flywithu/ntel-gpu-tools
1
0
Fork 0
mirror of https://github.com/tiagovignatti/intel-gpu-tools.git synced 2025-07-27 03:36:44 +00:00
Code Issues Releases Wiki Activity

igt/gem_busy: Discard the indirect batchbuffers

Browse Source 
The stack of batchbuffers is myth. In general there are only 2 levels of
stack for HEAD (ringbuffer, batchbuffer) and chaining up the batch
buffer just updates the lowest level of the stack. A BATCH_BUFFER_END at
any depth then returns to the ring.

So be creative and modify the batch buffer on the fly...

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This commit is contained in:
Chris Wilson 2016-04-22 19:08:10 +01:00
parent 71f4153287
commit 1e9a3acfd4
1 changed files with 57 additions and 95 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

152
tests/gem_busy.c
View File

@ -222,98 +222,49 @@ static void semaphore(int fd, unsigned ring, uint32_t flags)
gem_close(fd, handle[i]); gem_close(fd, handle[i]);
} }
static void create_indirect(int fd,
struct drm_i915_gem_exec_object2 *obj,
struct drm_i915_gem_exec_object2 *target)
{
const int gen = intel_gen(intel_get_drm_devid(fd));
const int nreloc = 128;
struct drm_i915_gem_relocation_entry *reloc;
uint32_t *batch;
int i, count;
reloc = calloc(nreloc, sizeof(*reloc));
obj->handle = gem_create(fd, 4096);
obj->relocs_ptr = (uintptr_t)reloc;
obj->relocation_count = nreloc;
batch = gem_mmap__cpu(fd, obj->handle, 0, 4096, PROT_WRITE);
gem_set_domain(fd, obj->handle,
I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
i = 0;
for (count = 0; count < nreloc; count++) {
reloc[count].target_handle = target->handle;
reloc[count].presumed_offset = target->offset;
reloc[count].offset = sizeof(uint32_t) * (i + 1);
reloc[count].delta = 0;
reloc[count].read_domains = I915_GEM_DOMAIN_COMMAND;
reloc[count].write_domain = 0;
batch[i] = MI_BATCH_BUFFER_START;
if (gen >= 8) {
batch[i] |= 1 << 8 | 1;
batch[++i] = target->offset;
batch[++i] = target->offset >> 32;
} else if (gen >= 6) {
batch[i] |= 1 << 8;
batch[++i] = target->offset;
} else {
batch[i] |= 2 << 6;
batch[++i] = target->offset;
if (gen < 4) {
batch[i] |= 1;
reloc[count].delta = 1;
}
}
i++;
}
batch[++i] = MI_BATCH_BUFFER_END;
igt_assert(i < 4096/sizeof(*batch));
munmap(batch, 4096);
}
static void store(int fd, unsigned ring, uint32_t flags) static void store(int fd, unsigned ring, uint32_t flags)
{ {
const int gen = intel_gen(intel_get_drm_devid(fd)); const int gen = intel_gen(intel_get_drm_devid(fd));
struct drm_i915_gem_exec_object2 obj[16]; struct drm_i915_gem_exec_object2 obj[2];
struct drm_i915_gem_relocation_entry store[1024]; #define SCRATCH 0
#define BATCH 1
struct drm_i915_gem_relocation_entry store[1024+1];
struct drm_i915_gem_execbuffer2 execbuf; struct drm_i915_gem_execbuffer2 execbuf;
unsigned size = ALIGN(ARRAY_SIZE(store)*16 + 4, 4096); unsigned size = ALIGN(ARRAY_SIZE(store)*16 + 4, 4096);
uint32_t read[2], write[2]; uint32_t read[2], write[2];
struct timespec tv; struct timespec tv;
uint32_t *batch; uint32_t *batch, *bbe;
int i, count, idx; int i, count;
gem_require_ring(fd, ring | flags); gem_require_ring(fd, ring | flags);
igt_skip_on_f(gen == 6 && (ring & ~(3<<13)) == I915_EXEC_BSD, igt_skip_on_f(gen == 6 && (ring & ~(3<<13)) == I915_EXEC_BSD,
"MI_STORE_DATA broken on gen6 bsd\n"); "MI_STORE_DATA broken on gen6 bsd\n");
gem_require_mmap_wc(fd);
gem_quiescent_gpu(fd); gem_quiescent_gpu(fd);
memset(&execbuf, 0, sizeof(execbuf)); memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)obj; execbuf.buffers_ptr = (uintptr_t)obj;
execbuf.buffer_count = 2; execbuf.buffer_count = 2;
execbuf.flags = ring | flags | LOCAL_EXEC_NO_RELOC; execbuf.flags = ring | flags;
if (gen < 6) if (gen < 6)
execbuf.flags |= I915_EXEC_SECURE; execbuf.flags |= I915_EXEC_SECURE;
memset(obj, 0, sizeof(obj)); memset(obj, 0, sizeof(obj));
obj[0].handle = gem_create(fd, 4096); obj[SCRATCH].handle = gem_create(fd, 4096);
obj[0].flags = EXEC_OBJECT_WRITE;
obj[1].handle = gem_create(fd, size);
