Skip to content

Commit

Permalink
ocl: discover device properties once and simplified setting up WG-size
Browse files Browse the repository at this point in the history 
* Store cl_device_id in c_dbcsr_acc_opencl_device_t and rely on it (instead of querying it repeatedly).
* Make maximum size of work-group (WG) and preferred multiple available as part of device structure.
* Discover subgroups. Avoid some compile-time decision (OpenCL 3.0).
* Simplify handling WG-size (and SG-size).
* Code cleanup (_DEBUG).
  • Loading branch information
@hfp
hfp committed Mar 26, 2024
1 parent d1157c4 commit 6db5b28
Show file tree
Hide file tree
Showing 5 changed files with 392 additions and 447 deletions.
144 changes: 74 additions & 70 deletions src/acc/opencl/acc_opencl.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -158,7 +158,7 @@ LIBXSMM_ATTRIBUTE_DTOR void c_dbcsr_acc_opencl_finalize(void) {
for (i = 0; i < ACC_OPENCL_MAXNDEVS; ++i) {
const cl_device_id device_id = c_dbcsr_acc_opencl_config.devices[i];
if (NULL != device_id) {
# if defined(CL_VERSION_1_2) && defined(_DEBUG)
# if defined(CL_VERSION_1_2)
ACC_OPENCL_EXPECT(EXIT_SUCCESS == clReleaseDevice(device_id));
# endif
/* c_dbcsr_acc_opencl_create_context scans for non-NULL devices */
Expand Down Expand Up @@ -665,13 +665,10 @@ int c_dbcsr_acc_finalize(void) {
assert(c_dbcsr_acc_opencl_config.ndevices < ACC_OPENCL_MAXNDEVS);
if (0 != c_dbcsr_acc_opencl_config.ndevices && NULL != cleanup) {
if (0 != c_dbcsr_acc_opencl_config.verbosity) {
cl_device_id device = NULL;
int d;
fprintf(stderr, "INFO ACC/OpenCL: pid=%u nthreads=%i", libxsmm_get_pid(), c_dbcsr_acc_opencl_config.nthreads);
if (NULL != c_dbcsr_acc_opencl_config.device.context &&
EXIT_SUCCESS ==
clGetContextInfo(c_dbcsr_acc_opencl_config.device.context, CL_CONTEXT_DEVICES, sizeof(cl_device_id), &device, NULL) &&
EXIT_SUCCESS == c_dbcsr_acc_opencl_device_id(device, NULL /*devid*/, &d))
EXIT_SUCCESS == c_dbcsr_acc_opencl_device_id(c_dbcsr_acc_opencl_config.device.id, NULL /*devid*/, &d))
{
fprintf(stderr, " device=%i", d);
}
Expand Down Expand Up @@ -969,24 +966,29 @@ int c_dbcsr_acc_opencl_create_context(cl_device_id active_id, cl_context* contex


int c_dbcsr_acc_opencl_set_active_device(ACC_OPENCL_LOCKTYPE* lock, int device_id) {
/* accessing devices is thread-safe (array is fixed after initialization) */
const cl_device_id active_id = c_dbcsr_acc_opencl_config.devices[device_id];
int result = EXIT_SUCCESS;
cl_device_id active_id = NULL, context_id = NULL;
assert(c_dbcsr_acc_opencl_config.ndevices < ACC_OPENCL_MAXNDEVS);
assert(0 <= device_id && device_id < c_dbcsr_acc_opencl_config.ndevices);
/* accessing devices is thread-safe (array is fixed after initialization) */
active_id = c_dbcsr_acc_opencl_config.devices[device_id];
if (NULL != active_id) {
cl_device_id context_id = NULL;
cl_context context = NULL;
if (NULL != lock) ACC_OPENCL_ACQUIRE(lock);
context = c_dbcsr_acc_opencl_config.device.context;
context_id = c_dbcsr_acc_opencl_config.device.id;
if (NULL != context) {
result = clGetContextInfo(context, CL_CONTEXT_DEVICES, sizeof(cl_device_id), &context_id, NULL);
