static void draw_rect_mmap_wc(int fd, struct buf_data *buf, struct rect *rect,
uint32_t color)
{
uint32_t *ptr;
uint32_t tiling, swizzle;
gem_set_domain(fd, buf->handle, I915_GEM_DOMAIN_GTT,
I915_GEM_DOMAIN_GTT);
gem_get_tiling(fd, buf->handle, &tiling, &swizzle);
/* We didn't implement suport for the older tiling methods yet. */
if (tiling != I915_TILING_NONE)
igt_require(intel_gen(intel_get_drm_devid(fd)) >= 5);
ptr = gem_mmap__wc(fd, buf->handle, 0, buf->size,
PROT_READ | PROT_WRITE);
switch (tiling) {
case I915_TILING_NONE:
draw_rect_ptr_linear(ptr, buf->stride, rect, color, buf->bpp);
break;
case I915_TILING_X:
draw_rect_ptr_tiled(ptr, buf->stride, swizzle, rect, color,
buf->bpp);
break;
default:
igt_assert(false);
break;
}
igt_assert(munmap(ptr, buf->size) == 0);
}
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 run_test(int fd, unsigned ring, unsigned flags)
{
const int gen = intel_gen(intel_get_drm_devid(fd));
const uint32_t bbe = MI_BATCH_BUFFER_END;
struct drm_i915_gem_exec_object2 obj[2];
struct drm_i915_gem_relocation_entry reloc[1024];
struct drm_i915_gem_execbuffer2 execbuf;
struct igt_hang_ring hang;
uint32_t *batch, *b;
int i;
gem_require_ring(fd, ring);
igt_skip_on_f(gen == 6 && (ring & ~(3<<13)) == I915_EXEC_BSD,
"MI_STORE_DATA broken on gen6 bsd\n");
gem_quiescent_gpu(fd);
memset(&execbuf, 0, sizeof(execbuf));
execbuf.buffers_ptr = (uintptr_t)obj;
execbuf.buffer_count = 2;
execbuf.flags = ring | (1 << 11);
if (gen < 6)
execbuf.flags |= I915_EXEC_SECURE;
memset(obj, 0, sizeof(obj));
obj[0].handle = gem_create(fd, 4096);
obj[0].flags |= EXEC_OBJECT_WRITE;
obj[1].handle = gem_create(fd, 1024*16 + 4096);
gem_write(fd, obj[1].handle, 0, &bbe, sizeof(bbe));
igt_require(__gem_execbuf(fd, &execbuf) == 0);
obj[1].relocs_ptr = (uintptr_t)reloc;
obj[1].relocation_count = 1024;
batch = gem_mmap__cpu(fd, obj[1].handle, 0, 16*1024 + 4096,
PROT_WRITE | PROT_READ);
gem_set_domain(fd, obj[1].handle,
I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
memset(reloc, 0, sizeof(reloc));
b = batch;
for (i = 0; i < 1024; i++) {
uint64_t offset;
reloc[i].target_handle = obj[0].handle;
reloc[i].presumed_offset = obj[0].offset;
reloc[i].offset = (b - batch + 1) * sizeof(*batch);
reloc[i].delta = i * sizeof(uint32_t);
reloc[i].read_domains = I915_GEM_DOMAIN_INSTRUCTION;
reloc[i].write_domain = I915_GEM_DOMAIN_INSTRUCTION;
offset = obj[0].offset + reloc[i].delta;
*b++ = MI_STORE_DWORD_IMM | (gen < 6 ? 1 << 22 : 0);
if (gen >= 8) {
*b++ = offset;
*b++ = offset >> 32;
} else if (gen >= 4) {
static void
test_huge_bo(int huge, int tiling)
{
uint64_t huge_object_size, last_offset;
char *ptr_cpu;
char *cpu_pattern;
uint32_t bo;
int i;
switch (huge) {
case -1:
huge_object_size = gem_mappable_aperture_size() / 2;
break;
case 0:
huge_object_size = gem_mappable_aperture_size() + PAGE_SIZE;
break;
default:
huge_object_size = gem_aperture_size(fd) + PAGE_SIZE;
