On Wed, Nov 16, 2016 at 04:07:33PM +0200, Abdiel Janulgue wrote:
>
>
> On 16.11.2016 15:56, Chris Wilson wrote:
> > On Wed, Nov 16, 2016 at 11:18:01PM +0200, Abdiel Janulgue wrote:
> >> A lot of igt testcases need some GPU workload to make sure a race
> >> window is big enough. Unfortunately having a fixed amount of
> >> workload leads to spurious test failures or overtly long runtimes
> >> on some fast/slow platforms. This library contains functionality
> >> to submit GPU workloads that should consume exactly a specific
> >> amount of time.
> >>
> >> v2 : Add recursive batch feature from Chris
> >> v3 : Drop auto-tuned stuff. Add bo dependecy to recursive batch
> >> by adding a dummy reloc to the bo as suggested by Ville.
> >> v4: Fix dependency reloc as write instead of read (Ville).
> >> Fix wrong handling of batchbuffer start on ILK causing
> >> test failure
> >> v5: Convert kms_busy to use this api
> >> v6: Add this library to docs
> >> v7: Document global use of batch, reuse defines
> >> Minor code cleanups.
> >> Rename igt_spin_batch and igt_post_spin_batch to
> >> igt_spin_batch_new and igt_spin_batch_free
> >> respectively (Tomeu Vizoso).
> >> Fix error in dependency relocation handling in HSW causing
> >> tests to fail.
> >>
> >> Cc: Daniel Vetter <daniel.vetter@xxxxxxxx>
> >> Cc: Ville Syrjälä <ville.syrjala@xxxxxxxxxxxxxxx>
> >> Cc: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> >> Cc: tomeu@xxxxxxxxxxxxxxx
> >> Signed-off-by: Abdiel Janulgue <abdiel.janulgue@xxxxxxxxxxxxxxx>
> >> ---
> >> .../intel-gpu-tools/intel-gpu-tools-docs.xml | 1 +
> >> lib/Makefile.sources | 2 +
> >> lib/igt.h | 1 +
> >> lib/igt_dummyload.c | 281 +++++++++++++++++++++
> >> lib/igt_dummyload.h | 43 ++++
> >> 5 files changed, 328 insertions(+)
> >> create mode 100644 lib/igt_dummyload.c
> >> create mode 100644 lib/igt_dummyload.h
> >>
> >> diff --git a/docs/reference/intel-gpu-tools/intel-gpu-tools-docs.xml b/docs/reference/intel-gpu-tools/intel-gpu-tools-docs.xml
> >> index c862f2a..55902ab 100644
> >> --- a/docs/reference/intel-gpu-tools/intel-gpu-tools-docs.xml
> >> +++ b/docs/reference/intel-gpu-tools/intel-gpu-tools-docs.xml
> >> @@ -32,6 +32,7 @@
> >> <xi:include href="xml/intel_io.xml"/>
> >> <xi:include href="xml/igt_vc4.xml"/>
> >> <xi:include href="xml/igt_vgem.xml"/>
> >> + <xi:include href="xml/igt_dummyload.xml"/>
> >> </chapter>
> >> <xi:include href="xml/igt_test_programs.xml"/>
> >>
> >> diff --git a/lib/Makefile.sources b/lib/Makefile.sources
> >> index e8e277b..7fc5ec2 100644
> >> --- a/lib/Makefile.sources
> >> +++ b/lib/Makefile.sources
> >> @@ -75,6 +75,8 @@ lib_source_list = \
> >> igt_draw.h \
> >> igt_pm.c \
> >> igt_pm.h \
> >> + igt_dummyload.c \
> >> + igt_dummyload.h \
> >> uwildmat/uwildmat.h \
> >> uwildmat/uwildmat.c \
> >> $(NULL)
> >> diff --git a/lib/igt.h b/lib/igt.h
> >> index d751f24..a0028d5 100644
> >> --- a/lib/igt.h
> >> +++ b/lib/igt.h
> >> @@ -32,6 +32,7 @@
> >> #include "igt_core.h"
> >> #include "igt_debugfs.h"
> >> #include "igt_draw.h"
> >> +#include "igt_dummyload.h"
> >> #include "igt_fb.h"
> >> #include "igt_gt.h"
> >> #include "igt_kms.h"
> >> diff --git a/lib/igt_dummyload.c b/lib/igt_dummyload.c
> >> new file mode 100644
> >> index 0000000..d266195
> >> --- /dev/null
> >> +++ b/lib/igt_dummyload.c
> >> @@ -0,0 +1,281 @@
> >> +/*
> >> + * Copyright © 2016 Intel Corporation
> >> + *
> >> + * Permission is hereby granted, free of charge, to any person obtaining a
> >> + * copy of this software and associated documentation files (the "Software"),
> >> + * to deal in the Software without restriction, including without limitation
> >> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> >> + * and/or sell copies of the Software, and to permit persons to whom the
> >> + * Software is furnished to do so, subject to the following conditions:
> >> + *
> >> + * The above copyright notice and this permission notice (including the next
> >> + * paragraph) shall be included in all copies or substantial portions of the
> >> + * Software.
> >> + *
> >> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> >> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> >> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> >> + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
> >> + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
> >> + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
> >> + * IN THE SOFTWARE.
> >> + *
> >> + */
> >> +
> >> +#include "igt.h"
> >> +#include "igt_dummyload.h"
> >> +#include <time.h>
> >> +#include <signal.h>
> >> +#include <sys/syscall.h>
> >> +
> >> +/**
> >> + * SECTION:igt_dummyload
> >> + * @short_description: Library for submitting GPU workloads
> >> + * @title: Dummyload
> >> + * @include: igt.h
> >> + *
> >> + * A lot of igt testcases need some GPU workload to make sure a race window is
> >> + * big enough. Unfortunately having a fixed amount of workload leads to
> >> + * spurious test failures or overly long runtimes on some fast/slow platforms.
