Upgrade pthreadpool to 029c88620802e1361ccf41d1970bd5b07fd6b7bb am: a85203c348 am: b4e66e0642 am: aeabd9ff75 am: b0d66a6829

Original change: https://android-review.googlesource.com/c/platform/external/pthreadpool/+/1361277

Change-Id: I8fc49776e53585086a622b1409d7b244f8ad6236

8 files changed, 2497 insertions, 159 deletions

diff --git a/src/fastpath.c b/src/fastpath.c
new file mode 100644
index 0000000..6abbebe
--- /dev/null
+++ b/src/fastpath.c
@@ -0,0 +1,1170 @@
+/* Standard C headers */
+#include <assert.h>
+#include <stdbool.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+
+#if PTHREADPOOL_USE_CPUINFO
+	#include <cpuinfo.h>
+#endif
+
+/* Dependencies */
+#include <fxdiv.h>
+
+/* Public library header */
+#include <pthreadpool.h>
+
+/* Internal library headers */
+#include "threadpool-atomics.h"
+#include "threadpool-common.h"
+#include "threadpool-object.h"
+#include "threadpool-utils.h"
+
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_1d_t task = (pthreadpool_task_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 */
+	size_t range_start = pthreadpool_load_relaxed_size_t(&thread->range_start);
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, range_start++);
+	}
+
+	/* 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 index = pthreadpool_decrement_fetch_relaxed_size_t(&other_thread->range_end);
+			task(argument, index);
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_1d_with_uarch_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_1d_with_id_t task = (pthreadpool_task_1d_with_id_t) pthreadpool_load_relaxed_void_p(&threadpool->task);
+	void *const argument = pthreadpool_load_relaxed_void_p(&threadpool->argument);
+
+	const uint32_t default_uarch_index = threadpool->params.parallelize_1d_with_uarch.default_uarch_index;
+	uint32_t uarch_index = default_uarch_index;
+	#if PTHREADPOOL_USE_CPUINFO
+		uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
+		if (uarch_index > threadpool->params.parallelize_1d_with_uarch.max_uarch_index) {
+			uarch_index = default_uarch_index;
+		}
+	#endif
+
+	const size_t threads_count = threadpool->threads_count.value;
+	const size_t range_threshold = -threads_count;
+
+	/* Process thread's own range of items */
+	size_t range_start = pthreadpool_load_relaxed_size_t(&thread->range_start);
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, uarch_index, range_start++);
+	}
+
+	/* 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 index = pthreadpool_decrement_fetch_relaxed_size_t(&other_thread->range_end);
+			task(argument, uarch_index, index);
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_1d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_1d_tile_1d_t task = (pthreadpool_task_1d_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 size_t tile = threadpool->params.parallelize_1d_tile_1d.tile;
+	size_t tile_start = range_start * tile;
+
+	const size_t range = threadpool->params.parallelize_1d_tile_1d.range;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, tile_start, min(range - tile_start, tile));
+		tile_start += tile;
+	}
+
+	/* 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 tile_index = pthreadpool_decrement_fetch_relaxed_size_t(&other_thread->range_end);
+			const size_t tile_start = tile_index * tile;
+			task(argument, tile_start, min(range - tile_start, tile));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_2d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_2d_t task = (pthreadpool_task_2d_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_j = threadpool->params.parallelize_2d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(range_start, range_j);
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j);
+		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_i_j = fxdiv_divide_size_t(linear_index, range_j);
+			task(argument, index_i_j.quotient, index_i_j.remainder);
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_2d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_2d_tile_1d_t task = (pthreadpool_task_2d_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_j = threadpool->params.parallelize_2d_tile_1d.tile_range_j;
+	const struct fxdiv_result_size_t tile_index_i_j = fxdiv_divide_size_t(range_start, tile_range_j);
+	const size_t tile_j = threadpool->params.parallelize_2d_tile_1d.tile_j;
+	size_t i = tile_index_i_j.quotient;
+	size_t start_j = tile_index_i_j.remainder * tile_j;
+
+	const size_t range_j = threadpool->params.parallelize_2d_tile_1d.range_j;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, start_j, min(range_j - start_j, tile_j));
+		start_j += tile_j;
+		if (start_j >= range_j) {
+			start_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_i_j = fxdiv_divide_size_t(linear_index, tile_range_j);
+			const size_t start_j = tile_index_i_j.remainder * tile_j;
+			task(argument, tile_index_i_j.quotient, start_j, min(range_j - start_j, tile_j));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_2d_tile_2d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_2d_tile_2d_t task = (pthreadpool_task_2d_tile_2d_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_j = threadpool->params.parallelize_2d_tile_2d.tile_range_j;
+	const struct fxdiv_result_size_t tile_index_i_j = fxdiv_divide_size_t(range_start, tile_range_j);
+	const size_t tile_i = threadpool->params.parallelize_2d_tile_2d.tile_i;
+	const size_t tile_j = threadpool->params.parallelize_2d_tile_2d.tile_j;
+	size_t start_i = tile_index_i_j.quotient * tile_i;
+	size_t start_j = tile_index_i_j.remainder * tile_j;
+
+	const size_t range_i = threadpool->params.parallelize_2d_tile_2d.range_i;
+	const size_t range_j = threadpool->params.parallelize_2d_tile_2d.range_j;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, start_i, start_j, min(range_i - start_i, tile_i), min(range_j - start_j, tile_j));
+		start_j += tile_j;
+		if (start_j >= range_j) {
+			start_j = 0;
+			start_i += tile_i;
+		}
+	}
+
+	/* 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_i_j = fxdiv_divide_size_t(linear_index, tile_range_j);
+			const size_t start_i = tile_index_i_j.quotient * tile_i;
+			const size_t start_j = tile_index_i_j.remainder * tile_j;
+			task(argument, start_i, start_j, min(range_i - start_i, tile_i), min(range_j - start_j, tile_j));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_2d_tile_2d_with_uarch_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_2d_tile_2d_with_id_t task = (pthreadpool_task_2d_tile_2d_with_id_t) pthreadpool_load_relaxed_void_p(&threadpool->task);
+	void *const argument = pthreadpool_load_relaxed_void_p(&threadpool->argument);
+
+	const uint32_t default_uarch_index = threadpool->params.parallelize_2d_tile_2d_with_uarch.default_uarch_index;
+	uint32_t uarch_index = default_uarch_index;
+	#if PTHREADPOOL_USE_CPUINFO
+		uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
+		if (uarch_index > threadpool->params.parallelize_2d_tile_2d_with_uarch.max_uarch_index) {
+			uarch_index = default_uarch_index;
+		}
+	#endif
+
+	const size_t threads_count = threadpool->threads_count.value;
+	const size_t range_threshold = -threads_count;
+
+	/* Process thread's own range of items */
+	const struct fxdiv_divisor_size_t tile_range_j = threadpool->params.parallelize_2d_tile_2d_with_uarch.tile_range_j;
+	const size_t range_start = pthreadpool_load_relaxed_size_t(&thread->range_start);
+	const struct fxdiv_result_size_t index = fxdiv_divide_size_t(range_start, tile_range_j);
+	const size_t range_i = threadpool->params.parallelize_2d_tile_2d_with_uarch.range_i;
+	const size_t tile_i = threadpool->params.parallelize_2d_tile_2d_with_uarch.tile_i;
+	const size_t range_j = threadpool->params.parallelize_2d_tile_2d_with_uarch.range_j;
+	const size_t tile_j = threadpool->params.parallelize_2d_tile_2d_with_uarch.tile_j;
+	size_t start_i = index.quotient * tile_i;
+	size_t start_j = index.remainder * tile_j;
+
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, uarch_index, start_i, start_j, min(range_i - start_i, tile_i), min(range_j - start_j, tile_j));
+		start_j += tile_j;
+		if (start_j >= range_j) {
+			start_j = 0;
+			start_i += tile_i;
+		}
+	}
+
+	/* 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_i_j = fxdiv_divide_size_t(linear_index, tile_range_j);
+			const size_t start_i = tile_index_i_j.quotient * tile_i;
+			const size_t start_j = tile_index_i_j.remainder * tile_j;
+			task(argument, uarch_index, start_i, start_j, min(range_i - start_i, tile_i), min(range_j - start_j, tile_j));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_3d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_3d_t task = (pthreadpool_task_3d_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_k = threadpool->params.parallelize_3d.range_k;
+	const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(range_start, range_k);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_3d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_ij_k.remainder;
+
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j, k);
+		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_ij_k = fxdiv_divide_size_t(linear_index, range_k);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_ij_k.remainder);
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_3d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_3d_tile_1d_t task = (pthreadpool_task_3d_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_k = threadpool->params.parallelize_3d_tile_1d.tile_range_k;
+	const struct fxdiv_result_size_t tile_index_ij_k = fxdiv_divide_size_t(range_start, tile_range_k);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_3d_tile_1d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_k.quotient, range_j);
+	const size_t tile_k = threadpool->params.parallelize_3d_tile_1d.tile_k;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t start_k = tile_index_ij_k.remainder * tile_k;
+
+	const size_t range_k = threadpool->params.parallelize_3d_tile_1d.range_k;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j, start_k, min(range_k - start_k, tile_k));
+		start_k += tile_k;
+		if (start_k >= range_k) {
+			start_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_ij_k = fxdiv_divide_size_t(linear_index, tile_range_k);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_k.quotient, range_j);
+			const size_t start_k = tile_index_ij_k.remainder * tile_k;
+			task(argument, index_i_j.quotient, index_i_j.remainder, start_k, min(range_k - start_k, tile_k));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_3d_tile_2d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_3d_tile_2d_t task = (pthreadpool_task_3d_tile_2d_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_k = threadpool->params.parallelize_3d_tile_2d.tile_range_k;
+	const struct fxdiv_result_size_t tile_index_ij_k = fxdiv_divide_size_t(range_start, tile_range_k);
+	const struct fxdiv_divisor_size_t tile_range_j = threadpool->params.parallelize_3d_tile_2d.tile_range_j;
+	const struct fxdiv_result_size_t tile_index_i_j = fxdiv_divide_size_t(tile_index_ij_k.quotient, tile_range_j);
+	const size_t tile_j = threadpool->params.parallelize_3d_tile_2d.tile_j;
+	const size_t tile_k = threadpool->params.parallelize_3d_tile_2d.tile_k;
+	size_t i = tile_index_i_j.quotient;
+	size_t start_j = tile_index_i_j.remainder * tile_j;
+	size_t start_k = tile_index_ij_k.remainder * tile_k;
+
+	const size_t range_k = threadpool->params.parallelize_3d_tile_2d.range_k;
+	const size_t range_j = threadpool->params.parallelize_3d_tile_2d.range_j;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, start_j, start_k, min(range_j - start_j, tile_j), min(range_k - start_k, tile_k));
+		start_k += tile_k;
+		if (start_k >= range_k) {
+			start_k = 0;
+			start_j += tile_j;
+			if (start_j >= range_j) {
+				start_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_ij_k = fxdiv_divide_size_t(linear_index, tile_range_k);
+			const struct fxdiv_result_size_t tile_index_i_j = fxdiv_divide_size_t(tile_index_ij_k.quotient, tile_range_j);
+			const size_t start_j = tile_index_i_j.remainder * tile_j;
+			const size_t start_k = tile_index_ij_k.remainder * tile_k;
+			task(argument, tile_index_i_j.quotient, start_j, start_k, min(range_j - start_j, tile_j), min(range_k - start_k, tile_k));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_3d_tile_2d_with_uarch_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_3d_tile_2d_with_id_t task = (pthreadpool_task_3d_tile_2d_with_id_t) pthreadpool_load_relaxed_void_p(&threadpool->task);
+	void *const argument = pthreadpool_load_relaxed_void_p(&threadpool->argument);
+
+	const uint32_t default_uarch_index = threadpool->params.parallelize_3d_tile_2d_with_uarch.default_uarch_index;
+	uint32_t uarch_index = default_uarch_index;
+	#if PTHREADPOOL_USE_CPUINFO
+		uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
+		if (uarch_index > threadpool->params.parallelize_3d_tile_2d_with_uarch.max_uarch_index) {
+			uarch_index = default_uarch_index;
+		}
+	#endif
+
+	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_k = threadpool->params.parallelize_3d_tile_2d_with_uarch.tile_range_k;
+	const struct fxdiv_result_size_t tile_index_ij_k = fxdiv_divide_size_t(range_start, tile_range_k);
+	const struct fxdiv_divisor_size_t tile_range_j = threadpool->params.parallelize_3d_tile_2d_with_uarch.tile_range_j;
+	const struct fxdiv_result_size_t tile_index_i_j = fxdiv_divide_size_t(tile_index_ij_k.quotient, tile_range_j);
+	const size_t tile_j = threadpool->params.parallelize_3d_tile_2d_with_uarch.tile_j;
+	const size_t tile_k = threadpool->params.parallelize_3d_tile_2d_with_uarch.tile_k;
+	size_t i = tile_index_i_j.quotient;
+	size_t start_j = tile_index_i_j.remainder * tile_j;
+	size_t start_k = tile_index_ij_k.remainder * tile_k;
+
+	const size_t range_k = threadpool->params.parallelize_3d_tile_2d_with_uarch.range_k;
+	const size_t range_j = threadpool->params.parallelize_3d_tile_2d_with_uarch.range_j;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, uarch_index, i, start_j, start_k, min(range_j - start_j, tile_j), min(range_k - start_k, tile_k));
+		start_k += tile_k;
+		if (start_k >= range_k) {
+			start_k = 0;
+			start_j += tile_j;
+			if (start_j >= range_j) {
+				start_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_ij_k = fxdiv_divide_size_t(linear_index, tile_range_k);
+			const struct fxdiv_result_size_t tile_index_i_j = fxdiv_divide_size_t(tile_index_ij_k.quotient, tile_range_j);
+			const size_t start_j = tile_index_i_j.remainder * tile_j;
+			const size_t start_k = tile_index_ij_k.remainder * tile_k;
+			task(argument, uarch_index, tile_index_i_j.quotient, start_j, start_k, min(range_j - start_j, tile_j), min(range_k - start_k, tile_k));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_4d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_4d_t task = (pthreadpool_task_4d_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_kl = threadpool->params.parallelize_4d.range_kl;
+	const struct fxdiv_result_size_t index_ij_kl = fxdiv_divide_size_t(range_start, range_kl);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_4d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_kl.quotient, range_j);
+	const struct fxdiv_divisor_size_t range_l = threadpool->params.parallelize_4d.range_l;
+	const struct fxdiv_result_size_t index_k_l = fxdiv_divide_size_t(index_ij_kl.remainder, range_l);
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_k_l.quotient;
+	size_t l = index_k_l.remainder;
+
+	const size_t range_k = threadpool->params.parallelize_4d.range_k;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j, k, l);
+		if (++l == range_l.value) {
+			l = 0;
+			if (++k == range_k) {
+				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_ij_kl = fxdiv_divide_size_t(linear_index, range_kl);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_kl.quotient, range_j);
