X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;ds=sidebyside;f=src%2Fintel%2Fvulkan%2Fanv_image.c;h=16a60833e022cf947e70bdbdfd39a219aaffa633;hb=f3621f4e71a68aae76093ee2232d4027057c2e19;hp=eb20b7179373efed4e2a174b65573b48518e899e;hpb=29e289fa655938f7814bdbb3de7996a8a0f04b60;p=mesa.git diff --git a/src/intel/vulkan/anv_image.c b/src/intel/vulkan/anv_image.c index eb20b717937..16a60833e02 100644 --- a/src/intel/vulkan/anv_image.c +++ b/src/intel/vulkan/anv_image.c @@ -41,9 +41,6 @@ choose_isl_surf_usage(VkImageUsageFlags vk_usage, { isl_surf_usage_flags_t isl_usage = 0; - /* FINISHME: Support aux surfaces */ - isl_usage |= ISL_SURF_USAGE_DISABLE_AUX_BIT; - if (vk_usage & VK_IMAGE_USAGE_SAMPLED_BIT) isl_usage |= ISL_SURF_USAGE_TEXTURE_BIT; @@ -56,27 +53,33 @@ choose_isl_surf_usage(VkImageUsageFlags vk_usage, if (vk_usage & VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT) isl_usage |= ISL_SURF_USAGE_CUBE_BIT; - if (vk_usage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { - switch (aspect) { - default: - unreachable("bad VkImageAspect"); - case VK_IMAGE_ASPECT_DEPTH_BIT: - isl_usage &= ~ISL_SURF_USAGE_DISABLE_AUX_BIT; - isl_usage |= ISL_SURF_USAGE_DEPTH_BIT; - break; - case VK_IMAGE_ASPECT_STENCIL_BIT: - isl_usage |= ISL_SURF_USAGE_STENCIL_BIT; - break; - } + /* Even if we're only using it for transfer operations, clears to depth and + * stencil images happen as depth and stencil so they need the right ISL + * usage bits or else things will fall apart. + */ + switch (aspect) { + case VK_IMAGE_ASPECT_DEPTH_BIT: + isl_usage |= ISL_SURF_USAGE_DEPTH_BIT; + break; + case VK_IMAGE_ASPECT_STENCIL_BIT: + isl_usage |= ISL_SURF_USAGE_STENCIL_BIT; + break; + case VK_IMAGE_ASPECT_COLOR_BIT: + break; + default: + unreachable("bad VkImageAspect"); } if (vk_usage & VK_IMAGE_USAGE_TRANSFER_SRC_BIT) { - /* Meta implements transfers by sampling from the source image. */ + /* blorp implements transfers by sampling from the source image. */ isl_usage |= ISL_SURF_USAGE_TEXTURE_BIT; } - if (vk_usage & VK_IMAGE_USAGE_TRANSFER_DST_BIT) { - /* Meta implements transfers by rendering into the destination image. */ + if (vk_usage & VK_IMAGE_USAGE_TRANSFER_DST_BIT && + aspect == VK_IMAGE_ASPECT_COLOR_BIT) { + /* blorp implements transfers by rendering into the destination image. + * Only request this with color images, as we deal with depth/stencil + * formats differently. */ isl_usage |= ISL_SURF_USAGE_RENDER_TARGET_BIT; } @@ -108,7 +111,7 @@ add_surface(struct anv_image *image, struct anv_surface *surf) surf->offset = align_u32(image->size, surf->isl.alignment); image->size = surf->offset + surf->isl.size; - image->alignment = MAX(image->alignment, surf->isl.alignment); + image->alignment = MAX2(image->alignment, surf->isl.alignment); } /** @@ -176,67 +179,82 @@ make_surface(const struct anv_device *dev, /* Add a HiZ surface to a depth buffer that will be used for rendering. */ - if (aspect == VK_IMAGE_ASPECT_DEPTH_BIT && - (image->usage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT)) { + if (aspect == VK_IMAGE_ASPECT_DEPTH_BIT) { + /* We don't advertise that depth buffers could be used as storage + * images. + */ + assert(!(image->usage & VK_IMAGE_USAGE_STORAGE_BIT)); /* Allow the user to control HiZ enabling. Disable by default on gen7 * because resolves are not currently implemented pre-BDW. */ - if (!env_var_as_boolean("INTEL_VK_HIZ", dev->info.gen >= 8)) { + if (!