static uint64_t submit_batch(int fd, unsigned ring_id)
{
const uint32_t batch[] = { MI_NOOP,
MI_BATCH_BUFFER_END };
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 exec;
uint64_t presumed_offset;
gem_require_ring(fd, ring_id);
exec.handle = gem_create(fd, 4096);
gem_write(fd, exec.handle, 0, batch, sizeof(batch));
exec.relocation_count = 0;
exec.relocs_ptr = 0;
exec.alignment = 0;
exec.offset = 0;
exec.flags = 0;
exec.rsvd1 = 0;
exec.rsvd2 = 0;
execbuf.buffers_ptr = (uintptr_t)&exec;
execbuf.buffer_count = 1;
execbuf.batch_start_offset = 0;
execbuf.batch_len = sizeof(batch);
execbuf.cliprects_ptr = 0;
execbuf.num_cliprects = 0;
execbuf.DR1 = 0;
execbuf.DR4 = 0;
execbuf.flags = ring_id;
i915_execbuffer2_set_context_id(execbuf, 0);
execbuf.rsvd2 = 0;
gem_execbuf(fd, &execbuf);
gem_sync(fd, exec.handle);
presumed_offset = exec.offset;
igt_set_stop_rings(igt_to_stop_ring_flag(ring_id));
gem_execbuf(fd, &execbuf);
gem_sync(fd, exec.handle);
igt_assert(igt_get_stop_rings() == STOP_RING_NONE);
igt_assert(presumed_offset == exec.offset);
gem_close(fd, exec.handle);
return exec.offset;
}
static void destroy_cairo_surface__blit(void *arg)
{
struct fb_blit_upload *blit = arg;
struct igt_fb *fb = blit->fb;
unsigned int obj_tiling = fb_mod_to_obj_tiling(fb->tiling);
munmap(blit->linear.map, blit->linear.size);
fb->cairo_surface = NULL;
gem_set_domain(blit->fd, blit->linear.handle,
I915_GEM_DOMAIN_GTT, 0);
igt_blitter_fast_copy__raw(blit->fd,
blit->linear.handle,
blit->linear.stride,
I915_TILING_NONE,
0, 0, /* src_x, src_y */
fb->width, fb->height,
fb->gem_handle,
fb->stride,
obj_tiling,
0, 0 /* dst_x, dst_y */);
gem_sync(blit->fd, blit->linear.handle);
gem_close(blit->fd, blit->linear.handle);
free(blit);
}
static void exec1(int fd, uint32_t handle, uint64_t reloc_ofs, unsigned flags, char *ptr)
{
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 gem_exec[1];
struct drm_i915_gem_relocation_entry gem_reloc[1];
gem_reloc[0].offset = reloc_ofs;
gem_reloc[0].delta = 0;
gem_reloc[0].target_handle = handle;
gem_reloc[0].read_domains = I915_GEM_DOMAIN_RENDER;
gem_reloc[0].write_domain = 0;
gem_reloc[0].presumed_offset = 0;
gem_exec[0].handle = handle;
gem_exec[0].relocation_count = 1;
gem_exec[0].relocs_ptr = (uintptr_t) gem_reloc;
gem_exec[0].alignment = 0;
gem_exec[0].offset = 0;
gem_exec[0].flags = 0;
gem_exec[0].rsvd1 = 0;
gem_exec[0].rsvd2 = 0;
execbuf.buffers_ptr = (uintptr_t)gem_exec;
execbuf.buffer_count = 1;
execbuf.batch_start_offset = 0;
execbuf.batch_len = 8;
execbuf.cliprects_ptr = 0;
execbuf.num_cliprects = 0;
execbuf.DR1 = 0;
execbuf.DR4 = 0;
execbuf.flags = flags;
i915_execbuffer2_set_context_id(execbuf, 0);
execbuf.rsvd2 = 0;
/* Avoid hitting slowpaths in the reloc processing which might yield a
* presumed_offset of -1. Happens when the batch is still busy from the
* last round. */
gem_sync(fd, handle);
gem_execbuf(fd, &execbuf);
igt_warn_on(gem_reloc[0].presumed_offset == -1);
if (use_64bit_relocs) {
uint64_t tmp;
if (ptr)
tmp = *(uint64_t *)(ptr+reloc_ofs);
else
gem_read(fd, handle, reloc_ofs, &tmp, sizeof(tmp));
igt_assert_eq(tmp, gem_reloc[0].presumed_offset);
} else {
uint32_t tmp;
if (ptr)
tmp = *(uint32_t *)(ptr+reloc_ofs);
else
gem_read(fd, handle, reloc_ofs, &tmp, sizeof(tmp));
igt_assert_eq(tmp, gem_reloc[0].presumed_offset);
}
}
static int loop(unsigned ring, int reps, unsigned flags)
{
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 gem_exec;
int fd;