obj[BATCH].handle = gem_create(fd, size);
obj[BATCH].relocs_ptr = (uintptr_t)store;
obj[BATCH].relocation_count = ARRAY_SIZE(store);
memset(store, 0, sizeof(store)); memset(store, 0, sizeof(store));
obj[1].relocs_ptr = (uintptr_t)store;
obj[1].relocation_count = ARRAY_SIZE(store);
batch = gem_mmap__cpu(fd, obj[1].handle, 0, size, PROT_WRITE); batch = gem_mmap__wc(fd, obj[BATCH].handle, 0, size, PROT_WRITE);
gem_set_domain(fd, obj[1].handle, gem_set_domain(fd, obj[BATCH].handle,
I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU); I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
i = 0; i = 0;
for (count = 0; count < ARRAY_SIZE(store); count++) { for (count = 0; count < 1024; count++) {
store[count].target_handle = obj[0].handle; store[count].target_handle = obj[SCRATCH].handle;
store[count].presumed_offset = -1; store[count].presumed_offset = -1;
store[count].offset = sizeof(uint32_t) * (i + 1); store[count].offset = sizeof(uint32_t) * (i + 1);
store[count].delta = sizeof(uint32_t) * count; store[count].delta = sizeof(uint32_t) * count;
@ -334,50 +285,61 @@ static void store(int fd, unsigned ring, uint32_t flags)
batch[++i] = count; batch[++i] = count;
i++; i++;
} }
batch[++i] = MI_BATCH_BUFFER_END;
bbe = &batch[i];
store[count].target_handle = obj[BATCH].handle; /* recurse */
store[count].presumed_offset = 0;
store[count].offset = sizeof(uint32_t) * (i + 1);
store[count].delta = 0;
store[count].read_domains = I915_GEM_DOMAIN_COMMAND;
store[count].write_domain = 0;
batch[i] = MI_BATCH_BUFFER_START;
if (gen >= 8) {
batch[i] |= 1 << 8 | 1;
batch[++i] = 0;
batch[++i] = 0;
} else if (gen >= 6) {
batch[i] |= 1 << 8;
batch[++i] = 0;
} else {
batch[i] |= 2 << 6;
batch[++i] = 0;
if (gen < 4) {
batch[i] |= 1;
store[count].delta = 1;
}
}
i++;
igt_assert(i < size/sizeof(*batch)); igt_assert(i < size/sizeof(*batch));
munmap(batch, size);
igt_require(__gem_execbuf(fd, &execbuf) == 0); igt_require(__gem_execbuf(fd, &execbuf) == 0);
do { __gem_busy(fd, obj[SCRATCH].handle, &read[SCRATCH], &write[SCRATCH]);
idx = execbuf.buffer_count++; __gem_busy(fd, obj[BATCH].handle, &read[BATCH], &write[BATCH]);
igt_require(idx < ARRAY_SIZE(obj));
create_indirect(fd, &obj[idx], &obj[idx-1]);
igt_require(__gem_execbuf(fd, &execbuf) == 0);
gem_execbuf(fd, &execbuf); *bbe = MI_BATCH_BUFFER_END;
__gem_busy(fd, obj[0].handle, &read[0], &write[0]); __sync_synchronize();
__gem_busy(fd, obj[idx].handle, &read[1], &write[1]);
igt_debug("After %d cycles: read[0]=%x read[1]=%x\n",
idx-1, read[0], read[1]);
} while (read[0] == 0 || read[1] == 0);
igt_assert_eq(write[0], ring); igt_assert_eq(write[SCRATCH], ring);
igt_assert_eq_u32(read[0], 1 << ring); igt_assert_eq_u32(read[SCRATCH], 1 << ring);
igt_assert_eq(write[1], 0); igt_assert_eq(write[BATCH], 0);
igt_assert_eq_u32(read[1], 1 << ring); igt_assert_eq_u32(read[BATCH], 1 << ring);
/* Calling busy in a loop should be enough to flush the rendering */ /* Calling busy in a loop should be enough to flush the rendering */
memset(&tv, 0, sizeof(tv)); memset(&tv, 0, sizeof(tv));
while (gem_busy(fd, obj[idx].handle)) while (gem_busy(fd, obj[BATCH].handle))
igt_assert(igt_seconds_elapsed(&tv) < 10); igt_assert(igt_seconds_elapsed(&tv) < 10);
igt_assert(!gem_busy(fd, obj[0].handle)); igt_assert(!gem_busy(fd, obj[SCRATCH].handle));
batch = gem_mmap__gtt(fd, obj[0].handle, 4096, PROT_READ); munmap(batch, size);
batch = gem_mmap__wc(fd, obj[SCRATCH].handle, 0, 4096, PROT_READ);
for (i = 0; i < 1024; i++) for (i = 0; i < 1024; i++)
igt_assert_eq_u32(batch[i], i); igt_assert_eq_u32(batch[i], i);
munmap(batch, 4096); munmap(batch, 4096);
for (i = 0; i <= idx; i++) { gem_close(fd, obj[BATCH].handle);
struct drm_i915_gem_relocation_entry *r; gem_close(fd, obj[SCRATCH].handle);
r = (struct drm_i915_gem_relocation_entry *)
(uintptr_t)obj[i].relocs_ptr;
if (r != store)
free(r);
gem_close(fd, obj[i].handle);
}
} }
static bool has_semaphores(int fd) static bool has_semaphores(int fd)