Re: [Mesa-dev] [PATCH] radv/winsys: replace bo list searchs with a hash table.

2018-01-30 Thread Thomas Helland 
2018-01-31 3:38 GMT+01:00 Dave Airlie :
> On 31 January 2018 at 12:27, Dieter Nützel  wrote:
>> Ping!
>
> I'm not sure this was much of a win, and if it makes things worse in some 
> cases,
> then it needs a lot more investigation, so probably consider it dead for now.
>
> Dave.
>

I have a new hash table implementation lying around on my local machine
that is specifically designed for storing pointers. It lowers memory usage,
and has much better cache locality and general performance is nice.
It has reduced the impact of hash tables quite noticeably with the testing
that I've done as of yet. I'll see if I can get it out the door soon,
as it might
come in handy for usecases like this.

>>
>> Am 11.01.2018 04:53, schrieb Dave Airlie:
>>>
>>> From: Dave Airlie 
>>>
>>> This should make the merging of cmd buffers less CPU intensive,
>>> note I said *should* :)
>>> ---
>>>  src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c | 47
>>> ---
>>>  1 file changed, 20 insertions(+), 27 deletions(-)
>>>
>>> diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
>>> b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
>>> index 0ee56f91447..9a39d237ae8 100644
>>> --- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
>>> +++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
>>> @@ -33,6 +33,7 @@
>>>  #include "radv_amdgpu_bo.h"
>>>  #include "sid.h"
>>>
>>> +#include "util/hash_table.h"
>>>
>>>  enum {
>>> VIRTUAL_BUFFER_HASH_TABLE_SIZE = 1024
>>> @@ -584,6 +585,9 @@ static int radv_amdgpu_create_bo_list(struct
>>> radv_amdgpu_winsys *ws,
>>> priorities[0] = 8;
>>> }
>>>
>>> +   struct hash_table *ht = _mesa_hash_table_create(NULL,
>>> _mesa_hash_pointer,
>>> +
>>> _mesa_key_pointer_equal);
>>> +
>>> for (unsigned i = 0; i < count + !!extra_cs; ++i) {
>>> struct radv_amdgpu_cs *cs;
>>>
>>> @@ -595,50 +599,39 @@ static int radv_amdgpu_create_bo_list(struct
>>> radv_amdgpu_winsys *ws,
>>> if (!cs->num_buffers)
>>> continue;
>>>
>>> -   if (unique_bo_count == 0) {
>>> -   memcpy(handles, cs->handles,
>>> cs->num_buffers * sizeof(amdgpu_bo_handle));
>>> -   memcpy(priorities, cs->priorities,
>>> cs->num_buffers * sizeof(uint8_t));
>>> -   unique_bo_count = cs->num_buffers;
>>> -   continue;
>>> -   }
>>> -   int unique_bo_so_far = unique_bo_count;
>>> for (unsigned j = 0; j < cs->num_buffers; ++j) {
>>> -   bool found = false;
>>> -   for (unsigned k = 0; k < unique_bo_so_far;
>>> ++k) {
>>> -   if (handles[k] == cs->handles[j])
>>> {
>>> -   found = true;
>>> -   priorities[k] =
>>> MAX2(priorities[k],
>>> -
>>> cs->priorities[j]);
>>> -   break;
>>> -   }
>>> -   }
>>> -   if (!found) {
>>> +   struct hash_entry *entry =
>>> _mesa_hash_table_search(ht, (void
>>> *)cs->handles[j]);
>>> +   if (!entry) {
>>> +   _mesa_hash_table_insert(ht, (void
>>> *)cs->handles[j], (void
>>> *)(uintptr_t)unique_bo_count);
>>> handles[unique_bo_count] =
>>> cs->handles[j];
>>> priorities[unique_bo_count] =
>>> cs->priorities[j];
>>> ++unique_bo_count;
>>> +   } else {
>>> +   int bo_idx = (uint32_t)(unsigned
>>> long)entry->data;
>>> +   priorities[bo_idx] =
>>> MAX2(priorities[bo_idx],
>>> +
>>> cs->priorities[j]);
>>> }
>>> }
>>> for (unsigned j = 0; j < cs->num_virtual_buffers;
>>> ++j) {
>>> struct radv_amdgpu_winsys_bo *virtual_bo =
>>> radv_amdgpu_winsys_bo(cs->virtual_buffers[j]);
>>> for(unsigned k = 0; k <
>>> virtual_bo->bo_count; ++k) {
>>> struct radv_amdgpu_winsys_bo *bo =
>>> virtual_bo->bos[k];
>>> -   bool found = false;
>>> -   for (unsigned m = 0; m <
>>> unique_bo_count; ++m) {
>>> -   if (handles[m] == bo->bo)
>>> {
>>> -   found = true;

