mesa.git
13 years agoi915g: Lazy emit dynamic state
Jakob Bornecrantz [Tue, 22 Feb 2011 23:12:08 +0000 (23:12 +0000)]
i915g: Lazy emit dynamic state

13 years agoi915g: Lazy emit immediate state
Jakob Bornecrantz [Mon, 21 Feb 2011 23:39:10 +0000 (23:39 +0000)]
i915g: Lazy emit immediate state

13 years agoi915g: Disable LIS7 state updates for now
Jakob Bornecrantz [Tue, 22 Feb 2011 22:07:03 +0000 (22:07 +0000)]
i915g: Disable LIS7 state updates for now

13 years agoi915g: Clean up in i915_state_immediate
Jakob Bornecrantz [Mon, 21 Feb 2011 23:09:43 +0000 (23:09 +0000)]
i915g: Clean up in i915_state_immediate

13 years agoi915g: Remove outdated comment
Jakob Bornecrantz [Mon, 21 Feb 2011 22:47:40 +0000 (22:47 +0000)]
i915g: Remove outdated comment

13 years agoi915g: Use dump function in sw winsys
Jakob Bornecrantz [Tue, 22 Feb 2011 00:20:39 +0000 (00:20 +0000)]
i915g: Use dump function in sw winsys

13 years agoi915g: Enable mirror repeat wrap mode
Jakob Bornecrantz [Wed, 23 Feb 2011 00:11:09 +0000 (00:11 +0000)]
i915g: Enable mirror repeat wrap mode

13 years agoi915g: Always set vbo to flush on flushes
Jakob Bornecrantz [Tue, 22 Feb 2011 22:28:06 +0000 (22:28 +0000)]
i915g: Always set vbo to flush on flushes

Reported-by Chris Wilson <chris@chris-wilson.co.uk>

13 years agointel: gen3 is particular sensitive to batch size
Chris Wilson [Wed, 23 Feb 2011 23:09:36 +0000 (23:09 +0000)]
intel: gen3 is particular sensitive to batch size

... and prefers a small batch whereas gen4+ prefer a large batch to
carry more state.

Tuning using openarena/padman indicate that a batch size of just 4096 is
best for those cases.

Bugzilla: https://bugs.freedesktop.org/process_bug.cgi
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi915: And remember assign the new value to the state reg...
Chris Wilson [Wed, 23 Feb 2011 22:09:12 +0000 (22:09 +0000)]
i915: And remember assign the new value to the state reg...

Fixes regression from 298ebb78de8a6b6edf0aa0fe8d784d00bbc2930e.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=34589
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoFix GLX_USE_TLS define.
Tom Fogal [Tue, 22 Feb 2011 05:32:18 +0000 (22:32 -0700)]
Fix GLX_USE_TLS define.

It was only getting set in the case of DRI drivers.

13 years agor600g: Request DWORD aligned vertex buffers.
Fabian Bieler [Mon, 14 Feb 2011 21:44:42 +0000 (22:44 +0100)]
r600g: Request DWORD aligned vertex buffers.

The spec says that the offsets in the vertex-fetch instructions need to be byte-aligned and makes no specification with regard to the required alignment of the offset and stride in the vertex resource constant register.

However, testing indicates that all three values need to be DWORD aligned.

13 years agost/mesa: fix computing the lowest address for interleaved attribs
Wiktor Janas [Wed, 23 Feb 2011 06:10:12 +0000 (07:10 +0100)]
st/mesa: fix computing the lowest address for interleaved attribs

Ptr can be very well NULL, so when there are two arrays, with one having
offset 0 (and thus NULL Ptr), and the other having a non-zero offset,
the non-zero value is taken as minimum (because of !low_addr ? start ...).
On 32-bit systems, this somehow works. On 64-bit systems, it leads to crashes.