+			const struct fxdiv_result_size_t index_k_l = fxdiv_divide_size_t(index_ij_kl.remainder, range_l);
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_k_l.quotient, index_k_l.remainder);
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_4d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_4d_tile_1d_t task = (pthreadpool_task_4d_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_kl = threadpool->params.parallelize_4d_tile_1d.tile_range_kl;
+	const struct fxdiv_result_size_t tile_index_ij_kl = fxdiv_divide_size_t(range_start, tile_range_kl);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_4d_tile_1d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_kl.quotient, range_j);
+	const struct fxdiv_divisor_size_t tile_range_l = threadpool->params.parallelize_4d_tile_1d.tile_range_l;
+	const struct fxdiv_result_size_t tile_index_k_l = fxdiv_divide_size_t(tile_index_ij_kl.remainder, tile_range_l);
+	const size_t tile_l = threadpool->params.parallelize_4d_tile_1d.tile_l;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = tile_index_k_l.quotient;
+	size_t start_l = tile_index_k_l.remainder * tile_l;
+
+	const size_t range_l = threadpool->params.parallelize_4d_tile_1d.range_l;
+	const size_t range_k = threadpool->params.parallelize_4d_tile_1d.range_k;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j, k, start_l, min(range_l - start_l, tile_l));
+		start_l += tile_l;
+		if (start_l >= range_l) {
+			start_l = 0;
+			if (++k == range_k) {
+				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_ij_kl = fxdiv_divide_size_t(linear_index, tile_range_kl);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_kl.quotient, range_j);
+			const struct fxdiv_result_size_t tile_index_k_l = fxdiv_divide_size_t(tile_index_ij_kl.remainder, tile_range_l);
+			const size_t start_l = tile_index_k_l.remainder * tile_l;
+			task(argument, index_i_j.quotient, index_i_j.remainder, tile_index_k_l.quotient, start_l, min(range_l - start_l, tile_l));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_4d_tile_2d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_4d_tile_2d_t task = (pthreadpool_task_4d_tile_2d_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_kl = threadpool->params.parallelize_4d_tile_2d.tile_range_kl;
+	const struct fxdiv_result_size_t tile_index_ij_kl = fxdiv_divide_size_t(range_start, tile_range_kl);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_4d_tile_2d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_kl.quotient, range_j);
+	const struct fxdiv_divisor_size_t tile_range_l = threadpool->params.parallelize_4d_tile_2d.tile_range_l;
+	const struct fxdiv_result_size_t tile_index_k_l = fxdiv_divide_size_t(tile_index_ij_kl.remainder, tile_range_l);
+	const size_t tile_k = threadpool->params.parallelize_4d_tile_2d.tile_k;
+	const size_t tile_l = threadpool->params.parallelize_4d_tile_2d.tile_l;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t start_k = tile_index_k_l.quotient * tile_k;
+	size_t start_l = tile_index_k_l.remainder * tile_l;
+
+	const size_t range_l = threadpool->params.parallelize_4d_tile_2d.range_l;
+	const size_t range_k = threadpool->params.parallelize_4d_tile_2d.range_k;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j, start_k, start_l, min(range_k - start_k, tile_k), min(range_l - start_l, tile_l));
+		start_l += tile_l;
+		if (start_l >= range_l) {
+			start_l = 0;
+			start_k += tile_k;
+			if (start_k >= range_k) {
+				start_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_ij_kl = fxdiv_divide_size_t(linear_index, tile_range_kl);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_kl.quotient, range_j);
+			const struct fxdiv_result_size_t tile_index_k_l = fxdiv_divide_size_t(tile_index_ij_kl.remainder, tile_range_l);
+			const size_t start_k = tile_index_k_l.quotient * tile_k;
+			const size_t start_l = tile_index_k_l.remainder * tile_l;
+			task(argument, index_i_j.quotient, index_i_j.remainder, start_k, start_l, min(range_k - start_k, tile_k), min(range_l - start_l, tile_l));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_4d_tile_2d_with_uarch_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_4d_tile_2d_with_id_t task = (pthreadpool_task_4d_tile_2d_with_id_t) pthreadpool_load_relaxed_void_p(&threadpool->task);
+	void *const argument = pthreadpool_load_relaxed_void_p(&threadpool->argument);
+
+	const uint32_t default_uarch_index = threadpool->params.parallelize_4d_tile_2d_with_uarch.default_uarch_index;
+	uint32_t uarch_index = default_uarch_index;
+	#if PTHREADPOOL_USE_CPUINFO
+		uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
+		if (uarch_index > threadpool->params.parallelize_4d_tile_2d_with_uarch.max_uarch_index) {
+			uarch_index = default_uarch_index;
+		}
+	#endif
+
+	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_kl = threadpool->params.parallelize_4d_tile_2d_with_uarch.tile_range_kl;
+	const struct fxdiv_result_size_t tile_index_ij_kl = fxdiv_divide_size_t(range_start, tile_range_kl);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_4d_tile_2d_with_uarch.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_kl.quotient, range_j);
+	const struct fxdiv_divisor_size_t tile_range_l = threadpool->params.parallelize_4d_tile_2d_with_uarch.tile_range_l;
+	const struct fxdiv_result_size_t tile_index_k_l = fxdiv_divide_size_t(tile_index_ij_kl.remainder, tile_range_l);
+	const size_t tile_k = threadpool->params.parallelize_4d_tile_2d_with_uarch.tile_k;
+	const size_t tile_l = threadpool->params.parallelize_4d_tile_2d_with_uarch.tile_l;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t start_k = tile_index_k_l.quotient * tile_k;
+	size_t start_l = tile_index_k_l.remainder * tile_l;
+
+	const size_t range_l = threadpool->params.parallelize_4d_tile_2d_with_uarch.range_l;
+	const size_t range_k = threadpool->params.parallelize_4d_tile_2d_with_uarch.range_k;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, uarch_index, i, j, start_k, start_l, min(range_k - start_k, tile_k), min(range_l - start_l, tile_l));
+		start_l += tile_l;
+		if (start_l >= range_l) {
+			start_l = 0;
+			start_k += tile_k;
+			if (start_k >= range_k) {
+				start_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_ij_kl = fxdiv_divide_size_t(linear_index, tile_range_kl);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_kl.quotient, range_j);
+			const struct fxdiv_result_size_t tile_index_k_l = fxdiv_divide_size_t(tile_index_ij_kl.remainder, tile_range_l);
+			const size_t start_k = tile_index_k_l.quotient * tile_k;
+			const size_t start_l = tile_index_k_l.remainder * tile_l;
+			task(argument, uarch_index, index_i_j.quotient, index_i_j.remainder, start_k, start_l, min(range_k - start_k, tile_k), min(range_l - start_l, tile_l));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_5d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_5d_t task = (pthreadpool_task_5d_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_lm = threadpool->params.parallelize_5d.range_lm;
+	const struct fxdiv_result_size_t index_ijk_lm = fxdiv_divide_size_t(range_start, range_lm);
+	const struct fxdiv_divisor_size_t range_k = threadpool->params.parallelize_5d.range_k;
+	const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(index_ijk_lm.quotient, range_k);
+	const struct fxdiv_divisor_size_t range_m = threadpool->params.parallelize_5d.range_m;
+	const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(index_ijk_lm.remainder, range_m);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_5d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+	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;
+
+	const size_t range_l = threadpool->params.parallelize_5d.range_l;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j, k, l, m);
+		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_lm = fxdiv_divide_size_t(linear_index, range_lm);
+			const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(index_ijk_lm.quotient, range_k);
+			const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(index_ijk_lm.remainder, range_m);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_ij_k.remainder, index_l_m.quotient, index_l_m.remainder);
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_5d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_5d_tile_1d_t task = (pthreadpool_task_5d_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_m = threadpool->params.parallelize_5d_tile_1d.tile_range_m;
+	const struct fxdiv_result_size_t tile_index_ijkl_m = fxdiv_divide_size_t(range_start, tile_range_m);
+	const struct fxdiv_divisor_size_t range_kl = threadpool->params.parallelize_5d_tile_1d.range_kl;
+	const struct fxdiv_result_size_t index_ij_kl = fxdiv_divide_size_t(tile_index_ijkl_m.quotient, range_kl);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_5d_tile_1d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_kl.quotient, range_j);
+	const struct fxdiv_divisor_size_t range_l = threadpool->params.parallelize_5d_tile_1d.range_l;
+	const struct fxdiv_result_size_t index_k_l = fxdiv_divide_size_t(index_ij_kl.remainder, range_l);
+	const size_t tile_m = threadpool->params.parallelize_5d_tile_1d.tile_m;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_k_l.quotient;
+	size_t l = index_k_l.remainder;
+	size_t start_m = tile_index_ijkl_m.remainder * tile_m;
+
+	const size_t range_m = threadpool->params.parallelize_5d_tile_1d.range_m;
+	const size_t range_k = threadpool->params.parallelize_5d_tile_1d.range_k;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j, k, l, start_m, min(range_m - start_m, tile_m));
+		start_m += tile_m;
+		if (start_m >= range_m) {
+			start_m = 0;
+			if (++l == range_l.value) {
+				l = 0;
+				if (++k == range_k) {
+					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_ijkl_m = fxdiv_divide_size_t(linear_index, tile_range_m);
+			const struct fxdiv_result_size_t index_ij_kl = fxdiv_divide_size_t(tile_index_ijkl_m.quotient, range_kl);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_kl.quotient, range_j);
+			const struct fxdiv_result_size_t index_k_l = fxdiv_divide_size_t(index_ij_kl.remainder, range_l);
+			size_t start_m = tile_index_ijkl_m.remainder * tile_m;
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_k_l.quotient, index_k_l.remainder, start_m,
+				min(range_m - start_m, tile_m));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_5d_tile_2d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_5d_tile_2d_t task = (pthreadpool_task_5d_tile_2d_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_lm = threadpool->params.parallelize_5d_tile_2d.tile_range_lm;
+	const struct fxdiv_result_size_t tile_index_ijk_lm = fxdiv_divide_size_t(range_start, tile_range_lm);
+	const struct fxdiv_divisor_size_t range_k = threadpool->params.parallelize_5d_tile_2d.range_k;
+	const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(tile_index_ijk_lm.quotient, range_k);
+	const struct fxdiv_divisor_size_t tile_range_m = threadpool->params.parallelize_5d_tile_2d.tile_range_m;
+	const struct fxdiv_result_size_t tile_index_l_m = fxdiv_divide_size_t(tile_index_ijk_lm.remainder, tile_range_m);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_5d_tile_2d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+	const size_t tile_l = threadpool->params.parallelize_5d_tile_2d.tile_l;
+	const size_t tile_m = threadpool->params.parallelize_5d_tile_2d.tile_m;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_ij_k.remainder;
+	size_t start_l = tile_index_l_m.quotient * tile_l;
+	size_t start_m = tile_index_l_m.remainder * tile_m;
+
+	const size_t range_m = threadpool->params.parallelize_5d_tile_2d.range_m;
+	const size_t range_l = threadpool->params.parallelize_5d_tile_2d.range_l;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j, k, start_l, start_m, min(range_l - start_l, tile_l), min(range_m - start_m, tile_m));
+		start_m += tile_m;
+		if (start_m >= range_m) {
+			start_m = 0;
+			start_l += tile_l;
+			if (start_l >= range_l) {
+				start_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_lm = fxdiv_divide_size_t(linear_index, tile_range_lm);
+			const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(tile_index_ijk_lm.quotient, range_k);
+			const struct fxdiv_result_size_t tile_index_l_m = fxdiv_divide_size_t(tile_index_ijk_lm.remainder, tile_range_m);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+			const size_t start_l = tile_index_l_m.quotient * tile_l;
+			const size_t start_m = tile_index_l_m.remainder * tile_m;
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_ij_k.remainder,
+				start_l, start_m, min(range_l - start_l, tile_l), min(range_m - start_m, tile_m));
+		}
+	}
+
+	/* 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)
+{
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_6d_tile_2d_t task = (pthreadpool_task_6d_tile_2d_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_mn = threadpool->params.parallelize_6d_tile_2d.tile_range_mn;
+	const struct fxdiv_result_size_t tile_index_ijkl_mn = fxdiv_divide_size_t(range_start, tile_range_mn);
+	const struct fxdiv_divisor_size_t range_kl = threadpool->params.parallelize_6d_tile_2d.range_kl;
+	const struct fxdiv_result_size_t index_ij_kl = fxdiv_divide_size_t(tile_index_ijkl_mn.quotient, range_kl);
+	const struct fxdiv_divisor_size_t tile_range_n = threadpool->params.parallelize_6d_tile_2d.tile_range_n;
+	const struct fxdiv_result_size_t tile_index_m_n = fxdiv_divide_size_t(tile_index_ijkl_mn.remainder, tile_range_n);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_6d_tile_2d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_kl.quotient, range_j);
+	const struct fxdiv_divisor_size_t range_l = threadpool->params.parallelize_6d_tile_2d.range_l;
+	const struct fxdiv_result_size_t index_k_l = fxdiv_divide_size_t(index_ij_kl.remainder, range_l);
+	const size_t tile_m = threadpool->params.parallelize_6d_tile_2d.tile_m;
+	const size_t tile_n = threadpool->params.parallelize_6d_tile_2d.tile_n;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_k_l.quotient;
+	size_t l = index_k_l.remainder;
+	size_t start_m = tile_index_m_n.quotient * tile_m;
+	size_t start_n = tile_index_m_n.remainder * tile_n;
+
+	const size_t range_n = threadpool->params.parallelize_6d_tile_2d.range_n;
+	const size_t range_m = threadpool->params.parallelize_6d_tile_2d.range_m;
+	const size_t range_k = threadpool->params.parallelize_6d_tile_2d.range_k;
+	while (pthreadpool_decrement_fetch_relaxed_size_t(&thread->range_length) < range_threshold) {
+		task(argument, i, j, k, l, start_m, start_n, min(range_m - start_m, tile_m), min(range_n - start_n, tile_n));
+		start_n += tile_n;
+		if (start_n >= range_n) {
+			start_n = 0;
+			start_m += tile_m;
+			if (start_m >= range_m) {
+				start_m = 0;
+				if (++l == range_l.value) {
+					l = 0;
+					if (++k == range_k) {
+						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_ijkl_mn = fxdiv_divide_size_t(linear_index, tile_range_mn);
+			const struct fxdiv_result_size_t index_ij_kl = fxdiv_divide_size_t(tile_index_ijkl_mn.quotient, range_kl);
+			const struct fxdiv_result_size_t tile_index_m_n = fxdiv_divide_size_t(tile_index_ijkl_mn.remainder, tile_range_n);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_kl.quotient, range_j);
+			const struct fxdiv_result_size_t index_k_l = fxdiv_divide_size_t(index_ij_kl.remainder, range_l);
+			const size_t start_m = tile_index_m_n.quotient * tile_m;
+			const size_t start_n = tile_index_m_n.remainder * tile_n;
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_k_l.quotient, index_k_l.remainder,
+				start_m, start_n, min(range_m - start_m, tile_m), min(range_n - start_n, tile_n));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
diff --git a/src/portable-api.c b/src/portable-api.c
index 84d6eda..ef36578 100644
--- a/src/portable-api.c
+++ b/src/portable-api.c
@@ -35,6 +35,7 @@ static void thread_parallelize_1d(struct pthreadpool* threadpool, struct thread_
 
