Re: [Mesa-dev] [PATCH 2/4] radeonsi: use compute shaders for clear_buffer & copy_buffer
It looks like there is a deadlock in LLVM when 2 shaders are compiled at the same time. I don't know why it doesn't happen with OpenGL. Marek On Thu, Oct 18, 2018 at 3:32 AM Michel Dänzer wrote: > On 2018-10-17 6:43 p.m., Marek Olšák wrote: > > Can you test the attached patch? > > Doesn't help, unfortunately. Backtraces with the patch attached. > > FWIW, this is on Bonaire. > > > -- > Earthling Michel Dänzer | http://www.amd.com > Libre software enthusiast | Mesa and X developer > ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH 2/4] radeonsi: use compute shaders for clear_buffer & copy_buffer
On Thu, 2018-10-18 at 09:32 +0200, Michel Dänzer wrote: > On 2018-10-17 6:43 p.m., Marek Olšák wrote: > > Can you test the attached patch? > > Doesn't help, unfortunately. Backtraces with the patch attached. > > FWIW, this is on Bonaire. Hi, fwiw, I don't see this hang on my carrizo/iceland machine. Jan > > > ___ > mesa-dev mailing list > mesa-dev@lists.freedesktop.org > https://lists.freedesktop.org/mailman/listinfo/mesa-dev signature.asc Description: This is a digitally signed message part ___ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH 2/4] radeonsi: use compute shaders for clear_buffer & copy_buffer
On 2018-10-17 6:43 p.m., Marek Olšák wrote: > Can you test the attached patch? Doesn't help, unfortunately. Backtraces with the patch attached. FWIW, this is on Bonaire. -- Earthling Michel Dänzer | http://www.amd.com Libre software enthusiast | Mesa and X developer Thread 22 (Thread 0x7f2f6f7fe700 (LWP 7364)): #0 0x7f2fbcd03e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e398d1f2f0) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 #1 __pthread_cond_wait_common (abstime=0x0, mutex=0x55e398d1f2a0, cond=0x55e398d1f2c8) at pthread_cond_wait.c:502 #2 __pthread_cond_wait (cond=0x55e398d1f2c8, mutex=0x55e398d1f2a0) at pthread_cond_wait.c:655 #3 0x7f2fbb69dcfa in cnd_wait (mtx=0x55e398d1f2a0, cond=0x55e398d1f2c8) at ../../../include/c11/threads_posix.h:155 #4 util_queue_thread_func (input=input@entry=0x55e398d227f0) at ../../../src/util/u_queue.c:270 #5 0x7f2fbb69d8e8 in impl_thrd_routine (p=) at ../../../include/c11/threads_posix.h:87 #6 0x7f2fbccfdf2a in start_thread (arg=0x7f2f6f7fe700) at pthread_create.c:463 #7 0x7f2fc1d79edf in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95 Thread 21 (Thread 0x7f2f6700 (LWP 7363)): #0 0x7f2fbcd03e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e398d1f2f0) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 #1 __pthread_cond_wait_common (abstime=0x0, mutex=0x55e398d1f2a0, cond=0x55e398d1f2c8) at pthread_cond_wait.c:502 #2 __pthread_cond_wait (cond=0x55e398d1f2c8, mutex=0x55e398d1f2a0) at pthread_cond_wait.c:655 #3 0x7f2fbb69dcfa in cnd_wait (mtx=0x55e398d1f2a0, cond=0x55e398d1f2c8) at ../../../include/c11/threads_posix.h:155 #4 util_queue_thread_func (input=input@entry=0x55e398d22640) at ../../../src/util/u_queue.c:270 #5 0x7f2fbb69d8e8 in impl_thrd_routine (p=) at ../../../include/c11/threads_posix.h:87 #6 0x7f2fbccfdf2a in start_thread (arg=0x7f2f6700) at pthread_create.c:463 #7 0x7f2fc1d79edf in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95 