Changeset: b02d018c0d3b for MonetDB
URL: http://dev.monetdb.org/hg/MonetDB?cmd=changeset;node=b02d018c0d3b
Modified Files:
monetdb5/extras/bwd/bwd.c
monetdb5/extras/bwd/operations.c
monetdb5/extras/bwd/utilities.c
Branch: bwd
Log Message:
* lots of bitmagic
* using byte-instructions instead of integer instructions on the device
because nvidia gpus don't support non-aligned int access
* made bwdevices report the endianness of the device
* printing debug info to stdout rather than gdkout
Unterschiede (gekürzt von 371 auf 300 Zeilen):
diff --git a/monetdb5/extras/bwd/bwd.c b/monetdb5/extras/bwd/bwd.c
--- a/monetdb5/extras/bwd/bwd.c
+++ b/monetdb5/extras/bwd/bwd.c
@@ -63,8 +63,10 @@ str deviceInfo(bat * resPlatform, bat *
{
clGetDeviceInfo(devices[j],CL_DEVICE_TYPE,sizeof(cl_ulong),&device_type,NULL);
char deviceName[1024];
+ cl_bool endianLittle;
if((err =
clGetDeviceInfo(devices[j],CL_DEVICE_NAME,sizeof(deviceName),&deviceName,NULL)))
die ("clGetDeviceInfo error: %s\n",clError(err));
- snprintf (deviceNameBuffer, 1024, "%d:
%s (%s)\n", j, deviceName, clDeviceString(device_type));
+ if((err =
clGetDeviceInfo(devices[j],CL_DEVICE_ENDIAN_LITTLE,sizeof(cl_bool),&endianLittle,NULL)))
die ("clGetDeviceInfo error: %s\n",clError(err));
+ snprintf (deviceNameBuffer, 1024, "%d:
%s (%s), endianness: %s\n", j, deviceName, clDeviceString(device_type),
(str[]){[CL_TRUE] = "little", [CL_FALSE] = "big"}[endianLittle]);
}
char* platformVersionBuffer;
{
diff --git a/monetdb5/extras/bwd/operations.c b/monetdb5/extras/bwd/operations.c
--- a/monetdb5/extras/bwd/operations.c
+++ b/monetdb5/extras/bwd/operations.c
@@ -18,10 +18,10 @@
str BWDLeftJoinApproximate(bat * res, bat * l, bat * r){
- ALGODEBUG THRprintf(GDKout, "#BWDfetchjoin: approximating;\n");
+ ALGODEBUG printf("#BWDfetchjoin: approximating;\n");
BAT* right = BATdescriptor(*r);
if(!batTailIsDecomposed(right)){
- THRprintf(GDKout, "bwd.%s : bat is not decomposed: %d, no
approximation can be provided", __func__, *r);
+ printf("bwd.%s : bat is not decomposed: %d, no approximation
can be provided", __func__, *r);
BBPreleaseref(right->batCacheid);
return MAL_SUCCEED;
}
@@ -54,17 +54,17 @@ str BWDLeftJoinApproximate(bat * res, ba
BBPreleaseref(right->batCacheid);
return MAL_SUCCEED;
} else
- THRprintf(GDKout, "bwd.BWDLeftJoinApproximate, %s
(BAThvoid(left): %d, BATtvoid(left): %d, BAThvoid(right): %d, right->tseqbase:
%ld)", "this case isn't implemented yet", BAThvoid(left), BATtvoid(left),
BATtvoid(right), left->tseqbase);
+ printf("bwd.BWDLeftJoinApproximate, %s (BAThvoid(left): %d,
BATtvoid(left): %d, BAThvoid(right): %d, right->tseqbase: %ld)", "this case
isn't implemented yet", BAThvoid(left), BATtvoid(left), BATtvoid(right),
left->tseqbase);
return MAL_SUCCEED;
};
str BWDLeftJoinRefine(bat * res, bat * l, bat * r, bat * approx){
if (!approx || !*approx) {
