<div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Sun, Sep 25, 2016 at 9:59 AM, Jason Ekstrand <<a href="mailto:jason@jlekstrand.net" target="_blank">jason@jlekstrand.net</a>> wrote: <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">This enables lossless render target compression but only for first mip level of the first slice and only within a render pass.Â At the end of the render pass it always does a full resolve. Signed-off-by: Jason Ekstrand <<a href="mailto:jason@jlekstrand.net">jason@jlekstrand.net</a>> --- Â src/intel/vulkan/anv_blorp.cÂ Â Â Â | 30 ++++++++++++++++++++++++++++ Â src/intel/vulkan/anv_device.cÂ Â Â | 40 ++++++++++++++++++++++++++++++++------ Â src/intel/vulkan/anv_image.cÂ Â Â Â |Â 9 +++++++++ Â src/intel/vulkan/anv_private.hÂ Â Â |Â 2 ++ Â src/intel/vulkan/genX_cmd_buffer.c | 17 ++++++++++++++++ Â 5 files changed, 92 insertions(+), 6 deletions(-) diff --git a/src/intel/vulkan/anv_blorp.c b/src/intel/vulkan/anv_blorp.c index 5cc19eb..74f1d80 100644 --- a/src/intel/vulkan/anv_blorp.c +++ b/src/intel/vulkan/anv_blorp.c @@ -929,3 +929,33 @@ anv_cmd_buffer_resolve_subpass(struct anv_cmd_buffer *cmd_buffer) Â Â blorp_batch_finish(&batch); Â } + +void +anv_cmd_buffer_resolve_framebuffer(struct anv_cmd_buffer *cmd_buffer, +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct anv_framebuffer *fb) +{ +Â Â struct blorp_batch batch; +Â Â blorp_batch_init(&cmd_buffer->device->blorp, &batch, cmd_buffer); + +Â Â for (uint32_t i = 0; i < fb->attachment_count; i++) { +Â Â Â struct anv_framebuffer_attachment *fb_att = &fb->attachments[i]; +Â Â Â const struct anv_image *image = fb_att->view->image; + +Â Â Â if (fb_att->aux_usage != ISL_AUX_USAGE_CCS_E && +Â Â Â Â Â fb_att->aux_usage != ISL_AUX_USAGE_CCS_D) +Â Â Â Â Â continue; + +Â Â Â struct blorp_surf surf; +Â Â Â get_blorp_surf_for_anv_image(image, VK_IMAGE_ASPECT_COLOR_BIT, &surf); +Â Â Â surf.aux_surf = &image->aux_surface.isl; +Â Â Â surf.aux_addr = (struct blorp_address) { +Â Â Â Â Â .buffer = image->bo, +Â Â Â Â Â .offset = image->offset + image->aux_surface.offset, +Â Â Â }; +Â Â Â surf.aux_usage = fb_att->aux_usage; + +Â Â Â blorp_ccs_resolve(&batch, &surf, fb_att->view->isl.format); +Â Â } + +Â Â blorp_batch_finish(&batch); +} diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c index 3d4b783..8c68e94 100644 --- a/src/intel/vulkan/anv_device.c +++ b/src/intel/vulkan/anv_device.c @@ -1765,6 +1765,30 @@ void anv_DestroySampler( Â Â anv_free2(&device->alloc, pAllocator, sampler); Â } +static enum isl_aux_usage +get_fb_attachment_aux_usage(struct anv_device *device, +Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct anv_framebuffer *fb, +Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct anv_framebuffer_attachment *fb_att) +{ +Â Â if (fb_att->view->image->aux_surface.isl.size == 0) +Â Â Â return ISL_AUX_USAGE_NONE; /* No aux surface */ + +Â Â /* This is a color buffer, it had better be CCS */ +Â Â assert(fb_att->view->image->aux_surface.isl.usage & ISL_SURF_USAGE_CCS_BIT); + +Â Â if (!isl_format_supports_lossless_compression(&device->info, +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â fb_att->view->isl.format)) +Â Â Â return ISL_AUX_USAGE_NONE; /* Unsupported format */ + +Â Â if (fb->layers > 1 || fb_att->view->isl.base_array_layer > 0) +Â Â Â return ISL_AUX_USAGE_NONE; /* Multi-layered CCS not yet supported */ + +Â Â if (fb_att->view->isl.base_level > 0) +Â Â Â return ISL_AUX_USAGE_NONE; /* Multi-LOD CCS not yet supported */ + +Â Â return ISL_AUX_USAGE_CCS_E; +} + Â VkResult anv_CreateFramebuffer( Â Â Â VkDeviceÂ Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â _device, Â Â Â const VkFramebufferCreateInfo*Â Â Â Â Â Â Â pCreateInfo, @@ -1783,6 +1807,10 @@ VkResult anv_CreateFramebuffer( Â Â if (framebuffer == NULL) Â Â Â Â return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); +Â Â framebuffer->width = pCreateInfo->width; +Â Â framebuffer->height = pCreateInfo->height; +Â Â framebuffer->layers = pCreateInfo->layers; + Â Â framebuffer->attachment_count = pCreateInfo->attachmentCount; Â Â