Skip to content

Multi-GPU for GEMM for macOS and GNU/Linux #350

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

Multi-GPU for GEMM for macOS and GNU/Linux #350

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
2cc67d2
Experimental Multi-GPU for GEMM
Oct 6, 2019
8bc42d8
Improving Stability to Check
Mar 14, 2021
2852281
Improving Stability to Check
Mar 14, 2021
c26a75b
Fix CMAKE min to 3.5
Oct 29, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion src/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,7 @@
# limitations under the License.
# ########################################################################

cmake_minimum_required(VERSION 2.8)
cmake_minimum_required(VERSION 3.7)

#User toggle-able options that can be changed on the command line with -D
option( BUILD_RUNTIME "Build the BLAS runtime library" ON )
Expand Down
14 changes: 10 additions & 4 deletions src/clBLAS.version.h.in
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,8 +15,14 @@
* ************************************************************************/


/* the configured version and settings for clblas
/*
the configured version and settings for clblas
*/
#define clblasVersionMajor @clBLAS_VERSION_MAJOR@
#define clblasVersionMinor @clBLAS_VERSION_MINOR@
#define clblasVersionPatch @clBLAS_VERSION_PATCH@

//#define clblasVersionMajor @clBLAS_VERSION_MAJOR@
//#define clblasVersionMinor @clBLAS_VERSION_MINOR@
//#define clblasVersionPatch @clBLAS_VERSION_PATCH@

static int clblasVersionMajor = 1;
static int clblasVersionMinor = 12;
static int clblasVersionPatch = 2;
303 changes: 151 additions & 152 deletions src/library/blas/specialCases/GemmSpecialCases.cpp
View file
Open in desktop

Large diffs are not rendered by default.

177 changes: 100 additions & 77 deletions src/library/blas/xgemm.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,9 +73,13 @@ static void force_gemm_column_major(
return static_cast<clblasStatus>(err);

const static unsigned int numGemmKernelArgs = 14;
void *gemmKernelArgs[numGemmKernelArgs];
size_t gemmKernelArgSizes[numGemmKernelArgs];

#if defined( _WIN32 )
__declspec( thread ) void *gemmKernelArgs[numGemmKernelArgs];
__declspec( thread ) size_t gemmKernelArgSizes[numGemmKernelArgs];
#else
__thread void *gemmKernelArgs[numGemmKernelArgs];
__thread size_t gemmKernelArgSizes[numGemmKernelArgs];
#endif

/******************************************************************************
* Is beta zero for optimization
Expand Down Expand Up @@ -115,7 +119,8 @@ static char *getKernelName(cl_kernel clKernel)
// The kernelNameLength turns out to be of proper length.
// CL_CHECK(err)

char *kernelName = new char[kernelNameLength];
char *kernelName;
kernelName = new char[kernelNameLength];
err = clGetKernelInfo(
clKernel,
CL_KERNEL_FUNCTION_NAME,
Expand All @@ -128,20 +133,14 @@ static char *getKernelName(cl_kernel clKernel)
}

typedef struct kernel_map_key_ {
cl_context context; // address of context
cl_device_id device; // address of device
cl_command_queue queue;
const char *kernelSource; // address of kernel source
} kernel_map_key;

bool operator<(const kernel_map_key & l, const kernel_map_key & r) {
if (l.context < r.context) {
return true;
} else if (r.context < l.context) {
return false;
}
if (l.device < r.device) {
if (l.queue < r.queue) {
return true;
} else if (r.device < l.device) {
} else if (r.queue < l.queue) {
return false;
}
if (l.kernelSource < r.kernelSource) {
Expand All @@ -166,30 +165,21 @@ void makeGemmKernel(
size_t *kernelBinarySize,
const char *binaryBuildOptions)
{

typedef std::map<kernel_map_key, cl_kernel> kernel_map_t;
#if defined( _WIN32 )
__declspec( thread ) static kernel_map_t *kernel_map = 0;
#if defined( _WIN32 )
/*__declspec( thread )*/ static kernel_map_t *kernel_map = 0;
#else
static __thread kernel_map_t *kernel_map = 0;
static /*__thread*/ kernel_map_t *kernel_map = 0;
#endif
if (!kernel_map) {
kernel_map = new kernel_map_t();
}

cl_context clContext;
cl_device_id clDevice;
cl_int err;

err = clGetCommandQueueInfo( clQueue, CL_QUEUE_CONTEXT, sizeof(clContext), &clContext, NULL);
CL_CHECK(err)
err = clGetCommandQueueInfo( clQueue, CL_QUEUE_DEVICE, sizeof(clDevice), &clDevice, NULL);
CL_CHECK(err)