Re: [Mesa-dev] [PATCH] radv/winsys: replace bo list searchs with a hash table.

2018-01-30 Thread Dieter Nützel 

Am 12.01.2018 01:31, schrieb Dieter Nützel:

Tested-by: Dieter Nützel 

This offer an additional 1-1,5% (~20 fps) boost with 'smoketest'.

6290 presents in 5.00045 seconds (FPS: 1257.89)
6291 presents in 5.00041 seconds (FPS: 1258.1)
6293 presents in 5.00065 seconds (FPS: 1258.44)
6289 presents in 5.0006 seconds (FPS: 1257.65)

But _only_ with 'ondemand' governor!!!

cpupower frequency-set -g ondemand => GOOD
cpupower frequency-set -g performance => BAD
cpupower frequency-set -g schedutil => BAD

performance: 100-110 fps _lower_
schedutil: 110-140 fps _lower_


Hello Dave,

if you refer to this part,

I'm not sure this was much of a win, and if it makes things worse in 
some
cases, then it needs a lot more investigation, so probably consider it 
dead

for now.

Dave.


it was _only_ additional finding on my side.
Boost is there, I'm running daily with it.

'New' finding for me was, that 'performance' and 'schedutil' is slower 
as 'ondemand' for Vulkan (only radv tested) opposed to OpenGL, where 
'performance' is the clear winner (in most games much more min and avg 
plus max fps).


With 'cpupower frequency-set -g performance' I get _more_ noise (bus?) 
from my system during 'smoketest'. (Apart from this patch.) And it is 
somewhat slower.


Your call.

Greetings,
Dieter



Am 11.01.2018 04:53, schrieb Dave Airlie:

From: Dave Airlie 

This should make the merging of cmd buffers less CPU intensive,
note I said *should* :)
---
 src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c | 47 
---

 1 file changed, 20 insertions(+), 27 deletions(-)

diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
index 0ee56f91447..9a39d237ae8 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
@@ -33,6 +33,7 @@
 #include "radv_amdgpu_bo.h"
 #include "sid.h"

+#include "util/hash_table.h"

 enum {
VIRTUAL_BUFFER_HASH_TABLE_SIZE = 1024
@@ -584,6 +585,9 @@ static int radv_amdgpu_create_bo_list(struct
radv_amdgpu_winsys *ws,
priorities[0] = 8;
}

+		struct hash_table *ht = _mesa_hash_table_create(NULL, 
_mesa_hash_pointer,

+   
_mesa_key_pointer_equal);
+
for (unsigned i = 0; i < count + !!extra_cs; ++i) {
struct radv_amdgpu_cs *cs;

@@ -595,50 +599,39 @@ static int radv_amdgpu_create_bo_list(struct
radv_amdgpu_winsys *ws,
if (!cs->num_buffers)
continue;

-   if (unique_bo_count == 0) {
-memcpy(handles, cs->handles, cs->num_buffers * 
sizeof(amdgpu_bo_handle));
-memcpy(priorities, cs->priorities, cs->num_buffers * 
sizeof(uint8_t));