fd = drm_open_driver(DRIVER_INTEL);
memset(&gem_exec, 0, sizeof(gem_exec));
gem_exec.handle = batch(fd);
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)&gem_exec;
execbuf.buffer_count = 1;
execbuf.flags = ring;
execbuf.flags |= LOCAL_I915_EXEC_HANDLE_LUT;
execbuf.flags |= LOCAL_I915_EXEC_NO_RELOC;
if (__gem_execbuf(fd, &execbuf)) {
execbuf.flags = ring;
if (__gem_execbuf(fd, &execbuf))
return 77;
}
while (reps--) {
struct timespec start, end;
unsigned count = 0;
gem_set_domain(fd, gem_exec.handle, I915_GEM_DOMAIN_GTT, 0);
sleep(1); /* wait for the hw to go back to sleep */
clock_gettime(CLOCK_MONOTONIC, &start);
do {
do_ioctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
count++;
if (flags & SYNC)
gem_sync(fd, gem_exec.handle);
clock_gettime(CLOCK_MONOTONIC, &end);
} while (elapsed(&start, &end) < 2.);
gem_sync(fd, gem_exec.handle);
clock_gettime(CLOCK_MONOTONIC, &end);
printf("%7.3f\n", 1e6*elapsed(&start, &end)/count);
}
return 0;
}
static void test_ring(int fd, unsigned ring, uint32_t flags)
{
uint32_t bbe = MI_BATCH_BUFFER_END;
uint32_t handle[3];
uint32_t read, write;
uint32_t active;
unsigned i;
gem_require_ring(fd, ring | flags);
handle[TEST] = gem_create(fd, 4096);
handle[BATCH] = gem_create(fd, 4096);
gem_write(fd, handle[BATCH], 0, &bbe, sizeof(bbe));
/* Create a long running batch which we can use to hog the GPU */
handle[BUSY] = busy_blt(fd);
/* Queue a batch after the busy, it should block and remain "busy" */
igt_assert(exec_noop(fd, handle, ring | flags, false));
igt_assert(still_busy(fd, handle[BUSY]));
__gem_busy(fd, handle[TEST], &read, &write);
igt_assert_eq(read, 1 << ring);
igt_assert_eq(write, 0);
/* Requeue with a write */
igt_assert(exec_noop(fd, handle, ring | flags, true));
igt_assert(still_busy(fd, handle[BUSY]));
__gem_busy(fd, handle[TEST], &read, &write);
igt_assert_eq(read, 1 << ring);
igt_assert_eq(write, ring);
/* Now queue it for a read across all available rings */
active = 0;
for (i = I915_EXEC_RENDER; i <= I915_EXEC_VEBOX; i++) {
if (exec_noop(fd, handle, i | flags, false))
active |= 1 << i;
}
igt_assert(still_busy(fd, handle[BUSY]));
__gem_busy(fd, handle[TEST], &read, &write);
igt_assert_eq(read, active);
igt_assert_eq(write, ring); /* from the earlier write */
/* Check that our long batch was long enough */
igt_assert(still_busy(fd, handle[BUSY]));
/* And make sure it becomes idle again */
gem_sync(fd, handle[TEST]);
__gem_busy(fd, handle[TEST], &read, &write);
igt_assert_eq(read, 0);
igt_assert_eq(write, 0);
for (i = TEST; i <= BATCH; i++)
gem_close(fd, handle[i]);
}
static void run_on_ring(int fd, unsigned ring_id, const char *ring_name)
{
uint32_t handle, handle_new;
uint64_t gtt_offset, gtt_offset_new;
uint32_t *batch_ptr, *batch_ptr_old;
unsigned split;
char buf[100];
int i;
gem_require_ring(fd, ring_id);
sprintf(buf, "testing %s cs tlb coherency: ", ring_name);
/* Shut up gcc, too stupid. */
batch_ptr_old = NULL;
handle = 0;
gtt_offset = 0;
for (split = 0; split < BATCH_SIZE/8 - 1; split += 2) {
igt_progress(buf, split, BATCH_SIZE/8 - 1);
handle_new = gem_create(fd, BATCH_SIZE);
batch_ptr = gem_mmap__cpu(fd, handle_new, 0, BATCH_SIZE,
PROT_READ | PROT_WRITE);
batch_ptr[split*2] = MI_BATCH_BUFFER_END;
for (i = split*2 + 2; i < BATCH_SIZE/8; i++)
batch_ptr[i] = 0xffffffff;
if (split > 0) {
gem_sync(fd, handle);
gem_close(fd, handle);
}
igt_assert_eq(exec(fd, handle_new, split, >t_offset_new, 0),
0);
if (split > 0) {