 	const pthreadpool_task_1d_t task = (pthreadpool_task_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 */
 	size_t range_start = pthreadpool_load_relaxed_size_t(&thread->range_start);
 	while (pthreadpool_try_decrement_relaxed_size_t(&thread->range_length)) {
@@ -69,7 +70,7 @@ static void thread_parallelize_1d_with_uarch(struct pthreadpool* threadpool, str
 	const uint32_t default_uarch_index = threadpool->params.parallelize_1d_with_uarch.default_uarch_index;
 	uint32_t uarch_index = default_uarch_index;
 	#if PTHREADPOOL_USE_CPUINFO
-		uarch_index = cpuinfo_get_current_uarch_index();
+		uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
 		if (uarch_index > threadpool->params.parallelize_1d_with_uarch.max_uarch_index) {
 			uarch_index = default_uarch_index;
 		}
@@ -280,7 +281,7 @@ static void thread_parallelize_2d_tile_2d_with_uarch(struct pthreadpool* threadp
 	const uint32_t default_uarch_index = threadpool->params.parallelize_2d_tile_2d_with_uarch.default_uarch_index;
 	uint32_t uarch_index = default_uarch_index;
 	#if PTHREADPOOL_USE_CPUINFO
-		uarch_index = cpuinfo_get_current_uarch_index();
+		uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
 		if (uarch_index > threadpool->params.parallelize_2d_tile_2d_with_uarch.max_uarch_index) {
 			uarch_index = default_uarch_index;
 		}
@@ -327,6 +328,106 @@ static void thread_parallelize_2d_tile_2d_with_uarch(struct pthreadpool* threadp
 	pthreadpool_fence_release();
 }
 
+static void thread_parallelize_3d(struct pthreadpool* threadpool, struct thread_info* thread) {
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_3d_t task = (pthreadpool_task_3d_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_k = threadpool->params.parallelize_3d.range_k;
+	const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(range_start, range_k);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_3d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_ij_k.remainder;
+
+	while (pthreadpool_try_decrement_relaxed_size_t(&thread->range_length)) {
+		task(argument, i, j, k);
+		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_ij_k = fxdiv_divide_size_t(linear_index, range_k);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_ij_k.remainder);
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+static void thread_parallelize_3d_tile_1d(struct pthreadpool* threadpool, struct thread_info* thread) {
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_3d_tile_1d_t task = (pthreadpool_task_3d_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_k = threadpool->params.parallelize_3d_tile_1d.tile_range_k;
+	const struct fxdiv_result_size_t tile_index_ij_k = fxdiv_divide_size_t(range_start, tile_range_k);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_3d_tile_1d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_k.quotient, range_j);
+	const size_t tile_k = threadpool->params.parallelize_3d_tile_1d.tile_k;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t start_k = tile_index_ij_k.remainder * tile_k;
+
+	const size_t range_k = threadpool->params.parallelize_3d_tile_1d.range_k;
+	while (pthreadpool_try_decrement_relaxed_size_t(&thread->range_length)) {
+		task(argument, i, j, start_k, min(range_k - start_k, tile_k));
+		start_k += tile_k;
+		if (start_k >= range_k) {
+			start_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_ij_k = fxdiv_divide_size_t(linear_index, tile_range_k);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_k.quotient, range_j);
+			const size_t start_k = tile_index_ij_k.remainder * tile_k;
+			task(argument, index_i_j.quotient, index_i_j.remainder, start_k, min(range_k - start_k, tile_k));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
 static void thread_parallelize_3d_tile_2d(struct pthreadpool* threadpool, struct thread_info* thread) {
 	assert(threadpool != NULL);
 	assert(thread != NULL);
@@ -393,7 +494,7 @@ static void thread_parallelize_3d_tile_2d_with_uarch(struct pthreadpool* threadp
 	const uint32_t default_uarch_index = threadpool->params.parallelize_3d_tile_2d_with_uarch.default_uarch_index;
 	uint32_t uarch_index = default_uarch_index;
 	#if PTHREADPOOL_USE_CPUINFO
-		uarch_index = cpuinfo_get_current_uarch_index();
+		uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
 		if (uarch_index > threadpool->params.parallelize_3d_tile_2d_with_uarch.max_uarch_index) {
 			uarch_index = default_uarch_index;
 		}
@@ -448,6 +549,122 @@ static void thread_parallelize_3d_tile_2d_with_uarch(struct pthreadpool* threadp
 	pthreadpool_fence_release();
 }
 
+static void thread_parallelize_4d(struct pthreadpool* threadpool, struct thread_info* thread) {
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_4d_t task = (pthreadpool_task_4d_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_kl = threadpool->params.parallelize_4d.range_kl;
+	const struct fxdiv_result_size_t index_ij_kl = fxdiv_divide_size_t(range_start, range_kl);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_4d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_kl.quotient, range_j);
+	const struct fxdiv_divisor_size_t range_l = threadpool->params.parallelize_4d.range_l;
+	const struct fxdiv_result_size_t index_k_l = fxdiv_divide_size_t(index_ij_kl.remainder, range_l);
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_k_l.quotient;
+	size_t l = index_k_l.remainder;
+
+	const size_t range_k = threadpool->params.parallelize_4d.range_k;
+	while (pthreadpool_try_decrement_relaxed_size_t(&thread->range_length)) {
+		task(argument, i, j, k, l);
+		if (++l == range_l.value) {
+			l = 0;
+			if (++k == range_k) {
+				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_ij_kl = fxdiv_divide_size_t(linear_index, range_kl);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_kl.quotient, range_j);
+			const struct fxdiv_result_size_t index_k_l = fxdiv_divide_size_t(index_ij_kl.remainder, range_l);
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_k_l.quotient, index_k_l.remainder);
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+static void thread_parallelize_4d_tile_1d(struct pthreadpool* threadpool, struct thread_info* thread) {
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_4d_tile_1d_t task = (pthreadpool_task_4d_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_kl = threadpool->params.parallelize_4d_tile_1d.tile_range_kl;
+	const struct fxdiv_result_size_t tile_index_ij_kl = fxdiv_divide_size_t(range_start, tile_range_kl);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_4d_tile_1d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_kl.quotient, range_j);
+	const struct fxdiv_divisor_size_t tile_range_l = threadpool->params.parallelize_4d_tile_1d.tile_range_l;
+	const struct fxdiv_result_size_t tile_index_k_l = fxdiv_divide_size_t(tile_index_ij_kl.remainder, tile_range_l);
+	const size_t tile_l = threadpool->params.parallelize_4d_tile_1d.tile_l;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = tile_index_k_l.quotient;
+	size_t start_l = tile_index_k_l.remainder * tile_l;
+
+	const size_t range_k = threadpool->params.parallelize_4d_tile_1d.range_k;
+	const size_t range_l = threadpool->params.parallelize_4d_tile_1d.range_l;
+	while (pthreadpool_try_decrement_relaxed_size_t(&thread->range_length)) {
+		task(argument, i, j, k, start_l, min(range_l - start_l, tile_l));
+		start_l += tile_l;
+		if (start_l >= range_l) {
+			start_l = 0;
+			if (++k == range_k) {
+				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_ij_kl = fxdiv_divide_size_t(linear_index, tile_range_kl);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(tile_index_ij_kl.quotient, range_j);
+			const struct fxdiv_result_size_t tile_index_k_l = fxdiv_divide_size_t(tile_index_ij_kl.remainder, tile_range_l);
+			const size_t start_l = tile_index_k_l.remainder * tile_l;
+			task(argument, index_i_j.quotient, index_i_j.remainder, tile_index_k_l.quotient, start_l, min(range_l - start_l, tile_l));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
 static void thread_parallelize_4d_tile_2d(struct pthreadpool* threadpool, struct thread_info* thread) {
 	assert(threadpool != NULL);
 	assert(thread != NULL);
@@ -521,7 +738,7 @@ static void thread_parallelize_4d_tile_2d_with_uarch(struct pthreadpool* threadp
 	const uint32_t default_uarch_index = threadpool->params.parallelize_4d_tile_2d_with_uarch.default_uarch_index;
 	uint32_t uarch_index = default_uarch_index;
 	#if PTHREADPOOL_USE_CPUINFO
-		uarch_index = cpuinfo_get_current_uarch_index();
+		uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
 		if (uarch_index > threadpool->params.parallelize_4d_tile_2d_with_uarch.max_uarch_index) {
 			uarch_index = default_uarch_index;
 		}
@@ -583,6 +800,137 @@ static void thread_parallelize_4d_tile_2d_with_uarch(struct pthreadpool* threadp
 	pthreadpool_fence_release();
 }
 
