summaryrefslogtreecommitdiffstats
path: root/src
Commit message (Collapse)AuthorAgeFilesLines
* i965: Use the upload BO for push constants on Gen9+.Kenneth Graunke2017-05-202-6/+11
| | | | | | | | | | | | | | | | | | | | | | | | | | Shaders can use quite a bit of uniform data. Better to put it in the upload buffers, like we do for client vertex data, rather than the batch buffer state area, which is primarly used for indirect state. This should free up batch space, allowing us to emit more commands in a batch before flushing. Because BRW_NEW_BATCH also causes a lot of state to be re-emitted, it may also reduce CPU overhead a little bit. We took this approach on Gen4-5, but switched to using the batch area on Gen6+ because buffer 0 is relative to Dynamic State Base Address by default, which is set to the start of the batch. On Gen9+, we already use a relocation due to a workaround, so this is trivial to change and has basically no downside. Unfortunately we can't change compute shader push constants because MEDIA_CURBE_LOAD always uses an offset from dynamic state base address. Improves performance in GLBenchmark 2.7/TRex Offscreen by: - Skylake GT4e: 0.52821% +/- 0.113402% (n = 190) - Apollolake: 0.510225% +/- 0.273064% (n = 70) Reviewed-by: Chris Forbes <[email protected]>
* i965: Drop BRW_NEW_PUSH_CONSTANT_ALLOCATION from CS packets.Kenneth Graunke2017-05-202-3/+1
| | | | | | | | | | | I don't think CS push constant uploading uses the section of L3 controlled by 3DSTATE_PUSH_CONSTANT_ALLOC_XS. So I don't think it needs to be re-emitted when that space is reallocated. The programming note in gen7_allocate_push_constants doesn't indicate this is necessary, at least. Reviewed-by: Chris Forbes <[email protected]>
* nvc0/ir: SHLADD's middle source must be an immediateIlia Mirkin2017-05-201-0/+2
| | | | | | | | The instruction encodings only allow for immediates. Don't try to replace a zero (which is dumb to have in that op in any case) with RZ. Signed-off-by: Ilia Mirkin <[email protected]> Cc: [email protected]
* android: add -Wl,--build-id=sha1 to LDFLAGS for libvulkan_intelTapani Pälli2017-05-201-0/+2
| | | | | | | Just like is done on desktop and what is expected by the build-id code. Signed-off-by: Tapani Pälli <[email protected]> Reviewed-by: Jason Ekstrand <[email protected]>
* automake: add SWR LLVM gen_builder.hpp workaroundEmil Velikov2017-05-201-30/+11
| | | | | | | | | | | | | | | | | | | | | As gen_builder.hpp file is generated, it contains information that is specific to the LLVM version it originates from. As suggested by Tim, the file seems to be forwards compatible. So in order to produce ship a file which will work everywhere we should be using earlies supported LLVM - 3.9. With this we're back on track and can build all of mesa without python/mako/flex and friends. In the long term we might want to see if the python generators can be updated to produce LLVM version agnostic files. At least within the range supported by SWR. Cc: <[email protected]> Cc: Chuck Atkins <[email protected]> Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Tim Rowley <[email protected]>
* st/mesa: don't mark the program as in cache_fallback when there is cache missTimothy Arceri2017-05-203-2/+9
| | | | | | | | | | | | | | | | | | When we fallback currently the gl_program objects are re-allocated. This is likely to change when the i965 cache lands, but for now this fixes a crash when using MESA_GLSL=cache_fb. This env var simulates the fallback path taken when a tgsi cache item doesn't exist due to being evicted previously or some kind of error. Unlike i965 we are always falling back at link time so it's safe to just re-allocate everything. We will be unnecessarily freeing and re-allocate a bunch of things here but it's probably not a huge deal, and can be changed when the i965 code lands. Fixes: 0e9991f957e2 ("glsl: don't reference shader prog data during cache fallback") Reviewed-by: Nicolai Hähnle <[email protected]>
