[RFC][PATCH] Userptr benchmark

[RFC][PATCH] Userptr benchmark

[Tvrtko Ursulin]

From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>

A simple userptr benchmark measuring creation and destruction of userptr
surfaces and also impact of having a different number of them in the
process address space.

Example test output from i7-4550U running Android is below.

Questions, comments and ideas are welcome.

IGT-Version: 1.5-NOT-GIT (android-ia) (Linux: 3.10.20-g667dce8-dirty x86_64)
create-destroy                =   258270 op/s
multi-create-destroy          =     2803 op/s
multi-create-destroy-random   =     2777 op/s
Subtest userptr-unsync: SUCCESS
malloc-free,                    0 bos =     7463 op/s
malloc-free-random              0 bos =     5911 op/s
malloc-realloc-free,            0 bos =     2427 op/s
malloc-realloc-free-random,     0 bos =      669 op/s
mmap-unmap,                     0 bos =      636 op/s
mmap-unmap-random,              0 bos =      561 op/s
malloc-free,                    1 bos =     7462 op/s
malloc-free-random              1 bos =     5915 op/s
malloc-realloc-free,            1 bos =     2444 op/s
malloc-realloc-free-random,     1 bos =      669 op/s
mmap-unmap,                     1 bos =      594 op/s
mmap-unmap-random,              1 bos =      531 op/s
malloc-free,                   10 bos =     7464 op/s
malloc-free-random             10 bos =     5915 op/s
malloc-realloc-free,           10 bos =     2444 op/s
malloc-realloc-free-random,    10 bos =      669 op/s
mmap-unmap,                    10 bos =      606 op/s
mmap-unmap-random,             10 bos =      542 op/s
malloc-free,                  100 bos =     7462 op/s
malloc-free-random            100 bos =     5913 op/s
malloc-realloc-free,          100 bos =     2444 op/s
malloc-realloc-free-random,   100 bos =      669 op/s
mmap-unmap,                   100 bos =      613 op/s
mmap-unmap-random,            100 bos =      546 op/s
malloc-free,                 1000 bos =     7437 op/s
malloc-free-random           1000 bos =     5908 op/s
malloc-realloc-free,         1000 bos =     2442 op/s
malloc-realloc-free-random,  1000 bos =      669 op/s
mmap-unmap,                  1000 bos =      600 op/s
mmap-unmap-random,           1000 bos =      535 op/s
Subtest userptr-impact-unsync: SUCCESS
create-destroy                =    67453 op/s
multi-create-destroy          =     1837 op/s
multi-create-destroy-random   =     1817 op/s
Subtest userptr-sync: SUCCESS
malloc-free,                    0 bos =     7362 op/s
malloc-free-random              0 bos =     5913 op/s
malloc-realloc-free,            0 bos =     2444 op/s
malloc-realloc-free-random,     0 bos =      670 op/s
mmap-unmap,                     0 bos =      512 op/s
mmap-unmap-random,              0 bos =      465 op/s
malloc-free,                    1 bos =     7463 op/s
malloc-free-random              1 bos =     5913 op/s
malloc-realloc-free,            1 bos =     2445 op/s
malloc-realloc-free-random,     1 bos =      669 op/s
mmap-unmap,                     1 bos =      447 op/s
mmap-unmap-random,              1 bos =      403 op/s
malloc-free,                   10 bos =     7455 op/s
malloc-free-random             10 bos =     5911 op/s
malloc-realloc-free,           10 bos =     2444 op/s
malloc-realloc-free-random,    10 bos =      669 op/s
mmap-unmap,                    10 bos =      452 op/s
mmap-unmap-random,             10 bos =      414 op/s
malloc-free,                  100 bos =     7462 op/s
malloc-free-random            100 bos =     5915 op/s
malloc-realloc-free,          100 bos =     2442 op/s
malloc-realloc-free-random,   100 bos =      668 op/s
mmap-unmap,                   100 bos =      453 op/s
mmap-unmap-random,            100 bos =      416 op/s
malloc-free,                 1000 bos =     7462 op/s
malloc-free-random           1000 bos =     5907 op/s
malloc-realloc-free,         1000 bos =     2441 op/s
malloc-realloc-free-random,  1000 bos =      667 op/s
mmap-unmap,                  1000 bos =      436 op/s
mmap-unmap-random,           1000 bos =      400 op/s
Subtest userptr-impact-sync: SUCCESS


Tvrtko Ursulin (1):
  tests/gem_userptr_benchmark: Benchmarking userptr surfaces and impact

 tests/.gitignore              |   1 +
 tests/Makefile.sources        |   1 +
 tests/gem_userptr_benchmark.c | 451 ++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 453 insertions(+)
 create mode 100644 tests/gem_userptr_benchmark.c

-- 
1.8.4.3

[PATCH] tests/gem_userptr_benchmark: Benchmarking userptr surfaces and impact

[Tvrtko Ursulin]

From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>

This adds a small benchmark for the new userptr functionality.

Apart from basic surface creation and destruction, also tested is the
impact of having userptr surfaces in the process address space. Reason
for that is the impact of MMU notifiers on common address space
operations like munmap() which is per process.

Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
---
 tests/.gitignore              |   1 +
 tests/Makefile.sources        |   1 +
 tests/gem_userptr_benchmark.c | 451 ++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 453 insertions(+)
 create mode 100644 tests/gem_userptr_benchmark.c