// is kernel already compiled?
kernel_map_key key;
key.queue = clQueue;
key.kernelSource = kernelSource;
key.context = clContext;
key.device = clDevice;
kernel_map_t::iterator idx = kernel_map->find(key);
if (idx == kernel_map->end()) {
*clKernel = NULL;
Expand All @@ -198,7 +188,16 @@ void makeGemmKernel(
return;
}

if (true /*!*clKernel*/) { // since kernel wasn't found in map
if (true/*!*clKernel*/) { // since kernel wasn't found in map
cl_context clContext;
cl_device_id clDevice;
cl_int err;

err = clGetCommandQueueInfo( clQueue, CL_QUEUE_CONTEXT, sizeof(clContext), &clContext, NULL);
CL_CHECK(err)
err = clGetCommandQueueInfo( clQueue, CL_QUEUE_DEVICE, sizeof(clDevice), &clDevice, NULL);
CL_CHECK(err)

// kernel has not been built, so build it (from binary, preferably)
cl_program clProgram;
cl_int clBinaryStatus;
Expand Down Expand Up @@ -258,15 +257,18 @@ void makeGemmKernel(
printf("\nBuild Log:\n\n");
printf("%s\n", buildLog);
//printf("\n\nKernel String:\n\n");
//printf("%s\n", kernelSource);
//printf("%s\n", *kernelSource);
//FIXME: The function should be exiting at this point
}

cl_uint numKernels = 0;

err = clCreateKernelsInProgram(
clProgram,
1, clKernel,
NULL );
&numKernels );
CL_CHECK(err)

err = clReleaseProgram(clProgram);
CL_CHECK(err)

Expand Down Expand Up @@ -309,6 +311,8 @@ void makeGemmKernel(
if (err != CL_SUCCESS)
return err;

clFinish(clQueue);

return CL_SUCCESS;
}

Expand Down Expand Up @@ -356,30 +360,41 @@ clblasGemm(
const cl_event *eventWaitList,
cl_event *events)
{


// cast types to opencl types
cl_mem A = iA;
cl_mem B = iB;
cl_uint M = static_cast<cl_uint>( iM );
cl_uint N = static_cast<cl_uint>( iN );
cl_uint K = static_cast<cl_uint>( iK );
cl_uint offA = static_cast<cl_uint>( iOffA );
cl_uint offB = static_cast<cl_uint>( iOffB );
cl_uint offC = static_cast<cl_uint>( iOffC );
cl_uint lda = static_cast<cl_uint>( iLda );
cl_uint ldb = static_cast<cl_uint>( iLdb );
cl_uint ldc = static_cast<cl_uint>( iLdc );

transA = correctTranspose<Precision>(transA);
transB = correctTranspose<Precision>(transB);
cl_mem A;
A = iA;
cl_mem B;
B = iB;
cl_uint M;
M = static_cast<cl_uint>( iM );
cl_uint N;
N = static_cast<cl_uint>( iN );
cl_uint K;
K = static_cast<cl_uint>( iK );
cl_uint offA;
offA = static_cast<cl_uint>( iOffA );
cl_uint offB;
offB = static_cast<cl_uint>( iOffB );
cl_uint offC;
offC = static_cast<cl_uint>( iOffC );
cl_uint lda;
lda = static_cast<cl_uint>( iLda );
cl_uint ldb;
ldb = static_cast<cl_uint>( iLdb );
cl_uint ldc;
ldc = static_cast<cl_uint>( iLdc );

clblasTranspose tA;
tA = correctTranspose<Precision>(transA);
clblasTranspose tB;
tB = correctTranspose<Precision>(transB);
// if debug build, validate input
// CHECK_QUEUES(numCommandQueues, commandQueues);
// CHECK_EVENTS(numEventsInWaitList, eventWaitList);
// CHECK_MATRIX_A(Precision, order, transA, A, M, K, offA, lda);
// CHECK_MATRIX_B(Precision, order, transB, B, K, N, offB, ldb);
// CHECK_MATRIX_A(Precision, order, tA, A, M, K, offA, lda);
// CHECK_MATRIX_B(Precision, order, tB, B, K, N, offB, ldb);
// CHECK_MATRIX_C(Precision, order, clblasNoTrans, C, M, N, offC, ldc);
force_gemm_column_major( order, transA, transB,
force_gemm_column_major( order, tA, tB,
M, N, offA, offB, lda, ldb, A, B );