* mesa: add an env var to force cache fallbackTimothy Arceri2017-05-203-1/+8
| | | | | | | | | | | | | | | | | For the gallium state tracker a tgsi binary may have been evicted from the cache to make space. In this case we would take the fallback path and recompile/link the shader. On i965 there are a number of reasons we can get to the program upload stage and have neither IR nor a valid cached binary. For example the binary may have been evicted from the cache or we need a variant that wasn't previously cached. This environment variable enables us to force the fallback path that would be taken in these cases and makes it easier to debug these otherwise hard to reproduce scenarios. Reviewed-by: Nicolai Hähnle <[email protected]>
* st/mesa: improve shader cache debug infoTimothy Arceri2017-05-201-2/+4
| | | | | | | | | | | This will explicitly state that we are following the fallback path when we find invalid/corrupt cache items. It will also output the fallback message when the fallback path is forced via an environment variable, the following patches will allow this. Reviewed-by: Nicolai Hähnle <[email protected]> Reviewed-by: Marek Olšák <[email protected]>
* st/xvmc: add DRI3 supportEmil Velikov2017-05-191-1/+3
| | | | | Signed-off-by: Emil Velikov <[email protected]> Tested-by: Ilia Mirkin <[email protected]>
* st/omx: add DRI3 supportEmil Velikov2017-05-191-1/+3
| | | | | | Cc: Christian König <[email protected]> Signed-off-by: Emil Velikov <[email protected]> Reviewed-and-Tested-by: Leo Liu <[email protected]>
* gallium/targets: link against XCB only as neededEmil Velikov2017-05-192-4/+12
| | | | | | | | OMX and VA can optionally use the X11 DRI2/DRI3, thus we should link only as required. Cc: <[email protected]> Signed-off-by: Emil Velikov <[email protected]>
* st/omx: fix building against X11-less setupsEmil Velikov2017-05-191-0/+6
| | | | | | | | | | The vl_*_screen_create API properly falls back to a NOP when we're building without specific platforms. So the only thing we need is to handle the lack of X11/Xlib.h and provide a dummy Display define. Cc: <[email protected]> Cc: Christian König <[email protected]> Signed-off-by: Emil Velikov <[email protected]>
* st/omx: remove unneeded X11 includeEmil Velikov2017-05-191-2/+0
| | | | | | | | | En route to a X11-less builds Cc: <[email protected]> Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Nayan Deshmukh <[email protected]> Reviewed-by: Christian König <[email protected]>
* st/omx: remove unused drm_driver.h includesEmil Velikov2017-05-192-2/+0
| | | | | | Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Nayan Deshmukh <[email protected]> Reviewed-by: Christian König <[email protected]>
* st/va: check if vl_*_screen_create has failed only onceEmil Velikov2017-05-191-4/+3
| | | | | | Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Nayan Deshmukh <[email protected]> Reviewed-by: Christian König <[email protected]>
* st/va: fix misplaced closing bracketEmil Velikov2017-05-191-1/+1
| | | | | | | | | | | It's been like this since the code was introduced. Fixes: 86eb4131a90 (st/va: add headless support, i.e. VA_DISPLAY_DRM) Cc: <[email protected]> Cc: Julien Isorce <[email protected]> Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Nayan Deshmukh <[email protected]> Reviewed-by: Christian König <[email protected]>
* st/va: move variable declaration to where its usedEmil Velikov2017-05-191-2/+1
| | | | | | | | ... and make it const, since we shouldn't tinker with it. Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Nayan Deshmukh <[email protected]> Reviewed-by: Christian König <[email protected]>
* auxiliary/vl: use vl_*_screen_create stubs when building w/o platformEmil Velikov2017-05-195-19/+35
