/*
* init_opencl_context_and_shmem
*
* We can have performance gain using asynchronous DMA transfer when data
* chunk it moved to OpenCL device from host machine, however, it requires
* preparations to ensure the memory region to be copied to/from is pinned
* on RAM; not swapped out. OpenCL provides an interface to map a certain
* host address area as pinned buffer object, even though its size is
* restricted to CL_DEVICE_MAX_MEM_ALLOC_SIZE parameter. Usually, it is
* much less than size of shared memory to be assigned to PG-Strom, around
* 500MB - 2GB in typical GPU/MIC device. So, we need to split a flat
* continuous memory into several 'zones' to pin it using OpenCL interface.
* Because it is a job of OpenCL intermediation server to collect properties
* of devices, and this server shall be launched post initialization stage,
* we also have to acquire and pin the shared memory region in the context
* of OpenCL intermediation server, not postmaster itself.
*/
static void
init_opencl_context_and_shmem(void)
{
Size zone_length = LONG_MAX;
cl_int i, rc;
/*
* Create an OpenCL context
*/
opencl_context = clCreateContext(NULL,
opencl_num_devices,
opencl_devices,
NULL,
NULL,
&rc);
if (rc != CL_SUCCESS)
elog(ERROR, "clCreateContext failed: %s", opencl_strerror(rc));
/*
* Create an OpenCL command queue for each device
*/
for (i=0; i < opencl_num_devices; i++)
{
const pgstrom_device_info *dev_info = pgstrom_get_device_info(i);
opencl_cmdq[i] =
clCreateCommandQueue(opencl_context,
opencl_devices[i],
CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE |
CL_QUEUE_PROFILING_ENABLE,
&rc);
if (rc != CL_SUCCESS)
elog(ERROR, "clCreateCommandQueue failed: %s",
opencl_strerror(rc));
if (zone_length > dev_info->dev_max_mem_alloc_size)
zone_length = (dev_info->dev_max_mem_alloc_size &
~((1UL << 20) - 1));
}
/* Lock shared memory of PG-Strom's private area */
pgstrom_setup_shmem(zone_length, on_shmem_zone_callback);
/* Lock shared memory of shared buffer area */
if (!on_shmem_zone_callback(BufferBlocks,
NBuffers * (Size) BLCKSZ,
"buffer", false))
{
Size total_size = NBuffers * (Size) BLCKSZ;
Size offset;
Assert((zone_length & (BLCKSZ - 1)) == 0);
for (offset = 0; offset < total_size; offset += zone_length)
{
on_shmem_zone_callback(BufferBlocks + offset,
Min(zone_length, total_size - offset),
"buffer", true);
}
}
}
Datum
pgstrom_opencl_device_info(PG_FUNCTION_ARGS)
{
FuncCallContext *fncxt;
Datum values[4];
bool isnull[4];
HeapTuple tuple;
uint32 dindex;
uint32 pindex;
const pgstrom_device_info *dinfo;
const char *key;
const char *value;
char buf[256];
int ofs = 0;
if (SRF_IS_FIRSTCALL())
{
TupleDesc tupdesc;
MemoryContext oldcxt;
fncxt = SRF_FIRSTCALL_INIT();
oldcxt = MemoryContextSwitchTo(fncxt->multi_call_memory_ctx);
tupdesc = CreateTemplateTupleDesc(4, false);
TupleDescInitEntry(tupdesc, (AttrNumber) 1, "dnum",
INT4OID, -1, 0);
TupleDescInitEntry(tupdesc, (AttrNumber) 2, "pnum",
INT4OID, -1, 0);
TupleDescInitEntry(tupdesc, (AttrNumber) 3, "property",
TEXTOID, -1, 0);
TupleDescInitEntry(tupdesc, (AttrNumber) 4, "value",
TEXTOID, -1, 0);
fncxt->tuple_desc = BlessTupleDesc(tupdesc);
fncxt->user_fctx = 0;
MemoryContextSwitchTo(oldcxt);
}
fncxt = SRF_PERCALL_SETUP();
dindex = fncxt->call_cntr / 55;
pindex = fncxt->call_cntr % 55;
if (dindex == pgstrom_get_device_nums())
SRF_RETURN_DONE(fncxt);
dinfo = pgstrom_get_device_info(dindex);
Assert(dinfo != NULL);
switch (pindex)
{
case 0:
key = "platform index";