/* Check that we've managed to collide in the tlb. */
igt_assert(gtt_offset == gtt_offset_new);
/* We hang onto the storage of the old batch by keeping
* the cpu mmap around. */
munmap(batch_ptr_old, BATCH_SIZE);
}
handle = handle_new;
gtt_offset = gtt_offset_new;
batch_ptr_old = batch_ptr;
}
}
static void exec(int fd, uint32_t handle)
{
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 gem_exec[1];
memset(gem_exec, 0, sizeof(gem_exec));
gem_exec[0].handle = handle;
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)gem_exec;
execbuf.buffer_count = 1;
execbuf.batch_start_offset = 0;
execbuf.batch_len = 4096;
gem_execbuf(fd, &execbuf);
gem_sync(fd, handle);
}
static int exec(int fd, uint32_t handle, int loops, unsigned ring_id)
{
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 gem_exec[1];
int ret = 0;
gem_exec[0].handle = handle;
gem_exec[0].relocation_count = 0;
gem_exec[0].relocs_ptr = 0;
gem_exec[0].alignment = 0;
gem_exec[0].offset = 0;
gem_exec[0].flags = 0;
gem_exec[0].rsvd1 = 0;
gem_exec[0].rsvd2 = 0;
execbuf.buffers_ptr = (uintptr_t)gem_exec;
execbuf.buffer_count = 1;
execbuf.batch_start_offset = 0;
execbuf.batch_len = 8;
execbuf.cliprects_ptr = 0;
execbuf.num_cliprects = 0;
execbuf.DR1 = 0;
execbuf.DR4 = 0;
execbuf.flags = ring_id;
i915_execbuffer2_set_context_id(execbuf, 0);
execbuf.rsvd2 = 0;
while (loops-- && ret == 0) {
ret = drmIoctl(fd,
DRM_IOCTL_I915_GEM_EXECBUFFER2,
&execbuf);
}
gem_sync(fd, handle);
return ret;
}
int main(int argc, char **argv)
{
struct timeval start, end;
uint8_t *buf;
uint32_t handle;
int size = OBJECT_SIZE;
int loop, i, tiling;
int fd;
if (argc > 1)
size = atoi(argv[1]);
if (size == 0) {
fprintf(stderr, "Invalid object size specified\n");
return 1;
}
buf = malloc(size);
memset(buf, 0, size);
fd = drm_open_any();
handle = gem_create(fd, size);
assert(handle);
for (tiling = I915_TILING_NONE; tiling <= I915_TILING_Y; tiling++) {
if (tiling != I915_TILING_NONE) {
printf("\nSetting tiling mode to %s\n",
tiling == I915_TILING_X ? "X" : "Y");
gem_set_tiling(fd, handle, tiling, 512);
}
if (tiling == I915_TILING_NONE) {
gem_set_domain(fd, handle,
I915_GEM_DOMAIN_CPU,
I915_GEM_DOMAIN_CPU);
{
uint32_t *base = gem_mmap__cpu(fd, handle, size, PROT_READ | PROT_WRITE);
volatile uint32_t *ptr = base;
int x = 0;
for (i = 0; i < size/sizeof(*ptr); i++)
x += ptr[i];
/* force overtly clever gcc to actually compute x */
ptr[0] = x;
munmap(base, size);
/* mmap read */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++) {
base = gem_mmap__cpu(fd, handle, size, PROT_READ | PROT_WRITE);
ptr = base;
x = 0;
for (i = 0; i < size/sizeof(*ptr); i++)
x += ptr[i];
/* force overtly clever gcc to actually compute x */
ptr[0] = x;
munmap(base, size);
}
gettimeofday(&end, NULL);
printf("Time to read %dk through a CPU map: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
/* mmap write */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++) {
base = gem_mmap__cpu(fd, handle, size, PROT_READ | PROT_WRITE);
ptr = base;
for (i = 0; i < size/sizeof(*ptr); i++)
ptr[i] = i;
munmap(base, size);
}
gettimeofday(&end, NULL);
printf("Time to write %dk through a CPU map: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++) {
base = gem_mmap__cpu(fd, handle, size, PROT_READ | PROT_WRITE);
memset(base, 0, size);
munmap(base, size);
}
gettimeofday(&end, NULL);
printf("Time to clear %dk through a CPU map: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
gettimeofday(&start, NULL);