diff --git a/tests/.gitignore b/tests/.gitignore
index 1a3b9d1..125bc88 100644
--- a/tests/.gitignore
+++ b/tests/.gitignore
@@ -93,6 +93,7 @@ gem_tiling_max_stride
 gem_unfence_active_buffers
 gem_unref_active_buffers
 gem_userptr_blits
+gem_userptr_benchmark
 gem_wait_render_timeout
 gem_write_read_ring_switch
 gen3_mixed_blits
diff --git a/tests/Makefile.sources b/tests/Makefile.sources
index c69e0e3..7be539d 100644
--- a/tests/Makefile.sources
+++ b/tests/Makefile.sources
@@ -117,6 +117,7 @@ TESTS_progs = \
 	gem_unfence_active_buffers \
 	gem_unref_active_buffers \
 	gem_userptr_blits \
+	gem_userptr_benchmark \
 	gem_wait_render_timeout \
 	gen3_mixed_blits \
 	gen3_render_linear_blits \
diff --git a/tests/gem_userptr_benchmark.c b/tests/gem_userptr_benchmark.c
new file mode 100644
index 0000000..32da02c
--- /dev/null
+++ b/tests/gem_userptr_benchmark.c
@@ -0,0 +1,451 @@
+/*
+ * Copyright © 2014 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ * Authors:
+ *    Tvrtko Ursulin <tvrtko.ursulin@intel.com>
+ *
+ */
+
+/** @file gem_userptr_benchmark.c
+ *
+ * Benchmark the userptr code and impact of having userptr surfaces
+ * in process address space on some normal operations.
+ *
+ */
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <fcntl.h>
+#include <inttypes.h>
+#include <errno.h>
+#include <sys/stat.h>
+#include <sys/time.h>
+#include <sys/mman.h>
+#include "drm.h"
+#include "i915_drm.h"
+#include "drmtest.h"
+#include "intel_bufmgr.h"
+#include "intel_batchbuffer.h"
+#include "intel_gpu_tools.h"
+
+#define WIDTH 128
+#define HEIGHT 128
+#define PAGE_SIZE 4096
+
+#define LOCAL_I915_GEM_USERPTR       0x34
+#define LOCAL_IOCTL_I915_GEM_USERPTR DRM_IOWR (DRM_COMMAND_BASE + LOCAL_I915_GEM_USERPTR, struct local_i915_gem_userptr)
+struct local_i915_gem_userptr {
+	uint64_t user_ptr;
+	uint64_t user_size;
+	uint32_t flags;
+#define I915_USERPTR_READ_ONLY (1<<0)
+#define I915_USERPTR_UNSYNCHRONIZED (1<<31)
+	uint32_t handle;
+};
+
+static uint32_t userptr_flags;
+
+static uint32_t linear[WIDTH*HEIGHT];
+
+static void gem_userptr_test_unsynchronized(void)
+{
+	userptr_flags = I915_USERPTR_UNSYNCHRONIZED;
+}
+
+static void gem_userptr_test_synchronized(void)
+{
+	userptr_flags = 0;
+}
+
+static int gem_userptr(int fd, void *ptr, int size, int read_only, uint32_t *handle)
+{
+	struct local_i915_gem_userptr userptr;
+	int ret;
+
+	userptr.user_ptr = (uintptr_t)ptr;
+	userptr.user_size = size;
+	userptr.flags = userptr_flags;
+	if (read_only)
+		userptr.flags |= I915_USERPTR_READ_ONLY;
+
+	ret = drmIoctl(fd, LOCAL_IOCTL_I915_GEM_USERPTR, &userptr);
+	if (ret)
+		ret = errno;
+	igt_skip_on_f(ret == ENODEV &&
+		      (userptr_flags & I915_USERPTR_UNSYNCHRONIZED) == 0,
+		      "Skipping, synchronized mappings with no kernel CONFIG_MMU_NOTIFIER?");
+	if (ret == 0)
+		*handle = userptr.handle;
+
+	return ret;
+}
+
+static uint32_t
+create_userptr(int fd, uint32_t val, uint32_t *ptr)
+{
+	uint32_t handle;
+	int i, ret;
+
+	ret = gem_userptr(fd, ptr, sizeof(linear), 0, &handle);
+	igt_assert(ret == 0);
+	igt_assert(handle != 0);
+
+	/* Fill the BO with dwords starting at val */
+	for (i = 0; i < WIDTH*HEIGHT; i++)
+		ptr[i] = val++;
+
+	return handle;
+}
+
+static void **handle_ptr_map;
+static unsigned int num_handle_ptr_map;
+
+static void add_handle_ptr(uint32_t handle, void *ptr)
+{
+	if (handle >= num_handle_ptr_map) {
+		handle_ptr_map = realloc(handle_ptr_map,
+					 (handle + 1000) * sizeof(void*));
+		num_handle_ptr_map = handle + 1000;
+	}
+
+	handle_ptr_map[handle] = ptr;
+}
+
+static void *get_handle_ptr(uint32_t handle)
+{
+	return handle_ptr_map[handle];
+}
+
+static void free_handle_ptr(uint32_t handle)
+{
+	igt_assert(handle < num_handle_ptr_map);
+	igt_assert(handle_ptr_map[handle]);
+
+	free(handle_ptr_map[handle]);
+	handle_ptr_map[handle] = NULL;
+}
+
+static uint32_t create_userptr_bo(int fd, int size)
+{
+	void *ptr;
+	uint32_t handle;
+	int ret;
+
+	ret = posix_memalign(&ptr, PAGE_SIZE, size);
+	igt_assert(ret == 0);
+
+	ret = gem_userptr(fd, (uint32_t *)ptr, size, 0, &handle);
+	igt_assert(ret == 0);
+	add_handle_ptr(handle, ptr);
+
+	return handle;
+}
+
+static void free_userptr_bo(int fd, uint32_t handle)
+{
+	gem_close(fd, handle);
+	free_handle_ptr(handle);
+}
+
+static int has_userptr(int fd)
+{
+	uint32_t handle = 0;
+	void *ptr;
+	uint32_t oldflags;
+	int ret;
+
+	assert(posix_memalign(&ptr, PAGE_SIZE, PAGE_SIZE) == 0);
+	oldflags = userptr_flags;
+	gem_userptr_test_unsynchronized();
+	ret = gem_userptr(fd, ptr, PAGE_SIZE, 0, &handle);
+	userptr_flags = oldflags;
+	if (ret != 0) {
+		free(ptr);
+		return 0;
+	}
+
+	gem_close(fd, handle);
+	free(ptr);
+
+	return handle != 0;
+}
+
+static const unsigned int nr_bos[] = {0, 1, 10, 100, 1000};
+static const unsigned int test_duration_sec = 3;
+
+static volatile unsigned int run_test;
+
+static void alarm_handler(int sig)
+{
+	assert(run_test == 1);
+	run_test = 0;
+}
+
+static void start_test(unsigned int duration)
+{
+	run_test = 1;
+	if (duration == 0)
+		duration = test_duration_sec;
+	signal(SIGALRM, alarm_handler);
+	alarm(duration);
+}
+
+static void exchange_ptr(void *array, unsigned i, unsigned j)
+{
+	void **arr, *tmp;
+	arr = (void **)array;
+
+	tmp = arr[i];
+	arr[i] = arr[j];
+	arr[j] = tmp;
+}
+
+static void test_malloc_free(int random)
+{
+	unsigned long iter = 0;
+	unsigned int i, tot = 1000;
+	void *ptr[tot];
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		for (i = 0; i < tot; i++) {
+			ptr[i] = malloc(1000);
+			assert(ptr[i]);
+		}
+		if (random)
+			igt_permute_array(ptr, tot, exchange_ptr);
+		for (i = 0; i < tot; i++)
+			free(ptr[i]);
+		iter++;
+	}
+
+	printf("%8lu op/s\n", iter / test_duration_sec);
+}
+
+static void test_malloc_realloc_free(int random)
+{
+	unsigned long iter = 0;
+	unsigned int i, tot = 1000;
+	void *ptr[tot];
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		for (i = 0; i < tot; i++) {
+			ptr[i] = malloc(1000);
+			assert(ptr[i]);
+		}
+		if (random)
+			igt_permute_array(ptr, tot, exchange_ptr);
+		for (i = 0; i < tot; i++) {
+			ptr[i] = realloc(ptr[i], 2000);
+			assert(ptr[i]);
+		}
+		if (random)
+			igt_permute_array(ptr, tot, exchange_ptr);
+		for (i = 0; i < tot; i++)
+			free(ptr[i]);
+		iter++;
+	}
+
+	printf("%8lu op/s\n", iter / test_duration_sec);
+}
+
+static void test_mmap_unmap(int random)
+{
+	unsigned long iter = 0;
+	unsigned int i, tot = 1000;
+	void *ptr[tot];
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		for (i = 0; i < tot; i++) {
+			ptr[i] = mmap(NULL, 1000, PROT_READ | PROT_WRITE,
+					MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+			assert(ptr[i] != MAP_FAILED);
+		}
+		if (random)
+			igt_permute_array(ptr, tot, exchange_ptr);
+		for (i = 0; i < tot; i++)
+			munmap(ptr[i], 1000);
+		iter++;
+	}
+
+	printf("%8lu op/s\n", iter / test_duration_sec);
+}
+
+static void test_impact(int fd)
+{
+	unsigned int total = sizeof(nr_bos) / sizeof(nr_bos[0]);
+	unsigned int subtest, i;
+	uint32_t handles[nr_bos[total-1]];
+
+	for (subtest = 0; subtest < total; subtest++) {
+		for (i = 0; i < nr_bos[subtest]; i++)
+			handles[i] = create_userptr_bo(fd, sizeof(linear));
+
+		printf("malloc-free,                %5u bos = ", nr_bos[subtest]);
+		test_malloc_free(0);
+		printf("malloc-free-random          %5u bos = ", nr_bos[subtest]);
+		test_malloc_free(1);
+
+		printf("malloc-realloc-free,        %5u bos = ", nr_bos[subtest]);
+		test_malloc_realloc_free(0);
+		printf("malloc-realloc-free-random, %5u bos = ", nr_bos[subtest]);
+		test_malloc_realloc_free(1);
+
+		printf("mmap-unmap,                 %5u bos = ", nr_bos[subtest]);
+		test_mmap_unmap(0);
+		printf("mmap-unmap-random,          %5u bos = ", nr_bos[subtest]);
+		test_mmap_unmap(1);
+
+		for (i = 0; i < nr_bos[subtest]; i++)
+			free_userptr_bo(fd, handles[i]);
+	}
+}
+
+static void test_single(int fd)
+{
+	char *ptr, *bo_ptr;
+	uint32_t handle = 0;
+	unsigned long iter = 0;
+	int ret;
+	unsigned long map_size = sizeof(linear) + PAGE_SIZE - 1;
+
+	ptr = mmap(NULL, map_size, PROT_READ | PROT_WRITE,
+			MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+	assert(ptr != MAP_FAILED);
+
+	bo_ptr = (char *)(((unsigned long)ptr + (PAGE_SIZE - 1))
+					& ~(PAGE_SIZE - 1));
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		ret = gem_userptr(fd, bo_ptr, sizeof(linear), 0, &handle);
+		assert(ret == 0);
+		gem_close(fd, handle);
+		iter++;
+	}
+
+	munmap(ptr, map_size);
+
+	printf("%8lu op/s\n", iter / test_duration_sec);
+}
+
+static void test_multiple(int fd, unsigned int batch, int random)
+{
+	char *ptr, *bo_ptr;
+	uint32_t handles[10000];
+	int map[10000];
+	unsigned long iter = 0;
+	int ret;
+	int i;
+	unsigned long map_size = batch * sizeof(linear) + PAGE_SIZE - 1;
+
+	assert(batch < (sizeof(handles) / sizeof(handles[0])));
+	assert(batch < (sizeof(map) / sizeof(map[0])));
+
+	ptr = mmap(NULL, map_size, PROT_READ | PROT_WRITE,
+			MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+	assert(ptr != MAP_FAILED);
+
+	bo_ptr = (char *)(((unsigned long)ptr + (PAGE_SIZE - 1))
+				& ~(PAGE_SIZE - 1));
+
+	for (i = 0; i < batch; i++)
+		map[i] = i;
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		if (random)
+			igt_permute_array(map, batch, igt_exchange_int);
+		for (i = 0; i < batch; i++) {
+			ret = gem_userptr(fd, bo_ptr + map[i] * sizeof(linear),
+						sizeof(linear),
+						0, &handles[i]);
+			assert(ret == 0);
+		}
+		if (random)
+			igt_permute_array(map, batch, igt_exchange_int);
+		for (i = 0; i < batch; i++)
+			gem_close(fd, handles[map[i]]);
+		iter++;
+	}
+
+	munmap(ptr, map_size);
+
+	printf("%8lu op/s\n", iter / test_duration_sec);
+}
+
+static void test_userptr(int fd)
+{
+	printf("create-destroy                = ");
+	test_single(fd);
+
+	printf("multi-create-destroy          = ");
+	test_multiple(fd, 100, 0);
+
+	printf("multi-create-destroy-random   = ");
+	test_multiple(fd, 100, 1);
+}
+
+int main(int argc, char **argv)
+{
+	int fd = -1, ret;
+
+	igt_skip_on_simulation();
+
+	igt_subtest_init(argc, argv);
+
+	fd = drm_open_any();
+	igt_assert(fd >= 0);
+
+	ret = has_userptr(fd);
+	igt_skip_on_f(ret == 0, "No userptr support - %s (%d)\n",
+			strerror(errno), ret);
+
+
+	gem_userptr_test_unsynchronized();
+
+	igt_subtest("userptr-unsync")
+		test_userptr(fd);
+
+	igt_subtest("userptr-impact-unsync")
+		test_impact(fd);
+
+	gem_userptr_test_synchronized();
+
+	igt_subtest("userptr-sync")
+		test_userptr(fd);
+
+	igt_subtest("userptr-impact-sync")
+		test_impact(fd);
+
+	igt_exit();
+
+	return 0;
+}
-- 
1.8.4.3

