#include <string.h>
#include <unistd.h>
#include <fcntl.h>
+#include <sys/mman.h>
#include "anv_private.h"
#include "util/debug.h"
.array_len = vk_info->arrayLayers,
.samples = vk_info->samples,
.min_alignment = 0,
- .min_pitch = anv_info->stride,
+ .row_pitch = anv_info->stride,
.usage = choose_isl_surf_usage(image->usage, aspect),
.tiling_flags = tiling_flags);
/* Add a HiZ surface to a depth buffer that will be used for rendering.
*/
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 (!(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 (dev->info.gen == 7) {
+ anv_perf_warn("Implement gen7 HiZ");
} else if (vk_info->mipLevels > 1) {
- anv_finishme("Test multi-LOD HiZ");
+ anv_perf_warn("Enable multi-LOD HiZ");
} else if (vk_info->arrayLayers > 1) {
- anv_finishme("Implement multi-arrayLayer HiZ clears and resolves");
+ anv_perf_warn("Implement multi-arrayLayer HiZ clears and resolves");
} else if (dev->info.gen == 8 && vk_info->samples > 1) {
- anv_finishme("Test gen8 multisampled HiZ");
- } else {
+ anv_perf_warn("Enable gen8 multisampled HiZ");
+ } else if (!unlikely(INTEL_DEBUG & DEBUG_NO_HIZ)) {
assert(image->aux_surface.isl.size == 0);
ok = isl_surf_get_hiz_surf(&dev->isl_dev, &image->depth_surface.isl,
&image->aux_surface.isl);
VkDeviceMemory _memory,
VkDeviceSize memoryOffset)
{
- ANV_FROM_HANDLE(anv_device, device, _device);
ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
ANV_FROM_HANDLE(anv_image, image, _image);
- if (mem) {
- image->bo = &mem->bo;
- image->offset = memoryOffset;
- } else {
+ if (mem == NULL) {
image->bo = NULL;
image->offset = 0;
+ return VK_SUCCESS;
}
- 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->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->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
- * not able to find space on the host to create a proper mapping.
- */
- if (map == NULL)
- return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
-
- memset(map, 0, image->aux_surface.isl.size);
-
- anv_gem_munmap(map, image->aux_surface.isl.size);
- }
+ image->bo = mem->bo;
+ image->offset = memoryOffset;
return VK_SUCCESS;
}
}
}
+/**
+ * This function determines the optimal buffer to use for a given
+ * VkImageLayout and other pieces of information needed to make that
+ * determination. This does not determine the optimal buffer to use
+ * during a resolve operation.
+ *
+ * @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 */
+ case VK_IMAGE_LAYOUT_RANGE_SIZE:
+ case VK_IMAGE_LAYOUT_MAX_ENUM:
+ unreachable("Invalid image layout.");
+
+ /* Undefined layouts
+ *
+ * The pre-initialized layout is equivalent to the undefined layout for
+ * optimally-tiled images. We can only do color compression (CCS or HiZ)
+ * on tiled images.
+ */
+ case VK_IMAGE_LAYOUT_UNDEFINED:
+ case VK_IMAGE_LAYOUT_PREINITIALIZED:
+ return ISL_AUX_USAGE_NONE;
+
+
+ /* 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)
{
iview->isl.usage = 0;
}
- /* If the HiZ buffer can be sampled from, set the constant clear color.
- * If it cannot, disable the isl aux usage flag.
- */
- float red_clear_color = 0.0f;
- enum isl_aux_usage surf_usage = image->aux_usage;
- if (image->aux_usage == ISL_AUX_USAGE_HIZ) {
- if (iview->aspect_mask & VK_IMAGE_ASPECT_DEPTH_BIT &&
- anv_can_sample_with_hiz(device->info.gen, image->samples)) {
- /* When a HiZ buffer is sampled on gen9+, ensure that
- * the constant fast clear value is set in the surface state.
- */
- if (device->info.gen >= 9)
- red_clear_color = ANV_HZ_FC_VAL;
- } else {
- surf_usage = ISL_AUX_USAGE_NONE;
- }
- }
-
/* 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
(image->usage & VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT &&
!(iview->aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT))) {
iview->sampler_surface_state = alloc_surface_state(device);
+ iview->no_aux_sampler_surface_state = alloc_surface_state(device);
+
+ /* Sampling is performed in one of two buffer configurations in anv: with
+ * an auxiliary buffer or without it. Sampler states aren't always needed
+ * for both configurations, but are currently created unconditionally for
+ * simplicity.
+ *
+ * TODO: Consider allocating each surface state only when necessary.
+ */
+
+ /* Create a sampler state with the optimal aux_usage for sampling. This
+ * may use the aux_buffer.
+ */
+ 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;
.aux_usage = surf_usage,
.mocs = device->default_mocs);
+ /* Create a sampler state that only uses the main buffer. */
+ isl_surf_fill_state(&device->isl_dev,
+ iview->no_aux_sampler_surface_state.map,
+ .surf = &surface->isl,
+ .view = &view,
+ .mocs = device->default_mocs);
+
anv_state_flush(device, iview->sampler_surface_state);
+ anv_state_flush(device, iview->no_aux_sampler_surface_state);
} else {
iview->sampler_surface_state.alloc_size = 0;
+ iview->no_aux_sampler_surface_state.alloc_size = 0;
}
/* NOTE: This one needs to go last since it may stomp isl_view.format */
.surf = &surface->isl,
.view = &view,
.aux_surf = &image->aux_surface.isl,
- .aux_usage = surf_usage,
+ .aux_usage = image->aux_usage,
.mocs = device->default_mocs);
if (isl_has_matching_typed_storage_image_format(&device->info,
.surf = &surface->isl,
.view = &view,
.aux_surf = &image->aux_surface.isl,
- .aux_usage = surf_usage,
+ .aux_usage = image->aux_usage,
.mocs = device->default_mocs);
} else {
anv_fill_buffer_surface_state(device, iview->storage_surface_state,
const uint32_t format_bs = isl_format_get_layout(view->format)->bpb / 8;
view->bo = buffer->bo;
view->offset = buffer->offset + pCreateInfo->offset;
- view->range = pCreateInfo->range == VK_WHOLE_SIZE ?
- buffer->size - pCreateInfo->offset : pCreateInfo->range;
+ view->range = anv_buffer_get_range(buffer, pCreateInfo->offset,
+ pCreateInfo->range);
view->range = align_down_npot_u32(view->range, format_bs);
if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT) {