value = psprintf("%u", dinfo->pl_info->pl_index);
break;
case 1:
key = "platform profile";
value = dinfo->pl_info->pl_profile;
break;
case 2:
key = "platform version";
value = dinfo->pl_info->pl_version;
break;
case 3:
key = "platform name";
value = dinfo->pl_info->pl_name;
break;
case 4:
key = "platform vendor";
value = dinfo->pl_info->pl_vendor;
break;
case 5:
key = "platform extensions";
value = dinfo->pl_info->pl_extensions;
break;
case 6:
key = "address bits";
value = psprintf("%u", dinfo->dev_address_bits);
break;
case 7:
key = "device available";
value = dinfo->dev_available ? "yes" : "no";
break;
case 8:
key = "compiler available";
value = dinfo->dev_compiler_available ? "yes" : "no";
break;
case 9:
key = "double fp config";
value = fp_config_to_cstring(dinfo->dev_double_fp_config);
break;
case 10:
key = "little endian";
value = dinfo->dev_endian_little ? "yes" : "no";
break;
case 11:
key = "error correction support";
value = dinfo->dev_error_correction_support ? "yes" : "no";
break;
case 12:
key = "execution capabilities";
if (dinfo->dev_execution_capabilities & CL_EXEC_KERNEL)
ofs += sprintf(buf + ofs, "OpenCL");
if (dinfo->dev_execution_capabilities & CL_EXEC_NATIVE_KERNEL)
ofs += sprintf(buf + ofs, "%sNative", ofs > 0 ? ", " : "");
value = buf;
break;
case 13:
key = "device extensions";
value = dinfo->dev_device_extensions;
break;
case 14:
key = "global mem cache size";
value = psprintf("%lu", dinfo->dev_global_mem_cache_size);
break;
case 15:
key = "global mem cache type";
switch (dinfo->dev_global_mem_cache_type)
{
case CL_NONE:
value = "none";
break;
case CL_READ_ONLY_CACHE:
value = "read only";
break;
case CL_READ_WRITE_CACHE:
value = "read write";
break;
default:
value = "???";
break;
}
break;
case 16:
key = "global mem cacheline size";
value = psprintf("%u", dinfo->dev_global_mem_cacheline_size);
break;
case 17:
key = "global mem size";
value = psprintf("%lu", dinfo->dev_global_mem_size);
break;
case 18:
key = "host unified memory";
value = dinfo->dev_host_unified_memory ? "yes" : "no";
break;
case 19:
key = "local mem size";
value = psprintf("%lu", dinfo->dev_local_mem_size);
break;
case 20:
key = "local mem type";
switch (dinfo->dev_local_mem_type)
{
case CL_LOCAL:
value = "local";
break;
case CL_GLOBAL:
value = "global";
break;
case CL_NONE:
value = "none";
break;
default:
value = "???";
break;
}
break;
case 21:
key = "max clock frequency";
value = psprintf("%u", dinfo->dev_max_clock_frequency);
break;
case 22:
key = "max compute units";
value = psprintf("%u", dinfo->dev_max_compute_units);
break;
case 23:
key = "max constant args";
value = psprintf("%u", dinfo->dev_max_constant_args);
break;
case 24:
key = "max constant buffer size";
value = psprintf("%lu", dinfo->dev_max_constant_buffer_size);
break;
case 25:
key = "max mem alloc size";
value = psprintf("%lu", dinfo->dev_max_mem_alloc_size);
break;
case 26:
key = "max parameter size";
value = psprintf("%lu", dinfo->dev_max_parameter_size);
break;
case 27:
key = "max samplers";
value = psprintf("%u", dinfo->dev_max_samplers);
break;
case 28:
key = "max work group size";
value = psprintf("%zu", dinfo->dev_max_work_group_size);
break;
case 29:
key = "max work group dimensions";
value = psprintf("%u", dinfo->dev_max_work_item_dimensions);
break;
case 30:
key = "max work item sizes";
value = psprintf("{%zu, %zu, %zu}",
dinfo->dev_max_work_item_sizes[0],
dinfo->dev_max_work_item_sizes[1],
dinfo->dev_max_work_item_sizes[2]);
break;