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Re: [RFC][PATCH] Userptr benchmark

[Chris Wilson]

On Wed, Feb 05, 2014 at 12:41:45PM +0000, Tvrtko Ursulin wrote:
> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> 
> A simple userptr benchmark measuring creation and destruction of userptr
> surfaces and also impact of having a different number of them in the
> process address space.
> 
> Example test output from i7-4550U running Android is below.
> 
> Questions, comments and ideas are welcome.

One silly idea I wanted to test was that speed of read/write access to
ptr was not affected by wrapping it up in a userptr.
 
> IGT-Version: 1.5-NOT-GIT (android-ia) (Linux: 3.10.20-g667dce8-dirty x86_64)
unsync vs sync
- unsync is much faster at creating userptr
  (not having to hook up and search the mmu-notifier)
- unsync is reasonably faster for destroying userptr
- there is no scaling issue with unsync, and minor (log(n)) scaling
  factor for sync

That seems in line with our expectations of i915_gem_userptr.c, which is
reassuring.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre

[PATCH] tests/gem_userptr_benchmark: Benchmarking userptr surfaces and impact

[Tvrtko Ursulin]

From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>

This adds a small benchmark for the new userptr functionality.

Apart from basic surface creation and destruction, also tested is the
impact of having userptr surfaces in the process address space. Reason
for that is the impact of MMU notifiers on common address space
operations like munmap() which is per process.

v2:
  * Moved to benchmarks.
  * Added pointer read/write tests.
  * Changed output to say iterations per second instead of
    operations per second.
  * Multiply result by batch size for multi-create* tests
    for a more comparable number with create-destroy test.

Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
---
 Android.mk                         |   3 +-
 benchmarks/.gitignore              |   1 +
 benchmarks/Android.mk              |  36 +++
 benchmarks/Makefile.am             |   7 +-
 benchmarks/Makefile.sources        |   6 +
 benchmarks/gem_userptr_benchmark.c | 513 +++++++++++++++++++++++++++++++++++++
 6 files changed, 558 insertions(+), 8 deletions(-)
 create mode 100644 benchmarks/Android.mk
 create mode 100644 benchmarks/Makefile.sources
 create mode 100644 benchmarks/gem_userptr_benchmark.c

diff --git a/Android.mk b/Android.mk
index 8aeb2d4..0c969b8 100644
--- a/Android.mk
+++ b/Android.mk
@@ -1,2 +1 @@
-include $(call all-named-subdir-makefiles, lib tests tools)
-
+include $(call all-named-subdir-makefiles, lib tests tools benchmarks)
diff --git a/benchmarks/.gitignore b/benchmarks/.gitignore
index ddea6f7..09e5bd8 100644
--- a/benchmarks/.gitignore
+++ b/benchmarks/.gitignore
@@ -1,3 +1,4 @@
+gem_userptr_benchmark
 intel_upload_blit_large
 intel_upload_blit_large_gtt
 intel_upload_blit_large_map
diff --git a/benchmarks/Android.mk b/benchmarks/Android.mk
new file mode 100644
index 0000000..5bb8ef5
--- /dev/null
+++ b/benchmarks/Android.mk
@@ -0,0 +1,36 @@
+LOCAL_PATH := $(call my-dir)
+
+include $(LOCAL_PATH)/Makefile.sources
+
+#================#
+
+define add_benchmark
+    include $(CLEAR_VARS)
+
+    LOCAL_SRC_FILES := $1.c
+
+    LOCAL_CFLAGS += -DHAVE_STRUCT_SYSINFO_TOTALRAM
+    LOCAL_CFLAGS += -DANDROID -UNDEBUG -include "check-ndebug.h"
+    LOCAL_CFLAGS += -std=c99
+    # FIXME: drop once Bionic correctly annotates "noreturn" on pthread_exit
+    LOCAL_CFLAGS += -Wno-error=return-type
+    # Excessive complaining for established cases. Rely on the Linux version warnings.
+    LOCAL_CFLAGS += -Wno-sign-compare
+
+    LOCAL_MODULE := $1
+    LOCAL_MODULE_TAGS := optional
+
+    LOCAL_STATIC_LIBRARIES := libintel_gpu_tools
+
+    LOCAL_SHARED_LIBRARIES := libpciaccess  \
+                              libdrm        \
+                              libdrm_intel
+
+    include $(BUILD_EXECUTABLE)
+endef
+
+#================#
+
+benchmark_list := $(bin_PROGRAMS)
+
+$(foreach item,$(benchmark_list),$(eval $(call add_benchmark,$(item))))
diff --git a/benchmarks/Makefile.am b/benchmarks/Makefile.am
index e2ad784..d173bf4 100644
--- a/benchmarks/Makefile.am
+++ b/benchmarks/Makefile.am
@@ -1,9 +1,4 @@
-
-bin_PROGRAMS = 				\
-	intel_upload_blit_large		\
-	intel_upload_blit_large_gtt	\
-	intel_upload_blit_large_map	\
-	intel_upload_blit_small
+include Makefile.sources
 