-   unique_bo_count = cs->num_buffers;
-   continue;
-   }
-   int unique_bo_so_far = unique_bo_count;
for (unsigned j = 0; j < cs->num_buffers; ++j) {
-   bool found = false;
-   for (unsigned k = 0; k < unique_bo_so_far; ++k) 
{
-   if (handles[k] == cs->handles[j]) {
-   found = true;
-   priorities[k] = 
MAX2(priorities[k],
-
cs->priorities[j]);
-   break;
-   }
-   }
-   if (!found) {
+   struct hash_entry *entry = 
_mesa_hash_table_search(ht, (void
*)cs->handles[j]);
+   if (!entry) {
+   _mesa_hash_table_insert(ht, (void 
*)cs->handles[j], (void
*)(uintptr_t)unique_bo_count);
handles[unique_bo_count] = 
cs->handles[j];
priorities[unique_bo_count] = 
cs->priorities[j];
++unique_bo_count;
+   } else {
+   int bo_idx = (uint32_t)(unsigned 
long)entry->data;
+   priorities[bo_idx] = 
MAX2(priorities[bo_idx],
+ 
cs->priorities[j]);
}
}
for (unsigned j = 0; j < cs->num_virtual_buffers; ++j) {
struct radv_amdgpu_winsys_bo *virtual_bo =
radv_amdgpu_winsys_bo(cs->virtual_buffers[j]);
for(unsigned k = 0; k < virtual_bo->bo_count; 
++k) {
struct

Re: [Mesa-dev] [PATCH] radv/winsys: replace bo list searchs with a hash table.

2018-01-30 Thread Dave Airlie 
On 31 January 2018 at 12:27, Dieter Nützel  wrote:
> Ping!

I'm not sure this was much of a win, and if it makes things worse in some cases,
then it needs a lot more investigation, so probably consider it dead for now.

Dave.

>
> Am 11.01.2018 04:53, schrieb Dave Airlie:
>>
>> From: Dave Airlie 
>>
>> This should make the merging of cmd buffers less CPU intensive,
>> note I said *should* :)
>> ---
>>  src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c | 47
>> ---
>>  1 file changed, 20 insertions(+), 27 deletions(-)
>>
>> diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
>> b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
>> index 0ee56f91447..9a39d237ae8 100644
>> --- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
>> +++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
>> @@ -33,6 +33,7 @@
>>  #include "radv_amdgpu_bo.h"
>>  #include "sid.h"
>>
>> +#include "util/hash_table.h"
>>
>>  enum {
>> VIRTUAL_BUFFER_HASH_TABLE_SIZE = 1024
>> @@ -584,6 +585,9 @@ static int radv_amdgpu_create_bo_list(struct
>> radv_amdgpu_winsys *ws,
>> priorities[0] = 8;
>> }
>>
>> +   struct hash_table *ht = _mesa_hash_table_create(NULL,
>> _mesa_hash_pointer,
>> +
>> _mesa_key_pointer_equal);
>> +
>> for (unsigned i = 0; i < count + !!extra_cs; ++i) {
>> struct radv_amdgpu_cs *cs;
>>
>> @@ -595,50 +599,39 @@ static int radv_amdgpu_create_bo_list(struct
>> radv_amdgpu_winsys *ws,
>> if (!cs->num_buffers)
>> continue;
>>
>> -   if (unique_bo_count == 0) {
>> -   memcpy(handles, cs->handles,
>> cs->num_buffers * sizeof(amdgpu_bo_handle));
>> -   memcpy(priorities, cs->priorities,
>> cs->num_buffers * sizeof(uint8_t));
>> -   unique_bo_count = cs->num_buffers;
>> -   continue;
>> -   }
>> -   int unique_bo_so_far = unique_bo_count;
>> for (unsigned j = 0; j < cs->num_buffers; ++j) {
>> -   bool found = false;
>> -   for (unsigned k = 0; k < unique_bo_so_far;
>> ++k) {
>> -   if (handles[k] == cs->handles[j])
>> {
>> -   found = true;
>> -   priorities[k] =
>> MAX2(priorities[k],
>> -
>> cs->priorities[j]);
>> -   break;
>> -   }
>> -   }
>> -   if (!found) {
>> +   struct hash_entry *entry =
>> _mesa_hash_table_search(ht, (void
>> *)cs->handles[j]);
>> +   if (!entry) {
>> +   _mesa_hash_table_insert(ht, (void
>> *)cs->handles[j], (void
>> *)(uintptr_t)unique_bo_count);
>> handles[unique_bo_count] =
>> cs->handles[j];
>> priorities[unique_bo_count] =
>> cs->priorities[j];
>> ++unique_bo_count;
>> +   } else {
>> +   int bo_idx = (uint32_t)(unsigned
>> long)entry->data;
>> +   priorities[bo_idx] =
>> MAX2(priorities[bo_idx],
>> +
>> cs->priorities[j]);
>> }
>> }
>> for (unsigned j = 0; j < cs->num_virtual_buffers;
>> ++j) {
>> struct radv_amdgpu_winsys_bo *virtual_bo =
>> radv_amdgpu_winsys_bo(cs->virtual_buffers[j]);
>> for(unsigned k = 0; k <
>> virtual_bo->bo_count; ++k) {
>> struct radv_amdgpu_winsys_bo *bo =
>> virtual_bo->bos[k];
>> -   bool found = false;
>> -   for (unsigned m = 0; m <
>> unique_bo_count; ++m) {
>> -   if (handles[m] == bo->bo)
>> {
>> -   found = true;
>> -   priorities[m] =
>> MAX2(priorities[m],
>> -
>> cs->virtual_buffer_priorities[j]);
>> -   break;
>> -   }
>> -   }
>> -   if (!found) {
>> +
>> +   struct hash_entry *entry =
>> _mesa_hash_table_search(ht, (void *)bo->bo);
>> +   if (!entry) {
>> +
>> _mesa_hash_table_insert(ht, (void