Signed-off-by: Marek Olšák <maraeo@gmail.com>
13 years agovbo: added vbo_check_buffers_are_unmapped() debug function
Brian Paul [Tue, 22 Feb 2011 21:31:15 +0000 (14:31 -0700)]
vbo: added vbo_check_buffers_are_unmapped() debug function

13 years agovbo: removed unused #defines, add comments
Brian Paul [Tue, 22 Feb 2011 21:23:50 +0000 (14:23 -0700)]
vbo: removed unused #defines, add comments

13 years agomesa: move comment, change debug code
Brian Paul [Tue, 22 Feb 2011 20:37:30 +0000 (13:37 -0700)]
mesa: move comment, change debug code

13 years agovbo: simplify NeedFlush flag clearing
Brian Paul [Tue, 22 Feb 2011 20:31:09 +0000 (13:31 -0700)]
vbo: simplify NeedFlush flag clearing

13 years agovbo: use ctx intstead of exec->ctx
Brian Paul [Tue, 22 Feb 2011 20:24:56 +0000 (13:24 -0700)]
vbo: use ctx intstead of exec->ctx

13 years agor300g: fix missing initializers warning
Brian Paul [Tue, 22 Feb 2011 19:44:42 +0000 (12:44 -0700)]
r300g: fix missing initializers warning

13 years agoi915g: remove extra semicolons
Brian Paul [Tue, 22 Feb 2011 19:44:10 +0000 (12:44 -0700)]
i915g: remove extra semicolons

13 years agoxlib: pass Display pointer to XMesaGarbageCollect()
Andy Skinner [Fri, 11 Feb 2011 15:31:25 +0000 (07:31 -0800)]
xlib: pass Display pointer to XMesaGarbageCollect()

Fixes an issue when different displays are used on different threads.

Signed-off-by: Brian Paul <brianp@vmware.com>
13 years agoi965: Increase Sandybridge point size clamp.
Kenneth Graunke [Sun, 20 Feb 2011 01:22:47 +0000 (17:22 -0800)]
i965: Increase Sandybridge point size clamp.

255.875 matches the hardware documentation.  Presumably this was a typo.

Found by inspection.  Not known to fix any issues.

Reviewed-by: Eric Anholt <eric@anholt.net>
13 years agoi965/fs: Correctly set up gl_FragCoord.w on Sandybridge.
Kenneth Graunke [Sun, 20 Feb 2011 00:48:24 +0000 (16:48 -0800)]
i965/fs: Correctly set up gl_FragCoord.w on Sandybridge.

pixel_w is the final result; wpos_w is used on gen4 to compute it.

NOTE: This is a candidate for the 7.10 branch.

Reviewed-by: Eric Anholt <eric@anholt.net>
13 years agoi965/fs: Refactor control flow stack handling.
Kenneth Graunke [Sun, 20 Feb 2011 00:12:28 +0000 (16:12 -0800)]
i965/fs: Refactor control flow stack handling.

We can't safely use fixed size arrays since Gen6+ supports unlimited
nesting of control flow.

NOTE: This is a candidate for the 7.10 branch.

Reviewed-by: Eric Anholt <eric@anholt.net>
13 years agoi965/fs: Avoid register coalescing away gen6 MATH workarounds.
Kenneth Graunke [Sat, 19 Feb 2011 09:05:11 +0000 (01:05 -0800)]
i965/fs: Avoid register coalescing away gen6 MATH workarounds.

The code that generates MATH instructions attempts to work around
the hardware ignoring source modifiers (abs and negate) by emitting
moves into temporaries.  Unfortunately, this pass coalesced those
registers, restoring the original problem.  Avoid doing that.

Fixes several OpenGL ES2 conformance failures on Sandybridge.

NOTE: This is a candidate for the 7.10 branch.

Reviewed-by: Eric Anholt <eric@anholt.net>
13 years agoi965/fs: Apply source modifier workarounds to POW as well.
Kenneth Graunke [Sat, 19 Feb 2011 09:03:08 +0000 (01:03 -0800)]
i965/fs: Apply source modifier workarounds to POW as well.

Single-operand math already had these workarounds, but POW (the only two
operand function) did not.  It needs them too - otherwise we can hit
assertion failures in brw_eu_emit.c when code is actually generated.

NOTE: This is a candidate for the 7.10 branch.