 AM_CPPFLAGS = -I$(top_srcdir) -I$(top_srcdir)/lib
 AM_CFLAGS = $(DRM_CFLAGS) $(CWARNFLAGS) $(CAIRO_CFLAGS)
diff --git a/benchmarks/Makefile.sources b/benchmarks/Makefile.sources
new file mode 100644
index 0000000..fd6c107
--- /dev/null
+++ b/benchmarks/Makefile.sources
@@ -0,0 +1,6 @@
+bin_PROGRAMS =                          \
+        intel_upload_blit_large         \
+        intel_upload_blit_large_gtt     \
+        intel_upload_blit_large_map     \
+        intel_upload_blit_small         \
+        gem_userptr_benchmark
diff --git a/benchmarks/gem_userptr_benchmark.c b/benchmarks/gem_userptr_benchmark.c
new file mode 100644
index 0000000..dc36f59
--- /dev/null
+++ b/benchmarks/gem_userptr_benchmark.c
@@ -0,0 +1,513 @@
+/*
+ * Copyright © 2014 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ * Authors:
+ *    Tvrtko Ursulin <tvrtko.ursulin@intel.com>
+ *
+ */
+
+/** @file gem_userptr_benchmark.c
+ *
+ * Benchmark the userptr code and impact of having userptr surfaces
+ * in process address space on some normal operations.
+ *
+ */
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <fcntl.h>
+#include <inttypes.h>
+#include <errno.h>
+#include <sys/stat.h>
+#include <sys/time.h>
+#include <sys/mman.h>
+#include "drm.h"
+#include "i915_drm.h"
+#include "drmtest.h"
+#include "intel_bufmgr.h"
+#include "intel_batchbuffer.h"
+#include "intel_gpu_tools.h"
+
+#define WIDTH 128
+#define HEIGHT 128
+#define PAGE_SIZE 4096
+
+#define LOCAL_I915_GEM_USERPTR       0x34
+#define LOCAL_IOCTL_I915_GEM_USERPTR DRM_IOWR (DRM_COMMAND_BASE + LOCAL_I915_GEM_USERPTR, struct local_i915_gem_userptr)
+struct local_i915_gem_userptr {
+	uint64_t user_ptr;
+	uint64_t user_size;
+	uint32_t flags;
+#define I915_USERPTR_READ_ONLY (1<<0)
+#define I915_USERPTR_UNSYNCHRONIZED (1<<31)
+	uint32_t handle;
+};
+
+static uint32_t userptr_flags;
+
+static uint32_t linear[WIDTH*HEIGHT];
+
+static void gem_userptr_test_unsynchronized(void)
+{
+	userptr_flags = I915_USERPTR_UNSYNCHRONIZED;
+}
+
+static void gem_userptr_test_synchronized(void)
+{
+	userptr_flags = 0;
+}
+
+static int gem_userptr(int fd, void *ptr, int size, int read_only, uint32_t *handle)
+{
+	struct local_i915_gem_userptr userptr;
+	int ret;
+
+	userptr.user_ptr = (uintptr_t)ptr;
+	userptr.user_size = size;
+	userptr.flags = userptr_flags;
+	if (read_only)
+		userptr.flags |= I915_USERPTR_READ_ONLY;
+
+	ret = drmIoctl(fd, LOCAL_IOCTL_I915_GEM_USERPTR, &userptr);
+	if (ret)
+		ret = errno;
+	igt_skip_on_f(ret == ENODEV &&
+		      (userptr_flags & I915_USERPTR_UNSYNCHRONIZED) == 0,
+		      "Skipping, synchronized mappings with no kernel CONFIG_MMU_NOTIFIER?");
+	if (ret == 0)
+		*handle = userptr.handle;
+
+	return ret;
+}
+
+static uint32_t
+create_userptr(int fd, uint32_t val, uint32_t *ptr)
+{
+	uint32_t handle;
+	int i, ret;
+
+	ret = gem_userptr(fd, ptr, sizeof(linear), 0, &handle);
+	igt_assert(ret == 0);
+	igt_assert(handle != 0);
+
+	/* Fill the BO with dwords starting at val */
+	for (i = 0; i < WIDTH*HEIGHT; i++)
+		ptr[i] = val++;
+
+	return handle;
+}
+
+static void **handle_ptr_map;
+static unsigned int num_handle_ptr_map;
+
+static void add_handle_ptr(uint32_t handle, void *ptr)
+{
+	if (handle >= num_handle_ptr_map) {
+		handle_ptr_map = realloc(handle_ptr_map,
+					 (handle + 1000) * sizeof(void*));
+		num_handle_ptr_map = handle + 1000;
+	}
+
+	handle_ptr_map[handle] = ptr;
+}
+
+static void *get_handle_ptr(uint32_t handle)
+{
+	return handle_ptr_map[handle];
+}
+
+static void free_handle_ptr(uint32_t handle)
+{
+	igt_assert(handle < num_handle_ptr_map);
+	igt_assert(handle_ptr_map[handle]);
+
+	free(handle_ptr_map[handle]);
+	handle_ptr_map[handle] = NULL;
+}
+
+static uint32_t create_userptr_bo(int fd, int size)
+{
+	void *ptr;
+	uint32_t handle;
+	int ret;
+
+	ret = posix_memalign(&ptr, PAGE_SIZE, size);
+	igt_assert(ret == 0);
+
+	ret = gem_userptr(fd, (uint32_t *)ptr, size, 0, &handle);
+	igt_assert(ret == 0);
+	add_handle_ptr(handle, ptr);
+
+	return handle;
+}
+
+static void free_userptr_bo(int fd, uint32_t handle)
+{
+	gem_close(fd, handle);
+	free_handle_ptr(handle);
+}
+
+static int has_userptr(int fd)
+{
+	uint32_t handle = 0;
+	void *ptr;
+	uint32_t oldflags;
+	int ret;
+
+	assert(posix_memalign(&ptr, PAGE_SIZE, PAGE_SIZE) == 0);
+	oldflags = userptr_flags;
+	gem_userptr_test_unsynchronized();
+	ret = gem_userptr(fd, ptr, PAGE_SIZE, 0, &handle);
+	userptr_flags = oldflags;
+	if (ret != 0) {
+		free(ptr);
+		return 0;
+	}
+
+	gem_close(fd, handle);
+	free(ptr);
+
+	return handle != 0;
+}
+
+static const unsigned int nr_bos[] = {0, 1, 10, 100, 1000};
+static const unsigned int test_duration_sec = 3;
+
+static volatile unsigned int run_test;
+
+static void alarm_handler(int sig)
+{
+	assert(run_test == 1);
+	run_test = 0;
+}
+
+static void start_test(unsigned int duration)
+{
+	run_test = 1;
+	if (duration == 0)
+		duration = test_duration_sec;
+	signal(SIGALRM, alarm_handler);
+	alarm(duration);
+}
+
+static void exchange_ptr(void *array, unsigned i, unsigned j)
+{
+	void **arr, *tmp;
+	arr = (void **)array;
+
+	tmp = arr[i];
+	arr[i] = arr[j];
+	arr[j] = tmp;
+}
+
+static void test_malloc_free(int random)
+{
+	unsigned long iter = 0;
+	unsigned int i, tot = 1000;
+	void *ptr[tot];
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		for (i = 0; i < tot; i++) {
+			ptr[i] = malloc(1000);
+			assert(ptr[i]);
+		}
+		if (random)
+			igt_permute_array(ptr, tot, exchange_ptr);
+		for (i = 0; i < tot; i++)
+			free(ptr[i]);
+		iter++;
+	}
+
+	printf("%8lu iter/s\n", iter / test_duration_sec);
+}
+
+static void test_malloc_realloc_free(int random)
+{
+	unsigned long iter = 0;
+	unsigned int i, tot = 1000;
+	void *ptr[tot];
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		for (i = 0; i < tot; i++) {
+			ptr[i] = malloc(1000);
+			assert(ptr[i]);
+		}
+		if (random)