Thread 20 (Thread 0x7f2f88ff9700 (LWP 7362)): #0 0x7f2fbcd03e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e398d1f2f0) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 #1 __pthread_cond_wait_common (abstime=0x0, mutex=0x55e398d1f2a0, cond=0x55e398d1f2c8) at pthread_cond_wait.c:502 #2 __pthread_cond_wait (cond=0x55e398d1f2c8, mutex=0x55e398d1f2a0) at pthread_cond_wait.c:655 #3 0x7f2fbb69dcfa in cnd_wait (mtx=0x55e398d1f2a0, cond=0x55e398d1f2c8) at ../../../include/c11/threads_posix.h:155 #4 util_queue_thread_func (input=input@entry=0x55e398d22490) at ../../../src/util/u_queue.c:270 #5 0x7f2fbb69d8e8 in impl_thrd_routine (p=) at ../../../include/c11/threads_posix.h:87 #6 0x7f2fbccfdf2a in start_thread (arg=0x7f2f88ff9700) at pthread_create.c:463 #7 0x7f2fc1d79edf in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95 Thread 19 (Thread 0x7f2f897fa700 (LWP 7361)): #0 0x7f2fbcd03e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e398d1f2f0) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 #1 __pthread_cond_wait_common (abstime=0x0, mutex=0x55e398d1f2a0, cond=0x55e398d1f2c8) at pthread_cond_wait.c:502 #2 __pthread_cond_wait (cond=0x55e398d1f2c8, mutex=0x55e398d1f2a0) at pthread_cond_wait.c:655 #3 0x7f2fbb69dcfa in cnd_wait (mtx=0x55e398d1f2a0, cond=0x55e398d1f2c8) at ../../../include/c11/threads_posix.h:155 #4 util_queue_thread_func (input=input@entry=0x55e398d222e0) at ../../../src/util/u_queue.c:270 #5 0x7f2fbb69d8e8 in impl_thrd_routine (p=) at ../../../include/c11/threads_posix.h:87 #6 0x7f2fbccfdf2a in start_thread (arg=0x7f2f897fa700) at pthread_create.c:463 #7 0x7f2fc1d79edf in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95 Thread 18 (Thread 0x7f2f89ffb700 (LWP 7360)): #0 0x7f2fbcd03e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e398d1f2f0) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 #1 __pthread_cond_wait_common (abstime=0x0, mutex=0x55e398d1f2a0, cond=0x55e398d1f2c8) at pthread_cond_wait.c:502 #2 __pthread_cond_wait (cond=0x55e398d1f2c8, mutex=0x55e398d1f2a0) at pthread_cond_wait.c:655 #3 0x7f2fbb69dcfa in cnd_wait (mtx=0x55e398d1f2a0, cond=0x55e398d1f2c8) at ../../../include/c11/threads_posix.h:155 #4 util_queue_thread_func (input=input@entry=0x55e398d238c0) at ../../../src/util/u_queue.c:270 #5 0x7f2fbb69d8e8 in impl_thrd_routine (p=) at ../../../include/c11/threads_posix.h:87 #6 0x7f2fbccfdf2a in start_thread (arg=0x7f2f89ffb700) at pthread_create.c:463 #7 0x7f2fc1d79edf in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95 Thread 17 (Thread 0x7f2f8a7fc700 (LWP 7359)): #0 0x7f2fbcd03e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e398d1ed70) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 #1
Re: [Mesa-dev] [PATCH 2/4] radeonsi: use compute shaders for clear_buffer & copy_buffer
Can you test the attached patch?
Marek
On Wed, Oct 17, 2018 at 9:31 AM Michel Dänzer wrote:
> On 2018-10-07 9:05 a.m., Marek Olšák wrote:
> > From: Marek Olšák
> >
> > Fast color clears should be much faster. Also, fast color clears on
> > evicted buffers should be 200x faster on GFX8 and older.
>
> Nice! Unfortunately, this broke clover with radeonsi. Everything using
> OpenCL seems to hang, see e.g. the attached backtraces from clinfo.
