xref: /third_party/vk-gl-cts/doc/testspecs/VK/sparse_resources.txt

Sparse resources tests

Tests:

dEQP-VK.sparse_resources.*

Includes:

1. Test fully resident buffer created with VK_BUFFER_CREATE_SPARSE_BINDING_BIT flag bit
2. Test fully resident image created with VK_IMAGE_CREATE_SPARSE_BINDING_BIT flag bit
3. Test partially resident buffer created with VK_BUFFER_CREATE_SPARSE_RESIDENCY_BIT flag bit
4. Test partially resident image created with VK_IMAGE_CREATE_SPARSE_RESIDENCY_BIT flag bit
5. Test partially resident image with mipmaps, put some mipmap levels in mip tail region
6. Test memory aliasing for fully resident buffer objects
7. Test memory aliasing for partially resident images
8. Test OpImageSparse* shader intrinsics

Description:

1. Test fully resident buffer created with VK_BUFFER_CREATE_SPARSE_BINDING_BIT flag bit

The test creates buffer object with VK_BUFFER_CREATE_SPARSE_BINDING_BIT flag bit. The size of the buffer is one
of the test parameters. The memory requirements of the buffer are being checked. Device memory is allocated
in chunks equal to the alignment parameter of buffer's memory requirements. The number of allocations is equal to
bufferRequirements.size / bufferRequirements.alignment.

The test creates two queues - one supporting sparse binding operations, the second one supporting compute and transfer operations.

First queue is used to perform binding of device memory to sparse buffer. The binding operation signals semaphore
used for synchronization.

The second queue is used to perform transfer operations. The test creates two non-sparse buffer objects,
one used as input and the second as output. The input buffer is used to transfer data to sparse buffer. The data is then
transfered further from sparse buffer to output buffer. The transer queue waits on a semaphore, before transfer operations
can be issued.

The validation part retrieves data back from output buffer to host memory. The data is then compared with reference data,
that was originally sent to input buffer. If the two data sets match, the test passes.

2. Test fully resident image created with VK_IMAGE_CREATE_SPARSE_BINDING_BIT flag bit

The test checks all supported types of images. It creates image with VK_IMAGE_CREATE_SPARSE_BINDING_BIT flag bit.
The memory requirements of the image are being checked. Device memory is allocated in chunks equal to the alignment parameter
of the image memory requirements. The number of allocations is equal to imageRequirements.size / imageRequirements.alignment.

The test creates two queues - one supporting sparse binding operations, the second one supporting compute and transfer operations.

First queue is used to perform binding of device memory to sparse image. The binding operation signals semaphore
used for synchronization.

The second queue is used to perform transfer operations. The test creates two non-sparse buffer objects,
one used as input and the second as output. The input buffer is used to transfer data to sparse image. The data is then
transfered further from sparse image to output buffer. The transfer queue waits on a semaphore, before transfer operations
can be issued.

The validation part retrieves data back from output buffer to host memory. The data is then compared with reference data,
that was originally sent to input buffer. If the two data sets match, the test passes.

3. Test partially resident buffer created with VK_BUFFER_CREATE_SPARSE_RESIDENCY_BIT flag bit

The test creates buffer object with VK_BUFFER_CREATE_SPARSE_RESIDENCY_BIT flag bit. The size of the buffer is one
of the test parameters. The sparse memory requirements of the buffer are being checked. Device memory is allocated
in chunks equal to the alignment parameter of buffer's memory requirements. Memory is bound to the buffer object leaving gaps
between bound blocks with the size equal to alignment.

The test creates two queues - one supporting sparse binding operations, the second one supporting compute and transfer operations.

First queue is used to perform binding of device memory to sparse buffer. The binding operation signals semaphore
used for synchronization.

The second queue is used to perform compute and transfer operations. A compute shader is invoked to fill the whole buffer with data.
Afterwards the data is transfered from sparse buffer to non-sparse output buffer.

The validation part retrieves data back from output buffer to host memory. The data is compared against the expected output
from compute shader. For parts of the data that correspond to the regions of sparse buffer that have device memory bound, the comparison is done
against expected output from compute shader. For parts that correspond to gaps, the data is random or should be filled with zeros if
residencyNonResidentStrict device sparse property is set to TRUE.

4. Test partially resident image created with VK_IMAGE_CREATE_SPARSE_RESIDENCY_BIT flag bit

The test creates image with VK_IMAGE_CREATE_SPARSE_RESIDENCY_BIT flag bit. The sparse memory requirements of the image are being checked.
Device memory is allocated in chunks equal to the alignment parameter of image's memory requirements.
Memory is bound to the image leaving gaps between bound blocks with the size equal to alignment.