(image->usage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT)) { + /* It will never be used as an attachment, HiZ is pointless. */ + } else if (image->usage & VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT) { + /* From the 1.0.37 spec: + * + * "An attachment used as an input attachment and depth/stencil + * attachment must be in either VK_IMAGE_LAYOUT_GENERAL or + * VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL." + * + * It will never have a layout of + * VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL, so HiZ is + * currently pointless. If transfer operations learn to use the HiZ + * buffer, we can enable HiZ for VK_IMAGE_LAYOUT_GENERAL and support + * input attachments. + */ + anv_finishme("Implement HiZ for input attachments"); + } else if (!env_var_as_boolean("INTEL_VK_HIZ", dev->info.gen >= 8)) { anv_finishme("Implement gen7 HiZ"); } else if (vk_info->mipLevels > 1) { anv_finishme("Test multi-LOD HiZ"); + } else if (vk_info->arrayLayers > 1) { + anv_finishme("Implement multi-arrayLayer HiZ clears and resolves"); } else if (dev->info.gen == 8 && vk_info->samples > 1) { anv_finishme("Test gen8 multisampled HiZ"); } else { - isl_surf_get_hiz_surf(&dev->isl_dev, &image->depth_surface.isl, - &image->hiz_surface.isl); - add_surface(image, &image->hiz_surface); + assert(image->aux_surface.isl.size == 0); + ok = isl_surf_get_hiz_surf(&dev->isl_dev, &image->depth_surface.isl, + &image->aux_surface.isl); + assert(ok); + add_surface(image, &image->aux_surface); + image->aux_usage = ISL_AUX_USAGE_HIZ; } - } - - return VK_SUCCESS; -} - -/** - * Parameter @a format is required and overrides VkImageCreateInfo::format. - */ -static VkImageUsageFlags -anv_image_get_full_usage(const VkImageCreateInfo *info, - VkImageAspectFlags aspects) -{ - VkImageUsageFlags usage = info->usage; - - if (info->samples > 1 && - (usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT)) { - /* Meta will resolve the image by binding it as a texture. */ - usage |= VK_IMAGE_USAGE_SAMPLED_BIT; - } - - if (usage & VK_IMAGE_USAGE_TRANSFER_SRC_BIT) { - /* Meta will transfer from the image by binding it as a texture. */ - usage |= VK_IMAGE_USAGE_SAMPLED_BIT; - } - - if (usage & VK_IMAGE_USAGE_TRANSFER_DST_BIT) { - /* For non-clear transfer operations, meta will transfer to the image by - * binding it as a color attachment, even if the image format is not - * a color format. - */ - usage |= VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT; - - if (aspects & (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)) { - /* vkCmdClearDepthStencilImage() only requires that - * VK_IMAGE_USAGE_TRANSFER_SRC_BIT be set. In particular, it does - * not require VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT. Meta - * clears the image, though, by binding it as a depthstencil - * attachment. - */ - usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT; + } else if (aspect == VK_IMAGE_ASPECT_COLOR_BIT && vk_info->samples == 1) { + if (!unlikely(INTEL_DEBUG & DEBUG_NO_RBC)) { + assert(image->aux_surface.isl.size == 0); + ok = isl_surf_get_ccs_surf(&dev->isl_dev, &anv_surf->isl, + &image->aux_surface.isl); + if (ok) { + add_surface(image, &image->aux_surface); + + /* For images created without MUTABLE_FORMAT_BIT set, we know that + * they will always be used with the original format. In + * particular, they will always be used with a format that + * supports color compression. If it's never used as a storage + * image, then it will only be used through the sampler or the as + * a render target. This means that it's safe to just leave + * compression on at all times for these formats. + */ + if (!(vk_info->usage & VK_IMAGE_USAGE_STORAGE_BIT) && + !(vk_info->flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) && + isl_format_supports_ccs_e(&dev->info, format)) { + image->aux_usage = ISL_AUX_USAGE_CCS_E; + } + } + } + } else if (aspect == VK_IMAGE_ASPECT_COLOR_BIT && vk_info->samples > 1) { + assert(image->aux_surface.isl.size == 0); + assert(!(vk_info->usage & VK_IMAGE_USAGE_STORAGE_BIT)); + ok = isl_surf_get_mcs_surf(&dev->isl_dev, &anv_surf->isl, + &image->aux_surface.isl); + if (ok) { + add_surface(image, &image->aux_surface); + image->aux_usage = ISL_AUX_USAGE_MCS; } } - return usage; + return VK_SUCCESS; } VkResult @@ -259,7 +277,7 @@ anv_image_create(VkDevice _device, anv_assert(pCreateInfo->extent.height > 0); anv_assert(pCreateInfo->extent.depth > 0); - image = anv_alloc2(&device->alloc, alloc, sizeof(*image), 8, + image = vk_alloc2(&device->alloc, alloc, sizeof(*image), 