Destroy resources even if there is a pipe_reference.
Signed-off-by: David Riley <davidriley@chromium.org>
Reviewed-By: Gert Wollny <gert.wollny@collabora.com>
Transfer strides are either set internally or coming from
virgl_renderer_transfer_{write,read}_iov. As far as I can tell,
Mesa always passes sane values. We also expect sane values in
places like vrend_renderer_transfer_write_iov,
vrend_transfer_send_readpixels, and vrend_transfer_send_getteximage
already. Let's reject bad strides.
This fixes, for example, transfers to 1D array (thus box->height is
1) with non-default stride.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
The origin change is already handled within Gallium, so it is only
necessary to deal with the Z range property.
v2: don't mark viewport as dirty when halfz is changed, this is already
handled by Gallium (Gurchetan)
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Just like all the other non-static vrend_shader functions.
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Ensure we can't create a texture resource with zero width, to
guarantee that we get a non-zero stride, since either:
1. A non-zero stride was specified in the transfer request, or
2. We will calculate a stride based on texture width which we guarantee
is not zero
Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
This is required to be able to properly handle transfers with
data layouts that are different from the resource layout.
Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
This commit improves bound checking by more accurately calculating
the expected transfer size. This improvement is achieved in two ways:
1. Using the image level (layer_)stride when the (layer_)stride is 0,
which matches the interpretation we use for performing the actual
read/write.
2. Calculating the transfer size by using the exact end offset in
the image, instead of using multiples of whole image rows.
The increased accuracy increases safety, and also allows us to support
some transfers with explicit strides that were previously rejected (see
added test).
Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
There is no need to fail such transfers, in the same way we don't fail
when we are given a stride for targets without height, especially since
both stride and layer_stride have a valid (albeit possibly not very
useful) interpretation for all targets types.
Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
We no longer fall back here, as we've already tried glGetTexImage and
failed in the case where that's possible.
Signed-off-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
These two snippets do the same, let's just place them in a helper and
call that instead.
Signed-off-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Looks like this was never really used for anything useful. Let's just
get rid of it.
Signed-off-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
Support the copy_transfer3d command, which transfers data to a host
resource by copying from another staging resource. This is used
by the guest to avoid waiting in case it needs to write to a busy
resource.
Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Add an internal mechanism to support synchronized transfers. This will
be used in upcoming commits to perform synchronized resource maps if
the transfer type requires it.
Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
VIRGL_BIND_STAGING is used for buffers that are backed only by guest iov
storage and are guaranteed to not be written to by the host, thus not
requiring additional synchronization. They will be used in upcoming
commits as staging buffers for copy transfers.
Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Rename VREND_RESOURCE_STORAGE_IOVEC to
VREND_RESOURCE_STORAGE_GUEST_ELSE_SYSTEM and introduce
VREND_RESOURCE_STORAGE_GUEST. The former is used as storage for query
buffers, and the latter will be used in upcoming commits for staging
buffers.
Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
These two members are written, but never read. Let's get rid of them.
This has the nice side-benefit of eliminating a warning, because of
an implicit cast.
Signed-off-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Reviewed-By: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
With 6a3cd2bd vrend: send list of readback-formats to guest
stricter format checks were introduced on GLES. This may lead to format
conversions in guest and host that can be avoided for additional
implementation specific readback formats. So when we build the format list
also check whether a bound texture format can also be read back from the
fbo and report the standard and additional formats also to the guest.
v2: Make things simpler and more correct (Erik) i.e.:
- default to false in the check for readback formats
- rename test function to indicate that we test color formats only
- don't pass the gl_version, instead use the epoxy functions
to get the platform info
v3: - check for DS readback support (Thanks to Erik for pointing out that this
is not available in unextended GLES)
- drop some superfluous braces (Erik)
v4: - move DS checks to separate commit (Erik)
v5: - split patch to move preparations that don't change
functionality to preceeding patches
- reduce the amount of error checking (Erik)
- only check readback for formats that can actually be used as
render targets
- remove some temporary variables (Erik)
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Also limit the scope of some variables.