if (EXIT_SUCCESS == result && active_id != context_id) {
assert(NULL != context_id);
assert(NULL != context_id);
if (active_id != context_id) {
# if defined(CL_VERSION_1_2)
ACC_OPENCL_EXPECT(EXIT_SUCCESS == clReleaseDevice(context_id));
# endif
result = clReleaseContext(context);
context_id = NULL;
context = NULL;
}
}
assert(NULL == context_id || active_id == context_id);
if (EXIT_SUCCESS == result && active_id != context_id) {
result = c_dbcsr_acc_opencl_create_context(active_id, &context);
assert(NULL != context || EXIT_SUCCESS != result);
Expand All @@ -1001,6 +1003,8 @@ int c_dbcsr_acc_opencl_set_active_device(ACC_OPENCL_LOCKTYPE* lock, int device_i
&c_dbcsr_acc_opencl_config.device.type);
if (EXIT_SUCCESS == result) {
char devname[ACC_OPENCL_BUFFERSIZE] = "";
const char* const sgexts[] = {"cl_intel_required_subgroup_size", "cl_intel_subgroups", "cl_khr_subgroups"};
size_t sgsizes[16], nbytes = 0, sgmin = (size_t)-1, i;
# if defined(ACC_OPENCL_CMDAGR)
ACC_OPENCL_STREAM_PROPERTIES_TYPE properties[4] = {
CL_QUEUE_PROPERTIES, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, 0 /* terminator */
Expand Down Expand Up @@ -1037,6 +1041,34 @@ int c_dbcsr_acc_opencl_set_active_device(ACC_OPENCL_LOCKTYPE* lock, int device_i
{
c_dbcsr_acc_opencl_config.device.unified = CL_FALSE;
}
if (EXIT_SUCCESS != clGetDeviceInfo(active_id, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t),
c_dbcsr_acc_opencl_config.device.wgsize, NULL))
{
c_dbcsr_acc_opencl_config.device.wgsize[0] = 1;
}
if (EXIT_SUCCESS != clGetDeviceInfo(active_id, 4199 /*CL_DEVICE_PREFERRED_WORK_GROUP_SIZE_MULTIPLE*/, sizeof(size_t),
c_dbcsr_acc_opencl_config.device.wgsize + 1, NULL)) /* CL_VERSION_3_0 */
{
c_dbcsr_acc_opencl_config.device.wgsize[1] = 1;
}
assert(0 == c_dbcsr_acc_opencl_config.device.wgsize[2]);
if (EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(active_id, sgexts, 2) &&
EXIT_SUCCESS ==
clGetDeviceInfo(active_id, 0x4108 /*CL_DEVICE_SUB_GROUP_SIZES_INTEL*/, sizeof(sgsizes), sgsizes, &nbytes))
{
for (i = 0; (i * sizeof(size_t)) < nbytes; ++i) {
const size_t sgsize = sgsizes[i];
if (sgsize < sgmin) sgmin = sgsize;
if (0 == (sgsize % c_dbcsr_acc_opencl_config.device.wgsize[1]) && c_dbcsr_acc_opencl_config.device.wgsize[2] < sgsize) {
if (c_dbcsr_acc_opencl_config.device.wgsize[1] < sgsize) c_dbcsr_acc_opencl_config.device.wgsize[1] = sgsize;
c_dbcsr_acc_opencl_config.device.wgsize[2] = sgsize;
}
}
if (0 != c_dbcsr_acc_opencl_config.device.wgsize[2]) c_dbcsr_acc_opencl_config.device.wgsize[2] = sgmin;
}
else {
c_dbcsr_acc_opencl_config.device.wgsize[2] = 0;
}
# if defined(ACC_OPENCL_MEM_DEVPTR)
if (0 != (4 & c_dbcsr_acc_opencl_config.xhints) && 2 <= *c_dbcsr_acc_opencl_config.device.std_level &&
0 != c_dbcsr_acc_opencl_config.device.intel && 0 == c_dbcsr_acc_opencl_config.device.unified &&
Expand Down Expand Up @@ -1076,7 +1108,12 @@ int c_dbcsr_acc_opencl_set_active_device(ACC_OPENCL_LOCKTYPE* lock, int device_i
c_dbcsr_acc_opencl_config.device.stream.queue = ACC_OPENCL_CREATE_COMMAND_QUEUE(context, active_id, properties, &result);
