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 *thread(void *data)
{
struct thread *t = data;
uint32_t bbe = MI_BATCH_BUFFER_END;
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 obj;
uint32_t *ctx;
memset(&obj, 0, sizeof(obj));
obj.handle = gem_create(t->fd, 4096);
gem_write(t->fd, obj.handle, 0, &bbe, sizeof(bbe));
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)&obj;
execbuf.buffer_count = 1;
ctx = malloc(t->num_ctx * sizeof(uint32_t));
igt_assert(ctx);
memcpy(ctx, t->all_ctx, t->num_ctx * sizeof(uint32_t));
igt_permute_array(ctx, t->num_ctx, xchg_int);
for (unsigned n = 0; n < t->num_ctx; n++) {
execbuf.rsvd1 = ctx[n];
gem_execbuf(t->fd, &execbuf);
}
free(ctx);
gem_close(t->fd, obj.handle);
return NULL;
}
static void run_on_ring(int fd, unsigned ring_id, const char *ring_name)
{
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 execobj;
struct {
uint32_t handle;
uint32_t *batch;
} obj[2];
unsigned i;
char buf[100];
gem_require_ring(fd, ring_id);
igt_require(has_softpin(fd));
for (i = 0; i < 2; i++) {
obj[i].handle = gem_create(fd, BATCH_SIZE);
obj[i].batch = mmap_coherent(fd, obj[i].handle, BATCH_SIZE);
memset(obj[i].batch, 0xff, BATCH_SIZE);
}
memset(&execobj, 0, sizeof(execobj));
execobj.handle = obj[0].handle;
obj[0].batch[0] = MI_BATCH_BUFFER_END;
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)&execobj;
execbuf.buffer_count = 1;
execbuf.flags = ring_id;
/* Execute once to allocate a gtt-offset */
gem_execbuf(fd, &execbuf);
execobj.flags = EXEC_OBJECT_PINNED;
sprintf(buf, "Testing %s cs tlb coherency: ", ring_name);
for (i = 0; i < BATCH_SIZE/64; i++) {
execobj.handle = obj[i&1].handle;
obj[i&1].batch[i*64/4] = MI_BATCH_BUFFER_END;
execbuf.batch_start_offset = i*64;
gem_execbuf(fd, &execbuf);
}
for (i = 0; i < 2; i++) {
gem_close(fd, obj[i].handle);
munmap(obj[i].batch, BATCH_SIZE);
}
}
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 void *gem_busyspin(void *arg)
{
const uint32_t bbe = MI_BATCH_BUFFER_END;
struct gem_busyspin *bs = arg;
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 obj;
unsigned engines[16];
unsigned nengine;
unsigned engine;
int fd;
fd = drm_open_driver(DRIVER_INTEL);
nengine = 0;
for_each_engine(fd, engine)
if (!ignore_engine(fd, engine)) engines[nengine++] = engine;
memset(&obj, 0, sizeof(obj));
obj.handle = gem_create(fd, 4096);
gem_write(fd, obj.handle, 0, &bbe, sizeof(bbe));
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)&obj;
execbuf.buffer_count = 1;
execbuf.flags |= LOCAL_I915_EXEC_HANDLE_LUT;
execbuf.flags |= LOCAL_I915_EXEC_NO_RELOC;
if (__gem_execbuf(fd, &execbuf)) {
execbuf.flags = 0;
gem_execbuf(fd, &execbuf);
}
while (!done) {
for (int n = 0; n < nengine; n++) {
execbuf.flags &= ~ENGINE_FLAGS;
execbuf.flags |= engines[n];
gem_execbuf(fd, &execbuf);
}
bs->count += nengine;
}
close(fd);
return NULL;
}
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 void make_busy(int fd, uint32_t handle)
{
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 gem_exec;
const uint32_t buf[] = {MI_BATCH_BUFFER_END};
gem_write(fd, handle, 0, buf, sizeof(buf));
memset(&gem_exec, 0, sizeof(gem_exec));
