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
scratch_buf_check(data_t *data, struct igt_buf *buf, int x, int y,
uint8_t color)
{
uint8_t val;
gem_read(data->drm_fd, buf->bo->handle, 0,
data->linear, sizeof(data->linear));
val = data->linear[y * WIDTH + x];
igt_assert_f(val == color,
"Expected 0x%02x, found 0x%02x at (%d,%d)\n",
color, val, x, y);
}
static void
check_bo(int fd, uint32_t handle, uint32_t val)
{
int i;
gem_read(fd, handle, 0, linear, sizeof(linear));
for (i = 0; i < WIDTH*HEIGHT; i++) {
igt_assert_f(linear[i] == val,
"Expected 0x%08x, found 0x%08x "
"at offset 0x%08x\n",
val, linear[i], i * 4);
val++;
}
}
static void
test_read(int fd)
{
void *dst;
uint32_t src;
/* copy from a fresh src to fresh dst to force pagefault on both */
dst = create_pointer(fd);
src = gem_create(fd, OBJECT_SIZE);
gem_read(fd, src, 0, dst, OBJECT_SIZE);
gem_close(fd, src);
munmap(dst, OBJECT_SIZE);
}
static void
check_bo(int fd, uint32_t handle, uint32_t val)
{
int i;
gem_read(fd, handle, 0, linear, sizeof(linear));
for (i = 0; i < WIDTH*HEIGHT; i++) {
if (linear[i] != val) {
fprintf(stderr, "Expected 0x%08x, found 0x%08x "
"at offset 0x%08x\n",
val, linear[i], i * 4);
abort();
}
val++;
}
}
static void
check_bo(int fd, drm_intel_bo *bo, uint32_t start_val)
{
int i;
gem_read(fd, bo->handle, 0, linear, sizeof(linear));
for (i = 0; i < 1024 * 1024 / 4; i++) {
if (linear[i] != start_val) {
fprintf(stderr, "Expected 0x%08x, found 0x%08x "
"at offset 0x%08x\n",
start_val, linear[i], i * 4);
abort();
}
start_val++;
}
}
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;
}
int main(int argc, char **argv)
{
data_t data = {0, };
struct intel_batchbuffer *batch = NULL;
struct igt_buf src, dst;
igt_render_copyfunc_t render_copy = NULL;
int opt_dump_aub = igt_aub_dump_enabled();
igt_simple_init_parse_opts(&argc, argv, "da", NULL, NULL,
opt_handler, NULL);
igt_fixture {
data.drm_fd = drm_open_any_render();
data.devid = intel_get_drm_devid(data.drm_fd);
data.bufmgr = drm_intel_bufmgr_gem_init(data.drm_fd, 4096);
igt_assert(data.bufmgr);
render_copy = igt_get_render_copyfunc(data.devid);
igt_require_f(render_copy,
"no render-copy function\n");
batch = intel_batchbuffer_alloc(data.bufmgr, data.devid);
igt_assert(batch);
}
scratch_buf_init(&data, &src, WIDTH, HEIGHT, STRIDE, SRC_COLOR);
scratch_buf_init(&data, &dst, WIDTH, HEIGHT, STRIDE, DST_COLOR);
scratch_buf_check(&data, &src, WIDTH / 2, HEIGHT / 2, SRC_COLOR);
scratch_buf_check(&data, &dst, WIDTH / 2, HEIGHT / 2, DST_COLOR);
if (opt_dump_png) {
scratch_buf_write_to_png(&src, "source.png");
scratch_buf_write_to_png(&dst, "destination.png");
}
if (opt_dump_aub) {
drm_intel_bufmgr_gem_set_aub_filename(data.bufmgr,
"rendercopy.aub");
drm_intel_bufmgr_gem_set_aub_dump(data.bufmgr, true);
}
/* This will copy the src to the mid point of the dst buffer. Presumably
* the out of bounds accesses will get clipped.