+			igt_permute_array(ptr, tot, exchange_ptr);
+		for (i = 0; i < tot; i++) {
+			ptr[i] = realloc(ptr[i], 2000);
+			assert(ptr[i]);
+		}
+		if (random)
+			igt_permute_array(ptr, tot, exchange_ptr);
+		for (i = 0; i < tot; i++)
+			free(ptr[i]);
+		iter++;
+	}
+
+	printf("%8lu iter/s\n", iter / test_duration_sec);
+}
+
+static void test_mmap_unmap(int random)
+{
+	unsigned long iter = 0;
+	unsigned int i, tot = 1000;
+	void *ptr[tot];
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		for (i = 0; i < tot; i++) {
+			ptr[i] = mmap(NULL, 1000, PROT_READ | PROT_WRITE,
+					MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+			assert(ptr[i] != MAP_FAILED);
+		}
+		if (random)
+			igt_permute_array(ptr, tot, exchange_ptr);
+		for (i = 0; i < tot; i++)
+			munmap(ptr[i], 1000);
+		iter++;
+	}
+
+	printf("%8lu iter/s\n", iter / test_duration_sec);
+}
+
+static void test_ptr_read(void *ptr)
+{
+	unsigned long iter = 0;
+	volatile unsigned long *p;
+	unsigned long i, loops;
+	register unsigned long v;
+
+	loops = sizeof(linear) / sizeof(unsigned long) / 4;
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		p = (unsigned long *)ptr;
+		for (i = 0; i < loops; i++) {
+			v = *p++;
+			v = *p++;
+			v = *p++;
+			v = *p++;
+		}
+		iter++;
+	}
+
+	printf("%8lu MB/s\n", iter / test_duration_sec * sizeof(linear) / 1000000);
+}
+
+static void test_ptr_write(void *ptr)
+{
+	unsigned long iter = 0;
+	volatile unsigned long *p;
+	register unsigned long i, loops;
+
+	loops = sizeof(linear) / sizeof(unsigned long) / 4;
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		p = (unsigned long *)ptr;
+		for (i = 0; i < loops; i++) {
+			*p++ = i;
+			*p++ = i;
+			*p++ = i;
+			*p++ = i;
+		}
+		iter++;
+	}
+
+	printf("%8lu MB/s\n", iter / test_duration_sec * sizeof(linear) / 1000000);
+}
+
+static void test_impact(int fd)
+{
+	unsigned int total = sizeof(nr_bos) / sizeof(nr_bos[0]);
+	unsigned int subtest, i;
+	uint32_t handles[nr_bos[total-1]];
+	void *ptr;
+	char buffer[sizeof(linear)];
+
+	for (subtest = 0; subtest < total; subtest++) {
+		for (i = 0; i < nr_bos[subtest]; i++)
+			handles[i] = create_userptr_bo(fd, sizeof(linear));
+
+		if (nr_bos[subtest] > 0)
+			ptr = get_handle_ptr(handles[0]);
+		else
+			ptr = buffer;
+
+		printf("ptr-read,                   %5u bos = ", nr_bos[subtest]);
+		test_ptr_read(ptr);
+
+		printf("ptr-write                   %5u bos = ", nr_bos[subtest]);
+		test_ptr_write(ptr);
+
+		printf("malloc-free,                %5u bos = ", nr_bos[subtest]);
+		test_malloc_free(0);
+		printf("malloc-free-random          %5u bos = ", nr_bos[subtest]);
+		test_malloc_free(1);
+
+		printf("malloc-realloc-free,        %5u bos = ", nr_bos[subtest]);
+		test_malloc_realloc_free(0);
+		printf("malloc-realloc-free-random, %5u bos = ", nr_bos[subtest]);
+		test_malloc_realloc_free(1);
+
+		printf("mmap-unmap,                 %5u bos = ", nr_bos[subtest]);
+		test_mmap_unmap(0);
+		printf("mmap-unmap-random,          %5u bos = ", nr_bos[subtest]);
+		test_mmap_unmap(1);
+
+		for (i = 0; i < nr_bos[subtest]; i++)
+			free_userptr_bo(fd, handles[i]);
+	}
+}
+
+static void test_single(int fd)
+{
+	char *ptr, *bo_ptr;
+	uint32_t handle = 0;
+	unsigned long iter = 0;
+	int ret;
+	unsigned long map_size = sizeof(linear) + PAGE_SIZE - 1;
+
+	ptr = mmap(NULL, map_size, PROT_READ | PROT_WRITE,
+			MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+	assert(ptr != MAP_FAILED);
+
+	bo_ptr = (char *)(((unsigned long)ptr + (PAGE_SIZE - 1))
+					& ~(PAGE_SIZE - 1));
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		ret = gem_userptr(fd, bo_ptr, sizeof(linear), 0, &handle);
+		assert(ret == 0);
+		gem_close(fd, handle);
+		iter++;
+	}
+
+	munmap(ptr, map_size);
+
+	printf("%8lu iter/s\n", iter / test_duration_sec);
+}
+
+static void test_multiple(int fd, unsigned int batch, int random)
+{
+	char *ptr, *bo_ptr;
+	uint32_t handles[10000];
+	int map[10000];
+	unsigned long iter = 0;
+	int ret;
+	int i;
+	unsigned long map_size = batch * sizeof(linear) + PAGE_SIZE - 1;
+
+	assert(batch < (sizeof(handles) / sizeof(handles[0])));
+	assert(batch < (sizeof(map) / sizeof(map[0])));
+
+	ptr = mmap(NULL, map_size, PROT_READ | PROT_WRITE,
+			MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+	assert(ptr != MAP_FAILED);
+
+	bo_ptr = (char *)(((unsigned long)ptr + (PAGE_SIZE - 1))
+				& ~(PAGE_SIZE - 1));
+
+	for (i = 0; i < batch; i++)
+		map[i] = i;
+
+	start_test(test_duration_sec);
+
+	while (run_test) {
+		if (random)
+			igt_permute_array(map, batch, igt_exchange_int);
+		for (i = 0; i < batch; i++) {
+			ret = gem_userptr(fd, bo_ptr + map[i] * sizeof(linear),
+						sizeof(linear),
+						0, &handles[i]);
+			assert(ret == 0);
+		}
+		if (random)
+			igt_permute_array(map, batch, igt_exchange_int);
+		for (i = 0; i < batch; i++)
+			gem_close(fd, handles[map[i]]);
+		iter++;
+	}
+
+	munmap(ptr, map_size);
+
+	printf("%8lu iter/s\n", iter * batch / test_duration_sec);
+}
+
+static void test_userptr(int fd)
+{
+	printf("create-destroy                = ");
+	test_single(fd);
+
+	printf("multi-create-destroy          = ");
+	test_multiple(fd, 100, 0);
+
+	printf("multi-create-destroy-random   = ");
+	test_multiple(fd, 100, 1);
+}
+
+int main(int argc, char **argv)
+{
+	int fd = -1, ret;
+
+	igt_skip_on_simulation();
+
+	igt_subtest_init(argc, argv);
+
+	fd = drm_open_any();
+	igt_assert(fd >= 0);
+
+	ret = has_userptr(fd);
+	igt_skip_on_f(ret == 0, "No userptr support - %s (%d)\n",
+			strerror(errno), ret);
+
+
+	gem_userptr_test_unsynchronized();
+
+	igt_subtest("userptr-unsync")
+		test_userptr(fd);
+
+	igt_subtest("userptr-impact-unsync")
+		test_impact(fd);
+
+	gem_userptr_test_synchronized();
+
+	igt_subtest("userptr-sync")
+		test_userptr(fd);
+
+	igt_subtest("userptr-impact-sync")
+		test_impact(fd);
+
+	igt_exit();
+
+	return 0;
+}
-- 
1.8.4.3