The test creates two queues - one supporting sparse binding operations, the second one supporting compute and transfer operations.

First queue is used to perform binding of device memory to sparse image. The binding operation signals semaphore
used for synchronization.

The second queue is used to perform compute and transfer operations. A compute shader is invoked to fill the whole image with data.
Afterwards the data is transfered from sparse image to non-sparse output buffer.

The validation part retrieves data back from output buffer to host memory. The data is compared against the expected output
from compute shader. For parts of the data that correspond to the regions of image that have device memory bound, the comparison is done
against expected output from compute shader. For parts that correspond to gaps, the data is random or should be filled with zeros if residencyNonResidentStrict
device sparse property is set to TRUE.

5. Test partially resident image with mipmaps, put some mipmap levels in mip tail region

The test creates image with maximum allowed number of mipmap levels. The sparse memory requirements of the image are being checked.
Each layer of each mipmap level receives a separate device memory binding. The mipmaps levels that end up in mip tail region receive one
binding for each mipmap level or one binding for all levels, depending on the value of VK_SPARSE_IMAGE_FORMAT_SINGLE_MIPTAIL_BIT.

A compute shader is invoked to fill each mipmap level with data. Afterwards the data is transfered to a non-sparse buffer object.

The validation part retrieves data back from output buffer to host memory. The data is compared against the expected output
from compute shader. The test passes if the data sets are equal.

6. Test memory aliasing for fully resident buffer objects

The test creates two fully resident buffers (READ and WRITE) with VK_BUFFER_CREATE_SPARSE_ALIASED_BIT
and VK_BUFFER_CREATE_SPARSE_BINDING_BIT flag bits. Both buffers have the same size.

The test creates two queues - one supporting sparse binding operations, the second one supporting compute and transfer operations.

First queue is used to perform binding of device memory to sparse buffers. One block of device memory is allocated
and bound to both buffers (buffers share memory).

The second queue is used to perform compute and transfer operations. A compute shader is invoked to fill the whole WRITE buffer with data.
Afterwards the data from READ buffer is being transfered to non-sparse output buffer.

The validation part retrieves data back from output buffer to host memory. The data is compared against the expected output
from compute shader. The test passes if the data sets are equal.

7. Test memory aliasing for partially resident images

The test creates two partially resident images (READ and WRITE) with VK_IMAGE_CREATE_SPARSE_ALIASED_BIT and VK_IMAGE_CREATE_SPARSE_RESIDENCY_BIT flag bits.
Both images have the same type, format and dimensions.

The test creates two queues - one supporting sparse binding operations, the second one supporting compute and transfer operations.

First queue is used to perform binding of device memory to sparse images. The memory bound via VkSparseImageMemoryBind is shared between
both images. The mipmap levels that land in the mip tail region have separate memory regions for both images.

The second queue is used to perform compute and transfer operations. The test creates two non-sparse buffer objects,
one used as input and the second as output. The input buffer is used to transfer data to READ sparse image to create some initial state.
Afterwards compute shaders are invoked to write data to each mipmap level of WRITE sparse image. The mipmap levels of READ image that share memory with
WRITE image should be overwritten by this operation, the mip tail region should be left intact. Next the data is copied from the READ image to the output buffer.

The validation part retrieves data back from output buffer to host memory. For each mipmap level that both images share memory for, the data is
compared against the expected output from compute shader. On the other hand for each mipmap level that landed in the mip tail region, the data is compared
against data stored in the input buffer (the compute shader could not have changed this data). The test passes if for each mipmap level
the comparison results in both data sets being the same.

8. Test OpImageSparse* shader intrinsics

The test creates sparse partially resident image. The memory is bound to the image every second mipmap level.

The test creates also a second non-sparse texels image with the same dimensions and format as the sparse one and
a third residency image with the same dimensions as the sparse one and unsigned int format.

For OpImageSparse* opcodes that are fed with float image coordinates the test creates a graphics queue, otherwise a compute queue is created.
In both cases the commands submited to queue have the purpose of copying the data from sparse image to texels and residency images using one
of the OpImageSparse* shader intrinsics. For graphics operations the data is copied via rendering to two color attachments, for compute operations
the data is copied via image load/store.

Data is retreived from the non-sparse images back to the CPU. Contents of the texels image are compared against the data originaly sent to the sparse image.
For mipmap levels of the sparse image that do not have backing device memory, the fetched data is compared against zeroed memory if residencyNonResidentStrict is set to VK_TRUE,
otherwise comparion for those mipmap levels is ommited. The data fetched from the residency image is checked, if for each mipmap level the OpImageSparseTexelsResident
returned correct residency information.