- THRprintf(GDKout, "bwd.%s : no approximation provided, falling
back to normal leftjoin", __func__);
+ printf("bwd.%s : no approximation provided, falling back to
normal leftjoin", __func__);
return ALGleftjoin(res, l, r);
}
- ALGODEBUG THRprintf(GDKout, "#BWDfetchjoin: approximating;\n");
+ ALGODEBUG printf("#BWDfetchjoin: approximating;\n");
BAT* left = BATdescriptor(*l);
BAT* right = BATdescriptor(*r);
@@ -93,7 +93,7 @@ static inline size_t clDeviceAddressByte
static size_t CL_DEVICE_ADDRESS_BYTES = 0;
if (!CL_DEVICE_ADDRESS_BYTES) {
cl_int err = clGetDeviceInfo(getDeviceID(),
CL_DEVICE_ADDRESS_BITS, sizeof(size_t), &CL_DEVICE_ADDRESS_BYTES, NULL);
- if(err) THRprintf(GDKout, "#%s, %s;\n", __func__, clError(err));
+ if(err) printf("#%s, %s;\n", __func__, clError(err));
CL_DEVICE_ADDRESS_BYTES /= 8;
}
return CL_DEVICE_ADDRESS_BYTES; //TODO: check
that CL_DEVICE_ADDRESS_BITS
@@ -143,24 +143,33 @@ str BWDThetauselectApproximate(bat *res,
void* zeros =
calloc((BATcount(data)+1)*(clDeviceAddressBytes() +
slot->approximationBits/8)+sizeof(clBAT), 1); //TODO: migrate buffer
initialization to GPU
clEnqueueWriteBuffer(getCommandQueue(),
slot->approximation, CL_TRUE, 0, (BATcount(data)+1)*(clDeviceAddressBytes() +
slot->approximationBits/8)+sizeof(clBAT), zeros, 0, NULL, NULL);
free(zeros);
- if(err) THRprintf(GDKout, "#%s, clCreateBuffer: %s;\n",
__func__, clError(err));
+ if(err) printf("#%s, clCreateBuffer: %s;\n", __func__,
clError(err));
}
const char* sourceCodeTemplate = "__kernel void uselect ("
"__global struct{int count; int padding; char
elements[];}* output,"
- "__global const void* approximation,"
+ "__global const char* approximation,"
"const %1$s operand"
") {"
/* " printf(\"%%d:\", get_global_id(0));" */
- " const %1$s value = (*(int*)(approximation +
get_global_id(0)*%3$d) << %4$d);"
+ " %1$s value = 0;"
+ " const size_t inputOffset = get_global_id(0)*%3$d;"
+ " {"
+ " for(int i = 0; i < %3$d; i++)"
+ " value += approximation[inputOffset + i] << i*8;"
+ " }"
" if(value %2$s operand)"
" {"
- " const int slotNumber =
atomic_inc(&(output->count));"
- " size_t* outputSlot = (size_t*)(output->elements +
slotNumber*(%3$d + sizeof(size_t)));"
- " *outputSlot |= get_global_id(0);"
- " outputSlot[1] |= (value >> %4$d);"
- " printf(\"value: %%d, value in slot(%%d) %%d,
global_id: %%d\\n\", value, slotNumber, outputSlot[1], get_global_id(0));"
+ " const size_t offset = atomic_inc(&(output->count))
* (sizeof(size_t) + %3$d);"
+ " for(int i = 0; i < sizeof(size_t); i++)"
+ " output->elements[offset+i] = get_global_id(0) >>
i*8;"
+ " for(int i = 0; i < sizeof(size_t); i++)"
+ " output->elements[offset+sizeof(size_t)+i] =
approximation[inputOffset + i];"
+ /* " size_t* outputSlot = (size_t*)(output->elements
+ slotNumber*(%3$d + sizeof(size_t)));" */
+ /* " *outputSlot |= get_global_id(0);" */