| | | | | | | | | | | | | | | | | | | | Provide a dummy stub when the user has opted w/o said platform, thus we can build the binaries without unnecessarily requiring X11/other headers. In order to avoid build and link-time issues, we remove the HAVE_DRI3 guards in the VA and VDPAU state-trackers. With this change st/va will return VA_STATUS_ERROR_ALLOCATION_FAILED instead of VA_STATUS_ERROR_UNIMPLEMENTED. That is fine since upstream users of libva such as vlc and mpv do little error checking, let alone distinguish between the two. Cc: Leo Liu <[email protected]> Cc: Guttula, Suresh <[email protected]> Cc: [email protected] Cc: Christian König <[email protected]> Signed-off-by: Emil Velikov <[email protected]>
* loader: build libloader_dri3_helper.la only with HAVE_PLATFORM_X11Emil Velikov2017-05-191-0/+2
| | | | | | | | Pretty much every other place does the same. Cc: [email protected] Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Eric Engestrom <[email protected]>
* vulkan: automake: remove unused VULKAN_LIB_DEPS variableEmil Velikov2017-05-191-6/+0
| | | | | Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Eric Engestrom <[email protected]>
* configure: check once for DRI3 dependenciesEmil Velikov2017-05-1912-7/+13
| | | | | | | | | | | | | | | | Currently we are having the XCB_DRI3 dependencies duplicated, partially. Just do a once-off check and add all of the respective CFLAGS/LIBS where needed. As a nice side effect this helps us solve a couple of FIXMEs. DRI3 is not a thing w/o X11 so disable it in such cases. Cc: [email protected] Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Eric Engestrom <[email protected]>
* configure: set HAVE_foo_PLATFORM as applicableEmil Velikov2017-05-192-6/+0
| | | | | | | | | | Rather than having multiple places that define the macros, do it just once in configure. Makes existing code a bit shorter and easier to manage as we fix the VL targets with follow-up commits. Cc: [email protected] Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Eric Engestrom <[email protected]>
* configure: update remaining --with-egl-platforms referencesEmil Velikov2017-05-191-1/+1
| | | | | | | | Rename the remaining references to omit the egl part. Cc: [email protected] Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Eric Engestrom <[email protected]>
* configure: rename remaining HAVE_EGL_PLATFORM_* guardsEmil Velikov2017-05-191-3/+3
| | | | | | | | | Analogous to others earlier, these will be used to control the platform for more than the EGL driver. Cc: <[email protected]> Signed-off-by: Emil Velikov <[email protected]> Reviewed-by: Eric Engestrom <[email protected]>
* Android: r600: fix build when LLVM is disabledRob Herring2017-05-191-0/+4
| | | | | | | | | | | | | | | | | | | | | | | | | There's still an error after my recent clean-up if LLVM is not patched to enable AMDGPU target: external/mesa3d/src/amd/common/ac_llvm_util.c:38:2: error: implicit declaration of function 'LLVMInitializeAMDGPUTargetInfo' is invalid in C99 [-Werror,-Wimplicit-function-declaration] LLVMInitializeAMDGPUTargetInfo(); ^ external/mesa3d/src/amd/common/ac_llvm_util.c:39:2: error: implicit declaration of function 'LLVMInitializeAMDGPUTarget' is invalid in C99 [-Werror,-Wimplicit-function-declaration] LLVMInitializeAMDGPUTarget(); ^ external/mesa3d/src/amd/common/ac_llvm_util.c:40:2: error: implicit declaration of function 'LLVMInitializeAMDGPUTargetMC' is invalid in C99 [-Werror,-Wimplicit-function-declaration] LLVMInitializeAMDGPUTargetMC(); ^ external/mesa3d/src/amd/common/ac_llvm_util.c:41:2: error: implicit declaration of function 'LLVMInitializeAMDGPUAsmPrinter' is invalid in C99 [-Werror,-Wimplicit-function-declaration] LLVMInitializeAMDGPUAsmPrinter(); ^ We need to drop libmesa_amd_common when LLVM is disabled, however there's still a dependency on include paths for ac_binary.h. So explicitly add the include path when LLVM is disabled. Signed-off-by: Rob Herring <[email protected]> Acked-by: Nicolai Hähnle <[email protected]>
* virgl: fix virgl_bo_transfer_{put, get} box struct copyRob Herring2017-05-191-2/+12