> >> + * This library contains functionality to submit GPU workloads that should
> >> + * consume exactly a specific amount of time.
> >> + */
> >> +
> >> +#define LOCAL_I915_EXEC_BSD_SHIFT (13)
> >> +#define LOCAL_I915_EXEC_BSD_MASK (3 << LOCAL_I915_EXEC_BSD_SHIFT)
> >> +
> >> +#define ENGINE_MASK (I915_EXEC_RING_MASK | LOCAL_I915_EXEC_BSD_MASK)
> >> +
> >> +static const int bo_size = 4096;
> >> +
> >> +static void
> >> +fill_object(struct drm_i915_gem_exec_object2 *obj, uint32_t gem_handle,
> >> + struct drm_i915_gem_relocation_entry *relocs, uint32_t count)
> >> +{
> >> + memset(obj, 0, sizeof(*obj));
> >> + obj->handle = gem_handle;
> >> + obj->relocation_count = count;
> >> + obj->relocs_ptr = (uintptr_t)relocs;
> >> +}
> >> +
> >> +static void
> >> +fill_reloc(struct drm_i915_gem_relocation_entry *reloc,
> >> + uint32_t gem_handle, uint32_t offset,
> >> + uint32_t read_domains, uint32_t write_domains)
> >> +{
> >> + reloc->target_handle = gem_handle;
> >> + reloc->delta = 0;
> >> + reloc->offset = offset * sizeof(uint32_t);
> >> + reloc->presumed_offset = 0;
> >> + reloc->read_domains = read_domains;
> >> + reloc->write_domain = write_domains;
> >> +}
> >> +
> >> +/*
> >> + * Needs to be global. Signal handlers don't accept arguments
> >> + */
> >> +static uint32_t *batch;
> >> +
> >> +static uint32_t emit_recursive_batch(int fd, int engine, unsigned dep_handle)
> >> +{
> >> + const int gen = intel_gen(intel_get_drm_devid(fd));
> >> + struct drm_i915_gem_exec_object2 obj[2];
> >> + struct drm_i915_gem_relocation_entry relocs[2];
> >> + struct drm_i915_gem_execbuffer2 execbuf;
> >> + unsigned engines[16];
> >> + unsigned nengine, handle;
> >> + int i = 0, reloc_count = 0, buf_count = 0;
> >> +
> >> + buf_count = 0;
> >> + nengine = 0;
> >> + if (engine < 0) {
> >> + for_each_engine(fd, engine)
> >> + if (engine)
> >> + engines[nengine++] = engine;
> >> + } else {
> >> + gem_require_ring(fd, engine);
> >> + engines[nengine++] = engine;
> >> + }
> >> + igt_require(nengine);
> >> +
> >> + memset(&execbuf, 0, sizeof(execbuf));
> >> + memset(obj, 0, sizeof(obj));
> >> + memset(relocs, 0, sizeof(relocs));
> >> +
> >> + execbuf.buffers_ptr = (uintptr_t) obj;
> >> + handle = gem_create(fd, bo_size);
> >> + batch = gem_mmap__gtt(fd, handle, bo_size, PROT_WRITE);
> >> + igt_assert(batch);
> >> + gem_set_domain(fd, handle,
> >> + I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
> >> +
> >> + if (gen >= 8) {
> >> + batch[i++] = MI_BATCH_BUFFER_START | 1 << 8 | 1;
> >> + /* recurse */
> >> + fill_reloc(&relocs[reloc_count], handle, i,
> >> + I915_GEM_DOMAIN_COMMAND, 0);
> >> + batch[i++] = 0;
> >> + batch[i++] = 0;
> >> + } else if (gen >= 6) {
> >> + batch[i++] = MI_BATCH_BUFFER_START | 1 << 8;
> >> + /* recurse */
> >> + fill_reloc(&relocs[reloc_count], handle, i,
> >> + I915_GEM_DOMAIN_COMMAND, 0);
> >> + batch[i++] = 0;
> >> + } else {
> >> + batch[i++] = MI_BATCH_BUFFER_START | 2 << 6 |
> >> + ((gen < 4) ? 1 : 0);
> >> + /* recurse */
> >> + fill_reloc(&relocs[reloc_count], handle, i,
> >> + I915_GEM_DOMAIN_COMMAND, 0);
> >> + batch[i++] = 0;
> >> + if (gen < 4)
> >> + relocs[reloc_count].delta = 1;
> >> + }
> >> + reloc_count++;
> >> +
> >> + if (dep_handle > 0) {
> >> + igt_assert(nengine == 1);
> >> + /* dummy write to dependency */
> >> + fill_object(&obj[buf_count], dep_handle, NULL, 0);
> >> + buf_count++;
> >> +
> >> + fill_reloc(&relocs[reloc_count], dep_handle, 256,
> >> + I915_GEM_DOMAIN_RENDER,
> >> + I915_GEM_DOMAIN_RENDER);
> >> + reloc_count++;
> >> +
> >
> > This is not the dummy load you were looking for. This is an infinite
> > walk with no termination condition.
>
> This is based on kms_busy's previous make_fb_busy() recursive batch
> setup. I'm wondering why that one works for making the bo busy? Or did I
> miss something?
Scatch is object 0, the batch is object 1.
In the current execbuffer API, the batch is always the last object in
the list. Your code executed the empty scratch object, which would then
quite happily walk through 2^48-4096 bytes before hitting the MI_BB_END
in the batch. (The relocation inside the scratch should just write
itself into the nopid register.)
-Chris
--
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
