index_lm_n.remainder * tile_n;
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_ij_k.remainder, index_l_m.quotient, index_l_m.remainder,
+				start_n, min(range_n - start_n, tile_n));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
 static void thread_parallelize_6d_tile_2d(struct pthreadpool* threadpool, struct thread_info* thread) {
 	assert(threadpool != NULL);
 	assert(thread != NULL);
@@ -2043,6 +2190,130 @@ void pthreadpool_parallelize_5d_tile_2d(
 	}
 }
 
+void pthreadpool_parallelize_6d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_6d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	size_t range_m,
+	size_t range_n,
+	uint32_t flags)
+{
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || (range_i | range_j | range_k | range_l | range_m | range_n) <= 1) {
+		/* No thread pool used: execute task sequentially on the calling thread */
+		struct fpu_state saved_fpu_state = { 0 };
+		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
+			saved_fpu_state = get_fpu_state();
+			disable_fpu_denormals();
+		}
+		for (size_t i = 0; i < range_i; i++) {
+			for (size_t j = 0; j < range_j; j++) {
+				for (size_t k = 0; k < range_k; k++) {
+					for (size_t l = 0; l < range_l; l++) {
+						for (size_t m = 0; m < range_m; m++) {
+							for (size_t n = 0; n < range_n; n++) {
+								task(argument, i, j, k, l, m, n);
+							}
+						}
+					}
+				}
+			}
+		}
+		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
+			set_fpu_state(saved_fpu_state);
+		}
+	} else {
+		const size_t range_lmn = range_l * range_m * range_n;
+		const size_t range = range_i * range_j * range_k * range_lmn;
+		const struct pthreadpool_6d_params params = {
+			.range_l = range_l,
+			.range_j = fxdiv_init_size_t(range_j),
+			.range_k = fxdiv_init_size_t(range_k),
+			.range_lmn = fxdiv_init_size_t(range_lmn),
+			.range_m = fxdiv_init_size_t(range_m),
+			.range_n = fxdiv_init_size_t(range_n),
+		};
+		thread_function_t parallelize_6d = &thread_parallelize_6d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (range < range_threshold) {
+				parallelize_6d = &pthreadpool_thread_parallelize_6d_fastpath;
+			}
+		#endif
+		pthreadpool_parallelize(
+			threadpool, parallelize_6d, &params, sizeof(params),
+			task, argument, range, flags);
+	}
+}
+
+void pthreadpool_parallelize_6d_tile_1d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_6d_tile_1d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	size_t range_m,
+	size_t range_n,
+	size_t tile_n,
+	uint32_t flags)
+{
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || ((range_i | range_j | range_k | range_l | range_m) <= 1 && range_n <= tile_n)) {
+		/* No thread pool used: execute task sequentially on the calling thread */
+		struct fpu_state saved_fpu_state = { 0 };
+		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
+			saved_fpu_state = get_fpu_state();
+			disable_fpu_denormals();
+		}
+		for (size_t i = 0; i < range_i; i++) {
+			for (size_t j = 0; j < range_j; j++) {
+				for (size_t k = 0; k < range_k; k++) {
+					for (size_t l = 0; l < range_l; l++) {
+						for (size_t m = 0; m < range_m; m++) {
+							for (size_t n = 0; n < range_n; n += tile_n) {
+								task(argument, i, j, k, l, m, n, min(range_n - n, tile_n));
+							}
+						}
+					}
+				}
+			}
+		}
+		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
+			set_fpu_state(saved_fpu_state);
+		}
+	} else {
+		const size_t tile_range_n = divide_round_up(range_n, tile_n);
+		const size_t tile_range_lmn = range_l * range_m * tile_range_n;
+		const size_t tile_range = range_i * range_j * range_k * tile_range_lmn;
+		const struct pthreadpool_6d_tile_1d_params params = {
+			.range_l = range_l,
+			.range_n = range_n,
+			.tile_n = tile_n,
+			.range_j = fxdiv_init_size_t(range_j),
+			.range_k = fxdiv_init_size_t(range_k),
+			.tile_range_lmn = fxdiv_init_size_t(tile_range_lmn),
+			.range_m = fxdiv_init_size_t(range_m),
+			.tile_range_n = fxdiv_init_size_t(tile_range_n),
+		};
+		thread_function_t parallelize_6d_tile_1d = &thread_parallelize_6d_tile_1d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_6d_tile_1d = &pthreadpool_thread_parallelize_6d_tile_1d_fastpath;
+			}
+		#endif
+		pthreadpool_parallelize(
+			threadpool, parallelize_6d_tile_1d, &params, sizeof(params),
+			task, argument, tile_range, flags);
+	}
+}
+
 void pthreadpool_parallelize_6d_tile_2d(
 	pthreadpool_t threadpool,
 	pthreadpool_task_6d_tile_2d_t task,
diff --git a/src/shim.c b/src/shim.c
index e90ac45..39ec884 100644
--- a/src/shim.c
+++ b/src/shim.c
@@ -383,6 +383,61 @@ void pthreadpool_parallelize_5d_tile_2d(
 	}
 }
 