* Resulting buffer should look like:
* _______
* |dst|dst|
* |dst|src|
* -------
*/
render_copy(batch, NULL,
&src, 0, 0, WIDTH, HEIGHT,
&dst, WIDTH / 2, HEIGHT / 2);
if (opt_dump_png)
scratch_buf_write_to_png(&dst, "result.png");
if (opt_dump_aub) {
drm_intel_gem_bo_aub_dump_bmp(dst.bo,
0, 0, WIDTH, HEIGHT,
AUB_DUMP_BMP_FORMAT_ARGB_8888,
STRIDE, 0);
drm_intel_bufmgr_gem_set_aub_dump(data.bufmgr, false);
} else if (check_all_pixels) {
uint32_t val;
int i, j;
gem_read(data.drm_fd, dst.bo->handle, 0,
data.linear, sizeof(data.linear));
for (i = 0; i < WIDTH; i++) {
for (j = 0; j < HEIGHT; j++) {
uint32_t color = DST_COLOR;
val = data.linear[j * WIDTH + i];
if (j >= HEIGHT/2 && i >= WIDTH/2)
color = SRC_COLOR;
igt_assert_f(val == color,
"Expected 0x%08x, found 0x%08x at (%d,%d)\n",
color, val, i, j);
}
}
} else {
scratch_buf_check(&data, &dst, 10, 10, DST_COLOR);
scratch_buf_check(&data, &dst, WIDTH - 10, HEIGHT - 10, SRC_COLOR);
}
igt_exit();
}
static void
check_gpu(int fd, uint32_t handle, uint32_t val)
{
gem_read(fd, handle, 0, linear, sizeof(linear));
check_cpu(linear, val);
}
/* Simulates SNA behaviour using negative self-relocations for
* STATE_BASE_ADDRESS command packets. If they wrap around (to values greater
* than the total size of the GTT), the GPU will hang.
* See https://bugs.freedesktop.org/show_bug.cgi?id=78533
*/
static int negative_reloc(int fd, unsigned flags)
{
struct drm_i915_gem_execbuffer2 execbuf;
struct drm_i915_gem_exec_object2 gem_exec[2];
struct drm_i915_gem_relocation_entry gem_reloc[1000];
uint64_t gtt_max = get_page_table_size(fd);
uint32_t buf[1024] = {MI_BATCH_BUFFER_END};
int i;
#define BIAS (256*1024)
igt_require(intel_gen(intel_get_drm_devid(fd)) >= 7);
memset(gem_exec, 0, sizeof(gem_exec));
gem_exec[0].handle = gem_create(fd, 4096);
gem_write(fd, gem_exec[0].handle, 0, buf, 8);
gem_reloc[0].offset = 1024;
gem_reloc[0].delta = 0;
gem_reloc[0].target_handle = gem_exec[0].handle;
gem_reloc[0].read_domains = I915_GEM_DOMAIN_COMMAND;
gem_exec[1].handle = gem_create(fd, 4096);
gem_write(fd, gem_exec[1].handle, 0, buf, 8);
gem_exec[1].relocation_count = 1;
gem_exec[1].relocs_ptr = (uintptr_t)gem_reloc;
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)gem_exec;
execbuf.buffer_count = 2;
execbuf.batch_len = 8;
do_or_die(drmIoctl(fd,
DRM_IOCTL_I915_GEM_EXECBUFFER2,
&execbuf));
gem_close(fd, gem_exec[1].handle);
igt_info("Found offset %lld for 4k batch\n", (long long)gem_exec[0].offset);
/*
* Ideally we'd like to be able to control where the kernel is going to
* place the buffer. We don't SKIP here because it causes the test
* to "randomly" flip-flop between the SKIP and PASS states.
*/
if (gem_exec[0].offset < BIAS) {
igt_info("Offset is below BIAS, not testing anything\n");
return 0;
}
memset(gem_reloc, 0, sizeof(gem_reloc));
for (i = 0; i < sizeof(gem_reloc)/sizeof(gem_reloc[0]); i++) {
gem_reloc[i].offset = 8 + 4*i;
gem_reloc[i].delta = -BIAS*i/1024;
gem_reloc[i].target_handle = flags & USE_LUT ? 0 : gem_exec[0].handle;
gem_reloc[i].read_domains = I915_GEM_DOMAIN_COMMAND;
}
gem_exec[0].relocation_count = sizeof(gem_reloc)/sizeof(gem_reloc[0]);
gem_exec[0].relocs_ptr = (uintptr_t)gem_reloc;
execbuf.buffer_count = 1;
execbuf.flags = flags & USE_LUT;
do_or_die(drmIoctl(fd,
DRM_IOCTL_I915_GEM_EXECBUFFER2,
&execbuf));
igt_info("Batch is now at offset %lld\n", (long long)gem_exec[0].offset);
gem_read(fd, gem_exec[0].handle, 0, buf, sizeof(buf));
gem_close(fd, gem_exec[0].handle);
for (i = 0; i < sizeof(gem_reloc)/sizeof(gem_reloc[0]); i++)
igt_assert(buf[2 + i] < gtt_max);
return 0;
}
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 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
}
int
main(int argc, char **argv)
{
int fd;
int i, iter = 100;
uint32_t tiling, swizzle;
uint32_t handle;
uint32_t devid;
fd = drm_open_any();
handle = create_bo(fd);
gem_get_tiling(fd, handle, &tiling, &swizzle);
devid = intel_get_drm_devid(fd);
if (IS_GEN2(devid)) {
tile_height = 16;
tile_width = 128;
tile_size = 2048;
} else {
tile_height = 8;
tile_width = 512;
tile_size = PAGE_SIZE;
}
/* Read a bunch of random subsets of the data and check that they come
* out right.