gem_exec.handle = handle;
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)&gem_exec;
execbuf.buffer_count = 1;
execbuf.flags |= LOCAL_I915_EXEC_HANDLE_LUT;
execbuf.flags |= LOCAL_I915_EXEC_NO_RELOC;
if (__gem_execbuf(fd, &execbuf)) {
execbuf.flags = 0;
gem_execbuf(fd, &execbuf);
}
}
static void
dontneed_before_exec(void)
{
int fd = drm_open_driver(DRIVER_INTEL);
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 exec;
uint32_t buf[] = { MI_BATCH_BUFFER_END, 0 };
memset(&execbuf, 0, sizeof(execbuf));
memset(&exec, 0, sizeof(exec));
exec.handle = gem_create(fd, OBJECT_SIZE);
gem_write(fd, exec.handle, 0, buf, sizeof(buf));
gem_madvise(fd, exec.handle, I915_MADV_DONTNEED);
execbuf.buffers_ptr = (uintptr_t)&exec;
execbuf.buffer_count = 1;
execbuf.batch_len = sizeof(buf);
gem_execbuf(fd, &execbuf);
gem_close(fd, exec.handle);
close(fd);
}
static void fill_ring(int fd, struct drm_i915_gem_execbuffer2 *execbuf)
{
int i;
/* The ring we've been using is 128k, and each rendering op
* will use at least 8 dwords:
*
* BATCH_START
* BATCH_START offset
* MI_FLUSH
* STORE_DATA_INDEX
* STORE_DATA_INDEX offset
* STORE_DATA_INDEX value
* MI_USER_INTERRUPT
* (padding)
*
* So iterate just a little more than that -- if we don't fill the ring
* doing this, we aren't likely to with this test.
*/
igt_debug("Executing execbuf %d times\n", 128*1024/(8*4));
for (i = 0; i < 128*1024 / (8 * 4); i++)
gem_execbuf(fd, execbuf);
}
static int negative_reloc_blt(int fd)
{
const int gen = intel_gen(intel_get_drm_devid(fd));
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 gem_exec[1024][2];
struct drm_i915_gem_relocation_entry gem_reloc;
uint32_t buf[1024], *b;
int i;
memset(&gem_reloc, 0, sizeof(gem_reloc));
gem_reloc.offset = 4 * sizeof(uint32_t);
gem_reloc.presumed_offset = ~0ULL;
gem_reloc.delta = -4096;
gem_reloc.target_handle = 0;
gem_reloc.read_domains = I915_GEM_DOMAIN_RENDER;
gem_reloc.write_domain = I915_GEM_DOMAIN_RENDER;
for (i = 0; i < 1024; i++) {
memset(gem_exec[i], 0, sizeof(gem_exec[i]));
gem_exec[i][0].handle = gem_create(fd, 4096);
gem_exec[i][0].flags = EXEC_OBJECT_NEEDS_FENCE;
b = buf;
*b++ = XY_COLOR_BLT_CMD_NOLEN |
((gen >= 8) ? 5 : 4) |
COLOR_BLT_WRITE_ALPHA | XY_COLOR_BLT_WRITE_RGB;
*b++ = 0xf0 << 16 | 1 << 25 | 1 << 24 | 4096;
*b++ = 1 << 16 | 0;
*b++ = 2 << 16 | 1024;
*b++ = ~0;
if (gen >= 8)
*b++ = ~0;
*b++ = 0xc0ffee ^ i;
*b++ = MI_BATCH_BUFFER_END;
if ((b - buf) & 1)
*b++ = 0;
gem_exec[i][1].handle = gem_create(fd, 4096);
gem_write(fd, gem_exec[i][1].handle, 0, buf, (b - buf) * sizeof(uint32_t));
gem_exec[i][1].relocation_count = 1;
gem_exec[i][1].relocs_ptr = (uintptr_t)&gem_reloc;
}
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffer_count = 2;
execbuf.batch_len = (b - buf) * sizeof(uint32_t);
execbuf.flags = USE_LUT;
if (gen >= 6)
execbuf.flags |= I915_EXEC_BLT;
for (i = 0; i < 1024; i++) {
execbuf.buffers_ptr = (uintptr_t)gem_exec[i];
gem_execbuf(fd, &execbuf);
}
for (i = 1024; i--;) {
gem_read(fd, gem_exec[i][0].handle,
i*sizeof(uint32_t), buf + i, sizeof(uint32_t));
gem_close(fd, gem_exec[i][0].handle);
gem_close(fd, gem_exec[i][1].handle);
