etnaviv: Set SE.CLIP registers, add margins for scissor/clip registers

This fixes rendering of full-screen quads (and other screen-filling
geometry, e.g. ioquake3 walls up-close) on gc3000. It should be a no-op
on other hardware.

- It looks like SE_CLIP registers were not set at all.
  I'm amazed that rendering worked without them. Emit them to
  avoid issues on gc3000.

- Define constants
  ETNA_SE_SCISSOR_MARGIN_RIGHT (0x1119)
  ETNA_SE_SCISSOR_MARGIN_BOTTOM (0x1111)
  ETNA_SE_CLIP_MARGIN_RIGHT (0xffff)
  ETNA_SE_CLIP_MARGIN_BOTTOM (0xffff)

  These demarcate the margin (fixp16) between the computed sizes and the
  value sent to the chip. I have set these to the numbers used by the
  Vivante driver for gc2000. I am not sure whether any old hardware was
  relying on the old numbers, or whether those were just a guess. But if
  so, these need to be moved to the _specs structure.

CC: <mesa-stable@lists.freedesktop.org>
Signed-off-by: Wladimir J. van der Laan <laanwj@gmail.com>
Acked-by: Christian Gmeiner <christian.gmeiner@gmail.com>

etnaviv: Generate new sin/cos instructions on GC3000

Shaders using sin/cos instructions were not working on GC3000.

The reason for this turns out to be that these chips implement sin/cos
in a different way (but using the same opcodes):

- Need their input scaled by 1/pi instead of 2/pi.

- Output an x and y component, which need to be multiplied to
  get the result.

- tex_amode needs to be set to 1.

Add a new bit to the compiler specs and generate these instructions
as necessary.

CC: <mesa-stable@lists.freedesktop.org>
Signed-off-by: Wladimir J. van der Laan <laanwj@gmail.com>
Acked-by: Christian Gmeiner <christian.gmeiner@gmail.com>

anv/cmd_buffer: Use the proper depth input attachment surface state

Commit 2852efcda40274acf3272611c6a3b7731523a72d moved the location of
the depth input attachment surface state from the render pass to the
image view, but failed to update the surface state location used when
emitting the binding table. Fix this by loading the surface state from
the correct location.

Fixes:
dEQP-VK.renderpass.formats.d16_unorm.input.*
dEQP-VK.renderpass.formats.d24_unorm_s8_uint.input.*
dEQP-VK.renderpass.formats.d32_sfloat.input.*
dEQP-VK.renderpass.formats.x8_d24_unorm_pack32.input.*
dEQP-VK.renderpass.attachment_allocation.input_output.93
dEQP-VK.renderpass.attachment_allocation.input_output.92
dEQP-VK.renderpass.attachment_allocation.input_output.82
dEQP-VK.renderpass.attachment_allocation.input_output.46

Cc: "17.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>

glsl: fix heap-buffer-overflow

The `end+1` skips the ']', whereas the `strlen+1` includes the final
'\0' in the move to terminate the string.

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

etnaviv: Cannot render to rb-swapped formats

Exposing rb swapped (or other swizzled) formats for rendering would
involve swizzing in the pixel shader. This is not the case at the
moment, so reject requests for creating such surfaces.

(GPUs that need an extra resolve step anyway due to multiple pixel
pipes, such as gc2000, might also do this swap in the resolve operation.
But this would be tricky to keep track of)

CC: <mesa-stable@lists.freedesktop.org>
Signed-off-by: Wladimir J. van der Laan <laanwj@gmail.com>
Acked-by: Christian Gmeiner <christian.gmeiner@gmail.com>

etnaviv: Avoid infinite loop in find_frame()

Use of unsigned loop control variable with '>= 0' would lead
to infinite loop.

Reported by clang:

etnaviv_compiler.c:1024:39: warning: comparison of unsigned expression
>= 0 is always true [-Wtautological-compare]
   for (unsigned sp = c->frame_sp; sp >= 0; sp--)
                                   ~~ ^  ~

v2: Simply use the same datatype as c->frame_sp is using.

CC: <mesa-stable@lists.freedesktop.org>
Reported-by: Rhys Kidd <rhyskidd@gmail.com>
Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Rhys Kidd <rhyskidd@gmail.com>

radv/ac: apply slice rounding to 1d arrays as well.