Expand All @@ -397,8 +412,8 @@ clblasGemm(
bool specialCaseHandled = false;

clblasStatus SpecialCaseStatus = GemmSpecialCases<Precision>(order,
transA,
transB,
tA,
tB,
M, N, K,
alpha,
A, offA, lda,
Expand Down Expand Up @@ -429,17 +444,20 @@ clblasGemm(
err = clGetDeviceInfo( clDevice, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(clDeviceNumCUs), &clDeviceNumCUs, NULL);
//CL_CHECK(err)
returnIfErr(err);
unsigned int deviceIdealNumThreads = (8 /*waves per CU*/)*(64 /*threads per wave*/)*clDeviceNumCUs;
float optimalNumElementsPerThread = ((float)M*N) / deviceIdealNumThreads;
unsigned int deviceIdealNumThreads;
deviceIdealNumThreads = (8 /*waves per CU*/)*(64 /*threads per wave*/)*clDeviceNumCUs;
float optimalNumElementsPerThread;
optimalNumElementsPerThread = ((float)M*N) / deviceIdealNumThreads;
//optimalNumElementsPerThread = 32;
bool betaNonZero = !isZero(beta);
bool betaNonZero;
betaNonZero = !isZero(beta);

#ifdef AUTOGEMM_PRINT_DEBUG
printf("%sgemm_%3s_%s%s_B%u_%llux%llux%llu\n",
getPrecision<Precision>(),
order==clblasColumnMajor ? "Col" : "Row",
transA==clblasNoTrans ? "N" : transA==clblasTrans ? "T" : "C",
transB==clblasNoTrans ? "N" : transB==clblasTrans ? "T" : "C",
tA==clblasNoTrans ? "N" : tA==clblasTrans ? "T" : "C",
tB==clblasNoTrans ? "N" : tB==clblasTrans ? "T" : "C",
betaNonZero ? 1 : 0,
iM, iN, iK );
#endif
Expand Down Expand Up @@ -470,8 +488,9 @@ clblasGemm(
unsigned int microTileNumRows;
unsigned int microTileNumCols;
unsigned int unroll;

gemmSelectKernel<Precision>(
order, transA, transB,
order, tA, tB,
iM, iN, iK,
betaNonZero,
optimalNumElementsPerThread,
Expand Down Expand Up @@ -502,15 +521,15 @@ clblasGemm(
if (!tileKernelSource) {
printf("ERROR: gemmSelectKernel() couldn't find kernel(s) for { order=%s, transA=%s, transB=%s, M=%u, N=%u, K=%u, beta=%u, onept=%f }\n",
order==clblasColumnMajor ? "ColMajor" : "RowMajor",
transA==clblasNoTrans ? "N" : transA==clblasTrans ? "T" : "C",
transB==clblasNoTrans ? "N" : transB==clblasTrans ? "T" : "C",
tA==clblasNoTrans ? "N" : tA==clblasTrans ? "T" : "C",
tB==clblasNoTrans ? "N" : tB==clblasTrans ? "T" : "C",
M, N, K,
betaNonZero ? 1 : 0,
optimalNumElementsPerThread );
gemmSelectKernel<Precision>(
order,
transA,
transB,
tA,
tB,
M,
N,
K,
Expand Down Expand Up @@ -542,14 +561,18 @@ clblasGemm(
return clblasNotImplemented;
}