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Re: [PATCH] tests/gem_userptr_benchmark: Benchmarking userptr surfaces and impact

[Daniel Vetter]

On Wed, Feb 05, 2014 at 05:33:06PM +0000, Tvrtko Ursulin wrote:
> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> 
> This adds a small benchmark for the new userptr functionality.
> 
> Apart from basic surface creation and destruction, also tested is the
> impact of having userptr surfaces in the process address space. Reason
> for that is the impact of MMU notifiers on common address space
> operations like munmap() which is per process.
> 
> v2:
>   * Moved to benchmarks.

I'd just keep it as an igt testcase, beating on the kernel a bit can't
hurt. And we have piles of other benchmark-like testcase already around.

Chris' comment on irc that I should create some better benchmark
infrastructure was probably more aimed to use all the microbenchmarks we
have to catch regressions. Atm we have zero infrastructure for tests to
expose performance numbers, so can't really do that. But that's work for
another day (month/year).
-Daniel

>   * Added pointer read/write tests.
>   * Changed output to say iterations per second instead of
>     operations per second.
>   * Multiply result by batch size for multi-create* tests
>     for a more comparable number with create-destroy test.
> 
> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> Cc: Chris Wilson <chris@chris-wilson.co.uk>
> ---
>  Android.mk                         |   3 +-
>  benchmarks/.gitignore              |   1 +
>  benchmarks/Android.mk              |  36 +++
>  benchmarks/Makefile.am             |   7 +-
>  benchmarks/Makefile.sources        |   6 +
>  benchmarks/gem_userptr_benchmark.c | 513 +++++++++++++++++++++++++++++++++++++
>  6 files changed, 558 insertions(+), 8 deletions(-)
>  create mode 100644 benchmarks/Android.mk
>  create mode 100644 benchmarks/Makefile.sources
>  create mode 100644 benchmarks/gem_userptr_benchmark.c
> 
> diff --git a/Android.mk b/Android.mk
> index 8aeb2d4..0c969b8 100644
> --- a/Android.mk
> +++ b/Android.mk
> @@ -1,2 +1 @@
> -include $(call all-named-subdir-makefiles, lib tests tools)
> -
> +include $(call all-named-subdir-makefiles, lib tests tools benchmarks)
> diff --git a/benchmarks/.gitignore b/benchmarks/.gitignore
> index ddea6f7..09e5bd8 100644
> --- a/benchmarks/.gitignore
> +++ b/benchmarks/.gitignore
> @@ -1,3 +1,4 @@
> +gem_userptr_benchmark
>  intel_upload_blit_large
>  intel_upload_blit_large_gtt
>  intel_upload_blit_large_map
> diff --git a/benchmarks/Android.mk b/benchmarks/Android.mk
> new file mode 100644
> index 0000000..5bb8ef5
> --- /dev/null
> +++ b/benchmarks/Android.mk
> @@ -0,0 +1,36 @@
> +LOCAL_PATH := $(call my-dir)
> +
> +include $(LOCAL_PATH)/Makefile.sources
> +
> +#================#
> +
> +define add_benchmark
> +    include $(CLEAR_VARS)
> +
> +    LOCAL_SRC_FILES := $1.c
> +
> +    LOCAL_CFLAGS += -DHAVE_STRUCT_SYSINFO_TOTALRAM
> +    LOCAL_CFLAGS += -DANDROID -UNDEBUG -include "check-ndebug.h"
> +    LOCAL_CFLAGS += -std=c99
> +    # FIXME: drop once Bionic correctly annotates "noreturn" on pthread_exit
> +    LOCAL_CFLAGS += -Wno-error=return-type
> +    # Excessive complaining for established cases. Rely on the Linux version warnings.
> +    LOCAL_CFLAGS += -Wno-sign-compare
> +
> +    LOCAL_MODULE := $1
> +    LOCAL_MODULE_TAGS := optional
> +
> +    LOCAL_STATIC_LIBRARIES := libintel_gpu_tools
> +
> +    LOCAL_SHARED_LIBRARIES := libpciaccess  \
> +                              libdrm        \
> +                              libdrm_intel
> +
> +    include $(BUILD_EXECUTABLE)
> +endef
> +
> +#================#
> +
> +benchmark_list := $(bin_PROGRAMS)
> +
> +$(foreach item,$(benchmark_list),$(eval $(call add_benchmark,$(item))))
> diff --git a/benchmarks/Makefile.am b/benchmarks/Makefile.am
> index e2ad784..d173bf4 100644
> --- a/benchmarks/Makefile.am
> +++ b/benchmarks/Makefile.am
> @@ -1,9 +1,4 @@
> -
> -bin_PROGRAMS = 				\
> -	intel_upload_blit_large		\
> -	intel_upload_blit_large_gtt	\
> -	intel_upload_blit_large_map	\
> -	intel_upload_blit_small
> +include Makefile.sources
>  
>  AM_CPPFLAGS = -I$(top_srcdir) -I$(top_srcdir)/lib
>  AM_CFLAGS = $(DRM_CFLAGS) $(CWARNFLAGS) $(CAIRO_CFLAGS)
> diff --git a/benchmarks/Makefile.sources b/benchmarks/Makefile.sources
> new file mode 100644
> index 0000000..fd6c107
> --- /dev/null
> +++ b/benchmarks/Makefile.sources
> @@ -0,0 +1,6 @@
> +bin_PROGRAMS =                          \
> +        intel_upload_blit_large         \
> +        intel_upload_blit_large_gtt     \
> +        intel_upload_blit_large_map     \
> +        intel_upload_blit_small         \
> +        gem_userptr_benchmark
> diff --git a/benchmarks/gem_userptr_benchmark.c b/benchmarks/gem_userptr_benchmark.c
> new file mode 100644
> index 0000000..dc36f59
> --- /dev/null
> +++ b/benchmarks/gem_userptr_benchmark.c
> @@ -0,0 +1,513 @@
> +/*
> + * Copyright © 2014 Intel Corporation
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice (including the next
> + * paragraph) shall be included in all copies or substantial portions of the
> + * Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
> + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
> + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
> + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
> + * IN THE SOFTWARE.
> + *
> + * Authors:
> + *    Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> + *
> + */
> +
> +/** @file gem_userptr_benchmark.c
> + *
> + * Benchmark the userptr code and impact of having userptr surfaces
> + * in process address space on some normal operations.
> + *
> + */
> +
> +#include <stdlib.h>
> +#include <stdio.h>
> +#include <string.h>
> +#include <fcntl.h>
> +#include <inttypes.h>
> +#include <errno.h>
> +#include <sys/stat.h>
> +#include <sys/time.h>
> +#include <sys/mman.h>
> +#include "drm.h"
> +#include "i915_drm.h"
> +#include "drmtest.h"
> +#include "intel_bufmgr.h"
> +#include "intel_batchbuffer.h"
> +#include "intel_gpu_tools.h"
> +
> +#define WIDTH 128
> +#define HEIGHT 128
> +#define PAGE_SIZE 4096
> +
> +#define LOCAL_I915_GEM_USERPTR       0x34