Fixes:
dEQP-VK.glsl.texture_functions.texture.*1darray*

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Cc: "17.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/geom: check if esgs and gsvs ring exists before filling geom rings

There are some corner cases where you end up with an esgs ring, but no
gsvs ring, test for both before dereferencing.

Fixes:
dEQP-VK.geometry.emit.points_emit_0_end_0

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: enable geometryShader and multiViewport capabilities.

This enables geometry shader support on radv.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: handle layer export from vs->fs properly

Fixes:
dEQP-VK.geometry.layered.1d_array.fragment_layer

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: emit esgs itemsize register.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: handle prim id inputs to fragment shader.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: emit geometry shaders to hardware

This emits the compiled geometry shader and other state registers.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: emit geometry ring size and pointers via preamble (v2)

This uses the scratch infrastructure to handle the esgs
and gsvs rings.

(this replaces the old code that did this with patching).

v2: fix correct ring sizes, reset sizes (Bas)

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: add gs ring size calculations to pipeline.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: add pipeline creation support for geometry shaders (v2.1)

This adds gs copy shader support to the pipeline cache, and few
geometry related changes.

v2: rebase for spill changes.
v2.1: fix incorrect pipeline destruction.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: handle primitive id

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: handle emitting vertex outputs to esgs ring.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: handle gs inputs

This handles geometry shader inputs written by the vertex (es) shader
to the esgs ring.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: add geom input support to get deref offset.

This just adds the API and fixes up the callers.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: handle invocation and primitive id intrinsics

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: handle geometry emit vertex and end prim intrinsics.

This handles emitting things to the gsvs ring, and sending the
correct GS msgs.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: handle emitting gs epilogue

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: add copy shader creation

This create the gs copy shader and compiles it.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: setup function parameters for vs as es and copy shader.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: pass some necessary gs info back to state handling.

We need this info to program some registers.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: emit vertex shader to correct hw block.

This emits the shader to the ES block in the correct case.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: propogate as_es flag into shader info from key.

This just places the flag into the shader info so we can use it from
the driver after we create the shader.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: extend shader stage code to cover geometry shaders.

This enables the paths for setting up user ptrs to vs/es and gs.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: start setting up the geom shader rings (v2)

This sets up the rings and adds the variables
needed to make them work.

v2: rework for sharing ring and scratch
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: handle geom shader sgpr/vgpr inputs

This just sets up the gpr inputs.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: add geom shader sendmsg defines.

This just adds some defines needed for geom shaders.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: add some geom shader info from nir->ac shader.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: move hw vertex shader emit to separate function

This is to later allow ES shaders to be emitted.

Review-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: fixup ia multi vgt param code to handle geom shaders.

This fixes up a few of the commented out blocks.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: add code to set gs_table_depth.

Review-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: add small helper to denote when a geom shader is in the pipeline.

Review-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: Prevent Coverity warning

Prevent Coverity seeing potential errors when src is
no initialized in the switch case.

Coverity-Id: 1396397
Signed-off-by: Robert Foss <robert.foss@collabora.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

mesa: add new MESA_GLSL flag for printing shader cache debug info

Reviewed-by: Eric Anholt <eric@anholt.net>

glsl: add cache to ctx and add sha1 string fields

We also add a flag for detecting shaders written to shader cache.

V2: dont leak cache

Signed-off-by: Timothy Arceri <timothy.arceri@collabora.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

glsl: add new uniform fields to be used to restore state from cache

Signed-off-by: Timothy Arceri <timothy.arceri@collabora.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

glsl: Switch to disable-by-default for the GLSL shader cache

The shader cache is expected to be developed incrementally over a
fairly long series of commits. For that period of instability, we
require users to opt into the shader cache by setting:

MESA_GLSL_CACHE_ENABLE=1

In the future, when the shader cache is complete, we can revert this
commit so that the cache will be on by default.

The user can always disable the cache with
MESA_GLSL_CACHE_DISABLE=1. That functionality is not affected by this
commit, (nor will it be affected by the future revert).

Reviewed-by: Eric Anholt <eric@anholt.net>

radv/ac: implement txs for buffer textures.

This fixes a bunch of buffer related:
dEQP-VK.memory.pipeline_barrier.*
tests, that were crashing in LLVM due to this being missing.

Reviewed-by: Andres Rodriguez<andresx7@gmail.com>
Cc: "17.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: handle nir irem opcode.