Reviewed-by: Eric Anholt <eric@anholt.net>
13 years agoi965: Fix shaders that write to gl_PointSize on Sandybridge.
Kenneth Graunke [Tue, 22 Feb 2011 18:04:18 +0000 (10:04 -0800)]
i965: Fix shaders that write to gl_PointSize on Sandybridge.

gl_PointSize (VERT_RESULT_PSIZ) doesn't take up a message register,
as it's part of the header.  Without this fix, writing to gl_PointSize
would cause the SF to read and use the wrong attributes, leading to all
kinds of random looking failure.

Reviewed-by: Eric Anholt <eric@anholt.net>
13 years agomesa: Avoid undeclared ffs function warning on mingw.
José Fonseca [Tue, 22 Feb 2011 14:59:09 +0000 (14:59 +0000)]
mesa: Avoid undeclared ffs function warning on mingw.

13 years agogallium: s/PIPE_TRANSFER_CPU_READ/PIPE_TRANSFER_READ/ in comments.
José Fonseca [Tue, 22 Feb 2011 14:14:45 +0000 (14:14 +0000)]
gallium: s/PIPE_TRANSFER_CPU_READ/PIPE_TRANSFER_READ/ in comments.

13 years agogallium/docs: Update PIPE_TRANSFER_xx docs. Reformat to use definitions.
José Fonseca [Tue, 22 Feb 2011 14:14:22 +0000 (14:14 +0000)]
gallium/docs: Update PIPE_TRANSFER_xx docs. Reformat to use definitions.

13 years agogallium: new transfer flag: DISCARD_WHOLE_RESOURCE
Keith Whitwell [Wed, 5 Jan 2011 17:33:43 +0000 (17:33 +0000)]
gallium: new transfer flag: DISCARD_WHOLE_RESOURCE

13 years agost/mesa: fix crash when using both user and vbo buffers with the same stride
Marek Olšák [Sun, 20 Feb 2011 17:05:24 +0000 (18:05 +0100)]
st/mesa: fix crash when using both user and vbo buffers with the same stride

If two buffers had the same stride where one buffer is a user one and
the other is a vbo, it was considered to be one interleaved buffer,
resulting in incorrect rendering and crashes.

This patch makes sure that the interleaved buffer is either user or vbo,
not both.

13 years agost/mesa: fix crash when DrawBuffer->_ColorDrawBuffers[0] is NULL
Marek Olšák [Sun, 20 Feb 2011 15:50:48 +0000 (16:50 +0100)]
st/mesa: fix crash when DrawBuffer->_ColorDrawBuffers[0] is NULL

This fixes the game Tiny and Big.

13 years agoi965: Trim the interleaved upload to the minimum number of vertices
Chris Wilson [Tue, 22 Feb 2011 11:17:39 +0000 (11:17 +0000)]
i965: Trim the interleaved upload to the minimum number of vertices

... should have no impact on a properly formatted draw operation.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: Reinstate max-index paranoia
Chris Wilson [Tue, 22 Feb 2011 11:19:32 +0000 (11:19 +0000)]
i965: Reinstate max-index paranoia

Don't trust the applications not to reference beyond the end of the
vertex buffers.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: Zero the offset into the vbo when uploading non-interleaved
Chris Wilson [Tue, 22 Feb 2011 11:18:25 +0000 (11:18 +0000)]
i965: Zero the offset into the vbo when uploading non-interleaved

Fixes regression from 559435d9152acc7162e4e60aae6591c7c6c8274b.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agost/dri: Track drawable context bindings
Jakob Bornecrantz [Wed, 1 Dec 2010 04:04:25 +0000 (05:04 +0100)]
st/dri: Track drawable context bindings

Needs to track this ourself since because we get into a race condition with
the dri_util.c code on make current when rendering to the front buffer.

This is what happens:
Old context is rendering to the front buffer.

App calls MakeCurrent with a new context. dri_util.c sets
drawable->driContextPriv to the new context and then calls the driver make
current. st/dri make current flushes the old context, which calls back into
st/dri via the flush frontbuffer hook. st/dri calls dri loader flush
frontbuffer, which calls invalidate buffer on the drawable into st/dri.