base = gem_mmap__cpu(fd, handle, size, PROT_READ | PROT_WRITE);
for (loop = 0; loop < 1000; loop++)
memset(base, 0, size);
munmap(base, size);
gettimeofday(&end, NULL);
printf("Time to clear %dk through a cached CPU map: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
}
/* CPU pwrite */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++)
gem_write(fd, handle, 0, buf, size);
gettimeofday(&end, NULL);
printf("Time to pwrite %dk through the CPU: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
/* CPU pread */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++)
gem_read(fd, handle, 0, buf, size);
gettimeofday(&end, NULL);
printf("Time to pread %dk through the CPU: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
}
/* prefault into gtt */
{
uint32_t *base = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE);
volatile uint32_t *ptr = base;
int x = 0;
for (i = 0; i < size/sizeof(*ptr); i++)
x += ptr[i];
/* force overtly clever gcc to actually compute x */
ptr[0] = x;
munmap(base, size);
}
/* mmap read */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++) {
uint32_t *base = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE);
volatile uint32_t *ptr = base;
int x = 0;
for (i = 0; i < size/sizeof(*ptr); i++)
x += ptr[i];
/* force overtly clever gcc to actually compute x */
ptr[0] = x;
munmap(base, size);
}
gettimeofday(&end, NULL);
printf("Time to read %dk through a GTT map: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
/* mmap write */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++) {
uint32_t *base = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE);
volatile uint32_t *ptr = base;
for (i = 0; i < size/sizeof(*ptr); i++)
ptr[i] = i;
munmap(base, size);
}
gettimeofday(&end, NULL);
printf("Time to write %dk through a GTT map: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
/* mmap clear */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++) {
uint32_t *base = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE);
memset(base, 0, size);
munmap(base, size);
}
gettimeofday(&end, NULL);
printf("Time to clear %dk through a GTT map: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
gettimeofday(&start, NULL);{
uint32_t *base = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE);
for (loop = 0; loop < 1000; loop++)
memset(base, 0, size);
munmap(base, size);
} gettimeofday(&end, NULL);
printf("Time to clear %dk through a cached GTT map: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
/* mmap read */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++) {
uint32_t *base = gem_mmap(fd, handle, size, PROT_READ | PROT_WRITE);
volatile uint32_t *ptr = base;
int x = 0;
for (i = 0; i < size/sizeof(*ptr); i++)
x += ptr[i];
/* force overtly clever gcc to actually compute x */
ptr[0] = x;
munmap(base, size);
}
gettimeofday(&end, NULL);
printf("Time to read %dk (again) through a GTT map: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
if (tiling == I915_TILING_NONE) {
/* GTT pwrite */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++)
gem_write(fd, handle, 0, buf, size);
gettimeofday(&end, NULL);
printf("Time to pwrite %dk through the GTT: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
/* GTT pread */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++)
gem_read(fd, handle, 0, buf, size);
gettimeofday(&end, NULL);
printf("Time to pread %dk through the GTT: %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