break;
}
intel_require_memory(1, huge_object_size, CHECK_RAM);
last_offset = huge_object_size - PAGE_SIZE;
cpu_pattern = malloc(PAGE_SIZE);
igt_assert(cpu_pattern);
for (i = 0; i < PAGE_SIZE; i++)
cpu_pattern[i] = i;
bo = gem_create(fd, huge_object_size);
/* Obtain CPU mapping for the object. */
ptr_cpu = gem_mmap__cpu(fd, bo, 0, huge_object_size,
PROT_READ | PROT_WRITE);
igt_assert(ptr_cpu);
gem_set_domain(fd, bo, I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
gem_close(fd, bo);
/* Write first page through the mapping and assert reading it back
* works. */
memcpy(ptr_cpu, cpu_pattern, PAGE_SIZE);
igt_assert(memcmp(ptr_cpu, cpu_pattern, PAGE_SIZE) == 0);
/* Write last page through the mapping and assert reading it back
* works. */
memcpy(ptr_cpu + last_offset, cpu_pattern, PAGE_SIZE);
igt_assert(memcmp(ptr_cpu + last_offset, cpu_pattern, PAGE_SIZE) == 0);
/* Cross check that accessing two simultaneous pages works. */
igt_assert(memcmp(ptr_cpu, ptr_cpu + last_offset, PAGE_SIZE) == 0);
munmap(ptr_cpu, huge_object_size);
free(cpu_pattern);
}
static void check_bo(int fd, uint32_t handle)
{
uint32_t *map;
int i;
igt_debug("Verifying result\n");
map = gem_mmap__cpu(fd, handle, 0, 4096, PROT_READ);
gem_set_domain(fd, handle, I915_GEM_DOMAIN_CPU, 0);
for (i = 0; i < 1024; i++)
igt_assert_eq(map[i], i);
munmap(map, 4096);
}
static void draw_rect_mmap_gtt(int fd, struct buf_data *buf, struct rect *rect,
uint32_t color)
{
uint32_t *ptr;
gem_set_domain(fd, buf->handle, I915_GEM_DOMAIN_GTT,
I915_GEM_DOMAIN_GTT);
ptr = gem_mmap__gtt(fd, buf->handle, buf->size, PROT_READ | PROT_WRITE);
draw_rect_ptr_linear(ptr, buf->stride, rect, color, buf->bpp);
igt_assert(munmap(ptr, buf->size) == 0);
}
static cairo_surface_t *get_cairo_surface(int fd, struct igt_fb *fb)
{
if (fb->cairo_surface == NULL) {
if (fb->tiling == LOCAL_I915_FORMAT_MOD_Y_TILED ||
fb->tiling == LOCAL_I915_FORMAT_MOD_Yf_TILED)
create_cairo_surface__blit(fd, fb);
else
create_cairo_surface__gtt(fd, fb);
}
gem_set_domain(fd, fb->gem_handle,
I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
igt_assert(cairo_surface_status(fb->cairo_surface) == CAIRO_STATUS_SUCCESS);
return fb->cairo_surface;
}
static void *
mmap_coherent(int fd, uint32_t handle, int size)
{
int domain;
void *ptr;
if (gem_has_llc(fd) || !gem_mmap__has_wc(fd)) {
domain = I915_GEM_DOMAIN_CPU;
ptr = gem_mmap__cpu(fd, handle, 0, size, PROT_WRITE);
} else {
domain = I915_GEM_DOMAIN_GTT;
ptr = gem_mmap__wc(fd, handle, 0, size, PROT_WRITE);
}
gem_set_domain(fd, handle, domain, domain);
return ptr;
}
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 uint32_t
create_bo(int fd)
{
uint32_t handle;
uint32_t *data;
int i;
handle = gem_create(fd, SIZE);
gem_set_tiling(fd, handle, I915_TILING_X, WIDTH * sizeof(uint32_t));
/* Write throught the fence to tiled the data.
* We then manually detile on reading back through the mmap(wc).