}
if (EXIT_SUCCESS == result) {
if (active_id != context_id) c_dbcsr_acc_opencl_config.device.context = context;
if (active_id != context_id) {
assert(active_id != c_dbcsr_acc_opencl_config.device.id);
c_dbcsr_acc_opencl_config.device.context = context;
c_dbcsr_acc_opencl_config.device.id = active_id;
}
assert(active_id == c_dbcsr_acc_opencl_config.device.id);
}
else memset(&c_dbcsr_acc_opencl_config.device, 0, sizeof(c_dbcsr_acc_opencl_config.device));
}
Expand Down Expand Up @@ -1117,41 +1154,8 @@ int c_dbcsr_acc_set_active_device(int device_id) {
}


int c_dbcsr_acc_opencl_wgsize(cl_device_id device, cl_kernel kernel, size_t* max_value, size_t* preferred_multiple) {
int result = (NULL != device && (NULL != preferred_multiple || NULL != max_value)) ? EXIT_SUCCESS : EXIT_FAILURE;
if (NULL != kernel) { /* kernel-specific */
if (NULL != max_value) {
ACC_OPENCL_CHECK(clGetKernelWorkGroupInfo(kernel, device, CL_KERNEL_WORK_GROUP_SIZE, sizeof(size_t), max_value, NULL),
"query maximum WG-size of kernel", result);
}
if (NULL != preferred_multiple) {
ACC_OPENCL_CHECK(clGetKernelWorkGroupInfo(
kernel, device, CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE, sizeof(size_t), preferred_multiple, NULL),
"query preferred multiple of WG-size of kernel", result);
}
}
else { /* device-specific */
if (NULL != max_value) {
ACC_OPENCL_CHECK(clGetDeviceInfo(device, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t), max_value, NULL),
"query maximum WG-size of device", result);
}
if (NULL != preferred_multiple) {
# if defined(CL_VERSION_3_0)
ACC_OPENCL_CHECK(
clGetDeviceInfo(device, CL_DEVICE_PREFERRED_WORK_GROUP_SIZE_MULTIPLE, sizeof(size_t), preferred_multiple, NULL),
"query preferred multiple of WG-size of device", result);
# else
*preferred_multiple = 1;
# endif
}
}
return result;
}


int c_dbcsr_acc_opencl_flags_atomics(const c_dbcsr_acc_opencl_device_t* devinfo, c_dbcsr_acc_opencl_atomic_fp_t kind,
const char* exts[], size_t* exts_maxlen, char flags[], size_t flags_maxlen) {
cl_device_id device_id = NULL;
size_t ext1, ext2;
int result = 0;
for (ext1 = 0; ext1 < (NULL != exts_maxlen ? *exts_maxlen : 0); ++ext1) {
Expand All @@ -1160,29 +1164,27 @@ int c_dbcsr_acc_opencl_flags_atomics(const c_dbcsr_acc_opencl_device_t* devinfo,
for (ext2 = ext1 + 1; ext2 < (NULL != exts_maxlen ? *exts_maxlen : 0); ++ext2) {
if (NULL == exts[ext2] || '\0' == *exts[ext2]) break;
}
if (NULL != devinfo && NULL != exts_maxlen && ext2 < *exts_maxlen &&
EXIT_SUCCESS == clGetContextInfo(devinfo->context, CL_CONTEXT_DEVICES, sizeof(cl_device_id), &device_id, NULL))
{
if (NULL != devinfo && NULL != exts_maxlen && ext2 < *exts_maxlen) {
const char* atomic_type = "";
switch (kind) {
case c_dbcsr_acc_opencl_atomic_fp_64: {
exts[ext1] = "cl_khr_fp64 cl_khr_int64_base_atomics cl_khr_int64_extended_atomics";
if (2 <= *devinfo->std_level && EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(device_id, exts, ext2)) {
if (2 <= *devinfo->std_level && EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(devinfo->id, exts, ext2)) {
atomic_type = "-DTA=long -DTA2=atomic_long -DTF=atomic_double";