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (!image) return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); @@ -272,8 +290,9 @@ anv_image_create(VkDevice _device, image->levels = pCreateInfo->mipLevels; image->array_size = pCreateInfo->arrayLayers; image->samples = pCreateInfo->samples; - image->usage = anv_image_get_full_usage(pCreateInfo, image->aspects); + image->usage = pCreateInfo->usage; image->tiling = pCreateInfo->tiling; + image->aux_usage = ISL_AUX_USAGE_NONE; uint32_t b; for_each_bit(b, image->aspects) { @@ -288,7 +307,7 @@ anv_image_create(VkDevice _device, fail: if (image) - anv_free2(&device->alloc, alloc, image); + vk_free2(&device->alloc, alloc, image); return r; } @@ -312,8 +331,12 @@ anv_DestroyImage(VkDevice _device, VkImage _image, const VkAllocationCallbacks *pAllocator) { ANV_FROM_HANDLE(anv_device, device, _device); + ANV_FROM_HANDLE(anv_image, image, _image); - anv_free2(&device->alloc, pAllocator, anv_image_from_handle(_image)); + if (!image) + return; + + vk_free2(&device->alloc, pAllocator, image); } VkResult anv_BindImageMemory( @@ -334,21 +357,23 @@ VkResult anv_BindImageMemory( image->offset = 0; } - if (anv_image_has_hiz(image)) { + if (image->aux_surface.isl.size > 0) { /* The offset and size must be a multiple of 4K or else the * anv_gem_mmap call below will return NULL. */ - assert((image->offset + image->hiz_surface.offset) % 4096 == 0); - assert(image->hiz_surface.isl.size % 4096 == 0); - - /* HiZ surfaces need to have their memory cleared to 0 before they - * can be used. If we let it have garbage data, it can cause GPU - * hangs on some hardware. + assert((image->offset + image->aux_surface.offset) % 4096 == 0); + assert(image->aux_surface.isl.size % 4096 == 0); + + /* Auxiliary surfaces need to have their memory cleared to 0 before they + * can be used. For CCS surfaces, this puts them in the "resolved" + * state so they can be used with CCS enabled before we ever touch it + * from the GPU. For HiZ, we need something valid or else we may get + * GPU hangs on some hardware and 0 works fine. */ void *map = anv_gem_mmap(device, image->bo->gem_handle, - image->offset + image->hiz_surface.offset, - image->hiz_surface.isl.size, + image->offset + image->aux_surface.offset, + image->aux_surface.isl.size, device->info.has_llc ? 0 : I915_MMAP_WC); /* If anv_gem_mmap returns NULL, it's likely that the kernel was @@ -357,9 +382,9 @@ VkResult anv_BindImageMemory( if (map == NULL) return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); - memset(map, 0, image->hiz_surface.isl.size); + memset(map, 0, image->aux_surface.isl.size); - anv_gem_munmap(map, image->hiz_surface.isl.size); + anv_gem_munmap(map, image->aux_surface.isl.size); } return VK_SUCCESS; @@ -412,15 +437,143 @@ void anv_GetImageSubresourceLayout( } } +/** + * This function determines the optimal buffer to use for device + * accesses given a VkImageLayout and other pieces of information needed to + * make that determination. This does not determine the optimal buffer to + * use during a resolve operation. + * + * NOTE: Some layouts do not support device access. + * + * @param devinfo The device information of the Intel GPU. + * @param image The image that may contain a collection of buffers. + * @param aspects The aspect(s) of the image to be accessed. + * @param layout The current layout of the image aspect(s). + * + * @return The primary buffer that should be used for the given layout. + */ +enum isl_aux_usage +anv_layout_to_aux_usage(const struct gen_device_info * const devinfo, + const struct anv_image * const image, + const VkImageAspectFlags aspects, + const VkImageLayout layout) +{ + /* Validate the inputs. */ + + /* The devinfo is needed as the optimal buffer varies across generations. */ + assert(devinfo != NULL); + + /* The layout of a NULL image is not properly defined. */ + assert(image != NULL); + + /* The aspects must be a subset of the image aspects. */ + assert(aspects & image->aspects && aspects <= image->aspects); + + /* Determine the optimal buffer. */ + + /* If there is no auxiliary