}
if (0) {
for (i = 0; i < 1024; i += 8)
igt_info("%08x %08x %08x %08x %08x %08x %08x %08x\n",
buf[i + 0], buf[i + 1], buf[i + 2], buf[i + 3],
buf[i + 4], buf[i + 5], buf[i + 6], buf[i + 7]);
}
for (i = 0; i < 1024; i++)
igt_assert_eq(buf[i], 0xc0ffee ^ i);
return 0;
}
static void processes(void)
{
const struct intel_execution_engine *e;
unsigned engines[16];
int num_engines;
struct rlimit rlim;
unsigned num_ctx;
uint32_t name;
int fd, *fds;
fd = drm_open_driver(DRIVER_INTEL);
num_ctx = get_num_contexts(fd);
num_engines = 0;
for (e = intel_execution_engines; e->name; e++) {
if (e->exec_id == 0)
continue;
if (!has_engine(fd, e))
continue;
if (e->exec_id == I915_EXEC_BSD) {
int is_bsd2 = e->flags != 0;
if (gem_has_bsd2(fd) != is_bsd2)
continue;
}
engines[num_engines++] = e->exec_id | e->flags;
if (num_engines == ARRAY_SIZE(engines))
break;
}
/* tweak rlimits to allow us to create this many files */
igt_assert(getrlimit(RLIMIT_NOFILE, &rlim) == 0);
if (rlim.rlim_cur < ALIGN(num_ctx + 1024, 1024)) {
rlim.rlim_cur = ALIGN(num_ctx + 1024, 1024);
if (rlim.rlim_cur > rlim.rlim_max)
rlim.rlim_max = rlim.rlim_cur;
igt_assert(setrlimit(RLIMIT_NOFILE, &rlim) == 0);
}
fds = malloc(num_ctx * sizeof(int));
igt_assert(fds);
for (unsigned n = 0; n < num_ctx; n++) {
fds[n] = drm_open_driver(DRIVER_INTEL);
if (fds[n] == -1) {
int err = errno;
for (unsigned i = n; i--; )
close(fds[i]);
free(fds);
errno = err;
igt_assert_f(0, "failed to create context %lld/%lld\n", (long long)n, (long long)num_ctx);
}
}
if (1) {
uint32_t bbe = MI_BATCH_BUFFER_END;
name = gem_create(fd, 4096);
gem_write(fd, name, 0, &bbe, sizeof(bbe));
name = gem_flink(fd, name);
}
igt_fork(child, NUM_THREADS) {
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 obj;
memset(&obj, 0, sizeof(obj));
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)&obj;
execbuf.buffer_count = 1;
igt_permute_array(fds, num_ctx, xchg_int);
for (unsigned n = 0; n < num_ctx; n++) {
obj.handle = gem_open(fds[n], name);
execbuf.flags = engines[n % num_engines];
gem_execbuf(fds[n], &execbuf);
gem_close(fds[n], obj.handle);
}
}
igt_waitchildren();
for (unsigned n = 0; n < num_ctx; n++)
close(fds[n]);
free(fds);
close(fd);
}
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 void
store_dword_loop(int fd, int ring, int count, int divider)
{
int i, val = 0;
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 obj[2];
struct drm_i915_gem_relocation_entry reloc[divider];
uint32_t handle[divider];
uint32_t *batch[divider];
uint32_t *target;
int gen = intel_gen(devid);
memset(obj, 0, sizeof(obj));
obj[0].handle = gem_create(fd, 4096);
target = mmap_coherent(fd, obj[0].handle, 4096);
memset(reloc, 0, sizeof(reloc));
for (i = 0; i < divider; i++) {
uint32_t *b;
handle[i] = gem_create(fd, 4096);
batch[i] = mmap_coherent(fd, handle[i], 4096);
gem_set_domain(fd, handle[i], coherent_domain, coherent_domain);
b = batch[i];
*b++ = MI_STORE_DWORD_IMM;
*b++ = 0;
*b++ = 0;
*b++ = 0;
*b++ = MI_BATCH_BUFFER_END;
reloc[i].target_handle = obj[0].handle;
reloc[i].offset = 4;
if (gen < 8)
reloc[i].offset += 4;
reloc[i].read_domains = I915_GEM_DOMAIN_INSTRUCTION;