This is where things gets wrong. st/dri grabs the context from the dri
drawable (which now points to the new context) and calls invalidate
framebuffer to the new context which has not yet set the new drawable as its
framebuffers since we have not called make current yet, it asserts.

13 years agoi965: Fix VB packet reuse when offset for the new buffer isn't stride aligned.
Eric Anholt [Tue, 22 Feb 2011 00:24:41 +0000 (16:24 -0800)]
i965: Fix VB packet reuse when offset for the new buffer isn't stride aligned.

Fixes regression in scissor-stencil-clear and 5 other tests.

13 years agoRevert "mesa: convert macros to inline functions"
Brian Paul [Tue, 22 Feb 2011 00:01:00 +0000 (17:01 -0700)]
Revert "mesa: convert macros to inline functions"

This reverts commit e9ff76aa81d9bd973d46b7e46f1e4ece2112a5b7.

Need to use macros so __FUNCTION__ reports the caller.

13 years agost/mesa: need to translate clear color according to surface's base format
Brian Paul [Mon, 21 Feb 2011 23:54:23 +0000 (16:54 -0700)]
st/mesa: need to translate clear color according to surface's base format

When clearing a GL_LUMINANCE_ALPHA buffer, for example, we need to convert
the clear color (R,G,B,A) to (R,R,R,A).  We were doing this for texture border
colors but not renderbuffers.  Move the translation function to st_format.c
and share it.

This fixes the piglit fbo-clear-formats test.

NOTE: This is a candidate for the 7.9 and 7.10 branches.

13 years agost/mesa: fix the default case in st_format_datatype()
Brian Paul [Mon, 21 Feb 2011 23:46:02 +0000 (16:46 -0700)]
st/mesa: fix the default case in st_format_datatype()

Part of the fix for piglit fbo-clear-formats

NOTE: This is a candidate for the 7.9 and 7.10 branches.

13 years agoi915g: add some throttling
Daniel Vetter [Mon, 21 Feb 2011 18:14:02 +0000 (19:14 +0100)]
i915g: add some throttling

Intel classic drivers switched to this, too, so it must be good.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
13 years agoi915g: s/bool/boolean/ style-fixup in winsys
Daniel Vetter [Mon, 21 Feb 2011 17:25:20 +0000 (18:25 +0100)]
i915g: s/bool/boolean/ style-fixup in winsys

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
13 years agoi915g: Fix warning
Jakob Bornecrantz [Mon, 21 Feb 2011 22:00:02 +0000 (22:00 +0000)]
i915g: Fix warning

13 years agoi915g: Add option to lie about caps
Jakob Bornecrantz [Sun, 20 Feb 2011 12:41:18 +0000 (13:41 +0100)]
i915g: Add option to lie about caps

13 years agoi915g: Move debug fields to screen
Jakob Bornecrantz [Sun, 20 Feb 2011 11:52:55 +0000 (12:52 +0100)]
i915g: Move debug fields to screen

13 years agoi915g: Use debug get once options
Jakob Bornecrantz [Sun, 20 Feb 2011 10:41:32 +0000 (11:41 +0100)]
i915g: Use debug get once options

13 years agoi915g: Rework texture tiling a bit
Jakob Bornecrantz [Sun, 20 Feb 2011 11:52:11 +0000 (12:52 +0100)]
i915g: Rework texture tiling a bit

13 years agoi915g: Anisotropic filtering works
Jakob Bornecrantz [Mon, 21 Feb 2011 21:27:05 +0000 (21:27 +0000)]
i915g: Anisotropic filtering works

13 years agoi915g: TODO about point sprites
Jakob Bornecrantz [Sun, 20 Feb 2011 13:00:03 +0000 (14:00 +0100)]
i915g: TODO about point sprites

13 years agoi915g: TODO about untested code hidden behind caps
Jakob Bornecrantz [Sun, 20 Feb 2011 12:58:11 +0000 (13:58 +0100)]
i915g: TODO about untested code hidden behind caps

Should be fairly easy to test and fix since you can look at
the code in the classic driver.