> +#define LOCAL_IOCTL_I915_GEM_USERPTR DRM_IOWR (DRM_COMMAND_BASE + LOCAL_I915_GEM_USERPTR, struct local_i915_gem_userptr)
> +struct local_i915_gem_userptr {
> +	uint64_t user_ptr;
> +	uint64_t user_size;
> +	uint32_t flags;
> +#define I915_USERPTR_READ_ONLY (1<<0)
> +#define I915_USERPTR_UNSYNCHRONIZED (1<<31)
> +	uint32_t handle;
> +};
> +
> +static uint32_t userptr_flags;
> +
> +static uint32_t linear[WIDTH*HEIGHT];
> +
> +static void gem_userptr_test_unsynchronized(void)
> +{
> +	userptr_flags = I915_USERPTR_UNSYNCHRONIZED;
> +}
> +
> +static void gem_userptr_test_synchronized(void)
> +{
> +	userptr_flags = 0;
> +}
> +
> +static int gem_userptr(int fd, void *ptr, int size, int read_only, uint32_t *handle)
> +{
> +	struct local_i915_gem_userptr userptr;
> +	int ret;
> +
> +	userptr.user_ptr = (uintptr_t)ptr;
> +	userptr.user_size = size;
> +	userptr.flags = userptr_flags;
> +	if (read_only)
> +		userptr.flags |= I915_USERPTR_READ_ONLY;
> +
> +	ret = drmIoctl(fd, LOCAL_IOCTL_I915_GEM_USERPTR, &userptr);
> +	if (ret)
> +		ret = errno;
> +	igt_skip_on_f(ret == ENODEV &&
> +		      (userptr_flags & I915_USERPTR_UNSYNCHRONIZED) == 0,
> +		      "Skipping, synchronized mappings with no kernel CONFIG_MMU_NOTIFIER?");
> +	if (ret == 0)
> +		*handle = userptr.handle;
> +
> +	return ret;
> +}
> +
> +static uint32_t
> +create_userptr(int fd, uint32_t val, uint32_t *ptr)
> +{
> +	uint32_t handle;
> +	int i, ret;
> +
> +	ret = gem_userptr(fd, ptr, sizeof(linear), 0, &handle);
> +	igt_assert(ret == 0);
> +	igt_assert(handle != 0);
> +
> +	/* Fill the BO with dwords starting at val */
> +	for (i = 0; i < WIDTH*HEIGHT; i++)
> +		ptr[i] = val++;
> +
> +	return handle;
> +}
> +
> +static void **handle_ptr_map;
> +static unsigned int num_handle_ptr_map;
> +
> +static void add_handle_ptr(uint32_t handle, void *ptr)
> +{
> +	if (handle >= num_handle_ptr_map) {
> +		handle_ptr_map = realloc(handle_ptr_map,
> +					 (handle + 1000) * sizeof(void*));
> +		num_handle_ptr_map = handle + 1000;
> +	}
> +
> +	handle_ptr_map[handle] = ptr;
> +}
> +
> +static void *get_handle_ptr(uint32_t handle)
> +{
> +	return handle_ptr_map[handle];
> +}
> +
> +static void free_handle_ptr(uint32_t handle)
> +{
> +	igt_assert(handle < num_handle_ptr_map);
> +	igt_assert(handle_ptr_map[handle]);
> +
> +	free(handle_ptr_map[handle]);
> +	handle_ptr_map[handle] = NULL;
> +}
> +
> +static uint32_t create_userptr_bo(int fd, int size)
> +{
> +	void *ptr;
> +	uint32_t handle;
> +	int ret;
> +
> +	ret = posix_memalign(&ptr, PAGE_SIZE, size);
> +	igt_assert(ret == 0);
> +
> +	ret = gem_userptr(fd, (uint32_t *)ptr, size, 0, &handle);
> +	igt_assert(ret == 0);
> +	add_handle_ptr(handle, ptr);
> +
> +	return handle;
> +}
> +
> +static void free_userptr_bo(int fd, uint32_t handle)
> +{
> +	gem_close(fd, handle);
> +	free_handle_ptr(handle);
> +}
> +
> +static int has_userptr(int fd)
> +{
> +	uint32_t handle = 0;
> +	void *ptr;
> +	uint32_t oldflags;
> +	int ret;
> +
> +	assert(posix_memalign(&ptr, PAGE_SIZE, PAGE_SIZE) == 0);
> +	oldflags = userptr_flags;
> +	gem_userptr_test_unsynchronized();
> +	ret = gem_userptr(fd, ptr, PAGE_SIZE, 0, &handle);
> +	userptr_flags = oldflags;
> +	if (ret != 0) {
> +		free(ptr);
> +		return 0;
> +	}
> +
> +	gem_close(fd, handle);
> +	free(ptr);
> +
> +	return handle != 0;
> +}
> +
> +static const unsigned int nr_bos[] = {0, 1, 10, 100, 1000};
> +static const unsigned int test_duration_sec = 3;
> +
> +static volatile unsigned int run_test;
> +
> +static void alarm_handler(int sig)
> +{
> +	assert(run_test == 1);
> +	run_test = 0;
> +}
> +
> +static void start_test(unsigned int duration)
> +{
> +	run_test = 1;
> +	if (duration == 0)
> +		duration = test_duration_sec;
> +	signal(SIGALRM, alarm_handler);
> +	alarm(duration);
> +}
> +
> +static void exchange_ptr(void *array, unsigned i, unsigned j)
> +{
> +	void **arr, *tmp;
> +	arr = (void **)array;
> +
> +	tmp = arr[i];
> +	arr[i] = arr[j];
> +	arr[j] = tmp;
> +}
> +
> +static void test_malloc_free(int random)
> +{
> +	unsigned long iter = 0;
> +	unsigned int i, tot = 1000;
> +	void *ptr[tot];
> +
> +	start_test(test_duration_sec);
> +
> +	while (run_test) {
> +		for (i = 0; i < tot; i++) {
> +			ptr[i] = malloc(1000);
> +			assert(ptr[i]);
> +		}
> +		if (random)
> +			igt_permute_array(ptr, tot, exchange_ptr);
> +		for (i = 0; i < tot; i++)
> +			free(ptr[i]);
> +		iter++;
> +	}
> +
> +	printf("%8lu iter/s\n", iter / test_duration_sec);
> +}
> +
> +static void test_malloc_realloc_free(int random)
> +{
> +	unsigned long iter = 0;
> +	unsigned int i, tot = 1000;
> +	void *ptr[tot];
> +
> +	start_test(test_duration_sec);
> +
> +	while (run_test) {
> +		for (i = 0; i < tot; i++) {
> +			ptr[i] = malloc(1000);
> +			assert(ptr[i]);
> +		}
> +		if (random)
> +			igt_permute_array(ptr, tot, exchange_ptr);
> +		for (i = 0; i < tot; i++) {
> +			ptr[i] = realloc(ptr[i], 2000);
> +			assert(ptr[i]);
> +		}
> +		if (random)
> +			igt_permute_array(ptr, tot, exchange_ptr);
> +		for (i = 0; i < tot; i++)
> +			free(ptr[i]);
> +		iter++;
> +	}
> +
> +	printf("%8lu iter/s\n", iter / test_duration_sec);
> +}
> +
> +static void test_mmap_unmap(int random)
> +{
> +	unsigned long iter = 0;
> +	unsigned int i, tot = 1000;
> +	void *ptr[tot];
> +
> +	start_test(test_duration_sec);
> +
> +	while (run_test) {
> +		for (i = 0; i < tot; i++) {
> +			ptr[i] = mmap(NULL, 1000, PROT_READ | PROT_WRITE,
> +					MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
> +			assert(ptr[i] != MAP_FAILED);
> +		}
> +		if (random)
> +			igt_permute_array(ptr, tot, exchange_ptr);
> +		for (i = 0; i < tot; i++)
> +			munmap(ptr[i], 1000);
> +		iter++;
> +	}
> +
> +	printf("%8lu iter/s\n", iter / test_duration_sec);
> +}
> +
> +static void test_ptr_read(void *ptr)
> +{
> +	unsigned long iter = 0;
> +	volatile unsigned long *p;
> +	unsigned long i, loops;
> +	register unsigned long v;
> +
> +	loops = sizeof(linear) / sizeof(unsigned long) / 4;
> +
> +	start_test(test_duration_sec);
> +
> +	while (run_test) {
> +		p = (unsigned long *)ptr;
> +		for (i = 0; i < loops; i++) {
> +			v = *p++;
> +			v = *p++;
> +			v = *p++;
> +			v = *p++;