Re: [Mesa-dev] [PATCH] radv/winsys: replace bo list searchs with a hash table.

2018-01-30 Thread Dieter Nützel 

Ping!

Am 11.01.2018 04:53, schrieb Dave Airlie:

From: Dave Airlie 

This should make the merging of cmd buffers less CPU intensive,
note I said *should* :)
---
 src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c | 47 
---

 1 file changed, 20 insertions(+), 27 deletions(-)

diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
index 0ee56f91447..9a39d237ae8 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
@@ -33,6 +33,7 @@
 #include "radv_amdgpu_bo.h"
 #include "sid.h"

+#include "util/hash_table.h"

 enum {
VIRTUAL_BUFFER_HASH_TABLE_SIZE = 1024
@@ -584,6 +585,9 @@ static int radv_amdgpu_create_bo_list(struct
radv_amdgpu_winsys *ws,
priorities[0] = 8;
}

+		struct hash_table *ht = _mesa_hash_table_create(NULL, 
_mesa_hash_pointer,

+   
_mesa_key_pointer_equal);
+
for (unsigned i = 0; i < count + !!extra_cs; ++i) {
struct radv_amdgpu_cs *cs;

@@ -595,50 +599,39 @@ static int radv_amdgpu_create_bo_list(struct
radv_amdgpu_winsys *ws,
if (!cs->num_buffers)
continue;

-   if (unique_bo_count == 0) {
-memcpy(handles, cs->handles, cs->num_buffers * 
sizeof(amdgpu_bo_handle));
-memcpy(priorities, cs->priorities, cs->num_buffers * 
sizeof(uint8_t));

-   unique_bo_count = cs->num_buffers;
-   continue;
-   }
-   int unique_bo_so_far = unique_bo_count;
for (unsigned j = 0; j < cs->num_buffers; ++j) {
-   bool found = false;
-   for (unsigned k = 0; k < unique_bo_so_far; ++k) 
{
-   if (handles[k] == cs->handles[j]) {
-   found = true;
-   priorities[k] = 
MAX2(priorities[k],
-
cs->priorities[j]);
-   break;
-   }
-   }
-   if (!found) {
+   struct hash_entry *entry = 
_mesa_hash_table_search(ht, (void
*)cs->handles[j]);
+   if (!entry) {
+   _mesa_hash_table_insert(ht, (void 
*)cs->handles[j], (void
*)(uintptr_t)unique_bo_count);
handles[unique_bo_count] = 
cs->handles[j];
priorities[unique_bo_count] = 
cs->priorities[j];
++unique_bo_count;
+   } else {
+   int bo_idx = (uint32_t)(unsigned 
long)entry->data;
+   priorities[bo_idx] = 
MAX2(priorities[bo_idx],
+ 
cs->priorities[j]);
}
}
for (unsigned j = 0; j < cs->num_virtual_buffers; ++j) {
struct radv_amdgpu_winsys_bo *virtual_bo =
radv_amdgpu_winsys_bo(cs->virtual_buffers[j]);
for(unsigned k = 0; k < virtual_bo->bo_count; 
++k) {
struct radv_amdgpu_winsys_bo *bo = 
virtual_bo->bos[k];
-   bool found = false;
-   for (unsigned m = 0; m < 
unique_bo_count; ++m) {
-   if (handles[m] == bo->bo) {
-   found = true;
-   priorities[m] = 
MAX2(priorities[m],
-   
cs->virtual_buffer_priorities[j]);
-   break;
-   }
-   }
-   if (!found) {
+
+	struct hash_entry *entry = _mesa_hash_table_search(ht, (void 
*)bo->bo);