13 years agoi915g: Reorg caps
Jakob Bornecrantz [Sun, 20 Feb 2011 10:45:48 +0000 (11:45 +0100)]
i915g: Reorg caps

13 years agost/mesa: fix incorrect texture size allocation in st_finalize_texture()
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
st/mesa: fix incorrect texture size allocation in st_finalize_texture()

If finalizing a non-POW mipmapped texture with an odd-sized base texture
image we were allocating the wrong size of gallium texture (off by one).
Need to be more careful about computing the base texture image size.

This fixes https://bugs.freedesktop.org/show_bug.cgi?id=34463

13 years agost/mesa: refactor guess_and_alloc_texture() code
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
st/mesa: refactor guess_and_alloc_texture() code

13 years agost/mesa: fix mipmap generation for non-POW textures
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
st/mesa: fix mipmap generation for non-POW textures

This is part of the fix for https://bugs.freedesktop.org/show_bug.cgi?id=34463

13 years agomesa: convert macros to inline functions
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
mesa: convert macros to inline functions

13 years agovbo: more comments
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
vbo: more comments

13 years agovbo: make vbo_exec_FlushVertices_internal() static
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
vbo: make vbo_exec_FlushVertices_internal() static

13 years agovbo: remove old debug code, add comments
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
vbo: remove old debug code, add comments

13 years agovbo: rename, document function params
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
vbo: rename, document function params

13 years agovbo: comments
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
vbo: comments

13 years agovbo: replace assert(0) with proper assertions
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
vbo: replace assert(0) with proper assertions

13 years agovbo: rename some vars, add new comments, fix formatting, etc.
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
vbo: rename some vars, add new comments, fix formatting, etc.

13 years agovbo: use ctx instead of exec->ctx
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
vbo: use ctx instead of exec->ctx

13 years agoradeon: add default switch case to silence unhandled enum warning
Brian Paul [Mon, 21 Feb 2011 22:11:44 +0000 (15:11 -0700)]
radeon: add default switch case to silence unhandled enum warning

13 years agoUse C-style system headers in C++ code to avoid issues with std:: namespace
Ian Romanick [Thu, 10 Feb 2011 18:26:42 +0000 (10:26 -0800)]
Use C-style system headers in C++ code to avoid issues with std:: namespace

13 years agointel: Fix insufficient integer width for upload buffer offset
Chris Wilson [Mon, 21 Feb 2011 20:56:06 +0000 (20:56 +0000)]
intel: Fix insufficient integer width for upload buffer offset

I was being overly miserly and gave the offset of the buffer into the bo
insufficient bits, distracted by the adjacency of the buffer[4096].

Ref: https://bugs.freedesktop.org/show_bug.cgi?id=34541
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agosvga: Remove some remaining fake S3TC rendering support.
José Fonseca [Mon, 21 Feb 2011 18:24:36 +0000 (18:24 +0000)]
svga: Remove some remaining fake S3TC rendering support.

13 years agoi965: Remove spurious duplicate ADVANCE_BATCH
Chris Wilson [Mon, 21 Feb 2011 16:02:26 +0000 (16:02 +0000)]
i965: Remove spurious duplicate ADVANCE_BATCH

... a leftover from a bad merge.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi915: Emit a single relocation per vbo
Chris Wilson [Thu, 25 Nov 2010 15:41:37 +0000 (15:41 +0000)]
i915: Emit a single relocation per vbo

Reducing the number of relocations has lots of nice knock-on effects,
not least including reducing batch buffer size, auxilliary array sizes
(vmalloced and copied into the kernel), processing of uncached
relocations etc.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi915: Suppress emission of redundant stencil updates
Chris Wilson [Fri, 26 Nov 2010 11:18:50 +0000 (11:18 +0000)]
i915: Suppress emission of redundant stencil updates

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi915: Separate BLEND from general context state.
Chris Wilson [Fri, 26 Nov 2010 10:57:06 +0000 (10:57 +0000)]
i915: Separate BLEND from general context state.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi915: Only flag context changes if the actual state is changed
Chris Wilson [Fri, 26 Nov 2010 10:25:23 +0000 (10:25 +0000)]
i915: Only flag context changes if the actual state is changed

