ntel-gpu-tools/tests/gem_exec_lut_handle.c
Chris Wilson e22525064d igt/gem_exec_lut_handle: Trim a few operations from submission
Move all the execbuffer construction upfront and do it once per round,
rather than per relocation pass. It helps reduce runtime, but more
importantly it removes the test overhead from out of the kernel
measurement.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
2015-04-08 21:14:46 +01:00

204 lines
5.8 KiB
C

/*
* Copyright © 2012 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.
*
* Authors:
* Chris Wilson <chris@chris-wilson.co.uk>
*
*/
/* Exercises the basic execbuffer using the handle LUT interface */
#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 "ioctl_wrappers.h"
#include "drmtest.h"
IGT_TEST_DESCRIPTION("Exercises the basic execbuffer using the handle LUT"
" interface.");
#define BATCH_SIZE (1024*1024)
#define LOCAL_I915_EXEC_NO_RELOC (1<<11)
#define LOCAL_I915_EXEC_HANDLE_LUT (1<<12)
#define MAX_NUM_EXEC 2048
#define MAX_NUM_RELOC 4096
#define SKIP_RELOC 0x1
#define NO_RELOC 0x2
struct drm_i915_gem_exec_object2 gem_exec[MAX_NUM_EXEC+1];
struct drm_i915_gem_relocation_entry gem_reloc[MAX_NUM_RELOC];
static uint32_t state = 0x12345678;
static uint32_t
hars_petruska_f54_1_random (void)
{
#define rol(x,k) ((x << k) | (x >> (32-k)))
return state = (state ^ rol (state, 5) ^ rol (state, 24)) + 0x37798849;
#undef rol
}
static int has_exec_lut(int fd)
{
struct drm_i915_gem_execbuffer2 execbuf;
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)(gem_exec + MAX_NUM_EXEC);
execbuf.buffer_count = 1;
execbuf.flags = LOCAL_I915_EXEC_HANDLE_LUT;
return drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf) == 0;
}
#define ELAPSED(a,b) (1e6*((b)->tv_sec - (a)->tv_sec) + ((b)->tv_usec - (a)->tv_usec))
igt_simple_main
{
uint32_t batch[2] = {MI_BATCH_BUFFER_END};
int fd, n, m, count;
const struct {
const char *name;
unsigned int flags;
} pass[] = {
{ .name = "relocation", .flags = 0 },
{ .name = "skip-relocs", .flags = SKIP_RELOC },
{ .name = "no-relocs", .flags = NO_RELOC },
{ .name = NULL },
}, *p;
igt_skip_on_simulation();
fd = drm_open_any();
memset(gem_exec, 0, sizeof(gem_exec));
for (n = 0; n < MAX_NUM_EXEC; n++)
gem_exec[n].handle = gem_create(fd, 4096);
gem_exec[n].handle = gem_create(fd, 4096);
gem_write(fd, gem_exec[n].handle, 0, batch, sizeof(batch));
memset(gem_reloc, 0, sizeof(gem_reloc));
for (n = 0; n < MAX_NUM_RELOC; n++) {
gem_reloc[n].offset = 1024;
gem_reloc[n].read_domains = I915_GEM_DOMAIN_RENDER;
}
igt_require(has_exec_lut(fd));
for (p = pass; p->name != NULL; p++) {
for (n = 1; n <= MAX_NUM_EXEC; n *= 2) {
double elapsed[16][2];
double s_x, s_y, s_xx, s_xy;
double A, B;
int i, j;
for (i = 0, m = 1; m <= MAX_NUM_RELOC; m *= 2, i++) {
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 *objects;
struct timeval start, end;
gem_exec[MAX_NUM_EXEC].relocation_count = m;
gem_exec[MAX_NUM_EXEC].relocs_ptr = (uintptr_t)gem_reloc;
objects = gem_exec + MAX_NUM_EXEC - n;
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)objects;
execbuf.buffer_count = n + 1;
execbuf.flags = LOCAL_I915_EXEC_HANDLE_LUT;
if (p->flags & NO_RELOC)
execbuf.flags |= LOCAL_I915_EXEC_NO_RELOC;
for (j = 0; j < m; j++) {
int target = hars_petruska_f54_1_random() % n;
gem_reloc[j].target_handle = target;
if (p->flags & SKIP_RELOC)
gem_reloc[j].presumed_offset = objects[target].offset;
else
gem_reloc[j].presumed_offset = 0;
}
gem_execbuf(fd,&execbuf);
gettimeofday(&start, NULL);
for (count = 0; count < 1000; count++)
gem_execbuf(fd, &execbuf);
gettimeofday(&end, NULL);
gem_sync(fd, gem_exec[MAX_NUM_EXEC].handle);
elapsed[i][1] = ELAPSED(&start, &end);
execbuf.flags &= ~LOCAL_I915_EXEC_HANDLE_LUT;
for (j = 0; j < m; j++) {
int target = gem_reloc[j].target_handle;
gem_reloc[j].target_handle = objects[target].handle;
if (p->flags & SKIP_RELOC)
gem_reloc[j].presumed_offset = objects[target].offset;
else
gem_reloc[j].presumed_offset = 0;
}
gem_execbuf(fd,&execbuf);
gettimeofday(&start, NULL);
for (count = 0; count < 1000; count++)
gem_execbuf(fd, &execbuf);
gettimeofday(&end, NULL);
gem_sync(fd, gem_exec[MAX_NUM_EXEC].handle);
elapsed[i][0] = ELAPSED(&start, &end);
}
igt_info("%s: buffers=%4d:", p->name, n);
s_x = s_y = s_xx = s_xy = 0;
for (j = 0; j < i; j++) {
int k = 1 << j;
s_x += k;
s_y += elapsed[j][0];
s_xx += k * k;
s_xy += k * elapsed[j][0];
}
B = (s_xy - s_x * s_y / j) / (s_xx - s_x * s_x / j);
A = s_y / j - B * s_x / j;
igt_info(" old=%7.0f + %.1f*reloc,", A, B);
s_x = s_y = s_xx = s_xy = 0;
for (j = 0; j < i; j++) {
int k = 1 << j;
s_x += k;
s_y += elapsed[j][1];
s_xx += k * k;
s_xy += k * elapsed[j][1];
}
B = (s_xy - s_x * s_y / j) / (s_xx - s_x * s_x / j);
A = s_y / j - B * s_x / j;
igt_info(" lut=%7.0f + %.1f*reloc (ns)", A, B);
igt_info("\n");
}
}
}