case 31:
key = "mem base address align";
value = psprintf("%u", dinfo->dev_mem_base_addr_align);
break;
case 32:
key = "device name";
value = dinfo->dev_name;
break;
case 33:
key = "native vector width (char)";
value = psprintf("%u", dinfo->dev_native_vector_width_char);
break;
case 34:
key = "native vector width (short)";
value = psprintf("%u", dinfo->dev_native_vector_width_short);
break;
case 35:
key = "native vector width (int)";
value = psprintf("%u", dinfo->dev_native_vector_width_int);
break;
case 36:
key = "native vector width (long)";
value = psprintf("%u", dinfo->dev_native_vector_width_long);
break;
case 37:
key = "native vector width (float)";
value = psprintf("%u", dinfo->dev_native_vector_width_float);
break;
case 38:
key = "native vector width (double)";
value = psprintf("%u", dinfo->dev_native_vector_width_double);
break;
case 39:
key = "opencl c version";
value = dinfo->dev_opencl_c_version;
break;
case 40:
key = "preferred vector width (char)";
value = psprintf("%u", dinfo->dev_preferred_vector_width_char);
break;
case 41:
key = "preferred vector width (short)";
value = psprintf("%u", dinfo->dev_preferred_vector_width_short);
break;
case 42:
key = "preferred vector width (int)";
value = psprintf("%u", dinfo->dev_preferred_vector_width_int);
break;
case 43:
key = "preferred vector width (long)";
value = psprintf("%u", dinfo->dev_preferred_vector_width_long);
break;
case 44:
key = "preferred vector width (float)";
value = psprintf("%u", dinfo->dev_preferred_vector_width_float);
break;
case 45:
key = "preferred vector width (double)";
value = psprintf("%u", dinfo->dev_preferred_vector_width_double);
break;
case 46:
key = "device profile";
value = dinfo->dev_profile;
break;
case 47:
key = "profiling timer resolution";
value = psprintf("%zu", dinfo->dev_profiling_timer_resolution);
break;
case 48:
key = "command queue properties";
if (dinfo->dev_queue_properties &
CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE)
ofs += sprintf(buf, "%sout of order", ofs > 0 ? ", " : "");
if (dinfo->dev_queue_properties & CL_QUEUE_PROFILING_ENABLE)
ofs += sprintf(buf, "%sprofiling", ofs > 0 ? ", " : "");
value = buf;
break;
case 49:
key = "single fp config";
value = fp_config_to_cstring(dinfo->dev_single_fp_config);
break;
case 50:
key = "device type";
if (dinfo->dev_type & CL_DEVICE_TYPE_CPU)
ofs += sprintf(buf, "%scpu", ofs > 0 ? ", " : "");
if (dinfo->dev_type & CL_DEVICE_TYPE_GPU)
ofs += sprintf(buf, "%sgpu", ofs > 0 ? ", " : "");
if (dinfo->dev_type & CL_DEVICE_TYPE_ACCELERATOR)
ofs += sprintf(buf, "%saccelerator", ofs > 0 ? ", " : "");
if (dinfo->dev_type & CL_DEVICE_TYPE_DEFAULT)
ofs += sprintf(buf, "%sdefault", ofs > 0 ? ", " : "");
if (dinfo->dev_type & CL_DEVICE_TYPE_CUSTOM)
ofs += sprintf(buf, "%scustom", ofs > 0 ? ", " : "");
value = buf;
break;
case 51:
key = "device vendor";
value = dinfo->dev_vendor;
break;
case 52:
key = "device vendor id";
value = psprintf("%u", dinfo->dev_vendor_id);
break;
case 53:
key = "device version";
value = dinfo->dev_version;
break;
case 54:
key = "driver version";
value = dinfo->driver_version;
break;
default:
elog(ERROR, "unexpected property index");
break;
}
memset(isnull, 0, sizeof(isnull));
values[0] = Int32GetDatum(dindex);
values[1] = Int32GetDatum(pindex);
values[2] = CStringGetTextDatum(key);
values[3] = CStringGetTextDatum(value);
tuple = heap_form_tuple(fncxt->tuple_desc, values, isnull);
SRF_RETURN_NEXT(fncxt, HeapTupleGetDatum(tuple));
}