}
else {
exts[ext1] = "cl_khr_fp64 cl_khr_int64_base_atomics";
if (EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(device_id, exts, ext2)) {
if (EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(devinfo->id, exts, ext2)) {
atomic_type = "-DTA=long";
}
else { /* fallback */
exts[ext1] = "cl_khr_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics";
if (2 <= *devinfo->std_level && EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(device_id, exts, ext2)) {
if (2 <= *devinfo->std_level && EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(devinfo->id, exts, ext2)) {
atomic_type = "-DATOMIC32_ADD64 -DTA=int -DTA2=atomic_int -DTF=atomic_double";
}
else {
exts[ext1] = "cl_khr_fp64 cl_khr_global_int32_base_atomics";
if (EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(device_id, exts, ext2)) {
if (EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(devinfo->id, exts, ext2)) {
atomic_type = "-DATOMIC32_ADD64 -DTA=int";
}
else kind = c_dbcsr_acc_opencl_atomic_fp_no;
Expand All @@ -1192,13 +1194,13 @@ int c_dbcsr_acc_opencl_flags_atomics(const c_dbcsr_acc_opencl_device_t* devinfo,
} break;
case c_dbcsr_acc_opencl_atomic_fp_32: {
exts[ext1] = "cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics";
if (2 <= *devinfo->std_level && EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(device_id, exts, ext2)) {
if (2 <= *devinfo->std_level && EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(devinfo->id, exts, ext2)) {
exts[ext2] = "cl_khr_int64_base_atomics cl_khr_int64_extended_atomics";
atomic_type = "-DTA=int -DTA2=atomic_int -DTF=atomic_float";
}
else {
exts[ext1] = "cl_khr_global_int32_base_atomics";
if (EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(device_id, exts, ext2)) {
if (EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(devinfo->id, exts, ext2)) {
exts[ext2] = "cl_khr_int64_base_atomics";
atomic_type = "-DTA=int";
}
Expand All @@ -1223,7 +1225,7 @@ int c_dbcsr_acc_opencl_flags_atomics(const c_dbcsr_acc_opencl_device_t* devinfo,
const int force_atomics = ((NULL == env_atomics || '\0' == *env_atomics) ? 0 : atoi(env_atomics));
if (NULL == env_atomics || '\0' == *env_atomics || 0 != force_atomics) {
cl_bitfield fp_atomics = 0;
if (EXIT_SUCCESS == clGetDeviceInfo(device_id,
if (EXIT_SUCCESS == clGetDeviceInfo(devinfo->id,
(cl_device_info)(c_dbcsr_acc_opencl_atomic_fp_64 == kind ? 0x4232 : 0x4231), sizeof(cl_bitfield),
&fp_atomics, NULL) &&
0 != (/*add*/ (1 << 1) & fp_atomics))
Expand Down Expand Up @@ -1329,11 +1331,7 @@ int c_dbcsr_acc_opencl_kernel(int source_is_file, const char source[], const cha
cl_kernel* kernel) {
char buffer[ACC_OPENCL_BUFFERSIZE] = "", buffer_name[ACC_OPENCL_MAXSTRLEN * 2];
int ok = EXIT_SUCCESS, source_is_cl = 1, nchar;
cl_device_id active_id = NULL;
int result = ((NULL != source && NULL != kernel_name && '\0' != *kernel_name)
? clGetContextInfo(
c_dbcsr_acc_opencl_config.device.context, CL_CONTEXT_DEVICES, sizeof(cl_device_id), &active_id, NULL)
: EXIT_FAILURE);
int result = ((NULL != source && NULL != kernel_name && '\0' != *kernel_name) ? EXIT_SUCCESS : EXIT_FAILURE);
