flywithu/ntel-gpu-tools
1
0
Fork 0
mirror of https://github.com/tiagovignatti/intel-gpu-tools.git synced 2025-06-24 00:06:09 +00:00
Code Issues Releases Wiki Activity

igt/gem_cs_prefetch: Replace explicit sync with implicit sync

Browse Source 
We can trade off the explicit sync (presumably to avoid some resource
starvation issue?) with the implicit sync of having to perform a
relocation. Using an implicit sync helps stress core kernel code,
besides being much faster!

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This commit is contained in:
Chris Wilson 2016-01-11 14:12:54 +00:00
parent a6090c7191
commit 38790b77d5
1 changed files with 75 additions and 69 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

144
tests/gem_cs_prefetch.c
View File

@ -36,23 +36,73 @@
* very last gtt pte.
*/
#include "igt.h"
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <fcntl.h>
#include <inttypes.h>
#include <errno.h>
#include <sys/stat.h>
#include <sys/time.h>
#include <drm.h>
#include "intel_bufmgr.h"
IGT_TEST_DESCRIPTION("Test the CS prefetch behaviour on batches.");
static drm_intel_bufmgr *bufmgr;
struct intel_batchbuffer *batch;
#define BATCH_SIZE 4096
struct shadow {
uint32_t handle;
struct drm_i915_gem_relocation_entry reloc;
};
static void setup(int fd, struct shadow *shadow)
{
int gen = intel_gen(intel_get_drm_devid(fd));
uint32_t *cpu;
int i = 0;
shadow->handle = gem_create(fd, 4096);
cpu = gem_mmap__cpu(fd, shadow->handle, 0, 4096, PROT_WRITE);
gem_set_domain(fd, shadow->handle,
I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
cpu[i++] = MI_STORE_DWORD_IMM;
if (gen >= 8) {
cpu[i++] = BATCH_SIZE - sizeof(uint32_t);
cpu[i++] = 0;
} else if (gen >= 4) {
cpu[i++] = 0;
cpu[i++] = BATCH_SIZE - sizeof(uint32_t);
} else {
cpu[i-1]--;
cpu[i++] = BATCH_SIZE - sizeof(uint32_t);
}
cpu[i++] = MI_BATCH_BUFFER_END;
cpu[i++] = MI_BATCH_BUFFER_END;
munmap(cpu, 4096);
memset(&shadow->reloc, 0, sizeof(shadow->reloc));
if (gen >= 8 || gen < 4)
shadow->reloc.offset = sizeof(uint32_t);
else
shadow->reloc.offset = 2*sizeof(uint32_t);
shadow->reloc.delta = BATCH_SIZE - sizeof(uint32_t);
shadow->reloc.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
shadow->reloc.write_domain = I915_GEM_DOMAIN_INSTRUCTION;
}
static uint32_t new_batch(int fd, struct shadow *shadow)
{
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 gem_exec[2];
memset(gem_exec, 0, sizeof(gem_exec));
gem_exec[0].handle = gem_create(fd, BATCH_SIZE);
gem_exec[1].handle = shadow->handle;
shadow->reloc.target_handle = gem_exec[0].handle;
gem_exec[1].relocs_ptr = (uintptr_t)&shadow->reloc;
gem_exec[1].relocation_count = 1;
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)gem_exec;
execbuf.buffer_count = 2;
gem_execbuf(fd, &execbuf);
return gem_exec[0].handle;
}
static void exec(int fd, uint32_t handle)
{
@ -66,77 +116,33 @@ static void exec(int fd, uint32_t handle)
execbuf.buffers_ptr = (uintptr_t)gem_exec;
execbuf.buffer_count = 1;
execbuf.batch_start_offset = 0;
execbuf.batch_len = 4096;
execbuf.batch_len = BATCH_SIZE;
gem_execbuf(fd, &execbuf);
gem_sync(fd, handle);
}
igt_simple_main
{
uint32_t batch_end[4] = {MI_BATCH_BUFFER_END, 0, 0, 0};
int fd, i, ret;
int count;
drm_intel_bo *sample_batch_bo;
struct shadow shadow;
uint64_t i, count;
int fd;
igt_skip_on_simulation();
fd = drm_open_driver(DRIVER_INTEL);
setup(fd, &shadow);
bufmgr = drm_intel_bufmgr_gem_init(fd, 4096);
igt_assert(bufmgr);
count = gem_aperture_size(fd) / BATCH_SIZE;
intel_require_memory(count, BATCH_SIZE, CHECK_RAM);
drm_intel_bufmgr_gem_enable_reuse(bufmgr);
count = gem_aperture_size(fd) / 4096;
intel_require_memory(count, 4096, CHECK_RAM);
batch = intel_batchbuffer_alloc(bufmgr, intel_get_drm_devid(fd));
igt_assert(batch);
sample_batch_bo = drm_intel_bo_alloc(bufmgr, "", 4096, 4096);
igt_assert(sample_batch_bo);
ret = drm_intel_bo_subdata(sample_batch_bo, 4096-sizeof(batch_end),
sizeof(batch_end), batch_end);
igt_assert(ret == 0);
/* fill the entire gart with batches and run them */
/* Fill the entire gart with batches and run them. */
for (i = 0; i < count; i++) {
drm_intel_bo *batch_bo;
batch_bo = drm_intel_bo_alloc(bufmgr, "", 4096, 4096);
igt_assert(batch_bo);
/* copy the sample batch with the gpu to the new one, so that we
* also test the unmappable part of the gtt. */
BLIT_COPY_BATCH_START(0);
OUT_BATCH((3 << 24) | /* 32 bits */
(0xcc << 16) | /* copy ROP */
4096);
OUT_BATCH(0); /* dst y1,x1 */
OUT_BATCH((1 << 16) | 1024);
OUT_RELOC_FENCED(batch_bo, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER, 0);
OUT_BATCH((0 << 16) | 0); /* src x1, y1 */
OUT_BATCH(4096);
OUT_RELOC_FENCED(sample_batch_bo, I915_GEM_DOMAIN_RENDER, 0, 0);
ADVANCE_BATCH();
intel_batchbuffer_flush(batch);
if (i % 100 == 0)
gem_sync(fd, batch_bo->handle);
drm_intel_bo_disable_reuse(batch_bo);
/* launch the newly created batch */
exec(fd, batch_bo->handle);
// leak buffers
//drm_intel_bo_unreference(batch_bo);
/* Launch the newly created batch... */
exec(fd, new_batch(fd, &shadow));
/* ...and leak the handle to consume the GTT */
igt_progress("gem_cs_prefetch: ", i, count);
}
igt_info("Test suceeded, cleanup up - this might take a while.\n");
drm_intel_bufmgr_destroy(bufmgr);
close(fd);
}