unsigned color_count = 0; Â Â for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) { @@ -1802,10 +1830,14 @@ VkResult anv_CreateFramebuffer( Â Â Â Â Â Â Â Â Â Â Â Â Â Â device->isl_dev.ss.align); Â Â unsigned color_idx = 0; Â Â for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) { -Â Â Â struct anv_image_view *iview = framebuffer->attachments[i].view; +Â Â Â struct anv_framebuffer_attachment *fb_att = &framebuffer->attachments[i]; +Â Â Â struct anv_image_view *iview = fb_att->view; + Â Â Â Â if (iview->image->aspects & VK_IMAGE_ASPECT_COLOR_BIT) { Â Â Â Â Â const uint16_t offset = (color_idx++) * device->isl_dev.ss.align; -Â Â Â Â Â framebuffer->attachments[i].rt_state_offset = offset; +Â Â Â Â Â fb_att->rt_state_offset = offset; +Â Â Â Â Â fb_att->aux_usage = +Â Â Â Â Â Â get_fb_attachment_aux_usage(device, framebuffer, fb_att); Â Â Â Â Â struct isl_view isl_view = iview->isl; Â Â Â Â Â isl_view.usage |= ISL_SURF_USAGE_RENDER_TARGET_BIT; @@ -1822,10 +1854,6 @@ VkResult anv_CreateFramebuffer( Â Â if (device->info.has_llc) Â Â Â Â anv_state_clflush(framebuffer->surface_states); -Â Â framebuffer->width = pCreateInfo->width; -Â Â framebuffer->height = pCreateInfo->height; -Â Â framebuffer->layers = pCreateInfo->layers; - Â Â *pFramebuffer = anv_framebuffer_to_handle(framebuffer); Â Â return VK_SUCCESS; diff --git a/src/intel/vulkan/anv_image.c b/src/intel/vulkan/anv_image.c index 405b923..ff99126 100644 --- a/src/intel/vulkan/anv_image.c +++ b/src/intel/vulkan/anv_image.c @@ -169,6 +169,15 @@ make_surface(const struct anv_device *dev, Â Â add_surface(image, anv_surf); +Â Â if (aspect == VK_IMAGE_ASPECT_COLOR_BIT) { +Â Â Â if (dev->info.gen >= 9 && vk_info->samples == 1) { +Â Â Â Â Â ok = isl_surf_get_ccs_surf(&dev->isl_dev, &anv_surf->isl, +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â &image->aux_surface.isl); +Â Â Â Â Â if (ok) +Â Â Â Â Â Â add_surface(image, &image->aux_surface); +Â Â Â } +Â Â } + Â Â return VK_SUCCESS; Â } diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h index 5c24c37..5f03ae0 100644 --- a/src/intel/vulkan/anv_private.h +++ b/src/intel/vulkan/anv_private.h @@ -1405,6 +1405,8 @@ anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer); Â void anv_cmd_buffer_clear_subpass(struct anv_cmd_buffer *cmd_buffer); Â void anv_cmd_buffer_resolve_subpass(struct anv_cmd_buffer *cmd_buffer); +void anv_cmd_buffer_resolve_framebuffer(struct anv_cmd_buffer *cmd_buffer, +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct anv_framebuffer *fb); Â const struct anv_image_view * Â anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer *cmd_buffer); diff --git a/src/intel/vulkan/genX_cmd_buffer.c b/src/intel/vulkan/genX_cmd_buffer.c index 90700e7..0962118 100644 --- a/src/intel/vulkan/genX_cmd_buffer.c +++ b/src/intel/vulkan/genX_cmd_buffer.c @@ -1346,6 +1346,23 @@ void genX(CmdEndRenderPass)( Â Â anv_cmd_buffer_resolve_subpass(cmd_buffer); +Â Â /* From the Sky Lake PRM Vol. 7, "Render Target Resolve": +Â Â * +Â Â *Â Â "When performing a render target resolve, PIPE_CONTROL with end of +Â Â *Â Â pipe sync must be delivered." +Â Â * +Â Â * By "end of pipe sync" it appears that they mean a pixel scoreboard stall +Â Â * This makes sense because the resolve operation probably needs the CCS to +Â Â * be fully valid before it looks at it. +Â Â */ +Â Â anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) { +Â Â Â pc.StallAtPixelScoreboardÂ Â Â = true; +Â Â Â pc.CommandStreamerStallEnable = true; +Â Â } </blockquote><div> </div><div>Topi, </div><div> </div><div>I found this little nugget in the PRM.Â It doesn't look like we're doing that in the GL driver.Â Maybe it has something to do with your problem? </div><div>--Jason </div><div>Â </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + +Â Â anv_cmd_buffer_resolve_framebuffer(cmd_buffer, +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â cmd_buffer->state.framebuffer); + Â #ifndef NDEBUG Â Â anv_dump_add_framebuffer(cmd_buffer, cmd_buffer->state.framebuffer); Â #endif -- 2.5.0.400.gff86faf </blockquote></div> </div></div>