| | | | | | | | | | | | | Commit 3dfe61ed6ec6 ("gallium: decrease the size of pipe_box - 24 -> 16 bytes") changed the size of pipe_box, but the virgl code was relying on pipe_box and drm_virtgpu_3d_box structs having the same size/layout doing a struct copy. Copy the fields one by one instead. Cc: Marek Olšák <[email protected]> Cc: Dave Airlie <[email protected]> Fixes: 3dfe61ed6ec ("gallium: decrease the size of pipe_box - 24 -> 16 bytes") Signed-off-by: Rob Herring <[email protected]> Reviewed-by: Marek Olšák <[email protected]>
* egl: add g_egldispatchstubs.h to the release tarballEmil Velikov2017-05-191-1/+2
| | | | | Fixes: ce562f9e3fa ("EGL: Implement the libglvnd interface for EGL (v3)") Signed-off-by: Emil Velikov <[email protected]>
* egl/android: fix segfault within swap_buffersTapani Pälli2017-05-191-1/+6
| | | | | | | | | | | | | | | | | | | | | | Function droid_swap_buffers may get called without dri2_surf->buffer set, in these cases we don't have a back buffer set either. Patch fixes segfault seen with 3DMark that uses android.opengl.GLSurfaceView for rendering it's UI. backtrace: #00 pc 00013f88 /system/lib/egl/libGLES_mesa.so (droid_swap_buffers+104) #01 pc 000117b2 /system/lib/egl/libGLES_mesa.so (dri2_swap_buffers+50) #02 pc 000058b2 /system/lib/egl/libGLES_mesa.so (eglSwapBuffers+386) #03 pc 00011329 /system/lib/libEGL.so (eglSwapBuffersWithDamageKHR+553) #04 pc 000118e7 /system/lib/libEGL.so (eglSwapBuffers+55) #05 pc 000754dc /system/lib/libandroid_runtime.so Note, this is v1 as v2 caused dEQP regressions. Fixes: 2acc69d ("EGL/Android: Add EGL_EXT_buffer_age extension") Signed-off-by: Tapani Pälli <[email protected]> Reviewed-by: Eric Engestrom <[email protected]> Acked-by: Emil Velikov <[email protected]> Cc: "17.1" <[email protected]>
* egl/wayland: Ensure we get a back bufferDaniel Stone2017-05-191-1/+9
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Commit 9ca6711faa03 changed the Wayland winsys to only block for the frame callback inside SwapBuffers, rather than get_back_bo. get_back_bo would perform a single non-blocking Wayland event dispatch, to try to find any release events which we had pulled off the wire but not actually processed. The blocking dispatch was moved to SwapBuffers. This removed a guarantee that we would've processed all events inside get_back_bo(), and introduced a failure whereby the server could've sent a buffer release event, but we wouldn't have read it. In clients unconstrained by SwapInterval (rendering ~as fast as possible), which were being displayed directly without composition (buffer release delayed), this could lead to get_back_bo() failing because there were no free buffers available to it. The drawing rightly failed, but this was papered over because of the path in eglSwapBuffers() which attempts to guarantee a BO, in order to support calling SwapBuffers twice in a row with no rendering actually having been performed. Since eglSwapBuffers will perform a blocking dispatch of Wayland events, a buffer release would have arrived by that point, and we could then choose a buffer to post to the server. The effect was that frames were displayed out-of-order, since we grabbed a frame with random past content to display to the compositor. Ideally get_back_bo() failing should store a failure flag inside the surface and cause the next SwapBuffers to fail, but for the meantime, restore the correct behaviour such that get_back_bo() no longer fails. Signed-off-by: Daniel Stone <[email protected]> Reported-by: Eero Tamminen <[email protected]> Acked-by: Pekka Paalanen <[email protected]> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98833 Fixes: 9ca6711faa03 ("Revert "wayland: Block for the frame callback in get_back_bo not dri2_swap_buffers"")
* egl/wayland: Use per-surface event queuesDaniel Stone2017-05-192-27/+68