reloc[i].write_domain = I915_GEM_DOMAIN_INSTRUCTION;
obj[1].relocation_count = 1;
}
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)obj;
execbuf.buffer_count = 2;
execbuf.flags = ring;
igt_info("running storedw loop on render with stall every %i batch\n", divider);
for (i = 0; i < SLOW_QUICK(0x2000, 0x10); i++) {
int j = i % divider;
gem_set_domain(fd, handle[j], coherent_domain, coherent_domain);
batch[j][3] = val;
obj[1].handle = handle[j];
obj[1].relocs_ptr = (uintptr_t)&reloc[j];
gem_execbuf(fd, &execbuf);
if (j == 0) {
gem_set_domain(fd, obj[0].handle, coherent_domain, 0);
igt_assert_f(*target == val,
"%d: value mismatch: stored 0x%08x, expected 0x%08x\n",
i, *target, val);
}
val++;
}
gem_set_domain(fd, obj[0].handle, coherent_domain, 0);
igt_info("completed %d writes successfully, current value: 0x%08x\n",
i, target[0]);
munmap(target, 4096);
gem_close(fd, obj[0].handle);
for (i = 0; i < divider; ++i) {
munmap(batch[i], 4096);
gem_close(fd, handle[i]);
}
}
static uint32_t busy_blt(int fd)
{
const int gen = intel_gen(intel_get_drm_devid(fd));
const int has_64bit_reloc = gen >= 8;
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 object[2];
struct drm_i915_gem_relocation_entry reloc[200], *r;
uint32_t read, write;
uint32_t *map;
int factor = 100;
int i = 0;
memset(object, 0, sizeof(object));
object[0].handle = gem_create(fd, 1024*1024);
object[1].handle = gem_create(fd, 4096);
r = memset(reloc, 0, sizeof(reloc));
map = gem_mmap__cpu(fd, object[1].handle, 0, 4096, PROT_WRITE);
gem_set_domain(fd, object[1].handle,
I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
#define COPY_BLT_CMD (2<<29|0x53<<22|0x6)
#define BLT_WRITE_ALPHA (1<<21)
#define BLT_WRITE_RGB (1<<20)
while (factor--) {
/* XY_SRC_COPY */
map[i++] = COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB;
if (has_64bit_reloc)
map[i-1] += 2;
map[i++] = 0xcc << 16 | 1 << 25 | 1 << 24 | (4*1024);
map[i++] = 0;
map[i++] = 256 << 16 | 1024;
r->offset = i * sizeof(uint32_t);
r->target_handle = object[0].handle;
r->read_domains = I915_GEM_DOMAIN_RENDER;
r->write_domain = I915_GEM_DOMAIN_RENDER;
r++;
map[i++] = 0;
if (has_64bit_reloc)
map[i++] = 0;
map[i++] = 0;
map[i++] = 4096;
r->offset = i * sizeof(uint32_t);
r->target_handle = object[0].handle;
r->read_domains = I915_GEM_DOMAIN_RENDER;
r->write_domain = 0;
r++;
map[i++] = 0;
if (has_64bit_reloc)
map[i++] = 0;
}
map[i++] = MI_BATCH_BUFFER_END;
igt_assert(i <= 4096/sizeof(uint32_t));
igt_assert(r - reloc <= ARRAY_SIZE(reloc));
munmap(map, 4096);
object[1].relocs_ptr = (uintptr_t)reloc;
object[1].relocation_count = r - reloc;
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (unsigned long)object;
execbuf.buffer_count = 2;
if (gen >= 6)
execbuf.flags = I915_EXEC_BLT;
gem_execbuf(fd, &execbuf);
__gem_busy(fd, object[0].handle, &read, &write);
igt_assert_eq(read, 1 << write);
igt_assert_eq(write, gen >= 6 ? I915_EXEC_BLT : I915_EXEC_RENDER);
igt_debug("Created busy handle %d\n", object[0].handle);
gem_close(fd, object[1].handle);
return object[0].handle;
}
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;
}