+ /* " outputSlot[1] |= (value >> %4$d);" */
+ /* " printf(\"value: %%d, value in slot(%%d) %%d,
global_id: %%d\\n\", value, slotNumber, outputSlot[1], get_global_id(0));" */
" }"
"}";
char* sourceCode = malloc(16384);
@@ -217,10 +226,9 @@ str BWDThetauselectRefine(bat *res, bat
BATseqbase(result, 0);
clBAT* resultClBAT = (clBAT*) malloc(approximationSize);
- /* Tloc(result, BUNfirst(result)); */
-
cl_int err = clEnqueueReadBuffer(getCommandQueue(),
batTailApproximation(approximation), CL_TRUE, 0, approximationSize,
resultClBAT, 0, NULL, NULL);
if(err) THRprintf(GDKout, "#%s, clEnqueueReadBuffer: %s;\n", __func__,
clError(err));
+ /* Tloc(result, BUNfirst(result)); */
int* resultRegion = (int*) Tloc(result, BUNfirst(result)); // type
specific
oid* positionRegion = (oid*) Hloc(result, BUNfirst(result)); // type
specific
size_t candidateCount = resultClBAT->count;
@@ -239,15 +247,25 @@ str BWDThetauselectRefine(bat *res, bat
if(deCompressedValue comparator *(int*)val) {
\
positionRegion[j] = index; \
resultRegion[j++] = deCompressedValue; \
- printf ("compressed value: %d, index: %d, value: %d\n",
compressedValue, index, resultRegion[j-1]); \
}\
}
+ /* printf ("compressed value: %d, index: %d, value: %d\n",
compressedValue, index, resultRegion[j-1]); \ */
switch (*OP[0]){
- case '<':
- refineLoop(<);
+ case '<':
+ switch((*OP)[1]){
+ case '\0':
+ refineLoop(<);
+ case '=':
+ refineLoop(<=);
+ }
case '>':
- refineLoop(>);
+ switch((*OP)[1]){
+ case '\0':
+ refineLoop(<=);
+ case '=':
+ refineLoop(<=);
+ }
case '=':
refineLoop(==);
}
diff --git a/monetdb5/extras/bwd/utilities.c b/monetdb5/extras/bwd/utilities.c
--- a/monetdb5/extras/bwd/utilities.c
+++ b/monetdb5/extras/bwd/utilities.c
@@ -18,97 +18,104 @@
#pragma mark OpenCL Utitily Functions
const char* clError(int err){
- static char* cl_Errors[64];
- if (err == -1001)
- return "CL_PLATFORM_NOT_FOUND_KHR";
- if (!cl_Errors[0]){
- cl_Errors[0] = "CL_SUCCESS";
- cl_Errors[1] = "CL_DEVICE_NOT_FOUND";
- cl_Errors[2] = "CL_DEVICE_NOT_AVAILABLE";
- cl_Errors[3] = "CL_COMPILER_NOT_AVAILABLE";
- cl_Errors[4] = "CL_MEM_OBJECT_ALLOCATION_FAILURE";
- cl_Errors[5] = "CL_OUT_OF_RESOURCES";
- cl_Errors[6] = "CL_OUT_OF_HOST_MEMORY";
- cl_Errors[7] = "CL_PROFILING_INFO_NOT_AVAILABLE";
- cl_Errors[8] = "CL_MEM_COPY_OVERLAP";
- cl_Errors[9] = "CL_IMAGE_FORMAT_MISMATCH";
- cl_Errors[10] = "CL_IMAGE_FORMAT_NOT_SUPPORTED";
- cl_Errors[11] = "CL_BUILD_PROGRAM_FAILURE";
- cl_Errors[12] = "CL_MAP_FAILURE";
-
- cl_Errors[30] = "CL_INVALID_VALUE";
- cl_Errors[31] = "CL_INVALID_DEVICE_TYPE";
- cl_Errors[32] = "CL_INVALID_PLATFORM";
- cl_Errors[33] = "CL_INVALID_DEVICE";
- cl_Errors[34] = "CL_INVALID_CONTEXT";
- cl_Errors[35] = "CL_INVALID_QUEUE_PROPERTIES";
- cl_Errors[36] = "CL_INVALID_COMMAND_QUEUE";
- cl_Errors[37] = "CL_INVALID_HOST_PTR";
- cl_Errors[38] = "CL_INVALID_MEM_OBJECT";