*/
data = gem_mmap__gtt(fd, handle, SIZE, PROT_READ | PROT_WRITE);
for (i = 0; i < WIDTH*HEIGHT; i++)
data[i] = i;
munmap(data, SIZE);
gem_set_domain(fd, handle, I915_GEM_DOMAIN_CPU, 0);
return 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);
}
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
set_domain_gtt(int fd, uint32_t handle)
{
gem_set_domain(fd, handle, I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
}
static void
test_huge_bo(int fd, int huge, int tiling)
{
uint32_t bo;
char *ptr;
char *tiled_pattern;
char *linear_pattern;
uint64_t size, last_offset;
int pitch = tiling == I915_TILING_Y ? 128 : 512;
int i;
switch (huge) {
case -1:
size = gem_mappable_aperture_size() / 2;
break;
case 0:
size = gem_mappable_aperture_size() + PAGE_SIZE;
break;
default:
size = gem_aperture_size(fd) + PAGE_SIZE;
break;
}
intel_require_memory(1, size, CHECK_RAM);
last_offset = size - PAGE_SIZE;
/* Create pattern */
bo = gem_create(fd, PAGE_SIZE);
if (tiling)
gem_set_tiling(fd, bo, tiling, pitch);
linear_pattern = gem_mmap__gtt(fd, bo, PAGE_SIZE,
PROT_READ | PROT_WRITE);
for (i = 0; i < PAGE_SIZE; i++)
linear_pattern[i] = i;
tiled_pattern = gem_mmap__cpu(fd, bo, 0, PAGE_SIZE, PROT_READ);
gem_set_domain(fd, bo, I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT, 0);
gem_close(fd, bo);
bo = gem_create(fd, size);
if (tiling)
gem_set_tiling(fd, bo, tiling, pitch);
/* Initialise first/last page through CPU mmap */
ptr = gem_mmap__cpu(fd, bo, 0, size, PROT_READ | PROT_WRITE);
memcpy(ptr, tiled_pattern, PAGE_SIZE);
memcpy(ptr + last_offset, tiled_pattern, PAGE_SIZE);
munmap(ptr, size);
/* Obtain mapping for the object through GTT. */
ptr = __gem_mmap__gtt(fd, bo, size, PROT_READ | PROT_WRITE);
igt_require_f(ptr, "Huge BO GTT mapping not supported.\n");
set_domain_gtt(fd, bo);
/* Access through GTT should still provide the CPU written values. */
igt_assert(memcmp(ptr , linear_pattern, PAGE_SIZE) == 0);
igt_assert(memcmp(ptr + last_offset, linear_pattern, PAGE_SIZE) == 0);
gem_set_tiling(fd, bo, I915_TILING_NONE, 0);
igt_assert(memcmp(ptr , tiled_pattern, PAGE_SIZE) == 0);
igt_assert(memcmp(ptr + last_offset, tiled_pattern, PAGE_SIZE) == 0);
munmap(ptr, size);
gem_close(fd, bo);
munmap(tiled_pattern, PAGE_SIZE);
munmap(linear_pattern, PAGE_SIZE);
}
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 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;
}
int main(int argc, char **argv)
{
int fd = drm_open_any();
int size = 0;
int busy = 0;
int reps = 13;
int c, n, s;
while ((c = getopt (argc, argv, "bs:r:")) != -1) {
switch (c) {
case 's':
size = atoi(optarg);
break;
case 'r':
reps = atoi(optarg);
if (reps < 1)
reps = 1;
break;
case 'b':
busy = true;
break;
default:
break;
}
}
if (size == 0) {
for (s = 4096; s <= OBJECT_SIZE; s <<= 1) {
igt_stats_t stats;
igt_stats_init_with_size(&stats, reps);
for (n = 0; n < reps; n++) {
struct timespec start, end;
uint64_t count = 0;
clock_gettime(CLOCK_MONOTONIC, &start);
do {
for (c = 0; c < 1000; c++) {
uint32_t handle;
