[Beignet] [PATCH] Refine the cl thread implement for queue.
He Junyan
junyan.he at inbox.com
Fri May 23 08:36:21 PDT 2014
I think it is hard to avoid using thread local data.
Because when the queue creating, we do not know how many
threads will use this queue later.
The GPGPU resources will be hold in the queue, but
at least every thread should have a local data to store
the index to find the GPGPU data in the queue.
And the thread should need to destroy the GPGPU resource
when the thread exit, while the queue's life time may be
much longer than the thread.
OpenCL spec fail to define the relationship between the
queue and the threads. This cause the dilemma.
Please give some good advices if any.
On Fri, 2014-05-23 at 16:33 +0800, Zhigang Gong wrote:
> Some minor comments as below.
>
> One thought about the usage of thread local data we are using here.
> The original reason why we want to use thread local data is to avoid
> lock as much as possible. But finally, we found to satisfy all the
> use scenario, we can't avoid lock any way. Now we introduce lock
> eventually. Then is there still good reason why we should use these
> thread local data any more?
>
> One possible question is as below:
>
> If one queue is used in another thread to enqueue task, does it make
> sense to create a thread local new gpgpu data and in this thread. Or
> we can just simply lock and wait for other thread to unlock the queue?
>
> On Tue, May 20, 2014 at 02:26:47PM +0800, junyan.he at inbox.com wrote:
> > From: Junyan He <junyan.he at linux.intel.com>
> >
> > Because the cl_command_queue can be used in several threads simultaneously but
> > without add ref to it, we now handle it like this:
> > Keep one threads_slot_array, every time the thread get gpgpu or batch buffer, if it
> > does not have a slot, assign it.
> > The resources are keeped in queue private, and resize it if needed.
> > When the thread exit, the slot will be set invalid.
> > When queue released, all the resources will be released. If user still enqueue, flush
> > or finish the queue after it has been released, the behavior is undefined.
> > TODO: Need to shrink the slot map.
> >
> > Signed-off-by: Junyan He <junyan.he at linux.intel.com>
> > ---
> > src/cl_command_queue.c | 6 +-
> > src/cl_command_queue_gen7.c | 2 +-
> > src/cl_context.c | 2 +-
> > src/cl_device_id.c | 2 +-
> > src/cl_thread.c | 261 +++++++++++++++++++++++++++++++++-----------
> > src/cl_thread.h | 6 +-
> > 6 files changed, 205 insertions(+), 74 deletions(-)
> >
> > diff --git a/src/cl_command_queue.c b/src/cl_command_queue.c
> > index 6a699c0..802d313 100644
> > --- a/src/cl_command_queue.c
> > +++ b/src/cl_command_queue.c
> > @@ -60,6 +60,7 @@ cl_command_queue_new(cl_context ctx)
> > /* The queue also belongs to its context */
> > cl_context_add_ref(ctx);
> >
> > +
> useless new line.
> > exit:
> > return queue;
> > error:
> > @@ -74,6 +75,7 @@ cl_command_queue_delete(cl_command_queue queue)
> > assert(queue);
> > if (atomic_dec(&queue->ref_n) != 1) return;
> >
> > + if (queue->ref_n < 0)
> weird if statement here? What's the purpose?