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
The name for the internal texture properies flags VIRGL_BIND_NEED-SWIZZLE
and VIRGL_BIND_CAN_TEXTURE_STORAGE is misleading, and the latter flag was
even mis-using the binding member of the texture format structure.
Rename these flags to make it clear that these are not binding flags and
correctly use the 'flags' member for CAN_TEXTURE_STORAGE. In addition, drop
adding the NEED_SWIZZLE flag to the binding in the alpha texture emulation.
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
To properly use compute shaders in the guest when running on a GL ES host
it is better to advertice GLSL level 430, because otherwise compute shaders
that don't explicitely add the extension ARB_compute_shaders to the shader
will not compile.
Fixes#99
v2: Add comment about why we want to advertise this feature level (Gurchetan)
v3: Don't bypass checks for level 400
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
PIPE_BIND_CUSTOM is used for query buffers and fences by the guest.
We can safely replace VREND_RESOURCE_STORAGE_SYSTEM by
VREND_RESOURCE_STORAGE_IOVEC.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
This allows a resource to use the guest pages as its storage when
the iovec is attached.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Check queries from within vrend_renderer_check_fences, after
glClientWaitSync but before write_fence. This makes sure query
results are available after the fence is waited for.
In other words, given this sequence of commands
BEGIN_QUERY > DRAW_VBO > END_QUERY > GET_QUERY_RESULT > FENCE
Waiting on the fence makes sure the query result becomes available
to the guest.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Handle these two command sequences,
VIRGL_CCMD_BEGIN_QUERY
VIRGL_CCMD_END_QUERY
VIRGL_CCMD_GET_QUERY_RESULT
VIRGL_CCMD_GET_QUERY_RESULT
or
VIRGL_CCMD_BEGIN_QUERY
VIRGL_CCMD_END_QUERY
VIRGL_CCMD_GET_QUERY_RESULT
(before the query result is available)
VIRGL_CCMD_BEGIN_QUERY
VIRGL_CCMD_END_QUERY
VIRGL_CCMD_GET_QUERY_RESULT
more carefully.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
This patch uses GL_EXT_framebuffer_fetch_non_coherent (preferred)
or GL_EXT_framebuffer_fetch to emulate the logiops in the fragement
shader. If neither of these extension are available then only
GL_COPY, GL_COPY_INVERTED, GL_CLEAR, and GL_SET are emulated.
Fixes piglit gl-1.0-logicop on GLES hosts.
v2: Use non_coherent access when possible
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Pass a mask of the non-array generic inputs that the next shader expects
is passed to the shader that is currently converted. If, after emitting
all generic outputs, some are missing, then these are also generated.
Limitations: This doesn't take care of input arrays that may not be emitted
as outputs, but since this problem seems to be only related to IO variables
that are implicitely declared this is not a problem.
Fixes piglit: glsl-routing
v2: rebase and update names to new naming
v3: Declare TCS outputs as arrays as required
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Tested-by: Elie Tournier <elie.tournier@collabora.com> (v1)
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
If a texture is bound in a GL context the texture object will not be
destroyed by the driver even if the program using the texture freed it.
At least with mesa Gallium drivers the effect that these texture objects
also hold references to sampler views that might have been created and
destroyed in sub-contexts. Now, when a new texture is allocated and
bound to the corresponding texture target, the old texture will finally
be destroyed at this point, and by doing so it will try to access these
already destroyed sampler views, resulting in a use after free.
Consequently, unbind the texture in context 0 after it has been allocated,
so that the driver doesn't hold an additional reference to it that inhibits
its destruction of the texture when the guest program requests this.
v2: Make sure that the texture is also unbound when the allocation function
bails out with an error (Chia-I Wu)
Closes#98
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
On one hand render_condition_enable might also be set when the
cond_render_gl_mode is empty, so no conditional rendering is
actually done and the copy path can be used. On the other hand,
on GLES rectangular textures are emulated, so the texture target
needs to be corrected, and for some reason different base formats
were required which doesn't make sense.
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
The GL fallback also must be used on GLES if the target surface is
a multi-sample surface or if the region blitted from a multi-sample
surface is not equal to the target region.