+static void thread_parallelize_5d(struct pthreadpool* threadpool, struct thread_info* thread) {
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_5d_t task = (pthreadpool_task_5d_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_lm = threadpool->params.parallelize_5d.range_lm;
+	const struct fxdiv_result_size_t index_ijk_lm = fxdiv_divide_size_t(range_start, range_lm);
+	const struct fxdiv_divisor_size_t range_k = threadpool->params.parallelize_5d.range_k;
+	const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(index_ijk_lm.quotient, range_k);
+	const struct fxdiv_divisor_size_t range_m = threadpool->params.parallelize_5d.range_m;
+	const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(index_ijk_lm.remainder, range_m);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_5d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+	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;
+
+	const size_t range_l = threadpool->params.parallelize_5d.range_l;
+	while (pthreadpool_try_decrement_relaxed_size_t(&thread->range_length)) {
+		task(argument, i, j, k, l, m);
+		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_lm = fxdiv_divide_size_t(linear_index, range_lm);
+			const struct fxdiv_result_size_t index_ij_k = fxdiv_divide_size_t(index_ijk_lm.quotient, range_k);
+			const struct fxdiv_result_size_t index_l_m = fxdiv_divide_size_t(index_ijk_lm.remainder, range_m);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_k.quotient, range_j);
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_ij_k.remainder, index_l_m.quotient, index_l_m.remainder);
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
+static void thread_parallelize_5d_tile_1d(struct pthreadpool* threadpool, struct thread_info* thread) {
+	assert(threadpool != NULL);
+	assert(thread != NULL);
+
+	const pthreadpool_task_5d_tile_1d_t task = (pthreadpool_task_5d_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_m = threadpool->params.parallelize_5d_tile_1d.tile_range_m;
+	const struct fxdiv_result_size_t tile_index_ijkl_m = fxdiv_divide_size_t(range_start, tile_range_m);
+	const struct fxdiv_divisor_size_t range_kl = threadpool->params.parallelize_5d_tile_1d.range_kl;
+	const struct fxdiv_result_size_t index_ij_kl = fxdiv_divide_size_t(tile_index_ijkl_m.quotient, range_kl);
+	const struct fxdiv_divisor_size_t range_j = threadpool->params.parallelize_5d_tile_1d.range_j;
+	const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_kl.quotient, range_j);
+	const struct fxdiv_divisor_size_t range_l = threadpool->params.parallelize_5d_tile_1d.range_l;
+	const struct fxdiv_result_size_t index_k_l = fxdiv_divide_size_t(index_ij_kl.remainder, range_l);
+	const size_t tile_m = threadpool->params.parallelize_5d_tile_1d.tile_m;
+	size_t i = index_i_j.quotient;
+	size_t j = index_i_j.remainder;
+	size_t k = index_k_l.quotient;
+	size_t l = index_k_l.remainder;
+	size_t start_m = tile_index_ijkl_m.remainder * tile_m;
+
+	const size_t range_m = threadpool->params.parallelize_5d_tile_1d.range_m;
+	const size_t range_k = threadpool->params.parallelize_5d_tile_1d.range_k;
+	while (pthreadpool_try_decrement_relaxed_size_t(&thread->range_length)) {
+		task(argument, i, j, k, l, start_m, min(range_m - start_m, tile_m));
+		start_m += tile_m;
+		if (start_m >= range_m) {
+			start_m = 0;
+			if (++l == range_l.value) {
+				l = 0;
+				if (++k == range_k) {
+					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_ijkl_m = fxdiv_divide_size_t(linear_index, tile_range_m);
+			const struct fxdiv_result_size_t index_ij_kl = fxdiv_divide_size_t(tile_index_ijkl_m.quotient, range_kl);
+			const struct fxdiv_result_size_t index_i_j = fxdiv_divide_size_t(index_ij_kl.quotient, range_j);
+			const struct fxdiv_result_size_t index_k_l = fxdiv_divide_size_t(index_ij_kl.remainder, range_l);
+			size_t start_m = tile_index_ijkl_m.remainder * tile_m;
+			task(argument, index_i_j.quotient, index_i_j.remainder, index_k_l.quotient, index_k_l.remainder, start_m,
+				min(range_m - start_m, tile_m));
+		}
+	}
+
+	/* Make changes by this thread visible to other threads */
+	pthreadpool_fence_release();
+}
+
 static void thread_parallelize_5d_tile_2d(struct pthreadpool* threadpool, struct thread_info* thread) {
 	assert(threadpool != NULL);
 	assert(thread != NULL);
@@ -740,7 +1088,8 @@ void pthreadpool_parallelize_1d(
 	size_t range,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || range <= 1) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || range <= 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) {
@@ -754,8 +1103,15 @@ void pthreadpool_parallelize_1d(
 			set_fpu_state(saved_fpu_state);
 		}
 	} else {
+		thread_function_t parallelize_1d = &thread_parallelize_1d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (range < range_threshold) {
+				parallelize_1d = &pthreadpool_thread_parallelize_1d_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_1d, NULL, 0,
+			threadpool, parallelize_1d, NULL, 0,
 			(void*) task, argument, range, flags);
 	}
 }
@@ -769,12 +1125,13 @@ void pthreadpool_parallelize_1d_with_uarch(
 	size_t range,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || range <= 1) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || range <= 1) {
 		/* No thread pool used: execute task sequentially on the calling thread */
 
 		uint32_t uarch_index = default_uarch_index;
 		#if PTHREADPOOL_USE_CPUINFO
-			uarch_index = cpuinfo_get_current_uarch_index();
+			uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
 			if (uarch_index > max_uarch_index) {
 				uarch_index = default_uarch_index;
 			}
@@ -796,8 +1153,15 @@ void pthreadpool_parallelize_1d_with_uarch(
 			.default_uarch_index = default_uarch_index,
 			.max_uarch_index = max_uarch_index,
 		};
+		thread_function_t parallelize_1d_with_uarch = &thread_parallelize_1d_with_uarch;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (range < range_threshold) {
+				parallelize_1d_with_uarch = &pthreadpool_thread_parallelize_1d_with_uarch_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_1d_with_uarch, &params, sizeof(params),
+			threadpool, parallelize_1d_with_uarch, &params, sizeof(params),
 			task, argument, range, flags);
 	}
 }
@@ -810,7 +1174,8 @@ void pthreadpool_parallelize_1d_tile_1d(
 	size_t tile,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || range <= tile) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || range <= tile) {
 		/* No thread pool used: execute task sequentially on the calling thread */
 		struct fpu_state saved_fpu_state = { 0 };
 		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
@@ -824,13 +1189,21 @@ void pthreadpool_parallelize_1d_tile_1d(
 			set_fpu_state(saved_fpu_state);
 		}
 	} else {
+		const size_t tile_range = divide_round_up(range, tile);
 		const struct pthreadpool_1d_tile_1d_params params = {
 			.range = range,
 			.tile = tile,
 		};
+		thread_function_t parallelize_1d_tile_1d = &thread_parallelize_1d_tile_1d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (range < range_threshold) {
+				parallelize_1d_tile_1d = &pthreadpool_thread_parallelize_1d_tile_1d_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_1d_tile_1d, &params, sizeof(params),
-			task, argument, divide_round_up(range, tile), flags);
+			threadpool, parallelize_1d_tile_1d, &params, sizeof(params),
+			task, argument, tile_range, flags);
 	}
 }
 
@@ -842,7 +1215,8 @@ void pthreadpool_parallelize_2d(
 	size_t range_j,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || (range_i | range_j) <= 1) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || (range_i | range_j) <= 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) {
@@ -858,12 +1232,20 @@ void pthreadpool_parallelize_2d(
 			set_fpu_state(saved_fpu_state);
 		}
 	} else {
+		const size_t range = range_i * range_j;
 		const struct pthreadpool_2d_params params = {
 			.range_j = fxdiv_init_size_t(range_j),
 		};
+		thread_function_t parallelize_2d = &thread_parallelize_2d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (range < range_threshold) {
+				parallelize_2d = &pthreadpool_thread_parallelize_2d_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_2d, &params, sizeof(params),
-			task, argument, range_i * range_j, flags);
+			threadpool, parallelize_2d, &params, sizeof(params),
+			task, argument, range, flags);
 	}
 }
 
@@ -876,7 +1258,8 @@ void pthreadpool_parallelize_2d_tile_1d(
 	size_t tile_j,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || (range_i <= 1 && range_j <= tile_j)) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || (range_i <= 1 && range_j <= tile_j)) {
 		/* No thread pool used: execute task sequentially on the calling thread */
 		struct fpu_state saved_fpu_state = { 0 };
 		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
@@ -893,14 +1276,22 @@ void pthreadpool_parallelize_2d_tile_1d(
 		}
 	} else {
 		const size_t tile_range_j = divide_round_up(range_j, tile_j);
+		const size_t tile_range = range_i * tile_range_j;
 		const struct pthreadpool_2d_tile_1d_params params = {
 			.range_j = range_j,
 			.tile_j = tile_j,
 			.tile_range_j = fxdiv_init_size_t(tile_range_j),
 		};
+		thread_function_t parallelize_2d_tile_1d = &thread_parallelize_2d_tile_1d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_2d_tile_1d = &pthreadpool_thread_parallelize_2d_tile_1d_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_2d_tile_1d, &params, sizeof(params),
-			task, argument, range_i * tile_range_j, flags);
+			threadpool, parallelize_2d_tile_1d, &params, sizeof(params),
+			task, argument, tile_range, flags);
 	}
 }
 
@@ -914,7 +1305,8 @@ void pthreadpool_parallelize_2d_tile_2d(
 	size_t tile_j,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || (range_i <= tile_i && range_j <= tile_j)) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || (range_i <= tile_i && range_j <= tile_j)) {
 		/* No thread pool used: execute task sequentially on the calling thread */
 		struct fpu_state saved_fpu_state = { 0 };
 		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
@@ -932,6 +1324,7 @@ void pthreadpool_parallelize_2d_tile_2d(
 	} else {
 		const size_t tile_range_i = divide_round_up(range_i, tile_i);
 		const size_t tile_range_j = divide_round_up(range_j, tile_j);
+		const size_t tile_range = tile_range_i * tile_range_j;
 		const struct pthreadpool_2d_tile_2d_params params = {
 			.range_i = range_i,
 			.tile_i = tile_i,
@@ -939,9 +1332,16 @@ void pthreadpool_parallelize_2d_tile_2d(
 			.tile_j = tile_j,
 			.tile_range_j = fxdiv_init_size_t(tile_range_j),
 		};
+		thread_function_t parallelize_2d_tile_2d = &thread_parallelize_2d_tile_2d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_2d_tile_2d = &pthreadpool_thread_parallelize_2d_tile_2d_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_2d_tile_2d, &params, sizeof(params),
-			task, argument, tile_range_i * tile_range_j, flags);
+			threadpool, parallelize_2d_tile_2d, &params, sizeof(params),
+			task, argument, tile_range, flags);
 	}
 }
 