> > /* Remove it from the list */
> > assert(queue->ctx);
> > pthread_mutex_lock(&queue->ctx->queue_lock);
> > @@ -89,7 +91,7 @@ cl_command_queue_delete(cl_command_queue queue)
> > queue->fulsim_out = NULL;
> > }
> >
> > - cl_thread_data_destroy(queue->thread_data);
> > + cl_thread_data_destroy(queue);
> > queue->thread_data = NULL;
> > cl_mem_delete(queue->perf);
> > cl_context_delete(queue->ctx);
> > @@ -430,7 +432,7 @@ cl_command_queue_flush(cl_command_queue queue)
> > LOCAL cl_int
> > cl_command_queue_finish(cl_command_queue queue)
> > {
> > - cl_gpgpu_sync(cl_get_thread_batch_buf());
> > + cl_gpgpu_sync(cl_get_thread_batch_buf(queue));
> > return CL_SUCCESS;
> > }
> >
> > diff --git a/src/cl_command_queue_gen7.c b/src/cl_command_queue_gen7.c
> > index 891d6f1..d875021 100644
> > --- a/src/cl_command_queue_gen7.c
> > +++ b/src/cl_command_queue_gen7.c
> > @@ -327,7 +327,7 @@ cl_command_queue_ND_range_gen7(cl_command_queue queue,
> > /* Start a new batch buffer */
> > batch_sz = cl_kernel_compute_batch_sz(ker);
> > cl_gpgpu_batch_reset(gpgpu, batch_sz);
> > - cl_set_thread_batch_buf(cl_gpgpu_ref_batch_buf(gpgpu));
> > + cl_set_thread_batch_buf(queue, cl_gpgpu_ref_batch_buf(gpgpu));
> > cl_gpgpu_batch_start(gpgpu);
> >
> > /* Issue the GPGPU_WALKER command */
> > diff --git a/src/cl_context.c b/src/cl_context.c
> > index 8190e6a..452afb8 100644
> > --- a/src/cl_context.c
> > +++ b/src/cl_context.c
> > @@ -203,7 +203,7 @@ cl_context_delete(cl_context ctx)
> > assert(ctx->buffers == NULL);
> > assert(ctx->drv);
> > cl_free(ctx->prop_user);
> > - cl_set_thread_batch_buf(NULL);
> > +// cl_set_thread_batch_buf(NULL);
> why not just remove the useless code?
>
> > cl_driver_delete(ctx->drv);
> > ctx->magic = CL_MAGIC_DEAD_HEADER; /* For safety */
> > cl_free(ctx);
> > diff --git a/src/cl_device_id.c b/src/cl_device_id.c
> > index 268789a..15a776d 100644
> > --- a/src/cl_device_id.c
> > +++ b/src/cl_device_id.c
> > @@ -106,7 +106,7 @@ LOCAL cl_device_id
> > cl_get_gt_device(void)
> > {
> > cl_device_id ret = NULL;
> > - cl_set_thread_batch_buf(NULL);
> > + // cl_set_thread_batch_buf(NULL);
> ditto?
>
> > const int device_id = cl_driver_get_device_id();
> > cl_device_id device = NULL;
> >
> > diff --git a/src/cl_thread.c b/src/cl_thread.c
> > index cadc3cd..1b517ac 100644
> > --- a/src/cl_thread.c
> > +++ b/src/cl_thread.c
> > @@ -15,113 +15,242 @@
> > * License along with this library. If not, see <http://www.gnu.org/licenses/>.
> > *
> > */
> > +#include <string.h>
> > +#include <stdio.h>
> >
> > #include "cl_thread.h"
> > #include "cl_alloc.h"
> > #include "cl_utils.h"
> >
> > -static __thread void* thread_batch_buf = NULL;
> > +/* Because the cl_command_queue can be used in several threads simultaneously but
> > + without add ref to it, we now handle it like this:
> > + Keep one threads_slot_array, every time the thread get gpgpu or batch buffer, if it
> > + does not have a slot, assign it.
> > + The resources are keeped in queue private, and resize it if needed.
> > + When the thread exit, the slot will be set invalid.
> > + When queue released, all the resources will be released. If user still enqueue, flush
> > + or finish the queue after it has been released, the behavior is undefined.
> > + TODO: Need to shrink the slot map.