+   if (!entry) {
+   _mesa_hash_table_insert(ht, (void 
*)bo->bo, (void
*)(uintptr_t)unique_bo_count);
handles[unique_bo_count] = 
bo->bo;
priorities[unique_bo_count] = 
cs->virtual_buffer_priorities[j];
++unique_bo_count;
+   }

Re: [Mesa-dev] [PATCH] radv/winsys: replace bo list searchs with a hash table.

2018-01-18 Thread Dieter Nützel 

Ping!
Any progress?

Am 11.01.2018 04:53, schrieb Dave Airlie:

From: Dave Airlie 

This should make the merging of cmd buffers less CPU intensive,
note I said *should* :)
---
 src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c | 47 
---

 1 file changed, 20 insertions(+), 27 deletions(-)

diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
index 0ee56f91447..9a39d237ae8 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
@@ -33,6 +33,7 @@
 #include "radv_amdgpu_bo.h"
 #include "sid.h"

+#include "util/hash_table.h"

 enum {
VIRTUAL_BUFFER_HASH_TABLE_SIZE = 1024
@@ -584,6 +585,9 @@ static int radv_amdgpu_create_bo_list(struct
radv_amdgpu_winsys *ws,
priorities[0] = 8;
}

+		struct hash_table *ht = _mesa_hash_table_create(NULL, 
_mesa_hash_pointer,

+   
_mesa_key_pointer_equal);
+
for (unsigned i = 0; i < count + !!extra_cs; ++i) {
struct radv_amdgpu_cs *cs;

@@ -595,50 +599,39 @@ static int radv_amdgpu_create_bo_list(struct
radv_amdgpu_winsys *ws,
if (!cs->num_buffers)
continue;

-   if (unique_bo_count == 0) {
-memcpy(handles, cs->handles, cs->num_buffers * 
sizeof(amdgpu_bo_handle));
-memcpy(priorities, cs->priorities, cs->num_buffers * 
sizeof(uint8_t));

-   unique_bo_count = cs->num_buffers;
-   continue;
-   }
-   int unique_bo_so_far = unique_bo_count;
for (unsigned j = 0; j < cs->num_buffers; ++j) {
-   bool found = false;
-   for (unsigned k = 0; k < unique_bo_so_far; ++k) 
{
-   if (handles[k] == cs->handles[j]) {
-   found = true;
-   priorities[k] = 
MAX2(priorities[k],
-
cs->priorities[j]);
-   break;
-   }
-   }
-   if (!found) {
+   struct hash_entry *entry = 
_mesa_hash_table_search(ht, (void
*)cs->handles[j]);
+   if (!entry) {
+   _mesa_hash_table_insert(ht, (void 
*)cs->handles[j], (void
*)(uintptr_t)unique_bo_count);
handles[unique_bo_count] = 
cs->handles[j];
priorities[unique_bo_count] = 
cs->priorities[j];
++unique_bo_count;
+   } else {
+   int bo_idx = (uint32_t)(unsigned 
long)entry->data;
+   priorities[bo_idx] = 
MAX2(priorities[bo_idx],
+ 
cs->priorities[j]);
}
}
for (unsigned j = 0; j < cs->num_virtual_buffers; ++j) {
struct radv_amdgpu_winsys_bo *virtual_bo =
radv_amdgpu_winsys_bo(cs->virtual_buffers[j]);
for(unsigned k = 0; k < virtual_bo->bo_count; 
++k) {
struct radv_amdgpu_winsys_bo *bo = 
virtual_bo->bos[k];
-   bool found = false;
-   for (unsigned m = 0; m < 
unique_bo_count; ++m) {
-   if (handles[m] == bo->bo) {
-   found = true;
-   priorities[m] = 
MAX2(priorities[m],
-   
cs->virtual_buffer_priorities[j]);
-   break;
-   }
-   }
-   if (!found) {
+
+	struct hash_entry *entry = _mesa_hash_table_search(ht, (void 
*)bo->bo);