unsigned int macroTileNumRows = workGroupNumRows*microTileNumRows;
unsigned int macroTileNumCols = workGroupNumCols*microTileNumCols;
bool needTileKernel = M/macroTileNumRows > 0
&& N/macroTileNumCols > 0;
bool needRowKernel = M%macroTileNumRows > 0 && N/macroTileNumCols > 0;
bool needColKernel = N%macroTileNumCols > 0 && M/macroTileNumRows > 0;
bool needCornerKernel = M%macroTileNumRows > 0 && N%macroTileNumCols > 0;
unsigned int macroTileNumRows;
macroTileNumRows = workGroupNumRows*microTileNumRows;
unsigned int macroTileNumCols;
macroTileNumCols = workGroupNumCols*microTileNumCols;
bool needTileKernel;
needTileKernel = M/macroTileNumRows > 0 && N/macroTileNumCols > 0;
bool needRowKernel;
needRowKernel = M%macroTileNumRows > 0 && N/macroTileNumCols > 0;
bool needColKernel;
needColKernel = N%macroTileNumCols > 0 && M/macroTileNumRows > 0;
bool needCornerKernel;
needCornerKernel = M%macroTileNumRows > 0 && N%macroTileNumCols > 0;
#if 0
printf("For M,N,K = %u,%u,%u and %u CUs selected tile is wg=%ux%u, microTile=%ux%u, macroTile=%ux%u kernelsNeeded=%u,%u,%u,%u\n",
M, N, K, clDeviceNumCUs,
Expand Down Expand Up @@ -603,11 +626,11 @@ clblasGemm(
if (needTileKernel) {
//printf("enqueueing tile kernel\n");
size_t globalWorkSize[2] = {(M/macroTileNumRows)*workGroupNumRows, (N/macroTileNumCols)*workGroupNumCols };
err = enqueueGemmKernel( commandQueues[numKernelsEnqueued%numCommandQueues], tileClKernel,
err = enqueueGemmKernel( commandQueues[0], tileClKernel,
gemmKernelArgs, gemmKernelArgSizes, numGemmKernelArgs,
globalWorkSize, localWorkSize,
numEventsInWaitList, eventWaitList,
&events[numKernelsEnqueued%numCommandQueues] );
events );
returnIfErr(err);
numKernelsEnqueued++;
}
Expand All @@ -618,11 +641,11 @@ clblasGemm(
if (needRowKernel) {
//printf("enqueueing row kernel\n");
size_t globalWorkSize[2] = {1*workGroupNumRows, (N/macroTileNumCols)*workGroupNumCols };
err = enqueueGemmKernel( commandQueues[numKernelsEnqueued%numCommandQueues], rowClKernel,
err = enqueueGemmKernel( commandQueues[0], rowClKernel,
gemmKernelArgs, gemmKernelArgSizes, numGemmKernelArgs,
globalWorkSize, localWorkSize,
numEventsInWaitList, eventWaitList,
&events[numKernelsEnqueued%numCommandQueues] );
events );
returnIfErr(err);
numKernelsEnqueued++;
}
Expand All @@ -633,11 +656,11 @@ clblasGemm(
if (needColKernel) {
//printf("enqueueing col kernel\n");
size_t globalWorkSize[2] = { (M/macroTileNumRows)*workGroupNumRows, 1*workGroupNumCols };
err = enqueueGemmKernel( commandQueues[numKernelsEnqueued%numCommandQueues], colClKernel,
err = enqueueGemmKernel( commandQueues[0], colClKernel,
gemmKernelArgs, gemmKernelArgSizes, numGemmKernelArgs,
globalWorkSize, localWorkSize,
numEventsInWaitList, eventWaitList,
&events[numKernelsEnqueued%numCommandQueues] );
events );
returnIfErr(err);
numKernelsEnqueued++;
}
Expand All @@ -648,11 +671,11 @@ clblasGemm(
if (needCornerKernel) {
//printf("enqueueing corner kernel\n");
size_t globalWorkSize[2] = { 1*workGroupNumRows, 1*workGroupNumCols };
err = enqueueGemmKernel( commandQueues[numKernelsEnqueued%numCommandQueues], cornerClKernel,
err = enqueueGemmKernel( commandQueues[0], cornerClKernel,
gemmKernelArgs, gemmKernelArgSizes, numGemmKernelArgs,
globalWorkSize, localWorkSize,
numEventsInWaitList, eventWaitList,
&events[numKernelsEnqueued%numCommandQueues] );
events );
returnIfErr(err);
numKernelsEnqueued++;
}
Expand Down
2 changes: 1 addition & 1 deletion src/library/tools/OCLBinaryGenerator/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,7 @@
# limitations under the License.
# ########################################################################

cmake_minimum_required(VERSION 2.6)
cmake_minimum_required(VERSION 3.5)
project(OCLBinaryGenerator C CXX)
ADD_DEFINITIONS(/D_CRT_SECURE_NO_WARNINGS)
ADD_EXECUTABLE(OCLBinaryGenerator OCLBinaryGenerator.cpp)
Expand Down
2 changes: 1 addition & 1 deletion src/library/tools/bingen/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,7 @@
# limitations under the License.
# ########################################################################

cmake_minimum_required(VERSION 2.6)
cmake_minimum_required(VERSION 3.5)
project(bingen C CXX)
ADD_DEFINITIONS(/D_CRT_SECURE_NO_WARNINGS)
ADD_EXECUTABLE(bingen bingen.cpp)
Expand Down
Loading