> > + */
> >
> > -typedef struct _cl_thread_spec_data {
> > +static int thread_array_num = 1;
> > +static int *thread_slot_map = NULL;
> > +static int thread_magic_num = 1;
> > +static pthread_mutex_t thread_queue_map_lock = PTHREAD_MUTEX_INITIALIZER;
> > +static pthread_key_t destroy_key;
> > +
> > +static __thread int thread_id = -1;
> > +static __thread int thread_magic = -1;
> > +
> > +typedef struct _thread_spec_data {
> > cl_gpgpu gpgpu ;
> > int valid;
> > -}cl_thread_spec_data;
> > + void* thread_batch_buf;
> > + int thread_magic;
> > +} thread_spec_data;
> >
> > -void cl_set_thread_batch_buf(void* buf) {
> > - if (thread_batch_buf) {
> > - cl_gpgpu_unref_batch_buf(thread_batch_buf);
> > - }
> > - thread_batch_buf = buf;
> > -}
> > +typedef struct _queue_thread_private {
> > + thread_spec_data** threads_data;
> > + int threads_data_num;
> > + pthread_mutex_t thread_data_lock;
> > +} queue_thread_private;
> >
> > -void* cl_get_thread_batch_buf(void) {
> > - return thread_batch_buf;
> > +static void thread_data_destructor(void *dummy) {
> > + pthread_mutex_lock(&thread_queue_map_lock);
> > + thread_slot_map[thread_id] = 0;
> > + pthread_mutex_unlock(&thread_queue_map_lock);
> > + free(dummy);
> > }
> >
> > -cl_gpgpu cl_get_thread_gpgpu(cl_command_queue queue)
> > +static thread_spec_data * __create_thread_spec_data(cl_command_queue queue, int create)
> > {
> > - pthread_key_t* key = queue->thread_data;
> > - cl_thread_spec_data* thread_spec_data = pthread_getspecific(*key);
> > -
> > - if (!thread_spec_data) {
> > - TRY_ALLOC_NO_ERR(thread_spec_data, CALLOC(struct _cl_thread_spec_data));
> > - if (pthread_setspecific(*key, thread_spec_data)) {
> > - cl_free(thread_spec_data);
> > - return NULL;
> > + queue_thread_private *thread_private = ((queue_thread_private *)(queue->thread_data));
> > + thread_spec_data* spec = NULL;
> > + int i = 0;
> > + int *old;
> > +
> > + if (thread_id == -1) {
> > + void * dummy = malloc(sizeof(int));
> > +
> > + pthread_mutex_lock(&thread_queue_map_lock);
> > + for (i = 0; i < thread_array_num; i++) {
> > + if (thread_slot_map[i] == 0) {
> > + thread_id = i;
> > + break;
> > + }
> > + }
> > +
> > + if (i == thread_array_num) {
> > + thread_array_num *= 2;
> > + old = thread_slot_map;
> > + thread_slot_map = malloc(sizeof(int) * thread_array_num);
> > + memset(thread_slot_map, 0, (sizeof(int) * thread_array_num));
> > + for (i = 0; i < thread_array_num/2; i++) {
> > + thread_slot_map[i] = old[i];
> > + }
> > +
> > + thread_id = thread_array_num/2;
> > + cl_free(old);
> why not just use a realloc() here?
> > + }
> > +
> > + thread_slot_map[thread_id] = 1;
> > +
> > + thread_magic = thread_magic_num++;
> > + pthread_mutex_unlock(&thread_queue_map_lock);
> > +
> > + pthread_setspecific(destroy_key, dummy);
> > + }
> > +
> > + pthread_mutex_lock(&thread_private->thread_data_lock);
> > + if (thread_array_num > thread_private->threads_data_num) {// just enlarge
> > + thread_spec_data ** old = thread_private->threads_data;
> > + int old_num = thread_private->threads_data_num;
> > + thread_private->threads_data_num = thread_array_num;
> > + thread_private->threads_data = malloc(thread_private->threads_data_num * sizeof(void *));
> > + memset(thread_private->threads_data, 0, sizeof(void*) * thread_private->threads_data_num);
> > + for (i = 0; i < old_num; i++) {
> > + thread_private->threads_data[i] = old[i];
> > }
> > + free(old);
> Another place which may better to use realloc;
> > }
> >
> > - if (!thread_spec_data->valid) {
> > - TRY_ALLOC_NO_ERR(thread_spec_data->gpgpu, cl_gpgpu_new(queue->ctx->drv));
> > - thread_spec_data->valid = 1;
> > + assert(thread_id != -1 && thread_id < thread_array_num);
> > + spec = thread_private->threads_data[thread_id];
> > + if (!spec && create) {
> > + spec = CALLOC(thread_spec_data);
> > + spec->thread_magic = thread_magic;
> > + thread_private->threads_data[thread_id] = spec;
> > }
> >
> > -error:
> > - return thread_spec_data->gpgpu;
> > + pthread_mutex_unlock(&thread_private->thread_data_lock);
> > +
> > + return spec;
> > }
> >
> > -void cl_invalid_thread_gpgpu(cl_command_queue queue)
> > +void* cl_thread_data_create(void)
> > {
> > - pthread_key_t* key = queue->thread_data;
> > - cl_thread_spec_data* thread_spec_data = pthread_getspecific(*key);
> > + queue_thread_private* thread_private = CALLOC(queue_thread_private);
> >
> > - if (!thread_spec_data) {
> > - return;
> > - }
> > + if (thread_private == NULL)
> > + return NULL;
> >
> > - if (!thread_spec_data->valid) {
> > - return;
> > + if (thread_slot_map == NULL) {
> > + pthread_mutex_lock(&thread_queue_map_lock);
> > + thread_slot_map = malloc(sizeof(int) * thread_array_num);
> > + memset(thread_slot_map, 0, (sizeof(int) * thread_array_num));
> calloc = malloc + memset to zero.
