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

[Dieter Nützel]

Tested-by: Dieter Nützel <Dieter at nuetzel-hh.de>

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 <airlied at redhat.com>
> 
> 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;
> +					} else {
> +						int bo_idx = (uint32_t)(unsigned long)entry->data;
> +						priorities[bo_idx] = MAX2(priorities[bo_idx],
> +									  cs->virtual_buffer_priorities[j]);
>  					}
>  				}
>  			}
>  		}
> +		_mesa_hash_table_destroy(ht, NULL);
> 
>  		if (unique_bo_count > 0) {
>  			r = amdgpu_bo_list_create(ws->dev, unique_bo_count, handles,