cl_program program = NULL;
FILE* file_src = NULL;
size_t size_src = 0;
Expand Down Expand Up @@ -1408,7 +1406,7 @@ int c_dbcsr_acc_opencl_kernel(int source_is_file, const char source[], const cha
else break;
}
# if !defined(NDEBUG)
if (EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(active_id, (const char* const*)&ext, 1))
if (EXIT_SUCCESS == c_dbcsr_acc_opencl_device_ext(c_dbcsr_acc_opencl_config.device.id, (const char* const*)&ext, 1))
# endif
{ /* NDEBUG: assume given extension is supported (confirmed upfront) */
if (NULL == line) { /* extension is not already part of source */
Expand Down Expand Up @@ -1504,7 +1502,8 @@ int c_dbcsr_acc_opencl_kernel(int source_is_file, const char source[], const cha
assert(NULL != program);
result = c_dbcsr_acc_opencl_flags(build_params, build_options, try_build_options, buffer, sizeof(buffer));
if (EXIT_SUCCESS == result) {
result = clBuildProgram(program, 1 /*num_devices*/, &active_id, buffer, NULL /*callback*/, NULL /*user_data*/);
result = clBuildProgram(
program, 1 /*num_devices*/, &c_dbcsr_acc_opencl_config.device.id, buffer, NULL /*callback*/, NULL /*user_data*/);
}
if (EXIT_SUCCESS != result && NULL != try_build_options && '\0' != *try_build_options) {
result = c_dbcsr_acc_opencl_flags(build_params, build_options, NULL /*try_build_options*/, buffer, sizeof(buffer));
Expand All @@ -1513,7 +1512,8 @@ int c_dbcsr_acc_opencl_kernel(int source_is_file, const char source[], const cha
program = clCreateProgramWithSource(c_dbcsr_acc_opencl_config.device.context, 1 /*nlines*/, &ext_source, NULL, &result);
assert(EXIT_SUCCESS != result || NULL != program);
if (EXIT_SUCCESS == result) {
result = clBuildProgram(program, 1 /*num_devices*/, &active_id, buffer, NULL /*callback*/, NULL /*user_data*/);
result = clBuildProgram(
program, 1 /*num_devices*/, &c_dbcsr_acc_opencl_config.device.id, buffer, NULL /*callback*/, NULL /*user_data*/);
}
}
ok = EXIT_FAILURE;
Expand Down Expand Up @@ -1568,14 +1568,15 @@ int c_dbcsr_acc_opencl_kernel(int source_is_file, const char source[], const cha
else
# endif
{
program = clCreateProgramWithBinary(c_dbcsr_acc_opencl_config.device.context, 1, &active_id, &size_src,
(const unsigned char**)&source, NULL /*binary_status*/, &result);
program = clCreateProgramWithBinary(c_dbcsr_acc_opencl_config.device.context, 1, &c_dbcsr_acc_opencl_config.device.id,
&size_src, (const unsigned char**)&source, NULL /*binary_status*/, &result);
}
if (EXIT_SUCCESS == result) {
assert(NULL != program);
result = c_dbcsr_acc_opencl_flags(build_params, build_options, try_build_options, buffer, sizeof(buffer));
if (EXIT_SUCCESS == result) {
result = clBuildProgram(program, 1 /*num_devices*/, &active_id, buffer, NULL /*callback*/, NULL /*user_data*/);
result = clBuildProgram(
program, 1 /*num_devices*/, &c_dbcsr_acc_opencl_config.device.id, buffer, NULL /*callback*/, NULL /*user_data*/);
}
if (EXIT_SUCCESS != result && NULL != try_build_options && '\0' != *try_build_options) {
result = c_dbcsr_acc_opencl_flags(build_params, build_options, NULL /*try_build_options*/, buffer, sizeof(buffer));