This fixes:
dEQP-VK.spirv_assembly.instruction.compute.opsrem.*

Reviewed-by: Andres Rodriguez <andresx7@gmail.com>
Cc: "17.0" <mesa-stable@lists.freedesktop.org"
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv/ac: fix multisample subpass image.

We weren't adding the fragment position properly.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Cc: "17.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radv: handle transfer_write as a dst flag.

It appears we can get image barriers like:
    srcStageMask:                   VkPipelineStageFlags = 4096 (VK_PIPELINE_STAGE_TRANSFER_BIT)
    dstStageMask:                   VkPipelineStageFlags = 4096 (VK_PIPELINE_STAGE_TRANSFER_BIT)
    dependencyFlags:                VkDependencyFlags = 0
    memoryBarrierCount:             uint32_t = 0
    pMemoryBarriers:                const VkMemoryBarrier* = NULL
    bufferMemoryBarrierCount:       uint32_t = 0
    pBufferMemoryBarriers:          const VkBufferMemoryBarrier* = NULL
    imageMemoryBarrierCount:        uint32_t = 1
    pImageMemoryBarriers:           const VkImageMemoryBarrier* = 0x7ffc882367b0
        pImageMemoryBarriers[0]:        const VkImageMemoryBarrier = 0x7ffc882367b0:
            sType:                          VkStructureType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER (45)
            pNext:                          const void* = NULL
            srcAccessMask:                  VkAccessFlags = 4096 (VK_ACCESS_TRANSFER_WRITE_BIT)
            dstAccessMask:                  VkAccessFlags = 4096 (VK_ACCESS_TRANSFER_WRITE_BIT)
            oldLayout:                      VkImageLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL (7)
            newLayout:                      VkImageLayout = VK_IMAGE_LAYOUT_GENERAL (1)
            srcQueueFamilyIndex:            uint32_t = 4294967295
            dstQueueFamilyIndex:            uint32_t = 4294967295
            image:                          VkImage = 0x2df55e0
            subresourceRange:               VkImageSubresourceRange = 0x7ffc882367e0:
                aspectMask:                     VkImageAspectFlags = 1 (VK_IMAGE_ASPECT_COLOR_BIT)
                baseMipLevel:                   uint32_t = 0
                levelCount:                     uint32_t = 1
                baseArrayLayer:                 uint32_t = 0
                layerCount:                     uint32_t = 1

This fixes all the CTS dEQP-VK.memory.pipeline_barrier.transfer_dst tests here,
not sure if this is a too large hammer.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Cc: "17.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>

r600: fix a compilation warning in r600_screen_create()

Should be r600_common_screen instead of r600_screen.

Fixes: 80157a2c20 ("gallium/radeon: clean up r600_query_init_backend_mask")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

gallium/radeon: merge dirty_fb_counter and dirty_tex_descriptor_counter

to simplify things in draw_vbo a little

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

winsys/radeon: clamp vram_vis_size to 256MB

the value from the kernel is wrong

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: handle count_from_stream_output in a few IA_MULTI_VGT_PARAM cases

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: don't invoke DCC decompression in update_all_texture_descriptors

This fixes a bug uncovered by the 17-part patch series, specifically:
  "gallium/radeon: merge dirty_fb_counter and dirty_tex_descriptor_counter"

If dirty_tex_counter has been updated and set_shader_image invokes DCC
decompression, the DCC decompression itself checks the counter and updates
descriptors, which in turn invokes the same DCC decompression. The blitter
can't handle the recursion and the driver eventually crashes.

Cc: 17.0 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: fold info->indirect conditionals into the last one in draw_vbo

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: atomize the scratch buffer state

The update frequency is very low.

Difference: Only account for the size when allocating a new one and when
            starting a new IB, and check for NULL. (v3)

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

r600: Fix stack overflow

Commit 7b5878ee0491e7a93914389a8369cd6752b9757d increased number of
outputs to 64, but left output array intact. This caused stack overflow
when number of outputs is bigger then 32. Found by ASAN.