> +		}
> +		iter++;
> +	}
> +
> +	printf("%8lu MB/s\n", iter / test_duration_sec * sizeof(linear) / 1000000);
> +}
> +
> +static void test_ptr_write(void *ptr)
> +{
> +	unsigned long iter = 0;
> +	volatile unsigned long *p;
> +	register unsigned long i, loops;
> +
> +	loops = sizeof(linear) / sizeof(unsigned long) / 4;
> +
> +	start_test(test_duration_sec);
> +
> +	while (run_test) {
> +		p = (unsigned long *)ptr;
> +		for (i = 0; i < loops; i++) {
> +			*p++ = i;
> +			*p++ = i;
> +			*p++ = i;
> +			*p++ = i;
> +		}
> +		iter++;
> +	}
> +
> +	printf("%8lu MB/s\n", iter / test_duration_sec * sizeof(linear) / 1000000);
> +}
> +
> +static void test_impact(int fd)
> +{
> +	unsigned int total = sizeof(nr_bos) / sizeof(nr_bos[0]);
> +	unsigned int subtest, i;
> +	uint32_t handles[nr_bos[total-1]];
> +	void *ptr;
> +	char buffer[sizeof(linear)];
> +
> +	for (subtest = 0; subtest < total; subtest++) {
> +		for (i = 0; i < nr_bos[subtest]; i++)
> +			handles[i] = create_userptr_bo(fd, sizeof(linear));
> +
> +		if (nr_bos[subtest] > 0)
> +			ptr = get_handle_ptr(handles[0]);
> +		else
> +			ptr = buffer;
> +
> +		printf("ptr-read,                   %5u bos = ", nr_bos[subtest]);
> +		test_ptr_read(ptr);
> +
> +		printf("ptr-write                   %5u bos = ", nr_bos[subtest]);
> +		test_ptr_write(ptr);
> +
> +		printf("malloc-free,                %5u bos = ", nr_bos[subtest]);
> +		test_malloc_free(0);
> +		printf("malloc-free-random          %5u bos = ", nr_bos[subtest]);
> +		test_malloc_free(1);
> +
> +		printf("malloc-realloc-free,        %5u bos = ", nr_bos[subtest]);
> +		test_malloc_realloc_free(0);
> +		printf("malloc-realloc-free-random, %5u bos = ", nr_bos[subtest]);
> +		test_malloc_realloc_free(1);
> +
> +		printf("mmap-unmap,                 %5u bos = ", nr_bos[subtest]);
> +		test_mmap_unmap(0);
> +		printf("mmap-unmap-random,          %5u bos = ", nr_bos[subtest]);
> +		test_mmap_unmap(1);
> +
> +		for (i = 0; i < nr_bos[subtest]; i++)
> +			free_userptr_bo(fd, handles[i]);
> +	}
> +}
> +
> +static void test_single(int fd)
> +{
> +	char *ptr, *bo_ptr;
> +	uint32_t handle = 0;
> +	unsigned long iter = 0;
> +	int ret;
> +	unsigned long map_size = sizeof(linear) + PAGE_SIZE - 1;
> +
> +	ptr = mmap(NULL, map_size, PROT_READ | PROT_WRITE,
> +			MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
> +	assert(ptr != MAP_FAILED);
> +
> +	bo_ptr = (char *)(((unsigned long)ptr + (PAGE_SIZE - 1))
> +					& ~(PAGE_SIZE - 1));
> +
> +	start_test(test_duration_sec);
> +
> +	while (run_test) {
> +		ret = gem_userptr(fd, bo_ptr, sizeof(linear), 0, &handle);
> +		assert(ret == 0);
> +		gem_close(fd, handle);
> +		iter++;
> +	}
> +
> +	munmap(ptr, map_size);
> +
> +	printf("%8lu iter/s\n", iter / test_duration_sec);
> +}
> +
> +static void test_multiple(int fd, unsigned int batch, int random)
> +{
> +	char *ptr, *bo_ptr;
> +	uint32_t handles[10000];
> +	int map[10000];
> +	unsigned long iter = 0;
> +	int ret;
> +	int i;
> +	unsigned long map_size = batch * sizeof(linear) + PAGE_SIZE - 1;
> +
> +	assert(batch < (sizeof(handles) / sizeof(handles[0])));
> +	assert(batch < (sizeof(map) / sizeof(map[0])));
> +
> +	ptr = mmap(NULL, map_size, PROT_READ | PROT_WRITE,
> +			MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
> +	assert(ptr != MAP_FAILED);
> +
> +	bo_ptr = (char *)(((unsigned long)ptr + (PAGE_SIZE - 1))
> +				& ~(PAGE_SIZE - 1));
> +
> +	for (i = 0; i < batch; i++)
> +		map[i] = i;
> +
> +	start_test(test_duration_sec);
> +
> +	while (run_test) {
> +		if (random)
> +			igt_permute_array(map, batch, igt_exchange_int);
> +		for (i = 0; i < batch; i++) {
> +			ret = gem_userptr(fd, bo_ptr + map[i] * sizeof(linear),
> +						sizeof(linear),
> +						0, &handles[i]);
> +			assert(ret == 0);
> +		}
> +		if (random)
> +			igt_permute_array(map, batch, igt_exchange_int);
> +		for (i = 0; i < batch; i++)
> +			gem_close(fd, handles[map[i]]);
> +		iter++;
> +	}
> +
> +	munmap(ptr, map_size);
> +
> +	printf("%8lu iter/s\n", iter * batch / test_duration_sec);
> +}
> +
> +static void test_userptr(int fd)
> +{
> +	printf("create-destroy                = ");
> +	test_single(fd);
> +
> +	printf("multi-create-destroy          = ");
> +	test_multiple(fd, 100, 0);
> +
> +	printf("multi-create-destroy-random   = ");
> +	test_multiple(fd, 100, 1);
> +}
> +
> +int main(int argc, char **argv)
> +{
> +	int fd = -1, ret;
> +
> +	igt_skip_on_simulation();
> +
> +	igt_subtest_init(argc, argv);
> +
> +	fd = drm_open_any();
> +	igt_assert(fd >= 0);
> +
> +	ret = has_userptr(fd);
> +	igt_skip_on_f(ret == 0, "No userptr support - %s (%d)\n",
> +			strerror(errno), ret);
> +
> +
> +	gem_userptr_test_unsynchronized();
> +
> +	igt_subtest("userptr-unsync")
> +		test_userptr(fd);
> +
> +	igt_subtest("userptr-impact-unsync")
> +		test_impact(fd);
> +
> +	gem_userptr_test_synchronized();
> +
> +	igt_subtest("userptr-sync")
> +		test_userptr(fd);
> +
> +	igt_subtest("userptr-impact-sync")
> +		test_impact(fd);
> +
> +	igt_exit();
> +
> +	return 0;
> +}
> -- 
> 1.8.4.3
> 
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

Re: [PATCH] tests/gem_userptr_benchmark: Benchmarking userptr surfaces and impact

[Tvrtko Ursulin]

On 02/05/2014 05:51 PM, Daniel Vetter wrote:
> On Wed, Feb 05, 2014 at 05:33:06PM +0000, Tvrtko Ursulin wrote:
>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
>>
>> This adds a small benchmark for the new userptr functionality.
>>
>> Apart from basic surface creation and destruction, also tested is the
>> impact of having userptr surfaces in the process address space. Reason
>> for that is the impact of MMU notifiers on common address space
>> operations like munmap() which is per process.
>>
>> v2:
>>    * Moved to benchmarks.
>
> I'd just keep it as an igt testcase, beating on the kernel a bit can't
> hurt. And we have piles of other benchmark-like testcase already around.

Are you sure? Ben suggested to move it there and I actually agree it 
makes more sense since it is mostly testing indirect effects on 
(seemingly) unrelated operations. Not to mention benchmark directory 
already exists and it is rather empty compared to tests...

Tvrtko