/* GTT pwrite, including clflush */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++) {
gem_write(fd, handle, 0, buf, size);
gem_sync(fd, handle);
}
gettimeofday(&end, NULL);
printf("Time to pwrite %dk through the GTT (clflush): %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
/* GTT pread, including clflush */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++) {
gem_sync(fd, handle);
gem_read(fd, handle, 0, buf, size);
}
gettimeofday(&end, NULL);
printf("Time to pread %dk through the GTT (clflush): %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
/* partial writes */
printf("Now partial writes.\n");
size /= 4;
/* partial GTT pwrite, including clflush */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++) {
gem_write(fd, handle, 0, buf, size);
gem_sync(fd, handle);
}
gettimeofday(&end, NULL);
printf("Time to pwrite %dk through the GTT (clflush): %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
/* partial GTT pread, including clflush */
gettimeofday(&start, NULL);
for (loop = 0; loop < 1000; loop++) {
gem_sync(fd, handle);
gem_read(fd, handle, 0, buf, size);
}
gettimeofday(&end, NULL);
printf("Time to pread %dk through the GTT (clflush): %7.3fµs\n",
size/1024, elapsed(&start, &end, loop));
size *= 4;
}
}
gem_close(fd, handle);
close(fd);
return 0;
}
static void big_exec(int fd, uint32_t handle, int ring)
{
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 *gem_exec;
uint32_t ctx_id1, ctx_id2;
int num_buffers = gem_available_aperture_size(fd) / 4096;
int i;
/* Make sure we only fill half of RAM with gem objects. */
igt_require(intel_get_total_ram_mb() * 1024 / 2 > num_buffers * 4);
gem_exec = calloc(num_buffers + 1, sizeof(*gem_exec));
igt_assert(gem_exec);
memset(gem_exec, 0, (num_buffers + 1) * sizeof(*gem_exec));
ctx_id1 = gem_context_create(fd);
ctx_id2 = gem_context_create(fd);
gem_exec[0].handle = handle;
execbuf.buffers_ptr = (uintptr_t)gem_exec;
execbuf.buffer_count = num_buffers + 1;
execbuf.batch_start_offset = 0;
execbuf.batch_len = 8;
execbuf.cliprects_ptr = 0;
execbuf.num_cliprects = 0;
execbuf.DR1 = 0;
execbuf.DR4 = 0;
execbuf.flags = ring;
execbuf.rsvd2 = 0;
execbuf.buffer_count = 1;
i915_execbuffer2_set_context_id(execbuf, ctx_id1);
do_ioctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
for (i = 0; i < num_buffers; i++) {
uint32_t tmp_handle = gem_create(fd, 4096);
gem_exec[i].handle = tmp_handle;
}
gem_exec[i].handle = handle;
execbuf.buffer_count = i + 1;
/* figure out how many buffers we can exactly fit */
while (drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2,
&execbuf) != 0) {
i--;
gem_close(fd, gem_exec[i].handle);
gem_exec[i].handle = handle;
execbuf.buffer_count--;
igt_info("trying buffer count %i\n", i - 1);
}
igt_info("reduced buffer count to %i from %i\n",
i - 1, num_buffers);
/* double check that it works */
do_ioctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
i915_execbuffer2_set_context_id(execbuf, ctx_id2);
do_ioctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
gem_sync(fd, handle);
}
static void create_cairo_surface__blit(int fd, struct igt_fb *fb)
{
struct fb_blit_upload *blit;
cairo_format_t cairo_format;
unsigned int obj_tiling = fb_mod_to_obj_tiling(fb->tiling);
int bpp, ret;
blit = malloc(sizeof(*blit));
igt_assert(blit);
/*
* We create a linear BO that we'll map for the CPU to write to (using
* cairo). This linear bo will be then blitted to its final
* destination, tiling it at the same time.