surface allocated, we must use the one and only + * main buffer. + */ + if (image->aux_surface.isl.size == 0) + return ISL_AUX_USAGE_NONE; + + /* All images that use an auxiliary surface are required to be tiled. */ + assert(image->tiling == VK_IMAGE_TILING_OPTIMAL); + + /* On BDW+, when clearing the stencil aspect of a depth stencil image, + * the HiZ buffer allows us to record the clear with a relatively small + * number of packets. Prior to BDW, the HiZ buffer provides no known benefit + * to the stencil aspect. + */ + if (devinfo->gen < 8 && aspects == VK_IMAGE_ASPECT_STENCIL_BIT) + return ISL_AUX_USAGE_NONE; + + const bool color_aspect = aspects == VK_IMAGE_ASPECT_COLOR_BIT; + + /* The following switch currently only handles depth stencil aspects. + * TODO: Handle the color aspect. + */ + if (color_aspect) + return image->aux_usage; + + switch (layout) { + + /* Invalid Layouts */ + + /* According to the Vulkan Spec, the following layouts are valid only as + * initial layouts in a layout transition and don't support device access. + */ + case VK_IMAGE_LAYOUT_UNDEFINED: + case VK_IMAGE_LAYOUT_PREINITIALIZED: + case VK_IMAGE_LAYOUT_RANGE_SIZE: + case VK_IMAGE_LAYOUT_MAX_ENUM: + unreachable("Invalid image layout for device access."); + + + /* Transfer Layouts + * + * This buffer could be a depth buffer used in a transfer operation. BLORP + * currently doesn't use HiZ for transfer operations so we must use the main + * buffer for this layout. TODO: Enable HiZ in BLORP. + */ + case VK_IMAGE_LAYOUT_GENERAL: + case VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL: + case VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL: + return ISL_AUX_USAGE_NONE; + + + /* Sampling Layouts */ + case VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL: + assert(!color_aspect); + /* Fall-through */ + case VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL: + if (anv_can_sample_with_hiz(devinfo, aspects, image->samples)) + return ISL_AUX_USAGE_HIZ; + else + return ISL_AUX_USAGE_NONE; + + case VK_IMAGE_LAYOUT_PRESENT_SRC_KHR: + assert(color_aspect); + + /* On SKL+, the render buffer can be decompressed by the presentation + * engine. Support for this feature has not yet landed in the wider + * ecosystem. TODO: Update this code when support lands. + * + * From the BDW PRM, Vol 7, Render Target Resolve: + * + * If the MCS is enabled on a non-multisampled render target, the + * render target must be resolved before being used for other + * purposes (display, texture, CPU lock) The clear value from + * SURFACE_STATE is written into pixels in the render target + * indicated as clear in the MCS. + * + * Pre-SKL, the render buffer must be resolved before being used for + * presentation. We can infer that the auxiliary buffer is not used. + */ + return ISL_AUX_USAGE_NONE; + + + /* Rendering Layouts */ + case VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL: + assert(color_aspect); + unreachable("Color images are not yet supported."); + + case VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL: + assert(!color_aspect); + return ISL_AUX_USAGE_HIZ; + } + + /* If the layout isn't recognized in the exhaustive switch above, the + * VkImageLayout value is not defined in vulkan.h. + */ + unreachable("layout is not a VkImageLayout enumeration member."); +} + + static struct anv_state -alloc_surface_state(struct anv_device *device, - struct anv_cmd_buffer *cmd_buffer) +alloc_surface_state(struct anv_device *device) { - if (cmd_buffer) { - return anv_cmd_buffer_alloc_surface_state(cmd_buffer); - } else { - return anv_state_pool_alloc(&device->surface_state_pool, 64, 64); - } + return anv_state_pool_alloc(&device->surface_state_pool, 64, 64); } static enum isl_channel_select @@ -442,14 +595,22 @@ remap_swizzle(VkComponentSwizzle swizzle, VkComponentSwizzle component, } } -void -anv_image_view_init(struct anv_image_view *iview, - struct anv_device *device, - const VkImageViewCreateInfo* pCreateInfo, - struct anv_cmd_buffer *cmd_buffer, - VkImageUsageFlags