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi915: suppress repeated sampler state emission
Chris Wilson [Thu, 25 Nov 2010 22:27:37 +0000 (22:27 +0000)]
i915: suppress repeated sampler state emission

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi915: Eliminate redundant CONSTANTS updates
Chris Wilson [Thu, 25 Nov 2010 21:39:21 +0000 (21:39 +0000)]
i915: Eliminate redundant CONSTANTS updates

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: Use compiler builtins when available
Chris Wilson [Tue, 8 Feb 2011 22:58:35 +0000 (22:58 +0000)]
i965: Use compiler builtins when available

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: Micro-optimise check_state
Chris Wilson [Sun, 20 Feb 2011 15:36:52 +0000 (15:36 +0000)]
i965: Micro-optimise check_state

Replace the intermediate tests due to the logical or with the bitwise
or.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agointel: use throttle ioctl for throttling
Chris Wilson [Thu, 30 Dec 2010 21:47:39 +0000 (21:47 +0000)]
intel: use throttle ioctl for throttling

Rather than waiting on the first batch after the last swapbuffers to be
retired, call into the kernel to wait upon the retirement of any request
less than 20ms old. This has the twofold advantage of (a) not blocking
any other clients from utilizing the device whilst we wait and (b) we
attain higher throughput without overloading the system.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: Remove unused 'next_free_page' member
Chris Wilson [Sat, 12 Feb 2011 11:28:25 +0000 (11:28 +0000)]
i965: Remove unused 'next_free_page' member

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agointel: Skip the flush before read-pixels via blit
Chris Wilson [Fri, 11 Feb 2011 00:03:48 +0000 (00:03 +0000)]
intel: Skip the flush before read-pixels via blit

As we will flush when reading the return values of the blit, we can forgo
the earlier flush.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agointel: extend current vertex buffers
Chris Wilson [Fri, 18 Feb 2011 10:37:43 +0000 (10:37 +0000)]
intel: extend current vertex buffers

If the next vertex arrays are a (discontiguous) continuation of the
current arrays, such that the new vertices are simply offset from the
start of the current vertex buffer definitions we can reuse those
defintions and avoid the overhead of relocations and invalidations.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agointel: Use specified alignment for writes into the upload buffer
Chris Wilson [Fri, 18 Feb 2011 12:30:37 +0000 (12:30 +0000)]
intel: Use specified alignment for writes into the upload buffer

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: Clean up brw_prepare_vertices()
Chris Wilson [Fri, 11 Feb 2011 15:29:26 +0000 (15:29 +0000)]
i965: Clean up brw_prepare_vertices()

Use a temporary glarray variable to replace the numerous input->glarray.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agointel: combine short memcpy using a temporary allocated buffer
Chris Wilson [Fri, 11 Feb 2011 19:40:08 +0000 (19:40 +0000)]
intel: combine short memcpy using a temporary allocated buffer

Using a temporary buffer for large discontiguous uploads into the common
buffer and a single buffered upload is faster than performing the
discontiguous copies through a mapping into the GTT.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: upload normal arrays as interleaved
Chris Wilson [Fri, 11 Feb 2011 14:14:18 +0000 (14:14 +0000)]
i965: upload normal arrays as interleaved

Upload the non-vbo arrays into a single interleaved buffer object, and
so need to just emit a single vertex buffer relocation.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: interleaved vbo
Chris Wilson [Fri, 11 Feb 2011 14:45:19 +0000 (14:45 +0000)]
i965: interleaved vbo

If the user passed in several arrays interleaved in the same vbo, only
emit a single vertex buffer and relocation.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: emit one vb packet per vbo
Chris Wilson [Fri, 11 Feb 2011 14:45:10 +0000 (14:45 +0000)]
i965: emit one vb packet per vbo

Track reuse of the vertex buffer objects and so minimise the number of
vertex buffers used by the hardware (and their relocations).

