int main(int argc, char **argv)
{
VkResult res;
int retval = EXIT_FAILURE;
VkInstance vk;
struct tut1_physical_device devs[MAX_DEVICES];
uint32_t dev_count = MAX_DEVICES;
/* Fire up Vulkan */
res = tut1_init(&vk);
if (res)
{
printf("Could not initialize Vulkan: %s\n", tut1_VkResult_string(res));
goto exit_bad_init;
}
printf("Vulkan is in the house.\n");
/* Take a look at what devices there are */
res = tut1_enumerate_devices(vk, devs, &dev_count);
if (res < 0)
{
printf("Could not enumerate devices: %s\n", tut1_VkResult_string(res));
goto exit_bad_enumerate;
}
else if (res == VK_INCOMPLETE)
{
print_surprise("", "you've got", "devices", "dream of");
printf("I have information on only %"PRIu32" of them:\n", dev_count);
}
else
printf("I detected the following %"PRIu32" device%s:\n", dev_count, dev_count == 1?"":"s");
/*
* Print out some of the information taken when enumerating physical devices. This is by no means an
* exhaustive printout, but to give you the idea.
*/
for (uint32_t i = 0; i < dev_count; ++i)
{
struct tut1_physical_device *dev = &devs[i];
VkPhysicalDeviceProperties *pr = &dev->properties;
printf(" - %s: %s (id: 0x%04X) from vendor 0x%04X [driver version: 0x%04X, API version: 0x%04X]\n",
tut1_VkPhysicalDeviceType_string(pr->deviceType), pr->deviceName,
pr->deviceID, pr->vendorID, pr->driverVersion, pr->apiVersion);
if (dev->queue_families_incomplete)
{
print_surprise(" ", "your device", "queue families", "imagine");
printf(" I have information on only %"PRIu32" of them:\n", dev->queue_family_count);
}
else
printf(" The device supports the following %"PRIu32" queue famil%s:\n", dev->queue_family_count, dev->queue_family_count == 1?"y":"ies");
for (uint32_t j = 0; j < dev->queue_family_count; ++j)
{
VkQueueFamilyProperties *qf = &dev->queue_families[j];
printf(" * %"PRIu32" queue%s with the following capabilit%s:\n", qf->queueCount, qf->queueCount == 1?"":"s",
qf->queueFlags && (qf->queueFlags & (qf->queueFlags - 1)) == 0?"y":"ies");
if (qf->queueFlags == 0)
printf(" None\n");
if ((qf->queueFlags & VK_QUEUE_GRAPHICS_BIT))
printf(" Graphics\n");
if ((qf->queueFlags & VK_QUEUE_COMPUTE_BIT))
printf(" Compute\n");
if ((qf->queueFlags & VK_QUEUE_TRANSFER_BIT))
printf(" Transfer\n");
if ((qf->queueFlags & VK_QUEUE_SPARSE_BINDING_BIT))
printf(" Sparse binding\n");
}
printf(" The device supports memories of the following types:\n");
for (uint32_t j = 0; j < dev->memories.memoryTypeCount; ++j)
{
printf(" *");
if (dev->memories.memoryTypes[j].propertyFlags == 0)
printf(" <no properties>");
if ((dev->memories.memoryTypes[j].propertyFlags & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT))
printf(" device-local");
if ((dev->memories.memoryTypes[j].propertyFlags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT))
printf(" host-visible");
if ((dev->memories.memoryTypes[j].propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
printf(" host-coherent");
if ((dev->memories.memoryTypes[j].propertyFlags & VK_MEMORY_PROPERTY_HOST_CACHED_BIT))
printf(" host-cached");
if ((dev->memories.memoryTypes[j].propertyFlags & VK_MEMORY_PROPERTY_LAZILY_ALLOCATED_BIT))
printf(" lazy");
printf(": Available in Heap of size %"PRIu64"MB\n", dev->memories.memoryHeaps[dev->memories.memoryTypes[j].heapIndex].size / (1024 * 1024));
}
}
/* Congratulations, you can now duplicate the `vulkaninfo` program. */
retval = 0;
/* Cleanup after yourself */
exit_bad_enumerate:
tut1_exit(vk);
exit_bad_init:
return retval;
}
int main(int argc, char **argv)
{
tut1_error res;
int retval = EXIT_FAILURE;
VkInstance vk;
struct tut1_physical_device phy_devs[MAX_DEVICES];
struct tut2_device devs[MAX_DEVICES];
uint32_t dev_count = MAX_DEVICES;
VkShaderModule shaders[MAX_DEVICES] = {NULL};
struct tut3_pipelines pipelines[MAX_DEVICES];
struct tut4_data test_data[MAX_DEVICES];
int success = 0;
/* How many threads to do the work on */
size_t thread_count = 8;
/* Whether the threads should take some CPU time as well */
bool busy_threads = false;
/* Default to 1MB of buffer data to work on */
size_t buffer_size = 1024 * 1024 / sizeof(float);
bool bad_args = false;
if (argc < 2)
bad_args = true;
if (argc > 2 && sscanf(argv[2], "%zu", &thread_count) != 1)
bad_args = true;
if (argc > 3)
{
int temp;