| | | | | | | | | | | | | | | | | | | | During display initialisation, we need a separate event queue to handle the registry events, which is correctly handled. But we also need separate per-surface event queues to handle swapchain-related events, such as surface frame events and buffer release events. This avoids two surfaces from the same EGLDisplay, both current on separate threads, dispatching each other's events. Create separate per-surface event queues, create wl_surface and wl_drm proxy wrapper objects per surface, so we eliminate the race around sending events to the wrong queue. swrast buffers do not need a dedicated proxy wrapper, as the wl_shm_pool used to create the wl_buffers, being transient, can itself be assigned to a queue. Signed-off-by: Daniel Stone <[email protected]> Reviewed-by: Lionel Landwerlin <[email protected]> Fixes: 36b9976e1f99 ("egl/wayland: Avoid race conditions when on non-main thread") Cc: [email protected]
* egl/wayland: Don't open-code roundtripDaniel Stone2017-05-191-25/+1
| | | | | | | | | | wl_display_roundtrip_queue() exists and can replace roundtrip(). The API was introduced with wayland 1.6, while we currently require 1.11. Signed-off-by: Daniel Stone <[email protected]> Reviewed-by: Emil Velikov <[email protected]> Reviewed-by: Lionel Landwerlin <[email protected]> Cc: [email protected]
* vulkan/wsi/wayland: Use proxy wrappers for swapchainDaniel Stone2017-05-191-16/+36
| | | | | | | | | | | | | | Though most swapchain operations used a queue, they were racy in that the object was created with the queue only set later, meaning that its event could potentially be dispatched from the default queue in between these two steps. Use proxy wrappers to avoid this race, also assigning wl_buffers created for the swapchain to the event queue. Signed-off-by: Daniel Stone <[email protected]> Reviewed-by: Lionel Landwerlin <[email protected]> Cc: [email protected]
* vulkan/wsi/wayland: Use per-display event queueDaniel Stone2017-05-191-12/+32
| | | | | | | | | | Calling random callbacks on the display's event queue is hostile, as we may call into client code when it least expects it. Create our own event queue, one per wsi_wl_display, and use that for the registry. Signed-off-by: Daniel Stone <[email protected]> Reviewed-by: Lionel Landwerlin <[email protected]> Cc: [email protected]
* vulkan/wsi/wayland: Remove roundtrip when creating imageDaniel Stone2017-05-191-1/+0
| | | | | | | | | | | | | There's no need to call wl_display_roundtrip() after trying to create a buffer through wl_drm; if it succeeds then everything is fine, and if it fails, then we get a fatal protocol error so can't recover anyway. Additionally, doing a roundtrip on the default / main application queue, is destructive anyway, so would need to be its own queue. Signed-off-by: Daniel Stone <[email protected]> Reviewed-by: Lionel Landwerlin <[email protected]> Cc: [email protected]
* vulkan: Fix Wayland uninitialised registryDaniel Stone2017-05-191-4/+5
| | | | | | | | | Untangle the exit cleanup paths so we don't try to use the registry variable before it's been initialised. Signed-off-by: Daniel Stone <[email protected]> Reviewed-by: Lionel Landwerlin <[email protected]> Cc: [email protected]
* i965/formats: Update the three-channel DXT1 mappingsNanley Chery2017-05-182-14/+5
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The procedure for decompressing an opaque DXT1 OpenGL format is dependant on the comparison of two colors stored in the first 32 bits of the compressed block. Here's the specified OpenGL behavior for reference: The RGB color for a texel at location (x,y) in the block is given by: RGB0, if color0 > color1 and code(x,y) == 0 RGB1, if color0 > color1 and code(x,y) == 1 (2*RGB0+RGB1)/3, if color0 > color1 and code(x,y) == 2 (RGB0+2*RGB1)/3, if color0 > color1 and code(x,y) == 3 RGB0, if color0 <= color1 and code(x,y) == 0 RGB1, if color0 <= color1 and code(x,y) == 1 (RGB0+RGB1)/2, if color0 <= color1 and code(x,y) == 2 BLACK, if color0 <= color1 and code(x,y) == 3 The sampling operation performed on an opaque DXT1 Intel format essentially hard-codes the comparison result of the two colors as color0 > color1. This means that the behavior is incompatible