>
>
> --
> Earthling Michel Dänzer | http://www.amd.com
> Libre software enthusiast | Mesa and X developer
>
From f0978b2afae808edf4ac281b14cd371305a5164b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Marek=20Ol=C5=A1=C3=A1k?=
Date: Wed, 17 Oct 2018 12:41:38 -0400
Subject: [PATCH] radeonsi: fix a deadlock due to partially-initialized context
on CI
---
src/gallium/drivers/radeonsi/si_pipe.c | 14 --
1 file changed, 8 insertions(+), 6 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 59e41c53300..06740bd0f5c 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -575,12 +575,6 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen,
>null_const_buf);
si_set_rw_buffer(sctx, SI_PS_CONST_SAMPLE_POSITIONS,
>null_const_buf);
-
- /* Clear the NULL constant buffer, because loads should return zeros. */
- uint32_t clear_value = 0;
- si_clear_buffer(sctx, sctx->null_const_buf.buffer, 0,
-sctx->null_const_buf.buffer->width0,
-_value, 4, SI_COHERENCY_SHADER);
}
uint64_t max_threads_per_block;
@@ -625,6 +619,14 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen,
/* this must be last */
si_begin_new_gfx_cs(sctx);
+
+ if (sctx->chip_class == CIK) {
+ /* Clear the NULL constant buffer, because loads should return zeros. */
+ uint32_t clear_value = 0;
+ si_clear_buffer(sctx, sctx->null_const_buf.buffer, 0,
+sctx->null_const_buf.buffer->width0,
+_value, 4, SI_COHERENCY_SHADER);
+ }
return >b;
fail:
fprintf(stderr, "radeonsi: Failed to create a context.\n");
--
2.17.1
___
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
Re: [Mesa-dev] [PATCH 2/4] radeonsi: use compute shaders for clear_buffer & copy_buffer
On 2018-10-07 9:05 a.m., Marek Olšák wrote: > From: Marek Olšák > > Fast color clears should be much faster. Also, fast color clears on > evicted buffers should be 200x faster on GFX8 and older. Nice! Unfortunately, this broke clover with radeonsi. Everything using OpenCL seems to hang, see e.g. the attached backtraces from clinfo. -- Earthling Michel Dänzer | http://www.amd.com Libre software enthusiast | Mesa and X developer (gdb) info threads Id Target Id Frame * 1Thread 0x7f63ecdb2740 (LWP 24202) "clinfo" syscall () at ../sysdeps/unix/sysv/linux/x86_64/syscall.S:38 2Thread 0x7f63e62bc700 (LWP 24203) "clinfo:rcs0" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915203af0) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 3Thread 0x7f63e5934700 (LWP 24204) "clinfo:disk$0" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915204768) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 4Thread 0x7f63e510a700 (LWP 24205) "clinfo:cs0" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915214aa0) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 5Thread 0x7f63d7fff700 (LWP 24206) "clinfo:disk$0" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e9152185a8) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 6Thread 0x7f63d77fe700 (LWP 24207) "clinfo:sh0" __lll_lock_wait () at ../sysdeps/unix/sysv/linux/x86_64/lowlevellock.S:135 7Thread 0x7f63d6ffd700 (LWP 24208) "clinfo:sh1" __lll_lock_wait () at ../sysdeps/unix/sysv/linux/x86_64/lowlevellock.S:135 8Thread 0x7f63c700 (LWP 24209) "clinfo:sh2" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915217d00) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 9Thread 0x7f63d67fc700 (LWP 24210) "clinfo:sh3" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915217d00) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 10 Thread 0x7f63d5ffb700 (LWP 24211) "clinfo:sh4" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915217d00) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 11 Thread 0x7f63d57fa700 (LWP 24212) "clinfo:sh5" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915217d00) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 12 Thread 0x7f63d4ff9700 (LWP 24213) "clinfo:sh6" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915217d00) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 13 Thread 0x7f63cf7fe700 (LWP 24214) "clinfo:sh7" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915217d00) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 14 Thread 0x7f63ceffd700 (LWP 24215) "clinfo:sh8" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915217d00) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 15 Thread 0x7f63ce7fc700 (LWP 24216) "clinfo:sh9" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915217d00) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 16 Thread 0x7f63cdffb700 (LWP 24217) "clinfo:sh10" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915217d00) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 17 Thread 0x7f63cd7fa700 (LWP 24218) "clinfo:sh11" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915217d00) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 18 Thread 0x7f63ccff9700 (LWP 24219) "clinfo:shlo0" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915218280) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 19 Thread 0x7f639bfff700 (LWP 24220) "clinfo:shlo1" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915218280) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 20 Thread 0x7f639b7fe700 (LWP 24221) "clinfo:shlo2" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915218280) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 21 Thread 0x7f639affd700 (LWP 24222) "clinfo:shlo3" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915218280) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 22 Thread 0x7f639a7fc700 (LWP 24223) "clinfo:shlo4" 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915218280) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 (gdb) thread apply all bt Thread 22 (Thread 0x7f639a7fc700 (LWP 24223)): #0 0x7f63e7e36e6c in futex_wait_cancelable (private=, expected=0, futex_word=0x55e915218280) at ../sysdeps/unix/sysv/linux/futex-internal.h:88 #1 __pthread_cond_wait_common
[Mesa-dev] [PATCH 2/4] radeonsi: use compute shaders for clear_buffer & copy_buffer
From: Marek Olšák
Fast color clears should be much faster. Also, fast color clears on
evicted buffers should be 200x faster on GFX8 and older.
