static int r600_get_compute_param(struct pipe_screen *screen,
enum pipe_compute_cap param,
void *ret)
{
struct r600_common_screen *rscreen = (struct r600_common_screen *)screen;
//TODO: select these params by asic
switch (param) {
case PIPE_COMPUTE_CAP_IR_TARGET: {
const char *gpu;
const char *triple;
if (rscreen->family <= CHIP_ARUBA || HAVE_LLVM < 0x0306) {
triple = "r600--";
} else {
triple = "amdgcn--";
}
switch(rscreen->family) {
/* Clang < 3.6 is missing Hainan in its list of
* GPUs, so we need to use the name of a similar GPU.
*/
#if HAVE_LLVM < 0x0306
case CHIP_HAINAN:
gpu = "oland";
break;
#endif
default:
gpu = r600_get_llvm_processor_name(rscreen->family);
break;
}
if (ret) {
sprintf(ret, "%s-%s", gpu, triple);
}
/* +2 for dash and terminating NIL byte */
return (strlen(triple) + strlen(gpu) + 2) * sizeof(char);
}
case PIPE_COMPUTE_CAP_GRID_DIMENSION:
if (ret) {
uint64_t *grid_dimension = ret;
grid_dimension[0] = 3;
}
return 1 * sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
if (ret) {
uint64_t *grid_size = ret;
grid_size[0] = 65535;
grid_size[1] = 65535;
grid_size[2] = 1;
}
return 3 * sizeof(uint64_t) ;
case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
if (ret) {
uint64_t *block_size = ret;
block_size[0] = 256;
block_size[1] = 256;
block_size[2] = 256;
}
return 3 * sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
if (ret) {
uint64_t *max_threads_per_block = ret;
*max_threads_per_block = 256;
}
return sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
if (ret) {
uint64_t *max_global_size = ret;
uint64_t max_mem_alloc_size;
r600_get_compute_param(screen,
PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE,
&max_mem_alloc_size);
/* In OpenCL, the MAX_MEM_ALLOC_SIZE must be at least
* 1/4 of the MAX_GLOBAL_SIZE. Since the
* MAX_MEM_ALLOC_SIZE is fixed for older kernels,
* make sure we never report more than
* 4 * MAX_MEM_ALLOC_SIZE.
*/
*max_global_size = MIN2(4 * max_mem_alloc_size,
rscreen->info.gart_size +
rscreen->info.vram_size);
}
return sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
if (ret) {
uint64_t *max_local_size = ret;
/* Value reported by the closed source driver. */
*max_local_size = 32768;
}
return sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE:
if (ret) {
uint64_t *max_input_size = ret;
/* Value reported by the closed source driver. */
*max_input_size = 1024;
}
return sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
if (ret) {
uint64_t *max_mem_alloc_size = ret;
/* XXX: The limit in older kernels is 256 MB. We
* should add a query here for newer kernels.
*/
*max_mem_alloc_size = 256 * 1024 * 1024;
}
return sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
if (ret) {
uint32_t *max_clock_frequency = ret;
*max_clock_frequency = rscreen->info.max_sclk;
}
return sizeof(uint32_t);
case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
if (ret) {
uint32_t *max_compute_units = ret;
*max_compute_units = rscreen->info.max_compute_units;
}
return sizeof(uint32_t);
case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
if (ret) {
uint32_t *images_supported = ret;
*images_supported = 0;
}
return sizeof(uint32_t);
case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE:
break; /* unused */
case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
if (ret) {
uint32_t *subgroup_size = ret;
*subgroup_size = r600_wavefront_size(rscreen->family);
}
return sizeof(uint32_t);
}
fprintf(stderr, "unknown PIPE_COMPUTE_CAP %d\n", param);
return 0;
}
static int r600_get_compute_param(struct pipe_screen *screen,
enum pipe_shader_ir ir_type,
enum pipe_compute_cap param,
void *ret)
{
struct r600_common_screen *rscreen = (struct r600_common_screen *)screen;
//TODO: select these params by asic
switch (param) {
case PIPE_COMPUTE_CAP_IR_TARGET: {
const char *gpu;
const char *triple = "r600--";
gpu = r600_get_llvm_processor_name(rscreen->family);
if (ret) {
sprintf(ret, "%s-%s", gpu, triple);
}
/* +2 for dash and terminating NIL byte */
return (strlen(triple) + strlen(gpu) + 2) * sizeof(char);
}
case PIPE_COMPUTE_CAP_GRID_DIMENSION:
if (ret) {
uint64_t *grid_dimension = ret;
grid_dimension[0] = 3;
}
return 1 * sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
if (ret) {
uint64_t *grid_size = ret;
grid_size[0] = 65535;
grid_size[1] = 65535;
grid_size[2] = 65535;
}
return 3 * sizeof(uint64_t) ;
case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
if (ret) {
uint64_t *block_size = ret;
unsigned threads_per_block = get_max_threads_per_block(rscreen, ir_type);
block_size[0] = threads_per_block;
block_size[1] = threads_per_block;
block_size[2] = threads_per_block;
}
return 3 * sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
if (ret) {
uint64_t *max_threads_per_block = ret;
*max_threads_per_block = get_max_threads_per_block(rscreen, ir_type);
}
return sizeof(uint64_t);
case PIPE_COMPUTE_CAP_ADDRESS_BITS:
if (ret) {
uint32_t *address_bits = ret;
address_bits[0] = 32;
}
return 1 * sizeof(uint32_t);
case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
if (ret) {
uint64_t *max_global_size = ret;
uint64_t max_mem_alloc_size;
r600_get_compute_param(screen, ir_type,
PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE,
&max_mem_alloc_size);
/* In OpenCL, the MAX_MEM_ALLOC_SIZE must be at least
* 1/4 of the MAX_GLOBAL_SIZE. Since the
* MAX_MEM_ALLOC_SIZE is fixed for older kernels,
* make sure we never report more than
* 4 * MAX_MEM_ALLOC_SIZE.
*/
*max_global_size = MIN2(4 * max_mem_alloc_size,
MAX2(rscreen->info.gart_size,
rscreen->info.vram_size));
}
return sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
if (ret) {
uint64_t *max_local_size = ret;
/* Value reported by the closed source driver. */
*max_local_size = 32768;
}
return sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE:
if (ret) {
uint64_t *max_input_size = ret;
/* Value reported by the closed source driver. */
*max_input_size = 1024;
}
return sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
if (ret) {
uint64_t *max_mem_alloc_size = ret;
*max_mem_alloc_size = rscreen->info.max_alloc_size;
}
return sizeof(uint64_t);
case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
if (ret) {
uint32_t *max_clock_frequency = ret;
*max_clock_frequency = rscreen->info.max_shader_clock;
}
return sizeof(uint32_t);
case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
if (ret) {
uint32_t *max_compute_units = ret;
*max_compute_units = rscreen->info.num_good_compute_units;
}
return sizeof(uint32_t);
case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
if (ret) {
uint32_t *images_supported = ret;
*images_supported = 0;
}
return sizeof(uint32_t);
case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE:
break; /* unused */
case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
if (ret) {
uint32_t *subgroup_size = ret;
*subgroup_size = r600_wavefront_size(rscreen->family);
}
return sizeof(uint32_t);
case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
if (ret) {
uint64_t *max_variable_threads_per_block = ret;
*max_variable_threads_per_block = 0;
}
return sizeof(uint64_t);
}
fprintf(stderr, "unknown PIPE_COMPUTE_CAP %d\n", param);
return 0;
}