> > + pthread_mutex_unlock(&thread_queue_map_lock);
> > +
> > + pthread_key_create(&destroy_key, thread_data_destructor);
> > }
> >
> > - assert(thread_spec_data->gpgpu);
> > - cl_gpgpu_delete(thread_spec_data->gpgpu);
> > - thread_spec_data->valid = 0;
> > + pthread_mutex_init(&thread_private->thread_data_lock, NULL);
> > +
> > + pthread_mutex_lock(&thread_private->thread_data_lock);
> > + thread_private->threads_data = malloc(thread_array_num * sizeof(void *));
> > + memset(thread_private->threads_data, 0, sizeof(void*) * thread_array_num);
> > + thread_private->threads_data_num = thread_array_num;
> > + pthread_mutex_unlock(&thread_private->thread_data_lock);
> > +
> > + return thread_private;
> > }
> >
> > -static void thread_data_destructor(void *data) {
> > - cl_thread_spec_data* thread_spec_data = (cl_thread_spec_data *)data;
> > +cl_gpgpu cl_get_thread_gpgpu(cl_command_queue queue)
> > +{
> > + thread_spec_data* spec = __create_thread_spec_data(queue, 1);
> >
> > - if (thread_batch_buf) {
> > - cl_gpgpu_unref_batch_buf(thread_batch_buf);
> > - thread_batch_buf = NULL;
> > + if (!spec->thread_magic && spec->thread_magic != thread_magic) {
> > + //We may get the slot from last thread. So free the resource.
> > + spec->valid = 0;
> > }
> >
> > - if (thread_spec_data->valid)
> > - cl_gpgpu_delete(thread_spec_data->gpgpu);
> > - cl_free(thread_spec_data);
> > + if (!spec->valid) {
> > + if (spec->thread_batch_buf) {
> > + cl_gpgpu_unref_batch_buf(spec->thread_batch_buf);
> > + spec->thread_batch_buf = NULL;
> > + }
> > + if (spec->gpgpu) {
> > + cl_gpgpu_delete(spec->gpgpu);
> > + spec->gpgpu = NULL;
> > + }
> > + TRY_ALLOC_NO_ERR(spec->gpgpu, cl_gpgpu_new(queue->ctx->drv));
> > + spec->valid = 1;
> > + }
> > +
> > + error:
> > + return spec->gpgpu;
> > }
> >
> > -/* Create the thread specific data. */
> > -void* cl_thread_data_create(void)
> > +void cl_set_thread_batch_buf(cl_command_queue queue, void* buf)
> > {
> > - int rc = 0;
> > + thread_spec_data* spec = __create_thread_spec_data(queue, 1);
> >
> > - pthread_key_t *thread_specific_key = CALLOC(pthread_key_t);
> > - if (thread_specific_key == NULL)
> > - return NULL;
> > + assert(spec && spec->thread_magic == thread_magic);
> > +
> > + if (spec->thread_batch_buf) {
> > + cl_gpgpu_unref_batch_buf(spec->thread_batch_buf);
> > + }
> > + spec->thread_batch_buf = buf;
> > +}
> >
> > - rc = pthread_key_create(thread_specific_key, thread_data_destructor);
> > +void* cl_get_thread_batch_buf(cl_command_queue queue) {
> > + thread_spec_data* spec = __create_thread_spec_data(queue, 1);
> >
> > - if (rc != 0)
> > - return NULL;
> > + assert(spec && spec->thread_magic == thread_magic);
> > +
> > + return spec->thread_batch_buf;
> > +}
> > +
> > +void cl_invalid_thread_gpgpu(cl_command_queue queue)
> > +{
> > + queue_thread_private *thread_private = ((queue_thread_private *)(queue->thread_data));
> > + thread_spec_data* spec = NULL;
> >
> > - return thread_specific_key;
> > + pthread_mutex_lock(&thread_private->thread_data_lock);
> > + spec = thread_private->threads_data[thread_id];
> > + assert(spec);
> > + pthread_mutex_unlock(&thread_private->thread_data_lock);
> > +
> > + if (!spec->valid) {
> > + return;