Expand All @@ -1587,12 +1588,13 @@ int c_dbcsr_acc_opencl_kernel(int source_is_file, const char source[], const cha
else
# endif
{
program = clCreateProgramWithBinary(c_dbcsr_acc_opencl_config.device.context, 1, &active_id, &size_src,
(const unsigned char**)&source, NULL /*binary_status*/, &result);
program = clCreateProgramWithBinary(c_dbcsr_acc_opencl_config.device.context, 1, &c_dbcsr_acc_opencl_config.device.id,
&size_src, (const unsigned char**)&source, NULL /*binary_status*/, &result);
}
assert(EXIT_SUCCESS != result || NULL != program);
if (EXIT_SUCCESS == result) {
result = clBuildProgram(program, 1 /*num_devices*/, &active_id, buffer, NULL /*callback*/, NULL /*user_data*/);
result = clBuildProgram(
program, 1 /*num_devices*/, &c_dbcsr_acc_opencl_config.device.id, buffer, NULL /*callback*/, NULL /*user_data*/);
}
}
ok = EXIT_FAILURE;
Expand Down Expand Up @@ -1624,7 +1626,9 @@ int c_dbcsr_acc_opencl_kernel(int source_is_file, const char source[], const cha
*kernel = NULL;
}
if (2 <= c_dbcsr_acc_opencl_config.verbosity || 0 > c_dbcsr_acc_opencl_config.verbosity) {
if (EXIT_SUCCESS == clGetProgramBuildInfo(program, active_id, CL_PROGRAM_BUILD_LOG, ACC_OPENCL_BUFFERSIZE, buffer, NULL)) {
if (EXIT_SUCCESS == clGetProgramBuildInfo(program, c_dbcsr_acc_opencl_config.device.id, CL_PROGRAM_BUILD_LOG,
ACC_OPENCL_BUFFERSIZE, buffer, NULL))
{
const char* info = buffer;
while ('\0' != *info && NULL != strchr("\n\r\t ", *info)) ++info; /* remove preceding newline etc. */
assert(NULL != kernel_name && '\0' != *kernel_name);
Expand Down
10 changes: 8 additions & 2 deletions src/acc/opencl/acc_opencl.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -262,8 +262,16 @@ typedef struct c_dbcsr_acc_opencl_device_t {
char std_flag[16];
/** OpenCL support-level (major and minor). */
cl_int std_level[2], std_clevel[2];
/**
* Maximum size of workgroup (WG), preferred multiple of WG-size (PM),
* and size of subgoup (SG) only if larger-equal than PM. SG is signaled
* smaller if an alternative SG-size exists (SG is zero if no support).
*/
size_t wgsize[3];
/** Kind of device (GPU, CPU, or other). */
cl_device_type type;
/** OpenCL device-ID. */
cl_device_id id;
/** Whether host memory is unified. */
cl_int unified;
/** Device-ID. */
Expand Down Expand Up @@ -381,8 +389,6 @@ int c_dbcsr_acc_opencl_device_ext(cl_device_id device, const char* const extname
int c_dbcsr_acc_opencl_create_context(cl_device_id device_id, cl_context* context);
/** Internal variant of c_dbcsr_acc_set_active_device. */
int c_dbcsr_acc_opencl_set_active_device(ACC_OPENCL_LOCKTYPE* lock, int device_id);
/** Get preferred multiple and max. size of workgroup (kernel- or device-specific). */
int c_dbcsr_acc_opencl_wgsize(cl_device_id device, cl_kernel kernel, size_t* max_value, size_t* preferred_multiple);
/**
* Build kernel from source with given kernel_name, build_params and build_options.
* The build_params are meant to instantiate the kernel (-D) whereas build_options
Expand Down
Loading

0 comments on commit 6db5b28

Please sign in to comment.