with OpenGL. This is stated in the SKL PRM, Vol 5: Memory Views: Opaque Textures (DXT1_RGB) Texture format DXT1_RGB is identical to DXT1, with the exception that the One-bit Alpha encoding is removed. Color 0 and Color 1 are not compared, and the resulting texel color is derived strictly from the Opaque Color Encoding. The alpha channel defaults to 1.0. Programming Note Context: Opaque Textures (DXT1_RGB) The behavior of this format is not compliant with the OGL spec. The opaque and non-opaque DXT1 OpenGL formats are specified to be decoded in exactly the same way except the BLACK value must have a transparent alpha channel in the latter. Use the four-channel BC1 Intel formats with the alpha set to 1 to provide the behavior required by the spec. Note that the alpha is already set to 1 for RGB formats in brw_get_texture_swizzle(). v2: Provide a more detailed commit message (Kenneth Graunke). v3: Ensure the alpha channel is set to 1 for DXT1 formats. Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100925 Cc: <[email protected]> Acked-by: Tapani Pälli <[email protected]> (v1) Reviewed-by: Kenneth Graunke <[email protected]> Signed-off-by: Nanley Chery <[email protected]>
* anv/formats: Update the three-channel BC1 mappingsNanley Chery2017-05-181-2/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The procedure for decompressing an opaque BC1 Vulkan format is dependant on the comparison of two colors stored in the first 32 bits of the compressed block. Here's the specified OpenGL (and Vulkan) behavior for reference: The RGB color for a texel at location (x,y) in the block is given by: RGB0, if color0 > color1 and code(x,y) == 0 RGB1, if color0 > color1 and code(x,y) == 1 (2*RGB0+RGB1)/3, if color0 > color1 and code(x,y) == 2 (RGB0+2*RGB1)/3, if color0 > color1 and code(x,y) == 3 RGB0, if color0 <= color1 and code(x,y) == 0 RGB1, if color0 <= color1 and code(x,y) == 1 (RGB0+RGB1)/2, if color0 <= color1 and code(x,y) == 2 BLACK, if color0 <= color1 and code(x,y) == 3 The sampling operation performed on an opaque DXT1 Intel format essentially hard-codes the comparison result of the two colors as color0 > color1. This means that the behavior is incompatible with OpenGL and Vulkan. This is stated in the SKL PRM, Vol 5: Memory Views: Opaque Textures (DXT1_RGB) Texture format DXT1_RGB is identical to DXT1, with the exception that the One-bit Alpha encoding is removed. Color 0 and Color 1 are not compared, and the resulting texel color is derived strictly from the Opaque Color Encoding. The alpha channel defaults to 1.0. Programming Note Context: Opaque Textures (DXT1_RGB) The behavior of this format is not compliant with the OGL spec. The opaque and non-opaque BC1 Vulkan formats are specified to be decoded in exactly the same way except the BLACK value must have a transparent alpha channel in the latter. Use the four-channel BC1 Intel formats with the alpha set to 1 to provide the behavior required by the spec. v2 (Kenneth Graunke): - Provide a more detailed commit message. Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100925 Cc: <[email protected]> Reviewed-by: Kenneth Graunke <[email protected]> Reviewed-by: Jason Ekstrand <[email protected]> Signed-off-by: Nanley Chery <[email protected]>
* anv: Add an option to abort on device lossJason Ekstrand2017-05-181-0/+5
| | | | | | | | | | | | | | | | This is mostly for running in our CI system to prevent dEQP from continuing on to the next test if we get a GPU hang. As it currently stands, dEQP uses the same VkDevice for almost all tests and if one of the tests hangs, we set the anv_device::device_lost flag and report VK_ERROR_DEVICE_LOST for all queue operations from that point forward without sending anything to the GPU. dEQP will happily continue trying to run tests and reporting failures until it eventually gets crash that forces the test runner to start over. This circumvents the problem by just aborting the process if we ever get a GPU hang. Since this is not the recommended behavior most of the time, we hide it behind an environment variable. Reviewed-by: Lionel Landwerlin <[email protected]>