- cl_Errors[39] = "CL_INVALID_IMAGE_FORMAT_DESCRIPTOR";
- cl_Errors[40] = "CL_INVALID_IMAGE_SIZE";
- cl_Errors[41] = "CL_INVALID_SAMPLER";
- cl_Errors[42] = "CL_INVALID_BINARY";
- cl_Errors[43] = "CL_INVALID_BUILD_OPTIONS";
- cl_Errors[44] = "CL_INVALID_PROGRAM";
- cl_Errors[45] = "CL_INVALID_PROGRAM_EXECUTABLE";
- cl_Errors[46] = "CL_INVALID_KERNEL_NAME";
- cl_Errors[47] = "CL_INVALID_KERNEL_DEFINITION";
- cl_Errors[48] = "CL_INVALID_KERNEL";
- cl_Errors[49] = "CL_INVALID_ARG_INDEX";
- cl_Errors[50] = "CL_INVALID_ARG_VALUE";
- cl_Errors[51] = "CL_INVALID_ARG_SIZE";
- cl_Errors[52] = "CL_INVALID_KERNEL_ARGS";
- cl_Errors[53] = "CL_INVALID_WORK_DIMENSION";
- cl_Errors[54] = "CL_INVALID_WORK_GROUP_SIZE";
- cl_Errors[55] = "CL_INVALID_WORK_ITEM_SIZE";
- cl_Errors[56] = "CL_INVALID_GLOBAL_OFFSET";
- cl_Errors[57] = "CL_INVALID_EVENT_WAIT_LIST";
- cl_Errors[58] = "CL_INVALID_EVENT";
- cl_Errors[59] = "CL_INVALID_OPERATION";
- cl_Errors[60] = "CL_INVALID_GL_OBJECT";
- cl_Errors[61] = "CL_INVALID_BUFFER_SIZE";
- cl_Errors[62] = "CL_INVALID_MIP_LEVEL";
- cl_Errors[63] = "CL_INVALID_GLOBAL_WORK_SIZE";
-
- }
- return cl_Errors[-err];
+ const str result = (str[]){
+ [-CL_SUCCESS] = "CL_SUCCESS",
+ [-CL_DEVICE_NOT_FOUND] = "CL_DEVICE_NOT_FOUND",
+ [-CL_DEVICE_NOT_AVAILABLE] = "CL_DEVICE_NOT_AVAILABLE",
+ [-CL_COMPILER_NOT_AVAILABLE] = "CL_COMPILER_NOT_AVAILABLE",
+ [-CL_MEM_OBJECT_ALLOCATION_FAILURE] =
"CL_MEM_OBJECT_ALLOCATION_FAILURE",
+ [-CL_OUT_OF_RESOURCES] = "CL_OUT_OF_RESOURCES",
+ [-CL_OUT_OF_HOST_MEMORY] = "CL_OUT_OF_HOST_MEMORY",
+ [-CL_PROFILING_INFO_NOT_AVAILABLE] =
"CL_PROFILING_INFO_NOT_AVAILABLE",
+ [-CL_MEM_COPY_OVERLAP] = "CL_MEM_COPY_OVERLAP",
+ [-CL_IMAGE_FORMAT_MISMATCH] = "CL_IMAGE_FORMAT_MISMATCH",
+ [-CL_IMAGE_FORMAT_NOT_SUPPORTED] =
"CL_IMAGE_FORMAT_NOT_SUPPORTED",
+ [-CL_BUILD_PROGRAM_FAILURE] = "CL_BUILD_PROGRAM_FAILURE",
+ [-CL_MAP_FAILURE] = "CL_MAP_FAILURE",
+ [-CL_INVALID_VALUE] = "CL_INVALID_VALUE",
+ [-CL_INVALID_DEVICE_TYPE] = "CL_INVALID_DEVICE_TYPE",
+ [-CL_INVALID_PLATFORM] = "CL_INVALID_PLATFORM",
+ [-CL_INVALID_DEVICE] = "CL_INVALID_DEVICE",
+ [-CL_INVALID_CONTEXT] = "CL_INVALID_CONTEXT",
+ [-CL_INVALID_QUEUE_PROPERTIES] = "CL_INVALID_QUEUE_PROPERTIES",
+ [-CL_INVALID_COMMAND_QUEUE] = "CL_INVALID_COMMAND_QUEUE",
+ [-CL_INVALID_HOST_PTR] = "CL_INVALID_HOST_PTR",
+ [-CL_INVALID_MEM_OBJECT] = "CL_INVALID_MEM_OBJECT",
+ [-CL_INVALID_IMAGE_FORMAT_DESCRIPTOR] =
"CL_INVALID_IMAGE_FORMAT_DESCRIPTOR",
+ [-CL_INVALID_IMAGE_SIZE] = "CL_INVALID_IMAGE_SIZE",
+ [-CL_INVALID_SAMPLER] = "CL_INVALID_SAMPLER",
+ [-CL_INVALID_BINARY] = "CL_INVALID_BINARY",
+ [-CL_INVALID_BUILD_OPTIONS] = "CL_INVALID_BUILD_OPTIONS",
+ [-CL_INVALID_PROGRAM] = "CL_INVALID_PROGRAM",
+ [-CL_INVALID_PROGRAM_EXECUTABLE] =
"CL_INVALID_PROGRAM_EXECUTABLE",
+ [-CL_INVALID_KERNEL_NAME] = "CL_INVALID_KERNEL_NAME",