@@ -957,12 +1357,13 @@ void pthreadpool_parallelize_2d_tile_2d_with_uarch(
 	size_t tile_j,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || (range_i <= tile_i && range_j <= tile_j)) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || (range_i <= tile_i && range_j <= tile_j)) {
 		/* No thread pool used: execute task sequentially on the calling thread */
 
 		uint32_t uarch_index = default_uarch_index;
 		#if PTHREADPOOL_USE_CPUINFO
-			uarch_index = cpuinfo_get_current_uarch_index();
+			uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
 			if (uarch_index > max_uarch_index) {
 				uarch_index = default_uarch_index;
 			}
@@ -984,6 +1385,7 @@ void pthreadpool_parallelize_2d_tile_2d_with_uarch(
 	} else {
 		const size_t tile_range_i = divide_round_up(range_i, tile_i);
 		const size_t tile_range_j = divide_round_up(range_j, tile_j);
+		const size_t tile_range = tile_range_i * tile_range_j;
 		const struct pthreadpool_2d_tile_2d_with_uarch_params params = {
 			.default_uarch_index = default_uarch_index,
 			.max_uarch_index = max_uarch_index,
@@ -993,9 +1395,112 @@ void pthreadpool_parallelize_2d_tile_2d_with_uarch(
 			.tile_j = tile_j,
 			.tile_range_j = fxdiv_init_size_t(tile_range_j),
 		};
+		thread_function_t parallelize_2d_tile_2d_with_uarch = &thread_parallelize_2d_tile_2d_with_uarch;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_2d_tile_2d_with_uarch = &pthreadpool_thread_parallelize_2d_tile_2d_with_uarch_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_2d_tile_2d_with_uarch, &params, sizeof(params),
-			task, argument, tile_range_i * tile_range_j, flags);
+			threadpool, parallelize_2d_tile_2d_with_uarch, &params, sizeof(params),
+			task, argument, tile_range, flags);
+	}
+}
+
+void pthreadpool_parallelize_3d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_3d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	uint32_t flags)
+{
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || (range_i | range_j | range_k) <= 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++) {
+					task(argument, i, j, k);
+				}
+			}
+		}
+		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
+			set_fpu_state(saved_fpu_state);
+		}
+	} else {
+		const size_t range = range_i * range_j * range_k;
+		const struct pthreadpool_3d_params params = {
+			.range_j = fxdiv_init_size_t(range_j),
+			.range_k = fxdiv_init_size_t(range_k),
+		};
+		thread_function_t parallelize_3d = &thread_parallelize_3d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (range < range_threshold) {
+				parallelize_3d = &pthreadpool_thread_parallelize_3d_fastpath;
+			}
+		#endif
+		pthreadpool_parallelize(
+			threadpool, parallelize_3d, &params, sizeof(params),
+			task, argument, range, flags);
+	}
+}
+
+void pthreadpool_parallelize_3d_tile_1d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_3d_tile_1d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t tile_k,
+	uint32_t flags)
+{
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || ((range_i | range_j) <= 1 && range_k <= tile_k)) {
+		/* 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 += tile_k) {
+					task(argument, i, j, k, min(range_k - k, tile_k));
+				}
+			}
+		}
+		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
+			set_fpu_state(saved_fpu_state);
+		}
+	} else {
+		const size_t tile_range_k = divide_round_up(range_k, tile_k);
+		const size_t tile_range = range_i * range_j * tile_range_k;
+		const struct pthreadpool_3d_tile_1d_params params = {
+			.range_k = range_k,
+			.tile_k = tile_k,
+			.range_j = fxdiv_init_size_t(range_j),
+			.tile_range_k = fxdiv_init_size_t(tile_range_k),
+		};
+		thread_function_t parallelize_3d_tile_1d = &thread_parallelize_3d_tile_1d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_3d_tile_1d = &pthreadpool_thread_parallelize_3d_tile_1d_fastpath;
+			}
+		#endif
+		pthreadpool_parallelize(
+			threadpool, parallelize_3d_tile_1d, &params, sizeof(params),
+			task, argument, tile_range, flags);
 	}
 }
 
@@ -1010,7 +1515,8 @@ void pthreadpool_parallelize_3d_tile_2d(
 	size_t tile_k,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || (range_i <= 1 && range_j <= tile_j && range_k <= tile_k)) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || (range_i <= 1 && range_j <= tile_j && range_k <= tile_k)) {
 		/* No thread pool used: execute task sequentially on the calling thread */
 		struct fpu_state saved_fpu_state = { 0 };
 		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
@@ -1030,6 +1536,7 @@ void pthreadpool_parallelize_3d_tile_2d(
 	} else {
 		const size_t tile_range_j = divide_round_up(range_j, tile_j);
 		const size_t tile_range_k = divide_round_up(range_k, tile_k);
+		const size_t tile_range = range_i * tile_range_j * tile_range_k;
 		const struct pthreadpool_3d_tile_2d_params params = {
 			.range_j = range_j,
 			.tile_j = tile_j,
@@ -1038,9 +1545,16 @@ void pthreadpool_parallelize_3d_tile_2d(
 			.tile_range_j = fxdiv_init_size_t(tile_range_j),
 			.tile_range_k = fxdiv_init_size_t(tile_range_k),
 		};
+		thread_function_t parallelize_3d_tile_2d = &thread_parallelize_3d_tile_2d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_3d_tile_2d = &pthreadpool_thread_parallelize_3d_tile_2d_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_3d_tile_2d, &params, sizeof(params),
-			task, argument, range_i * tile_range_j * tile_range_k, flags);
+			threadpool, parallelize_3d_tile_2d, &params, sizeof(params),
+			task, argument, tile_range, flags);
 	}
 }
 
@@ -1057,12 +1571,13 @@ void pthreadpool_parallelize_3d_tile_2d_with_uarch(
 	size_t tile_k,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || (range_i <= 1 && range_j <= tile_j && range_k <= tile_k)) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || (range_i <= 1 && range_j <= tile_j && range_k <= tile_k)) {
 		/* No thread pool used: execute task sequentially on the calling thread */
 
 		uint32_t uarch_index = default_uarch_index;
 		#if PTHREADPOOL_USE_CPUINFO
-			uarch_index = cpuinfo_get_current_uarch_index();
+			uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
 			if (uarch_index > max_uarch_index) {
 				uarch_index = default_uarch_index;
 			}
@@ -1086,6 +1601,7 @@ void pthreadpool_parallelize_3d_tile_2d_with_uarch(
 	} else {
 		const size_t tile_range_j = divide_round_up(range_j, tile_j);
 		const size_t tile_range_k = divide_round_up(range_k, tile_k);
+		const size_t tile_range = range_i * tile_range_j * tile_range_k;
 		const struct pthreadpool_3d_tile_2d_with_uarch_params params = {
 			.default_uarch_index = default_uarch_index,
 			.max_uarch_index = max_uarch_index,
@@ -1096,9 +1612,124 @@ void pthreadpool_parallelize_3d_tile_2d_with_uarch(
 			.tile_range_j = fxdiv_init_size_t(tile_range_j),
 			.tile_range_k = fxdiv_init_size_t(tile_range_k),
 		};
+		thread_function_t parallelize_3d_tile_2d_with_uarch = &thread_parallelize_3d_tile_2d_with_uarch;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_3d_tile_2d_with_uarch = &pthreadpool_thread_parallelize_3d_tile_2d_with_uarch_fastpath;
+			}
+		#endif
+		pthreadpool_parallelize(
+			threadpool, parallelize_3d_tile_2d_with_uarch, &params, sizeof(params),
+			task, argument, tile_range, flags);
+	}
+}
+
+void pthreadpool_parallelize_4d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_4d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	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) <= 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++) {
+						task(argument, i, j, k, l);
+					}
+				}
+			}
+		}
+		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
+			set_fpu_state(saved_fpu_state);
+		}
+	} else {
+		const size_t range_kl = range_k * range_l;
+		const size_t range = range_i * range_j * range_kl;
+		const struct pthreadpool_4d_params params = {
+			.range_k = range_k,
+			.range_j = fxdiv_init_size_t(range_j),
+			.range_kl = fxdiv_init_size_t(range_kl),
+			.range_l = fxdiv_init_size_t(range_l),
+		};
+		thread_function_t parallelize_4d = &thread_parallelize_4d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (range < range_threshold) {
+				parallelize_4d = &pthreadpool_thread_parallelize_4d_fastpath;
+			}
+		#endif
+		pthreadpool_parallelize(
+			threadpool, parallelize_4d, &params, sizeof(params),
+			task, argument, range, flags);
+	}
+}
+
+void pthreadpool_parallelize_4d_tile_1d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_4d_tile_1d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	size_t tile_l,
+	uint32_t flags)
+{
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || ((range_i | range_j | range_k) <= 1 && range_l <= tile_l)) {
+		/* 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 += tile_l) {
+						task(argument, i, j, k, l, min(range_l - l, tile_l));
+					}
+				}
+			}
+		}
+		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
+			set_fpu_state(saved_fpu_state);
+		}
+	} else {
+		const size_t tile_range_l = divide_round_up(range_l, tile_l);
+		const size_t tile_range_kl = range_k * tile_range_l;
+		const size_t tile_range = range_i * range_j * tile_range_kl;
+		const struct pthreadpool_4d_tile_1d_params params = {
+			.range_k = range_k,
+			.range_l = range_l,
+			.tile_l = tile_l,
+			.range_j = fxdiv_init_size_t(range_j),
+			.tile_range_kl = fxdiv_init_size_t(tile_range_kl),
+			.tile_range_l = fxdiv_init_size_t(tile_range_l),
+		};
+		thread_function_t parallelize_4d_tile_1d = &thread_parallelize_4d_tile_1d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_4d_tile_1d = &pthreadpool_thread_parallelize_4d_tile_1d_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_3d_tile_2d_with_uarch, &params, sizeof(params),
-			task, argument, range_i * tile_range_j * tile_range_k, flags);
+			threadpool, parallelize_4d_tile_1d, &params, sizeof(params),
+			task, argument, tile_range, flags);
 	}
 }
 
@@ -1114,7 +1745,8 @@ void pthreadpool_parallelize_4d_tile_2d(
 	size_t tile_l,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || ((range_i | range_j) <= 1 && range_k <= tile_k && range_l <= tile_l)) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || ((range_i | range_j) <= 1 && range_k <= tile_k && range_l <= tile_l)) {
 		/* No thread pool used: execute task sequentially on the calling thread */
 		struct fpu_state saved_fpu_state = { 0 };
 		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
@@ -1137,6 +1769,7 @@ void pthreadpool_parallelize_4d_tile_2d(
 	} else {
 		const size_t tile_range_l = divide_round_up(range_l, tile_l);
 		const size_t tile_range_kl = divide_round_up(range_k, tile_k) * tile_range_l;
+		const size_t tile_range = range_i * range_j * tile_range_kl;
 		const struct pthreadpool_4d_tile_2d_params params = {
 			.range_k = range_k,
 			.tile_k = tile_k,
@@ -1146,9 +1779,16 @@ void pthreadpool_parallelize_4d_tile_2d(
 			.tile_range_kl = fxdiv_init_size_t(tile_range_kl),
 			.tile_range_l = fxdiv_init_size_t(tile_range_l),
 		};
+		thread_function_t parallelize_4d_tile_2d = &thread_parallelize_4d_tile_2d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_4d_tile_2d = &pthreadpool_thread_parallelize_4d_tile_2d_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_4d_tile_2d, &params, sizeof(params),
-			task, argument, range_i * range_j * tile_range_kl, flags);
+			threadpool, parallelize_4d_tile_2d, &params, sizeof(params),
+			task, argument, tile_range, flags);
 	}
 }
 
@@ -1166,12 +1806,13 @@ void pthreadpool_parallelize_4d_tile_2d_with_uarch(
 	size_t tile_l,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || ((range_i | range_j) <= 1 && range_k <= tile_k && range_l <= tile_l)) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || ((range_i | range_j) <= 1 && range_k <= tile_k && range_l <= tile_l)) {
 		/* No thread pool used: execute task sequentially on the calling thread */
 