+   if (!entry) {
+   _mesa_hash_table_insert(ht, (void 
*)bo->bo, (void
*)(uintptr_t)unique_bo_count);
handles[unique_bo_count] = 
bo->bo;
priorities[unique_bo_count] = 
cs->virtual_buffer_priorities[j];
++unique_bo_count;
+

Re: [Mesa-dev] [PATCH] radv/winsys: replace bo list searchs with a hash table.

2018-01-11 Thread Dieter Nützel 

Tested-by: Dieter Nützel 

This offer an additional 1-1,5% (~20 fps) boost with 'smoketest'.

6290 presents in 5.00045 seconds (FPS: 1257.89)
6291 presents in 5.00041 seconds (FPS: 1258.1)
6293 presents in 5.00065 seconds (FPS: 1258.44)
6289 presents in 5.0006 seconds (FPS: 1257.65)

But _only_ with 'ondemand' governor!!!

cpupower frequency-set -g ondemand => GOOD
cpupower frequency-set -g performance => BAD
cpupower frequency-set -g schedutil => BAD

performance: 100-110 fps _lower_
schedutil: 110-140 fps _lower_

Cheers,
Dieter

Am 11.01.2018 04:53, schrieb Dave Airlie:

From: Dave Airlie 

This should make the merging of cmd buffers less CPU intensive,
note I said *should* :)
---
 src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c | 47 
---

 1 file changed, 20 insertions(+), 27 deletions(-)

diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
index 0ee56f91447..9a39d237ae8 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
@@ -33,6 +33,7 @@
 #include "radv_amdgpu_bo.h"
 #include "sid.h"

+#include "util/hash_table.h"

 enum {
VIRTUAL_BUFFER_HASH_TABLE_SIZE = 1024
@@ -584,6 +585,9 @@ static int radv_amdgpu_create_bo_list(struct
radv_amdgpu_winsys *ws,
priorities[0] = 8;
}

+		struct hash_table *ht = _mesa_hash_table_create(NULL, 
_mesa_hash_pointer,

+   
_mesa_key_pointer_equal);
+
for (unsigned i = 0; i < count + !!extra_cs; ++i) {
struct radv_amdgpu_cs *cs;

@@ -595,50 +599,39 @@ static int radv_amdgpu_create_bo_list(struct
radv_amdgpu_winsys *ws,
if (!cs->num_buffers)
continue;

-   if (unique_bo_count == 0) {
-memcpy(handles, cs->handles, cs->num_buffers * 
sizeof(amdgpu_bo_handle));
-memcpy(priorities, cs->priorities, cs->num_buffers * 
sizeof(uint8_t));

-   unique_bo_count = cs->num_buffers;
-   continue;
-   }
-   int unique_bo_so_far = unique_bo_count;
for (unsigned j = 0; j < cs->num_buffers; ++j) {
-   bool found = false;
-   for (unsigned k = 0; k < unique_bo_so_far; ++k) 
{
-   if (handles[k] == cs->handles[j]) {
-   found = true;
-   priorities[k] = 
MAX2(priorities[k],
-
cs->priorities[j]);
-   break;
-   }
-   }
-   if (!found) {
+   struct hash_entry *entry = 
_mesa_hash_table_search(ht, (void
*)cs->handles[j]);
+   if (!entry) {
+   _mesa_hash_table_insert(ht, (void 
*)cs->handles[j], (void
*)(uintptr_t)unique_bo_count);
handles[unique_bo_count] = 
cs->handles[j];
priorities[unique_bo_count] = 
cs->priorities[j];
++unique_bo_count;
+   } else {
+   int bo_idx = (uint32_t)(unsigned 
long)entry->data;
+   priorities[bo_idx] = 
MAX2(priorities[bo_idx],
+ 
cs->priorities[j]);
}
}
for (unsigned j = 0; j < cs->num_virtual_buffers; ++j) {
struct radv_amdgpu_winsys_bo *virtual_bo =
radv_amdgpu_winsys_bo(cs->virtual_buffers[j]);
for(unsigned k = 0; k < virtual_bo->bo_count; 
++k) {
struct radv_amdgpu_winsys_bo *bo = 
virtual_bo->bos[k];
-   bool found = false;
-   for (unsigned m = 0; m < 
unique_bo_count; ++m) {
-   if (handles[m] == bo->bo) {
-   found = true;
-   priorities[m] = 
MAX2(priorities[m],
-   
cs->virtual_buffer_priorities[j]);
-   break;
-   }
-   }
-   if (!found) {
+
+	struct