usage_mask) + +VkResult +anv_CreateImageView(VkDevice _device, + const VkImageViewCreateInfo *pCreateInfo, + const VkAllocationCallbacks *pAllocator, + VkImageView *pView) { + ANV_FROM_HANDLE(anv_device, device, _device); ANV_FROM_HANDLE(anv_image, image, pCreateInfo->image); + struct anv_image_view *iview; + + iview = vk_alloc2(&device->alloc, pAllocator, sizeof(*iview), 8, + VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); + if (iview == NULL) + return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); + const VkImageSubresourceRange *range = &pCreateInfo->subresourceRange; assert(range->layerCount > 0); @@ -509,14 +670,7 @@ anv_image_view_init(struct anv_image_view *iview, .depth = anv_minify(image->extent.depth , range->baseMipLevel), }; - if (image->type == VK_IMAGE_TYPE_3D && - usage_mask != VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) { - /* Meta renders to 3D texture slices. When it does so, it passes - * usage_mask == VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT. Since meta is the - * only thing that uses a non-zero usage_mask, this lets us easily - * detect the one case where we actually want an array range used for - * 3-D textures. - */ + if (pCreateInfo->viewType == VK_IMAGE_VIEW_TYPE_3D) { iview->isl.base_array_layer = 0; iview->isl.array_len = iview->extent.depth; } @@ -528,8 +682,28 @@ anv_image_view_init(struct anv_image_view *iview, iview->isl.usage = 0; } - if (image->usage & usage_mask & VK_IMAGE_USAGE_SAMPLED_BIT) { - iview->sampler_surface_state = alloc_surface_state(device, cmd_buffer); + /* Input attachment surfaces for color are allocated and filled + * out at BeginRenderPass time because they need compression information. + * Compression is not yet enabled for depth textures and stencil doesn't + * allow compression so we can just use the texture surface state from the + * view. + */ + if (image->usage & VK_IMAGE_USAGE_SAMPLED_BIT || + (image->usage & VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT && + !(iview->aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT))) { + iview->sampler_surface_state = alloc_surface_state(device); + + /* Select the optimal aux_usage for sampling. */ + const enum isl_aux_usage surf_usage = + anv_layout_to_aux_usage(&device->info, image, iview->aspect_mask, + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL); + + /* If this is a HiZ buffer we can sample from with a programmable clear + * value (SKL+), define the clear value to the optimal constant. + */ + const float red_clear_color = surf_usage == ISL_AUX_USAGE_HIZ && + device->info.gen >= 9 ? + ANV_HZ_FC_VAL : 0.0f; struct isl_view view = iview->isl; view.usage |= ISL_SURF_USAGE_TEXTURE_BIT; @@ -537,56 +711,50 @@ anv_image_view_init(struct anv_image_view *iview, iview->sampler_surface_state.map, .surf = &surface->isl, .view = &view, + .clear_color.f32 = { red_clear_color,}, + .aux_surf = &image->aux_surface.isl, + .aux_usage = surf_usage, .mocs = device->default_mocs); - if (!device->info.has_llc) - anv_state_clflush(iview->sampler_surface_state); + anv_state_flush(device, iview->sampler_surface_state); } else { iview->sampler_surface_state.alloc_size = 0; } - /* This is kind-of hackish. It is possible, due to get_full_usage above, - * to get a surface state with a non-renderable format but with - * VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT. This happens in particular for - * formats which aren't renderable but where we want to use Vulkan copy - * commands so VK_IMAGE_USAGE_TRANSFER_DST_BIT is set. In the case of a - * copy, meta will use a format that we can render to, but most of the rest - * of the time, we don't want to create those surface states. Once we - * start using blorp for copies, this problem will go away and we can - * remove a lot of hacks. - */ - if ((image->usage & usage_mask & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) && - isl_format_supports_rendering(&device->info, format.isl_format)) { - iview->color_rt_surface_state = alloc_surface_state(device, cmd_buffer); + /* NOTE: This one needs to go last since it may stomp isl_view.format */ + if (image->usage & VK_IMAGE_USAGE_STORAGE_BIT) { + iview->storage_surface_state = alloc_surface_state(device); + iview->writeonly_storage_surface_state = alloc_surface_state(device); struct isl_view view = iview->isl; - view.usage |= ISL_SURF_USAGE_RENDER_TARGET_BIT; + view.usage |= ISL_SURF_USAGE_STORAGE_BIT; + + /* Write-only accesses always used a typed write instruction and should + * therefore use the real format. + */ isl_surf_fill_state(&device->isl_dev, - iview->color_rt_surface_state.map, + iview->writeonly_storage_surface_state.map, .surf = &surface->isl, .view = &view, + .aux_surf = &image->aux_surface.isl, + .aux_usage = image->aux_usage, .mocs = device->default_mocs); - if (!device->info.has_llc) - anv_state_clflush(iview->color_rt_surface_state); - } else { - iview->color_rt_surface_state.alloc_size = 0; - } - - /* NOTE: This one needs to go last since it may stomp isl_view.format */ - if (image->usage & usage_mask & VK_IMAGE_USAGE_STORAGE_BIT) { - iview->storage_surface_state = alloc_surface_state(device, cmd_buffer); - if (isl_has_matching_typed_storage_image_format(&device->info, format.isl_format)) { - struct isl_view view = iview->isl; - view.usage |= ISL_SURF_USAGE_STORAGE_BIT; + /* Typed surface reads support a very limited subset of the shader + * image formats. Translate it into the closest format the hardware + * supports. + */ view.format = isl_lower_storage_image_format(&device->info, format.isl_format); + isl_surf_fill_state(&device->isl_dev, iview->storage_surface_state.map, .surf = &surface->isl, .view = &view, + .aux_surf = &image->aux_surface.isl, + .aux_usage = image->aux_usage, .mocs = device->default_mocs); } else { anv_fill_buffer_surface_state(device, iview->storage_surface_state, @@ -599,30 +767,14 @@ anv_image_view_init(struct anv_image_view *iview, &iview->storage_image_param, &surface->isl, &iview->isl); - if (!device->info.has_llc) - anv_state_clflush(iview->storage_surface_state); + anv_state_flush(device, iview->storage_surface_state); + anv_state_flush(device, iview->writeonly_storage_surface_state); } else { iview->storage_surface_state.alloc_size = 0; + iview->writeonly_storage_surface_state.alloc_size = 0; } -} -VkResult -anv_CreateImageView(VkDevice _device, - const VkImageViewCreateInfo *pCreateInfo, - const VkAllocationCallbacks *pAllocator, - VkImageView *pView) -{ - ANV_FROM_HANDLE(anv_device, device, _device); - struct anv_image_view *view; - - view = anv_alloc2(&device->alloc, pAllocator, sizeof(*view), 8, - VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); - if (view == NULL) - return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); - - anv_image_view_init(view, device, pCreateInfo, NULL, ~0); - - *pView = anv_image_view_to_handle(view); + *pView = anv_image_view_to_handle(iview); return VK_SUCCESS; } @@ -634,10 +786,8 @@ anv_DestroyImageView(VkDevice _device, VkImageView _iview, ANV_FROM_HANDLE(anv_device, device, _device); ANV_FROM_HANDLE(anv_image_view, iview, _iview); - if (iview->color_rt_surface_state.alloc_size > 0) { - anv_state_pool_free(&device->surface_state_pool, - iview->color_rt_surface_state); - } + if (!iview) + return; if (iview->sampler_surface_state.alloc_size > 0) { anv_state_pool_free(&device->surface_state_pool, @@ -649,16 +799,29 @@ anv_DestroyImageView(VkDevice _device, VkImageView _iview, iview->storage_surface_state); } - anv_free2(&device->alloc, pAllocator, iview); + if (iview->writeonly_storage_surface_state.alloc_size > 0) { + anv_state_pool_free(&device->surface_state_pool, + iview->writeonly_storage_surface_state); + } + + vk_free2(&device->alloc, pAllocator, iview); } -void anv_buffer_view_init(struct anv_buffer_view *view, - struct anv_device *device, - const VkBufferViewCreateInfo* pCreateInfo, - struct anv_cmd_buffer *cmd_buffer) +VkResult +anv_CreateBufferView(VkDevice _device, + const VkBufferViewCreateInfo *pCreateInfo, + const VkAllocationCallbacks *pAllocator, + VkBufferView *pView) { + ANV_FROM_HANDLE(anv_device, device, _device); ANV_FROM_HANDLE(anv_buffer, buffer, pCreateInfo->buffer); + struct anv_buffer_view *view; + + view = vk_alloc2(&device->alloc, pAllocator, sizeof(*view), 8, + VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); + if (!view) + return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); /* TODO: Handle the format swizzle? */ @@ -673,7 +836,7 @@ void anv_buffer_view_init(struct anv_buffer_view *view, view->range = align_down_npot_u32(view->range, format_bs); if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT) { - view->surface_state = alloc_surface_state(device, cmd_buffer); + view->surface_state = alloc_surface_state(device); anv_fill_buffer_surface_state(device, view->surface_state, view->format, @@ -683,7 +846,8 @@ void anv_buffer_view_init(struct anv_buffer_view *view, } if (buffer->usage & VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT) { - view->storage_surface_state = alloc_surface_state(device, cmd_buffer); + view->storage_surface_state = alloc_surface_state(device); + view->writeonly_storage_surface_state = alloc_surface_state(device); enum isl_format storage_format = isl_has_matching_typed_storage_image_format(&device->info, @@ -697,29 +861,19 @@ void anv_buffer_view_init(struct anv_buffer_view *view, (storage_format == ISL_FORMAT_RAW ? 1 : isl_format_get_layout(storage_format)->bpb / 8)); + /* Write-only accesses should use the original format. */ + anv_fill_buffer_surface_state(device, view->writeonly_storage_surface_state, + view->format, + view->offset, view->range, + isl_format_get_layout(view->format)->bpb / 8); + isl_buffer_fill_image_param(&device->isl_dev, &view->storage_image_param, view->format, view->range); } else { view->storage_surface_state = (struct anv_state){ 0 }; + view->writeonly_storage_surface_state = (struct anv_state){ 0 }; } -} - -VkResult -anv_CreateBufferView(VkDevice _device, - const VkBufferViewCreateInfo *pCreateInfo, - const VkAllocationCallbacks *pAllocator, - VkBufferView *pView) -{ - ANV_FROM_HANDLE(anv_device, device, _device); - struct anv_buffer_view *view; - - view = anv_alloc2(&device->alloc, pAllocator, sizeof(*view), 8, - VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); - if (!view) - return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); - - anv_buffer_view_init(view, device, pCreateInfo, NULL); *pView = anv_buffer_view_to_handle(view); @@ -733,6 +887,9 @@ anv_DestroyBufferView(VkDevice _device, VkBufferView bufferView, ANV_FROM_HANDLE(anv_device, device, _device); ANV_FROM_HANDLE(anv_buffer_view, view, bufferView); + if (!view) + return; + if (view->surface_state.alloc_size > 0) anv_state_pool_free(&device->surface_state_pool, view->surface_state); @@ -741,7 +898,11 @@ anv_DestroyBufferView(VkDevice _device, VkBufferView bufferView, anv_state_pool_free(&device->surface_state_pool, view->storage_surface_state); - anv_free2(&device->alloc, pAllocator, view); + if (view->writeonly_storage_surface_state.alloc_size > 0) + anv_state_pool_free(&device->surface_state_pool, + view->writeonly_storage_surface_state); + + vk_free2(&device->alloc, pAllocator, view); } const struct anv_surface * @@ -750,20 +911,8 @@ anv_image_get_surface_for_aspect_mask(const struct anv_image *image, { switch (aspect_mask) { case VK_IMAGE_ASPECT_COLOR_BIT: - /* Dragons will eat you. - * - * Meta attaches all destination surfaces as color render targets. Guess - * what surface the Meta Dragons really want. - */ - if (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT) { - return &image->depth_surface; - } else if (image->aspects & VK_IMAGE_ASPECT_STENCIL_BIT) { - return &image->stencil_surface; - } else { - assert(image->aspects == VK_IMAGE_ASPECT_COLOR_BIT); - return &image->color_surface; - } - break; + assert(image->aspects == VK_IMAGE_ASPECT_COLOR_BIT); + return &image->color_surface; case VK_IMAGE_ASPECT_DEPTH_BIT: assert(image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT); return &image->depth_surface;