[Mesa-dev] [PATCH v2 56/64] radeonsi: decompress resident textures/images before graphics/compute
Samuel Pitoiset
samuel.pitoiset at gmail.com
Wed Jun 7 07:22:50 UTC 2017
On 06/07/2017 01:46 AM, Marek Olšák wrote:
> On Wed, Jun 7, 2017 at 1:27 AM, Marek Olšák <maraeo at gmail.com> wrote:
>> On Tue, May 30, 2017 at 10:36 PM, Samuel Pitoiset
>> <samuel.pitoiset at gmail.com> wrote:
>>> Similar to the existing decompression code path except that it
>>> loops over the list of resident textures/images.
>>>
>>> v2: - store pipe_sampler_view instead of si_sampler_view
>>>
>>> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
>>> ---
>>> src/gallium/drivers/radeonsi/si_blit.c | 77 +++++++++++++++++++++++++--
>>> src/gallium/drivers/radeonsi/si_descriptors.c | 52 ++++++++++++++++++
>>> src/gallium/drivers/radeonsi/si_pipe.h | 3 ++
>>> 3 files changed, 129 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/src/gallium/drivers/radeonsi/si_blit.c b/src/gallium/drivers/radeonsi/si_blit.c
>>> index 343ca35736..a47f43958c 100644
>>> --- a/src/gallium/drivers/radeonsi/si_blit.c
>>> +++ b/src/gallium/drivers/radeonsi/si_blit.c
>>> @@ -22,6 +22,7 @@
>>> */
>>>
>>> #include "si_pipe.h"
>>> +#include "si_compute.h"
>>> #include "util/u_format.h"
>>> #include "util/u_surface.h"
>>>
>>> @@ -690,9 +691,6 @@ static void si_decompress_textures(struct si_context *sctx, unsigned shader_mask
>>> {
>>> unsigned compressed_colortex_counter, mask;
>>>
>>> - if (sctx->blitter->running)
>>> - return;
>>> -
>>
>> You can keep this.
>>
>>> /* Update the compressed_colortex_mask if necessary. */
>>> compressed_colortex_counter = p_atomic_read(&sctx->screen->b.compressed_colortex_counter);
>>> if (compressed_colortex_counter != sctx->b.last_compressed_colortex_counter) {
>>> @@ -719,14 +717,87 @@ static void si_decompress_textures(struct si_context *sctx, unsigned shader_mask
>>> si_check_render_feedback(sctx);
>>> }
>>>
>>> +static void si_decompress_resident_textures(struct si_context *sctx)
>>> +{
>>> + unsigned num_resident_tex_handles;
>>> + unsigned i;
>>> +
>>> + num_resident_tex_handles = sctx->resident_tex_handles.size /
>>> + sizeof(struct si_texture_handle *);
>>> +
>>> + for (i = 0; i < num_resident_tex_handles; i++) {
>>> + struct si_texture_handle *tex_handle =
>>> + *util_dynarray_element(&sctx->resident_tex_handles,
>>> + struct si_texture_handle *, i);
>>> + struct pipe_sampler_view *view = tex_handle->view;
>>> + struct si_sampler_view *sview = (struct si_sampler_view *)view;
>>> + struct r600_texture *tex;
>>> +
>>> + assert(view);
>>> + tex = (struct r600_texture *)view->texture;
>>> +
>>> + if (view->texture->target == PIPE_BUFFER)
>>> + continue;
>>> +
>>> + if (tex_handle->compressed_colortex)
>>> + si_decompress_color_texture(sctx, tex, view->u.tex.first_level,
>>> + view->u.tex.last_level);
>>> +
>>> + if (tex_handle->depth_texture)
>>> + si_flush_depth_texture(sctx, tex,
>>> + sview->is_stencil_sampler ? PIPE_MASK_S : PIPE_MASK_Z,
>>> + view->u.tex.first_level, view->u.tex.last_level,
>>> + 0, util_max_layer(&tex->resource.b.b, view->u.tex.first_level));
>>> + }
>>> +}
>>> +
>>> +static void si_decompress_resident_images(struct si_context *sctx)
>>> +{
>>> + unsigned num_resident_img_handles;
>>> + unsigned i;
>>> +
>>> + num_resident_img_handles = sctx->resident_img_handles.size /
>>> + sizeof(struct si_image_handle *);
>>> +
>>> + for (i = 0; i < num_resident_img_handles; i++) {
>>> + struct si_image_handle *img_handle =
>>> + *util_dynarray_element(&sctx->resident_img_handles,
>>> + struct si_image_handle *, i);
>>> + struct pipe_image_view *view = &img_handle->view;
>>> + struct r600_texture *tex;
>>> +
>>> + assert(view);
>>> + tex = (struct r600_texture *)view->resource;
>>> +
>>> + if (view->resource->target == PIPE_BUFFER)
>>> + continue;
>>> +
>>> + if (img_handle->compressed_colortex)
>>> + si_decompress_color_texture(sctx, tex, view->u.tex.level,
>>> + view->u.tex.level);
>>> + }
>>> +}
>>
>> The loops in the two functions above will destroy CPU performance,
>> because both functions are called for every draw call. We need to find
>> a better way.
>>
>> I suggest that si_resident_handles_update_compressed_colortex should
>> be rewritten to build 2 separate lists (for samples and images)
>> containing only references to bindless slots that need decompression,
>> and make_resident functions should update the separate lists too. Then
>> the two functions above can walk the separate lists instead of all
>> resident handles. The effect will be that only those slots that need
>> decompression will be checked. (ideally 0 of very few)
>
> Just saw patches 58-59. They certainly improve it a bit - at least
> apps not using bindless handles will not be affected.
>
> I guess the main optimization of this can be done in follow-up
> patches. My other comments below still apply here though.
This was the idea as well. This part can still be improved, but I would
prefer to do that in a separate patch. This series is, I think, big
enough. :)
>
>>
>>> +
>>> void si_decompress_graphics_textures(struct si_context *sctx)
>>> {
>>> + if (sctx->blitter->running)
>>> + return;
>>> +
>>> si_decompress_textures(sctx, u_bit_consecutive(0, SI_NUM_GRAPHICS_SHADERS));
>>> +
>>> + si_decompress_resident_textures(sctx);
>>> + si_decompress_resident_images(sctx);
>>
>> These changes...
>>
>>> }
>>>
>>> void si_decompress_compute_textures(struct si_context *sctx)
>>> {
>>> + if (sctx->blitter->running)
>>> + return;
>>> +
>>> si_decompress_textures(sctx, 1 << PIPE_SHADER_COMPUTE);
>>> +
>>> + si_decompress_resident_textures(sctx);
>>> + si_decompress_resident_images(sctx);
>>
>> ... and these changes can be moved into si_decompress_textures.
Sure.
Samuel
>>
>
> Marek
>
More information about the mesa-dev
mailing list