Cc: "12.0 13.0 17.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

gallium/radeon: add new HUD queries for monitoring the CP

There are even more counters in the CP_STAT register but I think
these ones are enough for now.

v2: only read (and expose) CP_STAT on VI+

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

gallium/radeon: add new GPU-sdma-busy HUD query

For simplicity, GPU-sdma-busy will return 0 on previous gens.

v2: only read SRBM_STATUS2 on Evergreen+

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

gallium/radeon: rename grbm to mmio in the gpu load path

We also want to monitor other MMIO counters like SRBM_STATUS2 in
order to know if SDMA is busy.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

winsys/amdgpu: add a fast exit path into amdgpu_cs_add_buffer

The time spent in the function dropped by 37% for torcs.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

winsys/amdgpu: do not iterate twice when adding fence dependencies

The perf difference is very small, 3.25->2.84% in amdgpu_cs_flush()
in the DXMD benchmark.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

winsys/amdgpu: add one likely() call in amdgpu_cs_flush()

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

hud: fix compilation warnings in hud_nic_graph_install()

v2: use PRId64 instead of PRIx64

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

st/mesa: make st_texture_get_sampler_view() static

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

gallium/radeon: remove r600_common_context::max_db

this cleanup is based on the vulkan driver, which seems to do the same thing

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

winsys/amdgpu: fix ADDR_REGISTER_VALUE::backendDisables

This would be a fix if the value was used anywhere.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

gallium/radeon: clean up r600_query_init_backend_mask

This just needs to be done for r600g in the screen.
We don't need an IB submission for every new context created for GCN.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: precompute IA_MULTI_VGT_PARAM values into a table

The perf difference is very small: 0.99% -> 0.40% for the time spent
in si_get_ia_multi_vgt_param when si_draw_vbo is 20%. Pretty much nothing.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: move VGT_VERTEX_REUSE_BLOCK_CNTL into shader states for Polaris

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: state atom IDs don't have to be off by one

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: use a bitmask for looping over dirty PM4 states

also move it to draw_vbo, because it should be 0 in most cases

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: atomize L2 prefetches

to move the big conditional statement out of draw_vbo

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: unbind disabled shader stages to prevent useless L2 prefetches

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: also prefetch compute shaders

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: update dirty_level_mask only after the first draw after FB change

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

gallium/radeon: allow VRAM-only placements again on APUs & recent amdgpu

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: don't set +fp64-denormals

it's the default and the name will change to +fp64-fp16-denormals.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

radeonsi: remove si_shader_context::param_tess_offchip

we don't use on-chip tess.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

etnaviv: force vertex buffers through the MMU

This fixes a vertex data corruption issue if some of the vertex streams
go through the MMU and some don't.

Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
Tested-by: Philipp Zabel <p.zabel@pengutronix.de>
Acked-by: Christian Gmeiner <christian.gmeiner@gmail.com>

radv: Expose VK_KHR_maintenance1

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: Fix vkCmdCopyImage for 2d slices into 3d Images

Previously the z offset of the destination image was being ignored. It
should be taken into account when copying into a 3d target.

Also, img_extent_el.depth was being incorrectly clamped to 1 due to the
source image being VK_IMAGE_TYPE_2D. This would result in the blit
failing to iterate over all the 3d slices. Instead we clamp to the
destination image type.

Fixes failures in CTS tests:
dEQP-VK.api.copy_and_blit.image_to_image.3d_images.*

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: Expose transfer format features.

Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>

radv: Don't allow any operations on non-supported depth/stencil formats.

We really use the depth block for the blits.

Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>

radv: use new error codes for AllocateDescriptorSets

There is a new error code in Maintenance1 that is more specific to the
situation: VK_ERROR_OUT_OF_POOL_MEMORY_KHR

Fixes CTS test case:
dEQP-VK.api.descriptor_pool.out_of_pool_memory

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: vkAllocateCommandBuffers should NULL all output handles

This is part of the spec and fixes CTS tests:
dEQP-VK.api.object_management.alloc_callback_fail_multiple.command_buffer_*

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: add trim command pool stub

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

i965: Support the force_glsl_version driconf option.

Gallium drivers have had this for a while.  It makes sense to support
it consistently across drivers, so expose it in i965 as well.

Cc: "17.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

i965: Fix check for negative pitch in can_do_fast_copy_blit().

At this point, the pitch is in bytes.  We haven't yet divided the pitch
by 4 for tiled surfaces, so abs(pitch) may be larger than 32K.  This
means the bit 15 trick won't work.

The caller now has signed integers anyway, so just pass those through
and do the obvious check.

Cc: "17.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

radv: Handle command buffers that need scratch memory.

v2: Create the descriptor BO with CPU access.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>

radv: Track scratch usage across pipelines & command buffers.