> > + }
> > +
> > + assert(spec->gpgpu);
> > + cl_gpgpu_delete(spec->gpgpu);
> > + spec->gpgpu = NULL;
> > + spec->valid = 0;
> > }
> >
> > /* The destructor for clean the thread specific data. */
> > -void cl_thread_data_destroy(void * data)
> > +void cl_thread_data_destroy(cl_command_queue queue)
> > {
> > - pthread_key_t *thread_specific_key = (pthread_key_t *)data;
> > -
> > - /* First release self spec data. */
> > - cl_thread_spec_data* thread_spec_data =
> > - pthread_getspecific(*thread_specific_key);
> > - if (thread_spec_data && thread_spec_data->valid) {
> > - cl_gpgpu_delete(thread_spec_data->gpgpu);
> > - if (thread_spec_data)
> > - cl_free(thread_spec_data);
> > + int i = 0;
> > + queue_thread_private *thread_private = ((queue_thread_private *)(queue->thread_data));
> > + int threads_data_num;
> > + thread_spec_data** threads_data;
> > +
> > + pthread_mutex_lock(&thread_private->thread_data_lock);
> > + assert(thread_private->threads_data_num == thread_array_num);
> > + threads_data_num = thread_private->threads_data_num;
> > + threads_data = thread_private->threads_data;
> > + thread_private->threads_data_num = 0;
> > + thread_private->threads_data = NULL;
> > + pthread_mutex_unlock(&thread_private->thread_data_lock);
> > + cl_free(thread_private);
> > + queue->thread_data = NULL;
> > +
> > + for (i = 0; i < threads_data_num; i++) {
> > + if (threads_data[i] != NULL && threads_data[i]->thread_batch_buf) {
> > + cl_gpgpu_unref_batch_buf(threads_data[i]->thread_batch_buf);
> > + threads_data[i]->thread_batch_buf = NULL;
> > + }
> > +
> > + if (threads_data[i] != NULL && threads_data[i]->valid) {
> > + cl_gpgpu_delete(threads_data[i]->gpgpu);
> > + threads_data[i]->gpgpu = NULL;
> > + threads_data[i]->valid = 0;
> > + }
> > + cl_free(threads_data[i]);
> > }
> >
> > - pthread_key_delete(*thread_specific_key);
> > - cl_free(thread_specific_key);
> > + cl_free(threads_data);
> > }
> > diff --git a/src/cl_thread.h b/src/cl_thread.h
> > index c8ab63c..bf855a2 100644
> > --- a/src/cl_thread.h
> > +++ b/src/cl_thread.h
> > @@ -27,7 +27,7 @@
> > void* cl_thread_data_create(void);
> >
> > /* The destructor for clean the thread specific data. */
> > -void cl_thread_data_destroy(void * data);
> > +void cl_thread_data_destroy(cl_command_queue queue);
> >
> > /* Used to get the gpgpu struct of each thread. */
> > cl_gpgpu cl_get_thread_gpgpu(cl_command_queue queue);
> > @@ -36,9 +36,9 @@ cl_gpgpu cl_get_thread_gpgpu(cl_command_queue queue);
> > void cl_invalid_thread_gpgpu(cl_command_queue queue);
> >
> > /* Used to set the batch buffer of each thread. */
> > -void cl_set_thread_batch_buf(void* buf);
> > +void cl_set_thread_batch_buf(cl_command_queue queue, void* buf);
> >
> > /* Used to get the batch buffer of each thread. */
> > -void* cl_get_thread_batch_buf(void);
> > +void* cl_get_thread_batch_buf(cl_command_queue queue);
> >
> > #endif /* __CL_THREAD_H__ */
> > --
> > 1.8.3.2
> >
> > _______________________________________________
> > Beignet mailing list
> > Beignet at lists.freedesktop.org
> > http://lists.freedesktop.org/mailman/listinfo/beignet
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet
More information about the Beignet
mailing list