 		uint32_t uarch_index = default_uarch_index;
 		#if PTHREADPOOL_USE_CPUINFO
-			uarch_index = cpuinfo_get_current_uarch_index();
+			uarch_index = cpuinfo_get_current_uarch_index_with_default(default_uarch_index);
 			if (uarch_index > max_uarch_index) {
 				uarch_index = default_uarch_index;
 			}
@@ -1198,6 +1839,7 @@ void pthreadpool_parallelize_4d_tile_2d_with_uarch(
 	} else {
 		const size_t tile_range_l = divide_round_up(range_l, tile_l);
 		const size_t tile_range_kl = divide_round_up(range_k, tile_k) * tile_range_l;
+		const size_t tile_range = range_i * range_j * tile_range_kl;
 		const struct pthreadpool_4d_tile_2d_with_uarch_params params = {
 			.default_uarch_index = default_uarch_index,
 			.max_uarch_index = max_uarch_index,
@@ -1209,9 +1851,132 @@ void pthreadpool_parallelize_4d_tile_2d_with_uarch(
 			.tile_range_kl = fxdiv_init_size_t(tile_range_kl),
 			.tile_range_l = fxdiv_init_size_t(tile_range_l),
 		};
+		thread_function_t parallelize_4d_tile_2d_with_uarch = &thread_parallelize_4d_tile_2d_with_uarch;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_4d_tile_2d_with_uarch = &pthreadpool_thread_parallelize_4d_tile_2d_with_uarch_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_4d_tile_2d_with_uarch, &params, sizeof(params),
-			task, argument, range_i * range_j * tile_range_kl, flags);
+			threadpool, parallelize_4d_tile_2d_with_uarch, &params, sizeof(params),
+			task, argument, tile_range, flags);
+	}
+}
+
+void pthreadpool_parallelize_5d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_5d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	size_t range_m,
+	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) {
+		/* 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++) {
+							task(argument, i, j, k, l, m);
+						}
+					}
+				}
+			}
+		}
+		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
+			set_fpu_state(saved_fpu_state);
+		}
+	} else {
+		const size_t range_lm = range_l * range_m;
+		const size_t range = range_i * range_j * range_k * range_lm;
+		const struct pthreadpool_5d_params params = {
+			.range_l = range_l,
+			.range_j = fxdiv_init_size_t(range_j),
+			.range_k = fxdiv_init_size_t(range_k),
+			.range_lm = fxdiv_init_size_t(range_lm),
+			.range_m = fxdiv_init_size_t(range_m),
+		};
+		thread_function_t parallelize_5d = &thread_parallelize_5d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (range < range_threshold) {
+				parallelize_5d = &pthreadpool_thread_parallelize_5d_fastpath;
+			}
+		#endif
+		pthreadpool_parallelize(
+			threadpool, parallelize_5d, &params, sizeof(params),
+			task, argument, range, flags);
+	}
+}
+
+void pthreadpool_parallelize_5d_tile_1d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_5d_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 tile_m,
+	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) <= 1 && range_m <= tile_m)) {
+		/* 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 += tile_m) {
+							task(argument, i, j, k, l, m, min(range_m - m, tile_m));
+						}
+					}
+				}
+			}
+		}
+		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
+			set_fpu_state(saved_fpu_state);
+		}
+	} else {
+		const size_t tile_range_m = divide_round_up(range_m, tile_m);
+		const size_t range_kl = range_k * range_l;
+		const size_t tile_range = range_i * range_j * range_kl * tile_range_m;
+		const struct pthreadpool_5d_tile_1d_params params = {
+			.range_k = range_k,
+			.range_m = range_m,
+			.tile_m = tile_m,
+			.range_j = fxdiv_init_size_t(range_j),
+			.range_kl = fxdiv_init_size_t(range_kl),
+			.range_l = fxdiv_init_size_t(range_l),
+			.tile_range_m = fxdiv_init_size_t(tile_range_m),
+		};
+		thread_function_t parallelize_5d_tile_1d = &thread_parallelize_5d_tile_1d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_5d_tile_1d = &pthreadpool_thread_parallelize_5d_tile_1d_fastpath;
+			}
+		#endif
+		pthreadpool_parallelize(
+			threadpool, parallelize_5d_tile_1d, &params, sizeof(params),
+			task, argument, tile_range, flags);
 	}
 }
 
@@ -1228,7 +1993,8 @@ void pthreadpool_parallelize_5d_tile_2d(
 	size_t tile_m,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || ((range_i | range_j | range_k) <= 1 && range_l <= tile_l && range_m <= tile_m)) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || ((range_i | range_j | range_k) <= 1 && range_l <= tile_l && range_m <= tile_m)) {
 		/* No thread pool used: execute task sequentially on the calling thread */
 		struct fpu_state saved_fpu_state = { 0 };
 		if (flags & PTHREADPOOL_FLAG_DISABLE_DENORMALS) {
@@ -1253,6 +2019,7 @@ void pthreadpool_parallelize_5d_tile_2d(
 	} else {
 		const size_t tile_range_m = divide_round_up(range_m, tile_m);
 		const size_t tile_range_lm = divide_round_up(range_l, tile_l) * tile_range_m;
+		const size_t tile_range = range_i * range_j * range_k * tile_range_lm;
 		const struct pthreadpool_5d_tile_2d_params params = {
 			.range_l = range_l,
 			.tile_l = tile_l,
@@ -1263,9 +2030,16 @@ void pthreadpool_parallelize_5d_tile_2d(
 			.tile_range_lm = fxdiv_init_size_t(tile_range_lm),
 			.tile_range_m = fxdiv_init_size_t(tile_range_m),
 		};
+		thread_function_t parallelize_5d_tile_2d = &thread_parallelize_5d_tile_2d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_5d_tile_2d = &pthreadpool_thread_parallelize_5d_tile_2d_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_5d_tile_2d, &params, sizeof(params),
-			task, argument, range_i * range_j * range_k * tile_range_lm, flags);
+			threadpool, parallelize_5d_tile_2d, &params, sizeof(params),
+			task, argument, tile_range, flags);
 	}
 }
 
@@ -1283,7 +2057,8 @@ void pthreadpool_parallelize_6d_tile_2d(
 	size_t tile_n,
 	uint32_t flags)
 {
-	if (threadpool == NULL || threadpool->threads_count.value <= 1 || ((range_i | range_j | range_k | range_l) <= 1 && range_m <= tile_m && range_n <= tile_n)) {
+	size_t threads_count;
+	if (threadpool == NULL || (threads_count = threadpool->threads_count.value) <= 1 || ((range_i | range_j | range_k | range_l) <= 1 && range_m <= tile_m && 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) {
@@ -1311,6 +2086,7 @@ void pthreadpool_parallelize_6d_tile_2d(
 		const size_t range_kl = range_k * range_l;
 		const size_t tile_range_n = divide_round_up(range_n, tile_n);
 		const size_t tile_range_mn = divide_round_up(range_m, tile_m) * tile_range_n;
+		const size_t tile_range = range_i * range_j * range_kl * tile_range_mn;
 		const struct pthreadpool_6d_tile_2d_params params = {
 			.range_k = range_k,
 			.range_m = range_m,
@@ -1323,8 +2099,15 @@ void pthreadpool_parallelize_6d_tile_2d(
 			.tile_range_mn = fxdiv_init_size_t(tile_range_mn),
 			.tile_range_n = fxdiv_init_size_t(tile_range_n),
 		};
+		thread_function_t parallelize_6d_tile_2d = &thread_parallelize_6d_tile_2d;
+		#if PTHREADPOOL_USE_FASTPATH
+			const size_t range_threshold = -threads_count;
+			if (tile_range < range_threshold) {
+				parallelize_6d_tile_2d = &pthreadpool_thread_parallelize_6d_tile_2d_fastpath;
+			}
+		#endif
 		pthreadpool_parallelize(
-			threadpool, &thread_parallelize_6d_tile_2d, &params, sizeof(params),
-			task, argument, range_i * range_j * range_kl * tile_range_mn, flags);
+			threadpool, parallelize_6d_tile_2d, &params, sizeof(params),
+			task, argument, tile_range, flags);
 	}
 }
diff --git a/src/pthreads.c b/src/pthreads.c
index 2d945a0..430ca79 100644
--- a/src/pthreads.c
+++ b/src/pthreads.c
@@ -108,8 +108,7 @@ static void wait_worker_threads(struct pthreadpool* threadpool) {
 
 	/* Spin-wait */
 	for (uint32_t i = PTHREADPOOL_SPIN_WAIT_ITERATIONS; i != 0; i--) {
-		/* This fence serves as a sleep instruction */
-		pthreadpool_fence_acquire();
+		pthreadpool_yield();
 
 		#if PTHREADPOOL_USE_FUTEX
 			has_active_threads = pthreadpool_load_acquire_uint32_t(&threadpool->has_active_threads);
@@ -151,8 +150,7 @@ static uint32_t wait_for_new_command(
 	if ((last_flags & PTHREADPOOL_FLAG_YIELD_WORKERS) == 0) {
 		/* Spin-wait loop */
 		for (uint32_t i = PTHREADPOOL_SPIN_WAIT_ITERATIONS; i != 0; i--) {
-			/* This fence serves as a sleep instruction */
-			pthreadpool_fence_acquire();
+			pthreadpool_yield();
 
 			command = pthreadpool_load_acquire_uint32_t(&threadpool->command);
 			if (command != last_command) {
diff --git a/src/shim.c b/src/shim.c
index 7bf378c..e90ac45 100644
--- a/src/shim.c
+++ b/src/shim.c
@@ -133,6 +133,43 @@ void pthreadpool_parallelize_2d_tile_2d_with_uarch(
 	}
 }
 
+void pthreadpool_parallelize_3d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_3d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	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++) {
+				task(argument, i, j, k);
+			}
+		}
+	}
+}
+
+void pthreadpool_parallelize_3d_tile_1d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_3d_tile_1d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t tile_k,
+	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 += tile_k) {
+				task(argument, i, j, k, min(range_k - k, tile_k));
+			}
+		}
+	}
+}
+
 void pthreadpool_parallelize_3d_tile_2d(
 	pthreadpool_t threadpool,
 	pthreadpool_task_3d_tile_2d_t task,
@@ -177,6 +214,49 @@ void pthreadpool_parallelize_3d_tile_2d_with_uarch(
 	}
 }
 
+void pthreadpool_parallelize_4d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_4d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	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++) {
+					task(argument, i, j, k, l);
+				}
+			}
+		}
+	}
+}
+
+void pthreadpool_parallelize_4d_tile_1d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_4d_tile_1d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	size_t tile_l,
+	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 += tile_l) {
+					task(argument, i, j, k, l, min(range_l - l, tile_l));
+				}
+			}
+		}
+	}
+}
+
 void pthreadpool_parallelize_4d_tile_2d(
 	pthreadpool_t threadpool,
 	pthreadpool_task_4d_tile_2d_t task,
@@ -227,6 +307,55 @@ void pthreadpool_parallelize_4d_tile_2d_with_uarch(
 	}
 }
 
+void pthreadpool_parallelize_5d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_5d_t task,
+	void* argument,
+	size_t range_i,
+	size_t range_j,
+	size_t range_k,
+	size_t range_l,
+	size_t range_m,
+	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++) {
+						task(argument, i, j, k, l, m);
+					}
+				}
+			}
+		}
+	}
+}
+
+void pthreadpool_parallelize_5d_tile_1d(
+	pthreadpool_t threadpool,
+	pthreadpool_task_5d_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 tile_m,
+	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 += tile_m) {
+						task(argument, i, j, k, l, m, min(range_m - m, tile_m));
+					}
+				}
+			}
+		}
+	}
+}
+
 void pthreadpool_parallelize_5d_tile_2d(
 	pthreadpool_t threadpool,
 	pthreadpool_task_5d_tile_2d_t task,
diff --git a/src/threadpool-atomics.h b/src/threadpool-atomics.h
index 474d12b..f0ddd89 100644
--- a/src/threadpool-atomics.h
+++ b/src/threadpool-atomics.h
@@ -4,16 +4,23 @@
 #include <stddef.h>
 #include <stdint.h>
 
+/* SSE-specific headers */
+#if defined(__i386__) || defined(__i686__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
+	#include <xmmintrin.h>
+#endif
+
+/* ARM-specific headers */
+#if defined(__ARM_ACLE)
+	#include <arm_acle.h>
+#endif
+
 /* MSVC-specific headers */
 #ifdef _MSC_VER
 	#include <intrin.h>
-	#if defined(_M_IX86) || defined(_M_X64) || defined(_M_AMD64)
-		#include <immintrin.h>
-	#endif
 #endif
 
 
-#if defined(__wasm__) && defined(__EMSCRIPTEN_PTHREADS__) && defined(__clang__)
+#if defined(__wasm__) && defined(__clang__)
 	/*
 	 * Clang for WebAssembly target lacks stdatomic.h header,
 	 * even though it supports the necessary low-level intrinsics.
@@ -123,239 +130,239 @@
 	static inline void pthreadpool_fence_release() {
 		__c11_atomic_thread_fence(__ATOMIC_RELEASE);
 	}
-#elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_AMD64))
-	typedef volatile uint32_t pthreadpool_atomic_uint32_t;
-	typedef volatile size_t   pthreadpool_atomic_size_t;
-	typedef void *volatile    pthreadpool_atomic_void_p;
+#elif defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)
+	#include <stdatomic.h>
+
+	typedef _Atomic(uint32_t) pthreadpool_atomic_uint32_t;
+	typedef _Atomic(size_t)   pthreadpool_atomic_size_t;
+	typedef _Atomic(void*)    pthreadpool_atomic_void_p;
 
 	static inline uint32_t pthreadpool_load_relaxed_uint32_t(
 		pthreadpool_atomic_uint32_t* address)
 	{
-		return *address;
+		return atomic_load_explicit(address, memory_order_relaxed);
 	}
 
 	static inline size_t pthreadpool_load_relaxed_size_t(
 		pthreadpool_atomic_size_t* address)
 	{
-		return *address;
+		return atomic_load_explicit(address, memory_order_relaxed);
 	}
 
 	static inline void* pthreadpool_load_relaxed_void_p(
 		pthreadpool_atomic_void_p* address)
 	{
-		return *address;
+		return atomic_load_explicit(address, memory_order_relaxed);
 	}
 
 	static inline uint32_t pthreadpool_load_acquire_uint32_t(
 		pthreadpool_atomic_uint32_t* address)
 	{
-		/* x86-64 loads always have acquire semantics; use only a compiler barrier */
-		const uint32_t value = *address;
-		_ReadBarrier();
-		return value;
+		return atomic_load_explicit(address, memory_order_acquire);
 	}
 