---
src/gallium/drivers/radeonsi/Makefile.sources | 1 +
src/gallium/drivers/radeonsi/meson.build | 1 +
src/gallium/drivers/radeonsi/si_clear.c | 10 +-
.../drivers/radeonsi/si_compute_blit.c| 285 ++
src/gallium/drivers/radeonsi/si_cp_dma.c | 180 +--
src/gallium/drivers/radeonsi/si_pipe.c| 22 +-
src/gallium/drivers/radeonsi/si_pipe.h| 51 ++--
src/gallium/drivers/radeonsi/si_test_dma.c| 3 +-
8 files changed, 350 insertions(+), 203 deletions(-)
create mode 100644 src/gallium/drivers/radeonsi/si_compute_blit.c
diff --git a/src/gallium/drivers/radeonsi/Makefile.sources
b/src/gallium/drivers/radeonsi/Makefile.sources
index abdc4e07f1e..aeb9b7982c4 100644
--- a/src/gallium/drivers/radeonsi/Makefile.sources
+++ b/src/gallium/drivers/radeonsi/Makefile.sources
@@ -4,20 +4,21 @@ GENERATED_SOURCES := \
C_SOURCES := \
$(GENERATED_SOURCES) \
cik_sdma.c \
driinfo_radeonsi.h \
si_blit.c \
si_buffer.c \
si_build_pm4.h \
si_clear.c \
si_compute.c \
si_compute.h \
+ si_compute_blit.c \
si_cp_dma.c \
si_debug.c \
si_descriptors.c \
si_dma.c \
si_dma_cs.c \
si_fence.c \
si_get.c \
si_gfx_cs.c \
si_gpu_load.c \
si_pipe.c \
diff --git a/src/gallium/drivers/radeonsi/meson.build
b/src/gallium/drivers/radeonsi/meson.build
index 4d6044f724b..2542f136d11 100644
--- a/src/gallium/drivers/radeonsi/meson.build
+++ b/src/gallium/drivers/radeonsi/meson.build
@@ -20,20 +20,21 @@
files_libradeonsi = files(
'cik_sdma.c',
'driinfo_radeonsi.h',
'si_blit.c',
'si_buffer.c',
'si_build_pm4.h',
'si_clear.c',
'si_compute.c',
'si_compute.h',
+ 'si_compute_blit.c',
'si_cp_dma.c',
'si_debug.c',
'si_descriptors.c',
'si_dma.c',
'si_dma_cs.c',
'si_fence.c',
'si_get.c',
'si_gfx_cs.c',
'si_gpu_load.c',
'si_perfcounter.c',
diff --git a/src/gallium/drivers/radeonsi/si_clear.c
b/src/gallium/drivers/radeonsi/si_clear.c
index 4e07de81bac..520e5b94f4a 100644
--- a/src/gallium/drivers/radeonsi/si_clear.c
+++ b/src/gallium/drivers/radeonsi/si_clear.c
@@ -249,21 +249,21 @@ void vi_dcc_clear_level(struct si_context *sctx,
* would be more efficient than separate per-layer clear
operations.
*/
assert(tex->buffer.b.b.nr_storage_samples <= 2 || num_layers ==
1);
dcc_offset += tex->surface.u.legacy.level[level].dcc_offset;
clear_size =
tex->surface.u.legacy.level[level].dcc_fast_clear_size *
num_layers;
}
si_clear_buffer(sctx, dcc_buffer, dcc_offset, clear_size,
- clear_value, SI_COHERENCY_CB_META);
+ _value, 4, SI_COHERENCY_CB_META);
}
/* Set the same micro tile mode as the destination of the last MSAA resolve.
* This allows hitting the MSAA resolve fast path, which requires that both
* src and dst micro tile modes match.
*/
static void si_set_optimal_micro_tile_mode(struct si_screen *sscreen,
struct si_texture *tex)
{
if (tex->buffer.b.is_shared ||
@@ -480,23 +480,24 @@ static void si_do_fast_color_clear(struct si_context
*sctx,
if (eliminate_needed && too_small)
continue;
/* DCC fast clear with MSAA should clear CMASK to 0xC.
*/
if (tex->buffer.b.b.nr_samples >= 2 &&
tex->cmask_buffer) {
/* TODO: This doesn't work with MSAA. */
if (eliminate_needed)
continue;
+ uint32_t clear_value = 0x;
si_clear_buffer(sctx, >cmask_buffer->b.b,
tex->cmask_offset,
tex->surface.cmask_size,
- 0x,
SI_COHERENCY_CB_META);
+ _value, 4,
SI_COHERENCY_CB_META);
need_decompress_pass = true;
}
vi_dcc_clear_level(sctx, tex, 0, reset_value);
if (eliminate_needed)
need_decompress_pass = true;
tex->separate_dcc_dirty = true;
} else {
@@ -511,23 +512,24 @@ static void si_do_fast_color_clear(struct si_context
*sctx,
/* RB+ doesn't work with CMASK fast clear on Stoney. */
if (sctx->family == CHIP_STONEY)
continue;