if (sscanf(argv[3], "%d", &temp) != 1)
bad_args = true;
else
busy_threads = temp;
}
if (argc > 4)
{
if (sscanf(argv[4], "%zu", &buffer_size) != 1)
bad_args = true;
else
buffer_size /= sizeof(float);
}
if (bad_args)
{
printf("Usage: %s shader_file [thread_count(8) [busy_threads(0) [buffer_size(1MB)]]]\n\n", argv[0]);
return EXIT_FAILURE;
}
/* Fire up Vulkan */
res = tut1_init(&vk);
if (!tut1_error_is_success(&res))
{
tut1_error_printf(&res, "Could not initialize Vulkan\n");
goto exit_bad_init;
}
/* Enumerate devices */
res = tut1_enumerate_devices(vk, phy_devs, &dev_count);
if (tut1_error_is_error(&res))
{
tut1_error_printf(&res, "Could not enumerate devices\n");
goto exit_bad_enumerate;
}
/* Set up devices */
for (uint32_t i = 0; i < dev_count; ++i)
{
res = tut2_setup(&phy_devs[i], &devs[i], VK_QUEUE_COMPUTE_BIT);
if (!tut1_error_is_success(&res))
{
tut1_error_printf(&res, "Could not setup logical device %u, command pools and queues\n", i);
goto exit_bad_setup;
}
}
/* Load our compute shader */
for (uint32_t i = 0; i < dev_count; ++i)
{
res = tut3_load_shader(&devs[i], argv[1], &shaders[i]);
if (!tut1_error_is_success(&res))
{
tut1_error_printf(&res, "Could not load shader on device %u\n", i);
goto exit_bad_shader;
}
}
/*
* Create the pipelines. There are as many pipelines created as command buffers (just for example). If
* there are not actually enough resources for them, as many as possible are created. In this test, we are
* not going to handle the case where some pipelines are not created.
*/
for (uint32_t i = 0; i < dev_count; ++i)
{
res = tut3_make_compute_pipeline(&devs[i], &pipelines[i], shaders[i]);
if (!tut1_error_is_success(&res))
{
tut1_error_printf(&res, "Could not allocate enough pipelines on device %u\n", i);
goto exit_bad_pipeline;
}
}
/*
* Prepare our test. Both the buffers and threads are divided near-equally among the physical devices, which
* are likely to be just 1 in your case, but who knows.
*/
for (uint32_t i = 0; i < dev_count; ++i)
{
size_t this_buffer_size = buffer_size / dev_count;
size_t this_thread_count = thread_count / dev_count;
/* Make sure the last device gets all the left-over */
if (i == dev_count - 1)
{
this_buffer_size = buffer_size - buffer_size / dev_count * (dev_count - 1);
this_thread_count = thread_count - thread_count / dev_count * (dev_count - 1);
}
res = tut4_prepare_test(&phy_devs[i], &devs[i], &pipelines[i], &test_data[i], this_buffer_size, this_thread_count);
if (!tut1_error_is_success(&res))
{
tut1_error_printf(&res, "Could not allocate resources on device %u\n", i);
goto exit_bad_test_prepare;
}
}
/*
* Ok, this was a LOT of initializing! But we are finally ready to run something. tut4_start_test() creates
* a test thread for us, which further spawns the corresponding device's thread_count threads that do the
* calculations. We then wait for the tests to finish with tut4_wait_test_end().
*/
for (uint32_t i = 0; i < dev_count; ++i)
{
if (tut4_start_test(&test_data[i], busy_threads))
{
printf("Could not start the test threads for device %u\n", i);
perror("Error");
}
}
printf("Running the tests...\n");
for (uint32_t i = 0; i < dev_count; ++i)
tut4_wait_test_end(&test_data[i]);
success = 1;
for (uint32_t i = 0; i < dev_count; ++i)
if (!test_data[i].success)
{
if (!tut1_error_is_success(&test_data[i].error))
tut1_error_printf(&test_data[i].error, "Error starting test on device %u\n", i);
else
printf("The test didn't produce expected results (device %u)\n", i);
success = 0;
}
if (success)
printf("Everything went well :) We just wasted your GPU doing something stupid\n");
/*
* You can time the execution of the program with time(1):
*
* $ time ./tut4/tut4 shaders/tut3.comp.spv <threads> ...
*
* Then try to play with different number of threads and see if the total execution time of the application
* changes and how!
*
* ...
*
* Did you try that? Already? Well, that was disappointing. More threads probably resulted in higher
* execution time, right? That actually makes sense. You see, we have N data to compute, and whether you tell
* the GPU to do N computations from one thread, or N/T computations each from T threads, you aren't actually
* doing any less computation. You probably just have more overhead from the threads.