handle = gem_create(fd, s);
gem_set_domain(fd, handle,
I915_GEM_DOMAIN_GTT,
I915_GEM_DOMAIN_GTT);
if (busy)
make_busy(fd, handle);
gem_close(fd, handle);
}
count += c;
clock_gettime(CLOCK_MONOTONIC, &end);
} while (end.tv_sec - start.tv_sec < 2);
igt_stats_push_float(&stats, count / elapsed(&start, &end));
}
printf("%f\n", igt_stats_get_trimean(&stats));
igt_stats_fini(&stats);
}
} else {
for (n = 0; n < reps; n++) {
struct timespec start, end;
uint64_t count = 0;
clock_gettime(CLOCK_MONOTONIC, &start);
do {
for (c = 0; c < 1000; c++) {
uint32_t handle;
handle = gem_create(fd, size);
gem_set_domain(fd, handle,
I915_GEM_DOMAIN_GTT,
I915_GEM_DOMAIN_GTT);
if (busy)
make_busy(fd, handle);
gem_close(fd, handle);
}
count += c;
clock_gettime(CLOCK_MONOTONIC, &end);
} while (end.tv_sec - start.tv_sec < 2);
printf("%f\n", count / elapsed(&start, &end));
}
}
return 0;
}
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]);
}
}
int main(int argc, char **argv)
{
int fd = drm_open_driver(DRIVER_INTEL);
enum map {CPU, GTT, WC} map = CPU;
enum dir {READ, WRITE, CLEAR, FAULT} dir = READ;
int tiling = I915_TILING_NONE;
struct timespec start, end;
void *buf = malloc(OBJECT_SIZE);
uint32_t handle;
void *ptr, *src, *dst;
int reps = 1;
int loops;
int c;
while ((c = getopt (argc, argv, "m:d:r:t:")) != -1) {
switch (c) {
case 'm':
if (strcmp(optarg, "cpu") == 0)
map = CPU;
else if (strcmp(optarg, "gtt") == 0)
map = GTT;
else if (strcmp(optarg, "wc") == 0)
map = WC;
else
abort();
break;
case 'd':
if (strcmp(optarg, "read") == 0)
dir = READ;
else if (strcmp(optarg, "write") == 0)
dir = WRITE;
else if (strcmp(optarg, "clear") == 0)
dir = CLEAR;
else if (strcmp(optarg, "fault") == 0)
dir = FAULT;
else
abort();
break;
case 't':
if (strcmp(optarg, "x") == 0)
tiling = I915_TILING_X;
else if (strcmp(optarg, "y") == 0)
tiling = I915_TILING_Y;
else if (strcmp(optarg, "none") == 0)
tiling = I915_TILING_NONE;
else
abort();
break;
case 'r':
reps = atoi(optarg);
if (reps < 1)
reps = 1;
break;
default:
break;
}
}
handle = gem_create(fd, OBJECT_SIZE);
switch (map) {
case CPU:
ptr = gem_mmap__cpu(fd, handle, 0, OBJECT_SIZE, PROT_WRITE);
gem_set_domain(fd, handle, I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
break;
case GTT:
ptr = gem_mmap__gtt(fd, handle, OBJECT_SIZE, PROT_WRITE);
gem_set_domain(fd, handle, I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
break;
case WC:
ptr = gem_mmap__wc(fd, handle, 0, OBJECT_SIZE, PROT_WRITE);
gem_set_domain(fd, handle, I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
break;
default:
abort();
}
gem_set_tiling(fd, handle, tiling, 512);
if (dir == READ) {
src = ptr;
dst = buf;
} else {
src = buf;
dst = ptr;
}
clock_gettime(CLOCK_MONOTONIC, &start);
switch (dir) {
case CLEAR:
case FAULT:
memset(dst, 0, OBJECT_SIZE);
break;
default:
memcpy(dst, src, OBJECT_SIZE);
break;
}
clock_gettime(CLOCK_MONOTONIC, &end);
loops = 2 / elapsed(&start, &end);
while (reps--) {
clock_gettime(CLOCK_MONOTONIC, &start);
for (c = 0; c < loops; c++) {
int page;
switch (dir) {