Re: [Mesa-dev] [PATCH] radv/winsys: replace bo list searchs with a hash table.

2018-01-11 Thread Samuel Pitoiset 
mmh, last time I tried to use a hash table for the amdgpu winsys of 
RadeonSI it was *worse*, so not sure if that really helps.


Though, I don't have any numbers to show you because it was a while ago.

On 01/11/2018 04:53 AM, Dave Airlie wrote:

From: Dave Airlie 

This should make the merging of cmd buffers less CPU intensive,
note I said *should* :)
---
  src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c | 47 ---
  1 file changed, 20 insertions(+), 27 deletions(-)

diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c 
b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
index 0ee56f91447..9a39d237ae8 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
@@ -33,6 +33,7 @@
  #include "radv_amdgpu_bo.h"
  #include "sid.h"
  
+#include "util/hash_table.h"
  
  enum {

VIRTUAL_BUFFER_HASH_TABLE_SIZE = 1024
@@ -584,6 +585,9 @@ static int radv_amdgpu_create_bo_list(struct 
radv_amdgpu_winsys *ws,
priorities[0] = 8;
}
  
+		struct hash_table *ht = _mesa_hash_table_create(NULL, _mesa_hash_pointer,

+   
_mesa_key_pointer_equal);
+
for (unsigned i = 0; i < count + !!extra_cs; ++i) {
struct radv_amdgpu_cs *cs;
  
@@ -595,50 +599,39 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,

if (!cs->num_buffers)
continue;
  
-			if (unique_bo_count == 0) {

-   memcpy(handles, cs->handles, cs->num_buffers * 
sizeof(amdgpu_bo_handle));
-   memcpy(priorities, cs->priorities, 
cs->num_buffers * sizeof(uint8_t));
-   unique_bo_count = cs->num_buffers;
-   continue;
-   }
-   int unique_bo_so_far = unique_bo_count;
for (unsigned j = 0; j < cs->num_buffers; ++j) {
-   bool found = false;
-   for (unsigned k = 0; k < unique_bo_so_far; ++k) 
{
-   if (handles[k] == cs->handles[j]) {
-   found = true;
-   priorities[k] = 
MAX2(priorities[k],
-
cs->priorities[j]);
-   break;
-   }
-   }
-   if (!found) {
+   struct hash_entry *entry = 
_mesa_hash_table_search(ht, (void *)cs->handles[j]);
+   if (!entry) {
+   _mesa_hash_table_insert(ht, (void 
*)cs->handles[j], (void *)(uintptr_t)unique_bo_count);
handles[unique_bo_count] = 
cs->handles[j];
priorities[unique_bo_count] = 
cs->priorities[j];
++unique_bo_count;
+   } else {
+   int bo_idx = (uint32_t)(unsigned 
long)entry->data;
+   priorities[bo_idx] = 
MAX2(priorities[bo_idx],
+ 
cs->priorities[j]);
}
}
for (unsigned j = 0; j < cs->num_virtual_buffers; ++j) {
struct radv_amdgpu_winsys_bo *virtual_bo = 
radv_amdgpu_winsys_bo(cs->virtual_buffers[j]);
for(unsigned k = 0; k < virtual_bo->bo_count; 
++k) {
struct radv_amdgpu_winsys_bo *bo = 
virtual_bo->bos[k];
-   bool found = false;
-   for (unsigned m = 0; m < 
unique_bo_count; ++m) {
-   if (handles[m] == bo->bo) {
-   found = true;
-   priorities[m] = 
MAX2(priorities[m],
-   
cs->virtual_buffer_priorities[j]);
-   break;
-   }
-   }
-   if (!found) {
+
+   struct hash_entry *entry = 
_mesa_hash_table_search(ht, (void *)bo->bo);
+   if (!entry) {
+   _mesa_hash_table_insert(ht, (void 
*)bo->bo, (void *)(uintptr_t)unique_bo_count);

[Mesa-dev] [PATCH] radv/winsys: replace bo list searchs with a hash table.