*
* So what's the deal with multi-threaded and Vulkan? Well, the problem is that this test was heavily
* GPU-bound, and as you have noticed, multi-CPU-threaded doesn't help. For this reason, this test has a
* little feature to "fake" some execution on the CPU threads as well. If you run the program like this:
*
* $ time ./tut4/tut4 shaders/tut3.comp.spv <threads> <fake> ...
*
* where <fake> can be either 0 (no CPU usage) or 1 (some fake CPU usage), and then experiment with different
* number of threads, you can see the benefit of multi-threading. In this case, while the GPU is working, the
* CPU thread spends time fake-doing something. If there is only one thread, the CPU cannot keep the GPU
* constantly busy, so the computation slows down. On the other hand, with multiple threads, the same amount
* of CPU work is spread out and done in parallel, so the threads together can feed the GPU with instructions
* faster.
*
* In this test, the total amount of time to waste is 3.2 seconds (32ms for each "render" operation, and there
* are a hundred of them). Depending on your GPU, you may notice that above a certain number of threads, there
* is no more any speedup. That is when the amount of time spent in each CPU thread becomes less than the time
* spent in the GPU for that thread's task, so whether the CPU spent time doing something before waiting for
* the GPU doesn't make a difference in the execution time.
*/
retval = 0;
/* Cleanup after yourself */
exit_bad_test_prepare:
for (uint32_t i = 0; i < dev_count; ++i)
tut4_free_test(&devs[i], &test_data[i]);
exit_bad_pipeline:
for (uint32_t i = 0; i < dev_count; ++i)
tut3_destroy_pipeline(&devs[i], &pipelines[i]);
exit_bad_shader:
for (uint32_t i = 0; i < dev_count; ++i)
tut3_free_shader(&devs[i], shaders[i]);
exit_bad_setup:
for (uint32_t i = 0; i < dev_count; ++i)
tut2_cleanup(&devs[i]);
exit_bad_enumerate:
tut1_exit(vk);
exit_bad_init:
return retval;
}
int main(int argc, char **argv)
{
tut1_error res;
int retval = EXIT_FAILURE;
VkInstance vk;
struct tut1_physical_device phy_devs[MAX_DEVICES];
struct tut2_device devs[MAX_DEVICES];
uint32_t dev_count = MAX_DEVICES;
VkShaderModule shaders[MAX_DEVICES] = {NULL};
struct tut3_pipelines pipelines[MAX_DEVICES];
if (argc < 2)
{
printf("Usage: %s shader_file\n\n", argv[0]);
return EXIT_FAILURE;
}
/* Fire up Vulkan */
res = tut1_init(&vk);
if (!tut1_error_is_success(&res))
{
tut1_error_printf(&res, "Could not initialize Vulkan\n");
goto exit_bad_init;
}
/* Enumerate devices */
res = tut1_enumerate_devices(vk, phy_devs, &dev_count);
if (tut1_error_is_error(&res))
{
tut1_error_printf(&res, "Could not enumerate devices\n");
goto exit_bad_enumerate;
}
/* Set up devices */
for (uint32_t i = 0; i < dev_count; ++i)
{
res = tut2_setup(&phy_devs[i], &devs[i], VK_QUEUE_COMPUTE_BIT);
if (!tut1_error_is_success(&res))
{
tut1_error_printf(&res, "Could not setup logical device %u, command pools and queues\n", i);
goto exit_bad_setup;
}
}
/* Load our compute shader */
for (uint32_t i = 0; i < dev_count; ++i)
{
res = tut3_load_shader(&devs[i], argv[1], &shaders[i]);
if (!tut1_error_is_success(&res))
{
tut1_error_printf(&res, "Could not load shader on device %u\n", i);
goto exit_bad_shader;
}
}
printf("Loaded the shader, awesome!\n");
/*
* Create the pipelines. There are as many pipelines created as command buffers (just for example). If
* there are not actually enough resources for them, as many as possible are created.
*/
for (uint32_t i = 0; i < dev_count; ++i)
tut3_make_compute_pipeline(&devs[i], &pipelines[i], shaders[i]);
/*
* Like tutorial 2, we have covered a lot of ground in this tutorial. Let's keep actual usage of our compute
* shader to the next tutorial, where we would see the effect of multiple threads on the processing speed.
*/
for (uint32_t i = 0; i < dev_count; ++i)
{
uint32_t count = 0;
for (uint32_t j = 0; j < pipelines[i].pipeline_count; ++j)
if (pipelines[i].pipelines[j].pipeline)
++count;
printf("Created %u pipeline%s on device %u\n", count, count == 1?"":"s", i);
}
retval = 0;
/* Cleanup after yourself */
for (uint32_t i = 0; i < dev_count; ++i)
tut3_destroy_pipeline(&devs[i], &pipelines[i]);
exit_bad_shader:
for (uint32_t i = 0; i < dev_count; ++i)
tut3_free_shader(&devs[i], shaders[i]);
exit_bad_setup:
for (uint32_t i = 0; i < dev_count; ++i)
tut2_cleanup(&devs[i]);
exit_bad_enumerate:
tut1_exit(vk);
exit_bad_init:
return retval;
}