 	static inline size_t pthreadpool_load_acquire_size_t(
 		pthreadpool_atomic_size_t* address)
 	{
-		/* x86-64 loads always have acquire semantics; use only a compiler barrier */
-		const size_t value = *address;
-		_ReadBarrier();
-		return value;
+		return atomic_load_explicit(address, memory_order_acquire);
 	}
 
 	static inline void pthreadpool_store_relaxed_uint32_t(
 		pthreadpool_atomic_uint32_t* address,
 		uint32_t value)
 	{
-		*address = value;
+		atomic_store_explicit(address, value, memory_order_relaxed);
 	}
 
 	static inline void pthreadpool_store_relaxed_size_t(
 		pthreadpool_atomic_size_t* address,
 		size_t value)
 	{
-		*address = value;
+		atomic_store_explicit(address, value, memory_order_relaxed);
 	}
 
 	static inline void pthreadpool_store_relaxed_void_p(
 		pthreadpool_atomic_void_p* address,
 		void* value)
 	{
-		*address = value;
+		atomic_store_explicit(address, value, memory_order_relaxed);
 	}
 
 	static inline void pthreadpool_store_release_uint32_t(
 		pthreadpool_atomic_uint32_t* address,
 		uint32_t value)
 	{
-		/* x86-64 stores always have release semantics; use only a compiler barrier */
-		_WriteBarrier();
-		*address = value;
+		atomic_store_explicit(address, value, memory_order_release);
 	}
 
 	static inline void pthreadpool_store_release_size_t(
 		pthreadpool_atomic_size_t* address,
 		size_t value)
 	{
-		/* x86-64 stores always have release semantics; use only a compiler barrier */
-		_WriteBarrier();
-		*address = value;
+		atomic_store_explicit(address, value, memory_order_release);
 	}
 
 	static inline size_t pthreadpool_decrement_fetch_relaxed_size_t(
 		pthreadpool_atomic_size_t* address)
 	{
-		return (size_t) _InterlockedDecrement64((volatile __int64*) address);
+		return atomic_fetch_sub_explicit(address, 1, memory_order_relaxed) - 1;
 	}
 
 	static inline size_t pthreadpool_decrement_fetch_release_size_t(
 		pthreadpool_atomic_size_t* address)
 	{
-		return (size_t) _InterlockedDecrement64((volatile __int64*) address);
+		return atomic_fetch_sub_explicit(address, 1, memory_order_release) - 1;
 	}
 
 	static inline bool pthreadpool_try_decrement_relaxed_size_t(
 		pthreadpool_atomic_size_t* value)
 	{
-		size_t actual_value = *value;
-		while (actual_value != 0) {
-			const size_t new_value = actual_value - 1;
-			const size_t expected_value = actual_value;
-			actual_value = _InterlockedCompareExchange64(
-				(volatile __int64*) value, (__int64) new_value, (__int64) expected_value);
-			if (actual_value == expected_value) {
-				return true;
+		#if defined(__clang__) && (defined(__arm__) || defined(__aarch64__))
+			size_t actual_value;
+			do {
+				actual_value = __builtin_arm_ldrex((const volatile size_t*) value);
+				if (actual_value == 0) {
+					__builtin_arm_clrex();
+					return false;
+				}
+			} while (__builtin_arm_strex(actual_value - 1, (volatile size_t*) value) != 0);
+			return true;
+		#else
+			size_t actual_value = pthreadpool_load_relaxed_size_t(value);
+			while (actual_value != 0) {
+				if (atomic_compare_exchange_weak_explicit(
+					value, &actual_value, actual_value - 1, memory_order_relaxed, memory_order_relaxed))
+				{
+					return true;
+				}
 			}
-		}
-		return false;
+			return false;
+		#endif
 	}
 
 	static inline void pthreadpool_fence_acquire() {
-		_mm_lfence();
-		_ReadBarrier();
+		atomic_thread_fence(memory_order_acquire);
 	}
 
 	static inline void pthreadpool_fence_release() {
-		_WriteBarrier();
-		_mm_sfence();
+		atomic_thread_fence(memory_order_release);
 	}
-#elif defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)
-	#include <stdatomic.h>
-
-	typedef _Atomic(uint32_t) pthreadpool_atomic_uint32_t;
-	typedef _Atomic(size_t)   pthreadpool_atomic_size_t;
-	typedef _Atomic(void*)    pthreadpool_atomic_void_p;
+#elif defined(_MSC_VER) && defined(_M_X64)
+	typedef volatile uint32_t pthreadpool_atomic_uint32_t;
+	typedef volatile size_t   pthreadpool_atomic_size_t;
+	typedef void *volatile    pthreadpool_atomic_void_p;
 
 	static inline uint32_t pthreadpool_load_relaxed_uint32_t(
 		pthreadpool_atomic_uint32_t* address)
 	{
-		return atomic_load_explicit(address, memory_order_relaxed);
+		return *address;
 	}
 
 	static inline size_t pthreadpool_load_relaxed_size_t(
 		pthreadpool_atomic_size_t* address)
 	{
-		return atomic_load_explicit(address, memory_order_relaxed);
+		return *address;
 	}
 
 	static inline void* pthreadpool_load_relaxed_void_p(
 		pthreadpool_atomic_void_p* address)
 	{
-		return atomic_load_explicit(address, memory_order_relaxed);
+		return *address;
 	}
 
 	static inline uint32_t pthreadpool_load_acquire_uint32_t(
 		pthreadpool_atomic_uint32_t* address)
 	{
-		return atomic_load_explicit(address, memory_order_acquire);
+		/* x86-64 loads always have acquire semantics; use only a compiler barrier */
+		const uint32_t value = *address;
+		_ReadBarrier();
+		return value;
 	}
 
 	static inline size_t pthreadpool_load_acquire_size_t(
 		pthreadpool_atomic_size_t* address)
 	{
-		return atomic_load_explicit(address, memory_order_acquire);
+		/* x86-64 loads always have acquire semantics; use only a compiler barrier */
+		const size_t value = *address;
+		_ReadBarrier();
+		return value;
 	}
 
 	static inline void pthreadpool_store_relaxed_uint32_t(
 		pthreadpool_atomic_uint32_t* address,
 		uint32_t value)
 	{
-		atomic_store_explicit(address, value, memory_order_relaxed);
+		*address = value;
 	}
 
 	static inline void pthreadpool_store_relaxed_size_t(
 		pthreadpool_atomic_size_t* address,
 		size_t value)
 	{
-		atomic_store_explicit(address, value, memory_order_relaxed);
+		*address = value;
 	}
 
 	static inline void pthreadpool_store_relaxed_void_p(
 		pthreadpool_atomic_void_p* address,
 		void* value)
 	{
-		atomic_store_explicit(address, value, memory_order_relaxed);
+		*address = value;
 	}
 
 	static inline void pthreadpool_store_release_uint32_t(
 		pthreadpool_atomic_uint32_t* address,
 		uint32_t value)
 	{
-		atomic_store_explicit(address, value, memory_order_release);
+		/* x86-64 stores always have release semantics; use only a compiler barrier */
+		_WriteBarrier();
+		*address = value;
 	}
 
 	static inline void pthreadpool_store_release_size_t(
 		pthreadpool_atomic_size_t* address,
 		size_t value)
 	{
-		atomic_store_explicit(address, value, memory_order_release);
+		/* x86-64 stores always have release semantics; use only a compiler barrier */
+		_WriteBarrier();
+		*address = value;
 	}
 
 	static inline size_t pthreadpool_decrement_fetch_relaxed_size_t(
 		pthreadpool_atomic_size_t* address)
 	{
-		return atomic_fetch_sub_explicit(address, 1, memory_order_relaxed) - 1;
+		return (size_t) _InterlockedDecrement64((volatile __int64*) address);
 	}
 
 	static inline size_t pthreadpool_decrement_fetch_release_size_t(
 		pthreadpool_atomic_size_t* address)
 	{
-		return atomic_fetch_sub_explicit(address, 1, memory_order_release) - 1;
+		return (size_t) _InterlockedDecrement64((volatile __int64*) address);
 	}
 
 	static inline bool pthreadpool_try_decrement_relaxed_size_t(
 		pthreadpool_atomic_size_t* value)
 	{
-		#if defined(__clang__) && (defined(__arm__) || defined(__aarch64__))
-			size_t actual_value;
-			do {
-				actual_value = __builtin_arm_ldrex((const volatile size_t*) value);
-				if (actual_value == 0) {
-					__builtin_arm_clrex();
-					return false;
-				}
-			} while (__builtin_arm_strex(actual_value - 1, (volatile size_t*) value) != 0);
-			return true;
-		#else
-			size_t actual_value = pthreadpool_load_relaxed_size_t(value);
-			while (actual_value != 0) {
-				if (atomic_compare_exchange_weak_explicit(
-					value, &actual_value, actual_value - 1, memory_order_relaxed, memory_order_relaxed))
-				{
-					return true;
-				}
+		size_t actual_value = *value;
+		while (actual_value != 0) {
+			const size_t new_value = actual_value - 1;
+			const size_t expected_value = actual_value;
+			actual_value = _InterlockedCompareExchange64(
+				(volatile __int64*) value, (__int64) new_value, (__int64) expected_value);
+			if (actual_value == expected_value) {
+				return true;
 			}
-			return false;
-		#endif
+		}
+		return false;
 	}
 
 	static inline void pthreadpool_fence_acquire() {
-		atomic_thread_fence(memory_order_acquire);
+		_mm_lfence();
+		_ReadBarrier();
 	}
 
 	static inline void pthreadpool_fence_release() {
-		atomic_thread_fence(memory_order_release);
+		_WriteBarrier();
+		_mm_sfence();
 	}
 #elif defined(_MSC_VER) && defined(_M_IX86)
 	typedef volatile uint32_t pthreadpool_atomic_uint32_t;
@@ -701,3 +708,21 @@
 #else
 	#error "Platform-specific implementation of threadpool-atomics.h required"
 #endif
+
+#if defined(__i386__) || defined(__i686__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
+	static inline void pthreadpool_yield() {
+		_mm_pause();
+	}
+#elif defined(__ARM_ACLE) || defined(_MSC_VER) && (defined(_M_ARM) || defined(_M_ARM64))
+	static inline void pthreadpool_yield() {
+		__yield();
+	}
+#elif defined(__GNUC__) && (defined(__ARM_ARCH) && (__ARM_ARCH >= 7) || (defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6KZ__)) && !defined(__thumb__))
+	static inline void pthreadpool_yield() {
+		__asm__ __volatile__("yield");
+	}
+#else
+	static inline void pthreadpool_yield() {
+		pthreadpool_fence_acquire();
+	}
+#endif
diff --git a/src/threadpool-object.h b/src/threadpool-object.h
index 239d116..9870e8a 100644
--- a/src/threadpool-object.h
+++ b/src/threadpool-object.h
@@ -179,6 +179,36 @@ struct pthreadpool_2d_tile_2d_with_uarch_params {
 	struct fxdiv_divisor_size_t tile_range_j;
 };
 
+struct pthreadpool_3d_params {
+	/**
+	 * FXdiv divisor for the range_j argument passed to the pthreadpool_parallelize_3d function.
+	 */
+	struct fxdiv_divisor_size_t range_j;
+	/**
+	 * FXdiv divisor for the range_k argument passed to the pthreadpool_parallelize_3d function.
+	 */
+	struct fxdiv_divisor_size_t range_k;
+};
+
+struct pthreadpool_3d_tile_1d_params {
+	/**
+	 * Copy of the range_k argument passed to the pthreadpool_parallelize_3d_tile_1d function.
+	 */
+	size_t range_k;
+	/**
+	 * Copy of the tile_k argument passed to the pthreadpool_parallelize_3d_tile_1d function.
+	 */
+	size_t tile_k;
+	/**
+	 * FXdiv divisor for the range_j argument passed to the pthreadpool_parallelize_3d_tile_1d function.
+	 */
+	struct fxdiv_divisor_size_t range_j;
+	/**
+	 * FXdiv divisor for the divide_round_up(range_k, tile_k) value.
+	 */
+	struct fxdiv_divisor_size_t tile_range_k;
+};
+
 struct pthreadpool_3d_tile_2d_params {
 	/**
 	 * Copy of the range_j argument passed to the pthreadpool_parallelize_3d_tile_2d function.
@@ -241,6 +271,52 @@ struct pthreadpool_3d_tile_2d_with_uarch_params {
 	struct fxdiv_divisor_size_t tile_range_k;
 };
 