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: upload transient indices into the same discontiguous buffer
Chris Wilson [Fri, 11 Feb 2011 00:18:21 +0000 (00:18 +0000)]
i965: upload transient indices into the same discontiguous buffer

As we now pack the indices into a common upload buffer, we can reuse a
single CMD_INDEX_BUFFER packet and translate each invocation with a
start vertex offset.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: suppress repeat-emission of identical vertex elements
Chris Wilson [Sun, 20 Feb 2011 13:37:00 +0000 (13:37 +0000)]
i965: suppress repeat-emission of identical vertex elements

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: Move repeat-instruction-suppression to batchbuffer core
Chris Wilson [Sun, 20 Feb 2011 13:23:47 +0000 (13:23 +0000)]
i965: Move repeat-instruction-suppression to batchbuffer core

Move the tracking of the last emitted instructions into the core
batchbuffer routines and take advantage of the shadow batch copy to
avoid extra memory allocations and copies.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agointel: use pwrite for batch
Chris Wilson [Thu, 10 Feb 2011 20:25:51 +0000 (20:25 +0000)]
intel: use pwrite for batch

It's faster. Not only is the memcpy more efficiently performed in the
kernel (making up for the system call overhead), but by not using mmap
we remove the greater overhead of tracking the vma of every batch.

And it means we can read back from the batch buffer without incurring
the cost of a uncached read through the GTT.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: drop state_bo references to batch_bo
Chris Wilson [Thu, 10 Feb 2011 18:31:13 +0000 (18:31 +0000)]
i965: drop state_bo references to batch_bo

As we use state relocations and we know that all the state belongs to
the same bo, we can drop the multiple references to the same bo.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: directly write wm state to batch
Chris Wilson [Thu, 10 Feb 2011 18:14:40 +0000 (18:14 +0000)]
i965: directly write wm state to batch

As we write directly into the batch in system memory, we do not need to
write first to the stack (as was to avoid read back through the GTT)

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: write cc straight to batch
Chris Wilson [Thu, 10 Feb 2011 18:11:58 +0000 (18:11 +0000)]
i965: write cc straight to batch

As we write directly into the batch in system memory, we do not need to
write first to the stack (as was to avoid read back through the GTT)

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: switch gen6 to use its own cc state bo
Chris Wilson [Thu, 10 Feb 2011 18:10:06 +0000 (18:10 +0000)]
i965: switch gen6 to use its own cc state bo

In preparation for a greater change, use the color_calc_state_bo already
provisioned for this purpose.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agointel: Buffered upload
Chris Wilson [Thu, 10 Feb 2011 00:25:17 +0000 (00:25 +0000)]
intel: Buffered upload

Rather than performing lots of little writes to update the common bo
upon each update, write those into a static buffer and flush that when
full (or at the end of the batch). Doing so gives a dramatic performance
improvement over and above using mmaped access.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agointel: Replace the bo for a complete update
Chris Wilson [Wed, 9 Feb 2011 10:00:09 +0000 (10:00 +0000)]
intel: Replace the bo for a complete update

Rather than performing a blit to completely overwrite a busy bo, simply
discard it and create a new one with the fresh data.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agoi965: Combine vb upload buffer with the general upload buffer
Chris Wilson [Tue, 8 Feb 2011 22:22:36 +0000 (22:22 +0000)]
i965: Combine vb upload buffer with the general upload buffer

Reuse the new common upload buffer for uploading temporary indices and
rebuilt vertex arrays.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agointel: Pack dynamic draws together
Chris Wilson [Tue, 8 Feb 2011 20:01:10 +0000 (20:01 +0000)]
intel: Pack dynamic draws together

Dynamic arrays have the tendency to be small and so allocating a bo for
each one is overkill and we can exploit many efficiency gains by packing
them together.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
13 years agointel: Use system memory for DYNAMIC_DRAW source objects
Chris Wilson [Tue, 8 Feb 2011 16:57:26 +0000 (16:57 +0000)]
intel: Use system memory for DYNAMIC_DRAW source objects

Dynamic draw buffers are used by clients for temporary arrays and for
uploading normal vertex arrays. By keeping the data in memory, we can
avoid reusing active buffer objects and reallocate them as they are
changed. This is important for Sandybridge which can not issue blits
within a batch and so ends up flushing the batch upon every update, that
is each batch only contains a single draw operation (if using dynamic
arrays or regular arrays from system memory).

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>