*/
bpp = igt_drm_format_to_bpp(fb->drm_format);
ret = create_bo_for_fb(fd, fb->width, fb->height, bpp,
LOCAL_DRM_FORMAT_MOD_NONE, 0,
&blit->linear.handle,
&blit->linear.size,
&blit->linear.stride);
igt_assert(ret == 0);
blit->fd = fd;
blit->fb = fb;
/* Copy fb content to linear BO */
gem_set_domain(fd, blit->linear.handle,
I915_GEM_DOMAIN_GTT, 0);
igt_blitter_fast_copy__raw(fd,
fb->gem_handle,
fb->stride,
obj_tiling,
0, 0, /* src_x, src_y */
fb->width, fb->height,
blit->linear.handle,
blit->linear.stride,
I915_TILING_NONE,
0, 0 /* dst_x, dst_y */);
gem_sync(fd, blit->linear.handle);
gem_set_domain(fd, blit->linear.handle,
I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
/* Setup cairo context */
blit->linear.map = gem_mmap__cpu(fd,
blit->linear.handle,
0,
blit->linear.size,
PROT_READ | PROT_WRITE);
cairo_format = drm_format_to_cairo(fb->drm_format);
fb->cairo_surface =
cairo_image_surface_create_for_data(blit->linear.map,
cairo_format,
fb->width, fb->height,
blit->linear.stride);
cairo_surface_set_user_data(fb->cairo_surface,
(cairo_user_data_key_t *)create_cairo_surface__blit,
blit, destroy_cairo_surface__blit);
}
static void execN(int fd, uint32_t handle, uint64_t batch_size, unsigned flags, char *ptr)
{
#define reloc_ofs(N, T) ((((N)+1) << 12) - 4*(1 + ((N) == ((T)-1))))
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 gem_exec[1];
struct drm_i915_gem_relocation_entry *gem_reloc;
uint64_t n, nreloc = batch_size >> 12;
gem_reloc = calloc(nreloc, sizeof(*gem_reloc));
igt_assert(gem_reloc);
for (n = 0; n < nreloc; n++) {
gem_reloc[n].offset = reloc_ofs(n, nreloc);
gem_reloc[n].target_handle = handle;
gem_reloc[n].read_domains = I915_GEM_DOMAIN_RENDER;
gem_reloc[n].presumed_offset = n ^ 0xbeefdeaddeadbeef;
if (ptr) {
if (use_64bit_relocs)
*(uint64_t *)(ptr + gem_reloc[n].offset) = gem_reloc[n].presumed_offset;
else
*(uint32_t *)(ptr + gem_reloc[n].offset) = gem_reloc[n].presumed_offset;
} else
gem_write(fd, handle, gem_reloc[n].offset, &gem_reloc[n].presumed_offset, 4*(1+use_64bit_relocs));
}
memset(gem_exec, 0, sizeof(gem_exec));
gem_exec[0].handle = handle;
gem_exec[0].relocation_count = nreloc;
gem_exec[0].relocs_ptr = (uintptr_t)gem_reloc;
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)gem_exec;
execbuf.buffer_count = 1;
execbuf.batch_start_offset = 0;
execbuf.batch_len = 8;
execbuf.flags = flags;
/* Avoid hitting slowpaths in the reloc processing which might yield a
* presumed_offset of -1. Happens when the batch is still busy from the
* last round. */
gem_sync(fd, handle);
gem_execbuf(fd, &execbuf);
for (n = 0; n < nreloc; n++)
igt_warn_on(gem_reloc[n].presumed_offset == -1);
if (use_64bit_relocs) {
for (n = 0; n < nreloc; n++) {
uint64_t tmp;
if (ptr)
tmp = *(uint64_t *)(ptr+reloc_ofs(n, nreloc));
else
gem_read(fd, handle, reloc_ofs(n, nreloc), &tmp, sizeof(tmp));
igt_assert_eq(tmp, gem_reloc[n].presumed_offset);
}
} else {
for (n = 0; n < nreloc; n++) {
uint32_t tmp;
if (ptr)
tmp = *(uint32_t *)(ptr+reloc_ofs(n, nreloc));
else
gem_read(fd, handle, reloc_ofs(n, nreloc), &tmp, sizeof(tmp));
igt_assert_eq(tmp, gem_reloc[n].presumed_offset);
}
}
free(gem_reloc);
#undef reloc_ofs
}
static int run(int object, int batch, int time, int reps)
{
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 exec[3];
struct drm_i915_gem_relocation_entry *reloc;
uint32_t *buf, handle, src, dst;
int fd, len, gen, size, nreloc;
int ring, count;
size = ALIGN(batch * 64, 4096);
reloc = malloc(sizeof(*reloc)*size/32*2);
fd = drm_open_driver(DRIVER_INTEL);
handle = gem_create(fd, size);
buf = gem_mmap__cpu(fd, handle, 0, size, PROT_WRITE);
gen = intel_gen(intel_get_drm_devid(fd));
has_64bit_reloc = gen >= 8;