* anv: Wrap the device lost error in vk_error in QueueSubmitJason Ekstrand2017-05-181-1/+1
| | | | | | | | We weren't wrapping this before because anv_cmd_buffer_execbuf may throw a more meaningful error message. However, we do change the error code into VK_ERROR_DEVICE_LOST, so we should print a new message. Reviewed-by: Lionel Landwerlin <[email protected]>
* radeonsi/gfx9: use CE RAM optimallyMarek Olšák2017-05-182-36/+134
| | | | | | | | | | | | | | On GFX9 with only 4K CE RAM, define the range of slots that will be allocated in CE RAM. All other slots will be uploaded directly. This will switch dynamically according to which slots are used by current shaders. GFX9 CE usage should now be similar to VI instead of being often disabled. Tested on VI by taking the GFX9 CE allocation codepath and setting num_ce_slots = 2 everywhere to get frequent switches between both modes. CE is still disabled on GFX9. Reviewed-by: Nicolai Hähnle <[email protected]>
* radeonsi: remove CE offset alignment restrictionMarek Olšák2017-05-181-2/+1
| | | | | | | This was only needed by LOAD_CONST_RAM, which is now only used to load whole CE. Reviewed-by: Nicolai Hähnle <[email protected]>
* radeonsi: only upload (dump to L2) those descriptors that are used by shadersMarek Olšák2017-05-184-24/+117
| | | | | | | This decreases the size of CE RAM dumps to L2, or the size of descriptor uploads without CE. Reviewed-by: Nicolai Hähnle <[email protected]>
* radeonsi: record which descriptor slots are used by shadersMarek Olšák2017-05-185-0/+41
| | | | Reviewed-by: Nicolai Hähnle <[email protected]>
* radeonsi: update si_ce_needed_cs_spaceMarek Olšák2017-05-181-8/+8
| | | | Reviewed-by: Nicolai Hähnle <[email protected]>
* radeonsi: do only 1 big CE dump at end of IBs and one reload in the preambleMarek Olšák2017-05-185-37/+37
| | | | | | | | A later commit will only upload descriptors used by shaders, so we won't do full dumps anymore, so the only way to have a complete mirror of CE RAM in memory is to do a separate dump after the last draw call. Reviewed-by: Nicolai Hähnle <[email protected]>
* radeonsi: remove early return in si_upload_descriptorsMarek Olšák2017-05-181-3/+0
| | | | | | | | All updates of descriptors_dirty also set dirty_mask, so the return is unnecessary. The next commit will want this function to be executed even if dirty_mask == 0. Reviewed-by: Nicolai Hähnle <[email protected]>
* radeonsi: clamp indirect index to the number of declared shader resourcesMarek Olšák2017-05-184-4/+15
| | | | | | | We'll do partial uploads of descriptor arrays, so we need to clamp against what shaders declare. Reviewed-by: Nicolai Hähnle <[email protected]>
* radeonsi: merge sampler and image descriptor lists into oneMarek Olšák2017-05-186-112/+99
| | | | | | | | | | | | Sampler slots: slot[8], .. slot[39] (ascending) Image slots: slot[7], .. slot[0] (descending) Each image occupies 1/2 of each slot, so there are 16 images in total, therefore the layout is: slot[15], .. slot[0]. (in 1/2 slot increments) Updating image slot 2n+i (i <= 1) also dirties and re-uploads slot 2n+!i. Reviewed-by: Nicolai Hähnle <[email protected]>
* radeonsi: merge constant and shader buffers descriptor lists into oneMarek Olšák2017-05-188-132/+152
| | | | | | | | | | Constant buffers: slot[16], .. slot[31] (ascending) Shader buffers: slot[15], .. slot[0] (descending) The idea is that if we have 4 constant buffers and 2 shader buffers, we only have to upload 6 slots. That optimization is left for a later commit. Reviewed-by: Nicolai Hähnle <[email protected]>
* gallium/u_threaded: add a fast path for unbinding shader buffersMarek Olšák2017-05-181-3/+9
| | | | | Reviewed-by: Samuel Pitoiset <[email protected]> Reviewed-by: Nicolai Hähnle <[email protected]>
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-29 11:30:02 +0000