+ [-CL_INVALID_KERNEL_DEFINITION] =
"CL_INVALID_KERNEL_DEFINITION",
+ [-CL_INVALID_KERNEL] = "CL_INVALID_KERNEL",
+ [-CL_INVALID_ARG_INDEX] = "CL_INVALID_ARG_INDEX",
+ [-CL_INVALID_ARG_VALUE] = "CL_INVALID_ARG_VALUE",
+ [-CL_INVALID_ARG_SIZE] = "CL_INVALID_ARG_SIZE",
+ [-CL_INVALID_KERNEL_ARGS] = "CL_INVALID_KERNEL_ARGS",
+ [-CL_INVALID_WORK_DIMENSION] = "CL_INVALID_WORK_DIMENSION",
+ [-CL_INVALID_WORK_GROUP_SIZE] = "CL_INVALID_WORK_GROUP_SIZE",
+ [-CL_INVALID_WORK_ITEM_SIZE] = "CL_INVALID_WORK_ITEM_SIZE",
+ [-CL_INVALID_GLOBAL_OFFSET] = "CL_INVALID_GLOBAL_OFFSET",
+ [-CL_INVALID_EVENT_WAIT_LIST] = "CL_INVALID_EVENT_WAIT_LIST",
+ [-CL_INVALID_EVENT] = "CL_INVALID_EVENT",
+ [-CL_INVALID_OPERATION] = "CL_INVALID_OPERATION",
+ [-CL_INVALID_GL_OBJECT] = "CL_INVALID_GL_OBJECT",
+ [-CL_INVALID_BUFFER_SIZE] = "CL_INVALID_BUFFER_SIZE",
+ [-CL_INVALID_MIP_LEVEL] = "CL_INVALID_MIP_LEVEL",
+ [-CL_INVALID_GLOBAL_WORK_SIZE] = "CL_INVALID_GLOBAL_WORK_SIZE"
+ }[-err];
+ if(result) return result;
+ return "unknown error";
}
cl_device_id getDeviceID(){
+ static int initialized = 0;
+ static cl_device_id result;
cl_platform_id platforms[4];
cl_uint foundPlatforms;
- if(clGetPlatformIDs(4,platforms,&foundPlatforms) != CL_SUCCESS)
THRprintf(GDKout, "problem when finding the platforms");
- if (!foundPlatforms) THRprintf(GDKout, "didn't find any OpenCL
Platforms");
+ if(clGetPlatformIDs(4,platforms,&foundPlatforms) != CL_SUCCESS)
printf("problem when finding the platforms");
+ if (!foundPlatforms) printf("didn't find any OpenCL Platforms");
- cl_device_id result;
- if (clGetDeviceIDs(*platforms,CL_DEVICE_TYPE_ALL, 1, &result, NULL) ==
CL_SUCCESS) return result;
- else THRprintf(GDKout, "error getting GPU device id\n");
-
- return NULL;
+ if (clGetDeviceIDs(*platforms,CL_DEVICE_TYPE_ALL, 1, &result, NULL) !=
CL_SUCCESS) printf("error getting GPU device id\n");
+ else initialized = 1;
+ return result;
}
-static cl_context clContextSingleton = NULL;
cl_context getCLContext(){
- if(clContextSingleton) return clContextSingleton;
+ static int initialized = 0;
+ static cl_context clContextSingleton;
+ if(initialized) {
+ printf("#%s, returning cl_context: %p;\n", __func__,
clContextSingleton);
+ return clContextSingleton;
+ }
int err;
clContextSingleton = clCreateContext(0,1,
(cl_device_id[]){getDeviceID()}, NULL, NULL, &err);
- if (err == CL_SUCCESS) return clContextSingleton;
- else THRprintf(GDKout, "failure when creating the context\n");
- return NULL;
+ if (err != CL_SUCCESS) printf("failure when creating the context\n");
+ else initialized= 1;
+ printf("#%s, returning cl_context: %p;\n", __func__,
clContextSingleton);
+ return clContextSingleton;
_______________________________________________
checkin-list mailing list
checkin-list@monetdb.org
https://www.monetdb.org/mailman/listinfo/checkin-list