+void pthreadpool_parallelize_6d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_6d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	size_t range_m,
+	size_t range_n,
+	uint32_t flags)
+{
+	for (size_t i = 0; i < range_i; i++) {
+		for (size_t j = 0; j < range_j; j++) {
+			for (size_t k = 0; k < range_k; k++) {
+				for (size_t l = 0; l < range_l; l++) {
+					for (size_t m = 0; m < range_m; m++) {
+						for (size_t n = 0; n < range_n; n++) {
+							task(argument, i, j, k, l, m, n);
+						}
+					}
+				}
+			}
+		}
+	}
+}
+
+void pthreadpool_parallelize_6d_tile_1d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_6d_tile_1d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	size_t range_m,
+	size_t range_n,
+	size_t tile_n,
+	uint32_t flags)
+{
+	for (size_t i = 0; i < range_i; i++) {
+		for (size_t j = 0; j < range_j; j++) {
+			for (size_t k = 0; k < range_k; k++) {
+				for (size_t l = 0; l < range_l; l++) {
+					for (size_t m = 0; m < range_m; m++) {
+						for (size_t n = 0; n < range_n; n += tile_n) {
+							task(argument, i, j, k, l, m, n, min(range_n - n, tile_n));
+						}
+					}
+				}
+			}
+		}
+	}
+}
+
 void pthreadpool_parallelize_6d_tile_2d(
 	pthreadpool_t threadpool,
 	pthreadpool_task_6d_tile_2d_t task,
diff --git a/src/threadpool-object.h b/src/threadpool-object.h
index 9870e8a..590dc96 100644
--- a/src/threadpool-object.h
+++ b/src/threadpool-object.h
@@ -476,6 +476,68 @@ struct pthreadpool_5d_tile_2d_params {
 	struct fxdiv_divisor_size_t tile_range_m;
 };
 
+struct pthreadpool_6d_params {
+	/**
+	 * Copy of the range_l argument passed to the pthreadpool_parallelize_6d function.
+	 */
+	size_t range_l;
+	/**
+	 * FXdiv divisor for the range_j argument passed to the pthreadpool_parallelize_6d function.
+	 */
+	struct fxdiv_divisor_size_t range_j;
+	/**
+	 * FXdiv divisor for the range_k argument passed to the pthreadpool_parallelize_6d function.
+	 */
+	struct fxdiv_divisor_size_t range_k;
+	/**
+	 * FXdiv divisor for the range_l * range_m * range_n value.
+	 */
+	struct fxdiv_divisor_size_t range_lmn;
+	/**
+	 * FXdiv divisor for the range_m argument passed to the pthreadpool_parallelize_6d function.
+	 */
+	struct fxdiv_divisor_size_t range_m;
+	/**
+	 * FXdiv divisor for the range_n argument passed to the pthreadpool_parallelize_6d function.
+	 */
+	struct fxdiv_divisor_size_t range_n;
+};
+
+struct pthreadpool_6d_tile_1d_params {
+	/**
+	 * Copy of the range_l argument passed to the pthreadpool_parallelize_6d_tile_1d function.
+	 */
+	size_t range_l;
+	/**
+	 * Copy of the range_n argument passed to the pthreadpool_parallelize_6d_tile_1d function.
+	 */
+	size_t range_n;
+	/**
+	 * Copy of the tile_n argument passed to the pthreadpool_parallelize_6d_tile_1d function.
+	 */
+	size_t tile_n;
+	/**
+	 * FXdiv divisor for the range_j argument passed to the pthreadpool_parallelize_6d_tile_1d function.
+	 */
+	struct fxdiv_divisor_size_t range_j;
+	/**
+	 * FXdiv divisor for the range_k argument passed to the pthreadpool_parallelize_6d_tile_1d function.
+	 */
+	struct fxdiv_divisor_size_t range_k;
+	/**
+	 * FXdiv divisor for the range_l * range_m * divide_round_up(range_n, tile_n) value.
+	 */
+	struct fxdiv_divisor_size_t tile_range_lmn;
+	/**
+	 * FXdiv divisor for the range_m argument passed to the pthreadpool_parallelize_6d_tile_1d function.
+	 */
+	struct fxdiv_divisor_size_t range_m;
+	/**
+	 * FXdiv divisor for the divide_round_up(range_n, tile_n) value.
+	 */
+	struct fxdiv_divisor_size_t tile_range_n;
+};
+
 struct pthreadpool_6d_tile_2d_params {
 	/**
 	 * Copy of the range_k argument passed to the pthreadpool_parallelize_6d_tile_2d function.
@@ -575,6 +637,8 @@ struct PTHREADPOOL_CACHELINE_ALIGNED pthreadpool {
 		struct pthreadpool_5d_params parallelize_5d;
 		struct pthreadpool_5d_tile_1d_params parallelize_5d_tile_1d;
 		struct pthreadpool_5d_tile_2d_params parallelize_5d_tile_2d;
+		struct pthreadpool_6d_params parallelize_6d;
+		struct pthreadpool_6d_tile_1d_params parallelize_6d_tile_1d;
 		struct pthreadpool_6d_tile_2d_params parallelize_6d_tile_2d;
 	} params;
 	/**
@@ -735,6 +799,14 @@ PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_5d_tile_2d_fastpath(
 	struct pthreadpool* threadpool,
 	struct thread_info* thread);
 
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_6d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_6d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
 PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_6d_tile_2d_fastpath(
 	struct pthreadpool* threadpool,
 	struct thread_info* thread);