src = gem_create(fd, object);
dst = gem_create(fd, object);
len = gem_linear_blt(fd, buf, 0, 0, 1, object, reloc);
if (has_64bit_reloc)
nreloc = len > 56 ? 4 : 2;
else
nreloc = len > 40 ? 4 : 2;
memset(exec, 0, sizeof(exec));
exec[0].handle = src;
exec[1].handle = dst;
exec[2].handle = handle;
exec[2].relocs_ptr = (uintptr_t)reloc;
exec[2].relocation_count = nreloc;
ring = 0;
if (gen >= 6)
ring = I915_EXEC_BLT;
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)exec;
execbuf.buffer_count = 3;
execbuf.batch_len = len;
execbuf.flags = ring;
execbuf.flags |= LOCAL_I915_EXEC_HANDLE_LUT;
if (__gem_execbuf(fd, &execbuf)) {
gem_set_domain(fd, handle, I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
len = gem_linear_blt(fd, buf, 0, src, dst, object, reloc);
igt_assert(len == execbuf.batch_len);
execbuf.flags = ring;
gem_execbuf(fd, &execbuf);
}
gem_sync(fd, handle);
if (batch > 1) {
if (execbuf.flags & LOCAL_I915_EXEC_HANDLE_LUT) {
src = 0;
dst = 1;
}
gem_set_domain(fd, handle, I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
for (int i = 1; i < batch; i++) {
len = gem_linear_blt(fd, buf, len - 8,
src, dst, object,
reloc + nreloc * i);
}
exec[2].relocation_count = nreloc * batch;
execbuf.batch_len = len;
gem_execbuf(fd, &execbuf);
gem_sync(fd, handle);
}
if (execbuf.flags & LOCAL_I915_EXEC_HANDLE_LUT)
execbuf.flags |= LOCAL_I915_EXEC_NO_RELOC;
/* Guess how many loops we need for 0.1s */
count = baseline((uint64_t)object * batch, 100);
while (reps--) {
double min = HUGE_VAL;
for (int s = 0; s <= time / 100; s++) {
struct timespec start, end;
double t;
clock_gettime(CLOCK_MONOTONIC, &start);
for (int loop = 0; loop < count; loop++)
gem_execbuf(fd, &execbuf);
gem_sync(fd, handle);
clock_gettime(CLOCK_MONOTONIC, &end);
t = elapsed(&start, &end);
if (t < min)
min = t;
}
printf("%7.3f\n", object/(1024*1024.)*batch*count/min);
}
close(fd);
return 0;
}
static void run(int object_size)
{
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 exec[3];
struct drm_i915_gem_relocation_entry reloc[4];
uint32_t buf[20];
uint32_t handle, src, dst;
int fd, len, count;
int ring;
fd = drm_open_any();
handle = gem_create(fd, 4096);
src = gem_create(fd, object_size);
dst = gem_create(fd, object_size);
len = gem_linear_blt(buf, src, dst, object_size, reloc);
gem_write(fd, handle, 0, buf, len);
exec[0].handle = src;
exec[0].relocation_count = 0;
exec[0].relocs_ptr = 0;
exec[0].alignment = 0;
exec[0].offset = 0;
exec[0].flags = 0;
exec[0].rsvd1 = 0;
exec[0].rsvd2 = 0;
exec[1].handle = dst;
exec[1].relocation_count = 0;
exec[1].relocs_ptr = 0;
exec[1].alignment = 0;
exec[1].offset = 0;
exec[1].flags = 0;
exec[1].rsvd1 = 0;
exec[1].rsvd2 = 0;
exec[2].handle = handle;
exec[2].relocation_count = len > 40 ? 4 : 2;
exec[2].relocs_ptr = (uintptr_t)reloc;
exec[2].alignment = 0;
exec[2].offset = 0;
exec[2].flags = 0;
exec[2].rsvd1 = 0;
exec[2].rsvd2 = 0;
ring = 0;
if (HAS_BLT_RING(intel_get_drm_devid(fd)))
ring = I915_EXEC_BLT;
execbuf.buffers_ptr = (uintptr_t)exec;
execbuf.buffer_count = 3;
execbuf.batch_start_offset = 0;
execbuf.batch_len = len;
execbuf.cliprects_ptr = 0;
execbuf.num_cliprects = 0;
execbuf.DR1 = 0;
execbuf.DR4 = 0;
execbuf.flags = ring;
i915_execbuffer2_set_context_id(execbuf, 0);
execbuf.rsvd2 = 0;
for (count = 1; count <= 1<<17; count <<= 1) {
struct timeval start, end;
gettimeofday(&start, NULL);
if (gem_exec(fd, &execbuf, count))
exit(1);
gem_sync(fd, handle);
gettimeofday(&end, NULL);
printf("Time to blt %d bytes x %6d: %7.3fµs, %s\n",
object_size, count,
elapsed(&start, &end, count),
bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
fflush(stdout);
}
gem_close(fd, handle);
close(fd);
}