This patch results in piglit reporing regressions with
multisample-blit * depth
on a GLES host. However, these piglits make bits to MS surfaces which
is not supported by glBlitFramebuffer on GLES. The old code took that
path, and mesa would actually report an error when compiled with -DDEBUG
butthe piglits still reported success. Unfortunately, resolving the MS
propperly is quite a challange for depth textures, and requires information
that is not available in the blit info (zNear and zFar), so getting these
piglits to act correcly on GLES hosts seems quite impossible.
v2: Also use fallback on GLES if the source is multi-sample and the formats
of source and dest are not equal
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
With that the srgb <-> linear handling can be simplified.
v2: Don't use texture views when TextureStorage can't be used to create
the texture, because then it is not immutable.
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
later we need this function in the blitter
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
v2: Don't use texture views if the texture of the format can't be
allocated by using TextureStorage, since then it is not immutable
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
On GL 4.4 (and GLES with EXT_sRGB_write_control) blits will decode and
encode sRGB when GL_FRAMEBUFFER_SRGB is supported, and they don't decode
when it isn't supported. Enabling or disabling this is signalled from
the guest by using the according surfaces.
In summary, there is no need to use the GL fallback for a blit decoding
sRGB, because this can also be handled in a normal blit.
Fixes piglits on a GL host:
blit texture srgb_to_linear downsample enabled render
blit renderbuffer srgb_to_linear msaa enabled render
blit renderbuffer srgb_to_linear msaa enabled clear
blit renderbuffer srgb_to_linear downsample enabled clear
v2: Fix typo in commit message (Erik)
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
The GL_EXT_color_buffer_float exptension spec defines that glReadPixels
can be done with the combination of GL_RGBA and GL_FLOAT, which matches
what OpenGL ES 3.2 allows. So let's enable the readback-format in this
case.
Thanks to Gert Wollny for pointing this out.
Signed-off-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Reviewed-By: Gert Wollny <gert.wollny@collabora.com>
On D-GL a program may contain a TE shader but no TC shader. On GLES either
both or none of TES and TCS need to be available. So if the guest sends a
shader program without a TCS, inject a passthrough shader using the
patch parameters given in the GL code.
Closes#84
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Given a source and destination image, consider the following sequence
of events:
1) glTexImage with src
2) virgl_resource_copy_region copies src to dst using GL
3) mesa attempts to readback dst, which returns the correct
result with vrend_transfer_send_getteximage only
Commands that reproduce this scenario on i965/qemu:
./bin/arb_copy_image-formats -auto
./bin/fbo-generatemipmap-formats GL_ARB_texture_float -auto
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Closes: #74
v2: * Enable enhanced layouts in the guest only when the host supports
it, because with indirect arrays enabled we re-emit the layout
the guest send.
* enable indirect arrays independently from the host capabilities
(following Eriks comments)
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-By: Gurchetan Singh <gurchetansingh@chromium.org>
Also check for GL_EXT_polygon_offset_clamp when deciding whether the
corresponding virgl feature is supported. This enables D3D11 in guest
Wine (and thus Steam/Proton), when running on GLES hosts that support
this extension.
Signed-off-by: Alexandros Frantzis <alexandros.frantzis@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
This is useful for knowing if we can expose
KHR_texture_compression_astc_sliced_3d in the guest.
Signed-off-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
The piglits and KHR tests that use TF3 (i.e. multiple output
streams) tend to hang, and since the GLES host doesn't support
this but we lie about the GLSL version it is better to signal
the guest correctly that this is not supported.
v2: Set flag only on caps v2 code path, otherwise we might read
from uninitialized memory
v3: Increment feature check version to indicate that the TF3
feature flag is valid.
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
In order to properly propagate the output masks to the inputs of the next
shader the shader slection has do be done in the order the shaders will be
called later.
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
This is useful to avoid unsupported format-conversions on GLES.
Signed-off-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
This code-path is never going to work on GLES, so let's avoid trying.
Signed-off-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Reviewed-by: Gurchetan Singh <gurchetansingh@chromium.org>
David Riley
Chia-I Wu
Gert Wollny
Emil Velikov
Alexandros Frantzis
Erik Faye-Lund
Gurchetan Singh