2018-01-10 Thread Dave Airlie 
From: Dave Airlie 

This should make the merging of cmd buffers less CPU intensive,
note I said *should* :)
---
 src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c | 47 ---
 1 file changed, 20 insertions(+), 27 deletions(-)

diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c 
b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
index 0ee56f91447..9a39d237ae8 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
@@ -33,6 +33,7 @@
 #include "radv_amdgpu_bo.h"
 #include "sid.h"
 
+#include "util/hash_table.h"
 
 enum {
VIRTUAL_BUFFER_HASH_TABLE_SIZE = 1024
@@ -584,6 +585,9 @@ static int radv_amdgpu_create_bo_list(struct 
radv_amdgpu_winsys *ws,
priorities[0] = 8;
}
 
+   struct hash_table *ht = _mesa_hash_table_create(NULL, 
_mesa_hash_pointer,
+   
_mesa_key_pointer_equal);
+
for (unsigned i = 0; i < count + !!extra_cs; ++i) {
struct radv_amdgpu_cs *cs;
 
@@ -595,50 +599,39 @@ static int radv_amdgpu_create_bo_list(struct 
radv_amdgpu_winsys *ws,
if (!cs->num_buffers)
continue;
 
-   if (unique_bo_count == 0) {
-   memcpy(handles, cs->handles, cs->num_buffers * 
sizeof(amdgpu_bo_handle));
-   memcpy(priorities, cs->priorities, 
cs->num_buffers * sizeof(uint8_t));
-   unique_bo_count = cs->num_buffers;
-   continue;
-   }
-   int unique_bo_so_far = unique_bo_count;
for (unsigned j = 0; j < cs->num_buffers; ++j) {
-   bool found = false;
-   for (unsigned k = 0; k < unique_bo_so_far; ++k) 
{
-   if (handles[k] == cs->handles[j]) {
-   found = true;
-   priorities[k] = 
MAX2(priorities[k],
-
cs->priorities[j]);
-   break;
-   }
-   }
-   if (!found) {
+   struct hash_entry *entry = 
_mesa_hash_table_search(ht, (void *)cs->handles[j]);
+   if (!entry) {
+   _mesa_hash_table_insert(ht, (void 
*)cs->handles[j], (void *)(uintptr_t)unique_bo_count);
handles[unique_bo_count] = 
cs->handles[j];
priorities[unique_bo_count] = 
cs->priorities[j];
++unique_bo_count;
+   } else {
+   int bo_idx = (uint32_t)(unsigned 
long)entry->data;
+   priorities[bo_idx] = 
MAX2(priorities[bo_idx],
+ 
cs->priorities[j]);
}
}
for (unsigned j = 0; j < cs->num_virtual_buffers; ++j) {
struct radv_amdgpu_winsys_bo *virtual_bo = 
radv_amdgpu_winsys_bo(cs->virtual_buffers[j]);
for(unsigned k = 0; k < virtual_bo->bo_count; 
++k) {
struct radv_amdgpu_winsys_bo *bo = 
virtual_bo->bos[k];
-   bool found = false;
-   for (unsigned m = 0; m < 
unique_bo_count; ++m) {
-   if (handles[m] == bo->bo) {
-   found = true;
-   priorities[m] = 
MAX2(priorities[m],
-   
cs->virtual_buffer_priorities[j]);
-   break;
-   }
-   }
-   if (!found) {
+
+   struct hash_entry *entry = 
_mesa_hash_table_search(ht, (void *)bo->bo);
+   if (!entry) {
+   _mesa_hash_table_insert(ht, 
(void *)bo->bo, (void *)(uintptr_t)unique_bo_count);
handles[unique_bo_count] = 
bo->bo;
priorities[unique_bo_count] = 
cs->virtual_buffer_priorities[j];