*/
for (i = 0; i < iter; i++) {
int size = WIDTH * HEIGHT * 4;
int offset = (random() % size) & ~3;
int len = (random() % size) & ~3;
int j;
if (len == 0)
len = 4;
if (offset + len > size)
len = size - offset;
if (i == 0) {
offset = 0;
len = size;
}
gem_read(fd, handle, offset, linear, len);
/* Translate from offsets in the read buffer to the swizzled
* address that it corresponds to. This is the opposite of
* what Mesa does (calculate offset to be read given the linear
* offset it's looking for).
*/
for (j = offset; j < offset + len; j += 4) {
uint32_t expected_val, found_val;
int swizzled_offset;
const char *swizzle_str;
switch (swizzle) {
case I915_BIT_6_SWIZZLE_NONE:
swizzled_offset = j;
swizzle_str = "none";
break;
case I915_BIT_6_SWIZZLE_9:
swizzled_offset = j ^
swizzle_bit(9, j);
swizzle_str = "bit9";
break;
case I915_BIT_6_SWIZZLE_9_10:
swizzled_offset = j ^
swizzle_bit(9, j) ^
swizzle_bit(10, j);
swizzle_str = "bit9^10";
break;
case I915_BIT_6_SWIZZLE_9_11:
swizzled_offset = j ^
swizzle_bit(9, j) ^
swizzle_bit(11, j);
swizzle_str = "bit9^11";
break;
case I915_BIT_6_SWIZZLE_9_10_11:
swizzled_offset = j ^
swizzle_bit(9, j) ^
swizzle_bit(10, j) ^
swizzle_bit(11, j);
swizzle_str = "bit9^10^11";
break;
default:
fprintf(stderr, "Bad swizzle bits; %d\n",
swizzle);
abort();
}
expected_val = calculate_expected(swizzled_offset);
found_val = linear[(j - offset) / 4];
if (expected_val != found_val) {
fprintf(stderr,
"Bad read [%d]: %d instead of %d at 0x%08x "
"for read from 0x%08x to 0x%08x, swizzle=%s\n",
i, found_val, expected_val, j,
offset, offset + len,
swizzle_str);
abort();
}
}
}
close(fd);
return 0;
}
int main(int argc, char **argv)
{
int fd = drm_open_driver(DRIVER_INTEL);
int domain = I915_GEM_DOMAIN_GTT;
enum dir { READ, WRITE } dir = READ;
void *buf = malloc(OBJECT_SIZE);
uint32_t handle;
int reps = 13;
int c, size;
while ((c = getopt (argc, argv, "D:d:r:")) != -1) {
switch (c) {
case 'd':
if (strcmp(optarg, "cpu") == 0)
domain = I915_GEM_DOMAIN_CPU;
else if (strcmp(optarg, "gtt") == 0)
domain = I915_GEM_DOMAIN_GTT;
break;
case 'D':
if (strcmp(optarg, "read") == 0)
dir = READ;
else if (strcmp(optarg, "write") == 0)
dir = WRITE;
else
abort();
break;
case 'r':
reps = atoi(optarg);
if (reps < 1)
reps = 1;
break;
default:
break;
}
}
handle = gem_create(fd, OBJECT_SIZE);
for (size = 1; size <= OBJECT_SIZE; size <<= 1) {
igt_stats_t stats;
int n;
igt_stats_init_with_size(&stats, reps);
for (n = 0; n < reps; n++) {
struct timespec start, end;
gem_set_domain(fd, handle, domain, domain);
clock_gettime(CLOCK_MONOTONIC, &start);
if (dir == READ)
gem_read(fd, handle, 0, buf, size);
else
gem_write(fd, handle, 0, buf, size);
clock_gettime(CLOCK_MONOTONIC, &end);
igt_stats_push(&stats, elapsed(&start, &end));
}
printf("%7.3f\n", igt_stats_get_trimean(&stats)/1000);
igt_stats_fini(&stats);
}
return 0;
}