+struct pthreadpool_4d_params {
+	/**
+	 * Copy of the range_k argument passed to the pthreadpool_parallelize_4d function.
+	 */
+	size_t range_k;
+	/**
+	 * FXdiv divisor for the range_j argument passed to the pthreadpool_parallelize_4d function.
+	 */
+	struct fxdiv_divisor_size_t range_j;
+	/**
+	 * FXdiv divisor for the range_k * range_l value.
+	 */
+	struct fxdiv_divisor_size_t range_kl;
+	/**
+	 * FXdiv divisor for the range_l argument passed to the pthreadpool_parallelize_4d function.
+	 */
+	struct fxdiv_divisor_size_t range_l;
+};
+
+struct pthreadpool_4d_tile_1d_params {
+	/**
+	 * Copy of the range_k argument passed to the pthreadpool_parallelize_4d_tile_1d function.
+	 */
+	size_t range_k;
+	/**
+	 * Copy of the range_l argument passed to the pthreadpool_parallelize_4d_tile_1d function.
+	 */
+	size_t range_l;
+	/**
+	 * Copy of the tile_l argument passed to the pthreadpool_parallelize_4d_tile_1d function.
+	 */
+	size_t tile_l;
+	/**
+	 * FXdiv divisor for the range_j argument passed to the pthreadpool_parallelize_4d_tile_1d function.
+	 */
+	struct fxdiv_divisor_size_t range_j;
+	/**
+	 * FXdiv divisor for the range_k * divide_round_up(range_l, tile_l) value.
+	 */
+	struct fxdiv_divisor_size_t tile_range_kl;
+	/**
+	 * FXdiv divisor for the divide_round_up(range_l, tile_l) value.
+	 */
+	struct fxdiv_divisor_size_t tile_range_l;
+};
+
 struct pthreadpool_4d_tile_2d_params {
 	/**
 	 * Copy of the range_k argument passed to the pthreadpool_parallelize_4d_tile_2d function.
@@ -311,6 +387,60 @@ struct pthreadpool_4d_tile_2d_with_uarch_params {
 	struct fxdiv_divisor_size_t tile_range_l;
 };
 
+struct pthreadpool_5d_params {
+	/**
+	 * Copy of the range_l argument passed to the pthreadpool_parallelize_5d function.
+	 */
+	size_t range_l;
+	/**
+	 * FXdiv divisor for the range_j argument passed to the pthreadpool_parallelize_5d function.
+	 */
+	struct fxdiv_divisor_size_t range_j;
+	/**
+	 * FXdiv divisor for the range_k argument passed to the pthreadpool_parallelize_5d function.
+	 */
+	struct fxdiv_divisor_size_t range_k;
+	/**
+	 * FXdiv divisor for the range_l * range_m value.
+	 */
+	struct fxdiv_divisor_size_t range_lm;
+	/**
+	 * FXdiv divisor for the range_m argument passed to the pthreadpool_parallelize_5d function.
+	 */
+	struct fxdiv_divisor_size_t range_m;
+};
+
+struct pthreadpool_5d_tile_1d_params {
+	/**
+	 * Copy of the range_k argument passed to the pthreadpool_parallelize_5d_tile_1d function.
+	 */
+	size_t range_k;
+	/**
+	 * Copy of the range_m argument passed to the pthreadpool_parallelize_5d_tile_1d function.
+	 */
+	size_t range_m;
+	/**
+	 * Copy of the tile_m argument passed to the pthreadpool_parallelize_5d_tile_1d function.
+	 */
+	size_t tile_m;
+	/**
+	 * FXdiv divisor for the range_j argument passed to the pthreadpool_parallelize_5d_tile_1d function.
+	 */
+	struct fxdiv_divisor_size_t range_j;
+	/**
+	 * FXdiv divisor for the range_k * range_l value.
+	 */
+	struct fxdiv_divisor_size_t range_kl;
+	/**
+	 * FXdiv divisor for the range_l argument passed to the pthreadpool_parallelize_5d_tile_1d function.
+	 */
+	struct fxdiv_divisor_size_t range_l;
+	/**
+	 * FXdiv divisor for the divide_round_up(range_m, tile_m) value.
+	 */
+	struct fxdiv_divisor_size_t tile_range_m;
+};
+
 struct pthreadpool_5d_tile_2d_params {
 	/**
 	 * Copy of the range_l argument passed to the pthreadpool_parallelize_5d_tile_2d function.
@@ -434,10 +564,16 @@ struct PTHREADPOOL_CACHELINE_ALIGNED pthreadpool {
 		struct pthreadpool_2d_tile_1d_params parallelize_2d_tile_1d;
 		struct pthreadpool_2d_tile_2d_params parallelize_2d_tile_2d;
 		struct pthreadpool_2d_tile_2d_with_uarch_params parallelize_2d_tile_2d_with_uarch;
+		struct pthreadpool_3d_params parallelize_3d;
+		struct pthreadpool_3d_tile_1d_params parallelize_3d_tile_1d;
 		struct pthreadpool_3d_tile_2d_params parallelize_3d_tile_2d;
 		struct pthreadpool_3d_tile_2d_with_uarch_params parallelize_3d_tile_2d_with_uarch;
+		struct pthreadpool_4d_params parallelize_4d;
+		struct pthreadpool_4d_tile_1d_params parallelize_4d_tile_1d;
 		struct pthreadpool_4d_tile_2d_params parallelize_4d_tile_2d;
 		struct pthreadpool_4d_tile_2d_with_uarch_params parallelize_4d_tile_2d_with_uarch;
+		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_tile_2d_params parallelize_6d_tile_2d;
 	} params;
@@ -526,3 +662,79 @@ PTHREADPOOL_INTERNAL void pthreadpool_parallelize(
 	void* context,
 	size_t linear_range,
 	uint32_t flags);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_1d_with_uarch_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_1d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_2d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_2d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_2d_tile_2d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_2d_tile_2d_with_uarch_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_3d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_3d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_3d_tile_2d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_3d_tile_2d_with_uarch_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_4d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_4d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_4d_tile_2d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_4d_tile_2d_with_uarch_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_5d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_5d_tile_1d_fastpath(
+	struct pthreadpool* threadpool,
+	struct thread_info* thread);
+
+PTHREADPOOL_INTERNAL void pthreadpool_thread_parallelize_5d_tile_2d_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);
diff --git a/src/threadpool-utils.h b/src/threadpool-utils.h
index 24fee43..91e2445 100644
--- a/src/threadpool-utils.h
+++ b/src/threadpool-utils.h
@@ -4,25 +4,22 @@
 #include <stddef.h>
 
 /* SSE-specific headers */
-#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
+#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
 	#include <xmmintrin.h>
 #endif
 
 /* MSVC-specific headers */
-#if defined(_MSC_VER) && _MSC_VER >= 1920
+#if defined(_MSC_VER)
 	#include <intrin.h>
-	#if defined(_M_IX86) || defined(_M_X64) || defined(_M_AMD64)
-		#include <immintrin.h>
-	#endif
 #endif
 
 
 struct fpu_state {
-#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
+#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
 	uint32_t mxcsr;
-#elif defined(__arm__) && defined(__ARM_FP) && (__ARM_FP != 0)
+#elif defined(__GNUC__) && defined(__arm__) && defined(__ARM_FP) && (__ARM_FP != 0) || defined(_MSC_VER) && defined(_M_ARM)
 	uint32_t fpscr;
-#elif defined(__aarch64__)
+#elif defined(__GNUC__) && defined(__aarch64__) || defined(_MSC_VER) && defined(_M_ARM64)
 	uint64_t fpcr;
 #else
 	char unused;
@@ -31,37 +28,63 @@ struct fpu_state {
 
 static inline struct fpu_state get_fpu_state() {
 	struct fpu_state state = { 0 };
-#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
+#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
 	state.mxcsr = (uint32_t) _mm_getcsr();
-#elif defined(__arm__) && defined(__ARM_FP) && (__ARM_FP != 0)
+#elif defined(_MSC_VER) && defined(_M_ARM)
+	state.fpscr = (uint32_t) _MoveFromCoprocessor(10, 7, 1, 0, 0);
+#elif defined(_MSC_VER) && defined(_M_ARM64)
+	state.fpcr = (uint64_t) _ReadStatusReg(0x5A20);
+#elif defined(__GNUC__) && defined(__arm__) && defined(__ARM_FP) && (__ARM_FP != 0)
 	__asm__ __volatile__("VMRS %[fpscr], fpscr" : [fpscr] "=r" (state.fpscr));
-#elif defined(__aarch64__)
+#elif defined(__GNUC__) && defined(__aarch64__)
 	__asm__ __volatile__("MRS %[fpcr], fpcr" : [fpcr] "=r" (state.fpcr));
 #endif
 	return state;
 }
 
 static inline void set_fpu_state(const struct fpu_state state) {
-#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
+#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
 	_mm_setcsr((unsigned int) state.mxcsr);
-#elif defined(__arm__) && defined(__ARM_FP) && (__ARM_FP != 0)
+#elif defined(_MSC_VER) && defined(_M_ARM)
+	_MoveToCoprocessor((int) state.fpscr, 10, 7, 1, 0, 0);
+#elif defined(_MSC_VER) && defined(_M_ARM64)
+	_WriteStatusReg(0x5A20, (__int64) state.fpcr);
+#elif defined(__GNUC__) && defined(__arm__) && defined(__ARM_FP) && (__ARM_FP != 0)
 	__asm__ __volatile__("VMSR fpscr, %[fpscr]" : : [fpscr] "r" (state.fpscr));
-#elif defined(__aarch64__)
+#elif defined(__GNUC__) && defined(__aarch64__)
 	__asm__ __volatile__("MSR fpcr, %[fpcr]" : : [fpcr] "r" (state.fpcr));
 #endif
 }
 
 static inline void disable_fpu_denormals() {
-#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
+#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
 	_mm_setcsr(_mm_getcsr() | 0x8040);
-#elif defined(__arm__) && defined(__ARM_FP) && (__ARM_FP != 0)
+#elif defined(_MSC_VER) && defined(_M_ARM)
+	int fpscr = _MoveFromCoprocessor(10, 7, 1, 0, 0);
+	fpscr |= 0x1000000;
+	_MoveToCoprocessor(fpscr, 10, 7, 1, 0, 0);
+#elif defined(_MSC_VER) && defined(_M_ARM64)
+	__int64 fpcr = _ReadStatusReg(0x5A20);
+	fpcr |= 0x1080000;
+	_WriteStatusReg(0x5A20, fpcr);
+#elif defined(__GNUC__) && defined(__arm__) && defined(__ARM_FP) && (__ARM_FP != 0)
 	uint32_t fpscr;
-	__asm__ __volatile__(
-			"VMRS %[fpscr], fpscr\n"
-			"ORR %[fpscr], #0x1000000\n"
-			"VMSR fpscr, %[fpscr]\n"
-		: [fpscr] "=r" (fpscr));
-#elif defined(__aarch64__)
+	#if defined(__thumb__) && !defined(__thumb2__)
+		__asm__ __volatile__(
+				"VMRS %[fpscr], fpscr\n"
+				"ORRS %[fpscr], %[bitmask]\n"
+				"VMSR fpscr, %[fpscr]\n"
+			: [fpscr] "=l" (fpscr)
+			: [bitmask] "l" (0x1000000)
+			: "cc");
+	#else
+		__asm__ __volatile__(
+				"VMRS %[fpscr], fpscr\n"
+				"ORR %[fpscr], #0x1000000\n"
+				"VMSR fpscr, %[fpscr]\n"
+			: [fpscr] "=r" (fpscr));
+	#endif
+#elif defined(__GNUC__) && defined(__aarch64__)
 	uint64_t fpcr;
 	__asm__ __volatile__(
 			"MRS %[fpcr], fpcr\n"
diff --git a/src/windows.c b/src/windows.c
index 19e534f..c9b88f7 100644
--- a/src/windows.c
+++ b/src/windows.c
@@ -35,8 +35,7 @@ static void wait_worker_threads(struct pthreadpool* threadpool, uint32_t event_i
 
 	/* Spin-wait */
 	for (uint32_t i = PTHREADPOOL_SPIN_WAIT_ITERATIONS; i != 0; i--) {
-		/* This fence serves as a sleep instruction */
-		pthreadpool_fence_acquire();
+		pthreadpool_yield();
 
 		active_threads = pthreadpool_load_acquire_size_t(&threadpool->active_threads);
 		if (active_threads == 0) {
@@ -63,8 +62,7 @@ static uint32_t wait_for_new_command(
 	if ((last_flags & PTHREADPOOL_FLAG_YIELD_WORKERS) == 0) {
 		/* Spin-wait loop */
 		for (uint32_t i = PTHREADPOOL_SPIN_WAIT_ITERATIONS; i != 0; i--) {
-			/* This fence serves as a sleep instruction */
-			pthreadpool_fence_acquire();
+			pthreadpool_yield();
 
 			command = pthreadpool_load_acquire_uint32_t(&threadpool->command);
 			if (command != last_command) {