Based on code written by Dave Airlie.

Signed-off-by: Bas Nieuwenhuizen <basni@oogle.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>

radv/ac: Add compiler support for spilling.

Based on code written by Dave Airlie.

Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>

radv/amdgpu: Support a preamble CS.

Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>

i965: add assert to while_jumps_before_offset()

jip should always be negative here as its the result of
do instruction - while instruction.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: fix up asserts in brw_inst_set_jip()

We are casting from a signed 32bit int to an unsigned 16bit int
so shift 15 bits rather than 16.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

llvmpipe: Use LLVMDumpModule, not DumpModule.

Forgot the prefix ...

Fixes: 0fca80b3db64dc1d004f78e22b9de86a07e9de96
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>

various: Fix missing DumpModule with recent LLVM.

Since LLVM revision 293359 DumpModule gets only implemented when
either a debug build or LLVM_ENABLE_DUMP is set.

This patch adds a direct replacement for the function for radv and
radeonsi, However, as I don't know a good place to put common LLVM
code for all three I inlined the implementation for LLVMPipe.

v2: Use the new code for LLVM 3.4+ instead of LLVM 5+ & fixed indentation

Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>

r600g: use ieee variants of multiplication instructions

This matches the behavior of most other drivers, including nouveau,
radeonsi, and i965.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

r600g: add support for optionally using non-IEEE mul ops

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

vc4: Coalesce into TLB writes as well as VPM/tex.

This generally cuts an instruction when blending is enabled and we thus
have a single instruction generating the color value.

total instructions in shared programs: 91759 -> 91634 (-0.14%)
instructions in affected programs:     5338 -> 5213 (-2.34%)

vc4: Avoid an extra temporary and mov in ffloor/ffract/fceil.

shader-db results:

total instructions in shared programs: 92611 -> 91764 (-0.91%)
instructions in affected programs:     27417 -> 26570 (-3.09%)

The star is one shader in glmark2's terrain (drops 16% of its
instructions), but there are also wins in mupen64plus and glb2.7.

vc4: Flip the switch to run the GLSL compiler optimization loop once.

This has almost no effect on shader-db:

total instructions in shared programs: 92572 -> 92611 (0.04%)
instructions in affected programs:     4486 -> 4525 (0.87%)

Looking at 2 of the 7 different shaders that were hurt (all of which were
in mupen64), they all appear to be just differences in order of
instructions at the NIR level.

The advantage is that this should significantly reduce time in the compiler.

i965: Unbind deleted shaders from brw_context, fixing malloc heisenbug.

Applications may delete a shader program, create a new one, and bind it
before the next draw.  With terrible luck, malloc may randomly return a
chunk of memory for the new gl_program that happened to be the exact
same pointer as our previously bound gl_program.  In this case, our
logic to detect new programs in brw_upload_pipeline_state() would break:

      if (brw->vertex_program != ctx->VertexProgram._Current) {
         brw->vertex_program = ctx->VertexProgram._Current;
         brw->ctx.NewDriverState |= BRW_NEW_VERTEX_PROGRAM;
      }

Because the pointer is the same, we'd think it was the same program.
But it could be wildly different - a different stage altogether,
different sets of resources, and so on.  This causes utter chaos.

As unlikely as this seems, I believe I hit this when running a subset
of the CTS in a loop, in a group of tests that churns through simple
programs, deleting and rebuilding them.  Presumably malloc uses a
bucketing cache of sorts, and so freeing up a gl_program and allocating
a new one fairly quickly causes it to reuse that memory.

The result was that brw->vertex_program->info.num_ssbos claimed the
program had SSBOs, while brw->vs.base.prog_data.binding_table claimed
that there were none.  This was crazy, because the binding table is
calculated from info.num_ssbos - the shader info appeared to change
between shader compile time and draw time.  Careful use of watchpoints
revealed that it was being clobbered by rzalloc's memset when building
an entirely different program...

Fortunately, our 0xd0d0d0d0 canary for unused binding table entries
caused us to crash out of bounds when trying to upload SSBOs, or we
may have never discovered this heisenbug.

Fixes crashes in GL45-CTS.compute_shader.sso-case2 when using a hacked
cts-runner that only runs GL45-CTS.compute_shader.s* in EGL config ID 5
at 64x64 in a loop with 100 iterations.

Cc: "17.0 13.0 12.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>