case CLEAR:
memset(dst, 0, OBJECT_SIZE);
break;
case FAULT:
munmap(ptr, OBJECT_SIZE);
switch (map) {
case CPU:
ptr = gem_mmap__cpu(fd, handle, 0, OBJECT_SIZE, PROT_WRITE);
break;
case GTT:
ptr = gem_mmap__gtt(fd, handle, OBJECT_SIZE, PROT_WRITE);
break;
case WC:
ptr = gem_mmap__wc(fd, handle, 0, OBJECT_SIZE, PROT_WRITE);
break;
default:
abort();
}
for (page = 0; page < OBJECT_SIZE; page += 4096) {
uint32_t *x = (uint32_t *)ptr + page/4;
__asm__ __volatile__("": : :"memory");
page += *x; /* should be zero! */
}
break;
default:
memcpy(dst, src, OBJECT_SIZE);
break;
}
}
clock_gettime(CLOCK_MONOTONIC, &end);
printf("%7.3f\n", OBJECT_SIZE / elapsed(&start, &end) * loops / (1024*1024));
}
return 0;
}
static void gem_userptr_sync(int fd, uint32_t handle)
{
gem_set_domain(fd, handle, I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
}
int main(int argc, char **argv)
{
int fd = drm_open_driver(DRIVER_INTEL);
enum map {CPU, GTT, WC} map = CPU;
enum dir {READ, WRITE, CLEAR, FAULT} dir = READ;
int tiling = I915_TILING_NONE;
void *buf = malloc(OBJECT_SIZE);
uint32_t handle;
void *ptr, *src, *dst;
int reps = 13;
int c, size;
while ((c = getopt (argc, argv, "m:d:r:t:")) != -1) {
switch (c) {
case 'm':
if (strcmp(optarg, "cpu") == 0)
map = CPU;
else if (strcmp(optarg, "gtt") == 0)
map = GTT;
else if (strcmp(optarg, "wc") == 0)
map = WC;
else
abort();
break;
case 'd':
if (strcmp(optarg, "read") == 0)
dir = READ;
else if (strcmp(optarg, "write") == 0)
dir = WRITE;
else if (strcmp(optarg, "clear") == 0)
dir = CLEAR;
else if (strcmp(optarg, "fault") == 0)
dir = FAULT;
else
abort();
break;
case 't':
if (strcmp(optarg, "x") == 0)
tiling = I915_TILING_X;
else if (strcmp(optarg, "y") == 0)
tiling = I915_TILING_Y;
else if (strcmp(optarg, "none") == 0)
tiling = I915_TILING_NONE;
else
abort();
break;
case 'r':
reps = atoi(optarg);
if (reps < 1)
reps = 1;
break;
default:
break;
}
}
handle = gem_create(fd, OBJECT_SIZE);
switch (map) {
case CPU:
ptr = gem_mmap__cpu(fd, handle, 0, OBJECT_SIZE, PROT_WRITE);
gem_set_domain(fd, handle, I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
break;
case GTT:
ptr = gem_mmap__gtt(fd, handle, OBJECT_SIZE, PROT_WRITE);
gem_set_domain(fd, handle, I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
break;
case WC:
ptr = gem_mmap__wc(fd, handle, 0, OBJECT_SIZE, PROT_WRITE);
gem_set_domain(fd, handle, I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
break;
default:
abort();
}
gem_set_tiling(fd, handle, tiling, 512);
if (dir == READ) {
src = ptr;
dst = buf;
} else {
src = buf;
dst = ptr;
}
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;
int page;
clock_gettime(CLOCK_MONOTONIC, &start);
switch (dir) {
case CLEAR:
memset(dst, 0, size);
break;
case FAULT:
for (page = 0; page < OBJECT_SIZE; page += 4096) {
uint32_t *x = (uint32_t *)ptr + page/4;
page += *x; /* should be zero! */
}
break;
default:
memcpy(dst, src, size);
break;
}
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;
}
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;
}
