// // Copyright (c) 2017 The Khronos Group Inc. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // // http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. // #include "Utility.h" #include #include "FunctionList.h" #if defined( __APPLE__ ) #include #endif int TestFunc_Float_Float(const Func *f, MTdata); int TestFunc_Double_Double(const Func *f, MTdata); #if defined( __cplusplus) extern "C" #endif const vtbl _unary = { "unary", TestFunc_Float_Float, TestFunc_Double_Double }; static int BuildKernel( const char *name, int vectorSize, cl_uint kernel_count, cl_kernel *k, cl_program *p ); static int BuildKernelDouble( const char *name, int vectorSize, cl_uint kernel_count, cl_kernel *k, cl_program *p ); static int BuildKernel( const char *name, int vectorSize, cl_uint kernel_count, cl_kernel *k, cl_program *p ) { const char *c[] = { "__kernel void math_kernel", sizeNames[vectorSize], "( __global float", sizeNames[vectorSize], "* out, __global float", sizeNames[vectorSize], "* in)\n" "{\n" " int i = get_global_id(0);\n" " out[i] = ", name, "( in[i] );\n" "}\n" }; const char *c3[] = { "__kernel void math_kernel", sizeNames[vectorSize], "( __global float* out, __global float* in)\n" "{\n" " size_t i = get_global_id(0);\n" " if( i + 1 < get_global_size(0) )\n" " {\n" " float3 f0 = vload3( 0, in + 3 * i );\n" " f0 = ", name, "( f0 );\n" " vstore3( f0, 0, out + 3*i );\n" " }\n" " else\n" " {\n" " size_t parity = i & 1; // Figure out how many elements are left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two buffer size \n" " float3 f0;\n" " switch( parity )\n" " {\n" " case 1:\n" " f0 = (float3)( in[3*i], NAN, NAN ); \n" " break;\n" " case 0:\n" " f0 = (float3)( in[3*i], in[3*i+1], NAN ); \n" " break;\n" " }\n" " f0 = ", name, "( f0 );\n" " switch( parity )\n" " {\n" " case 0:\n" " out[3*i+1] = f0.y; \n" " // fall through\n" " case 1:\n" " out[3*i] = f0.x; \n" " break;\n" " }\n" " }\n" "}\n" }; const char **kern = c; size_t kernSize = sizeof(c)/sizeof(c[0]); if( sizeValues[vectorSize] == 3 ) { kern = c3; kernSize = sizeof(c3)/sizeof(c3[0]); } char testName[32]; snprintf( testName, sizeof( testName ) -1, "math_kernel%s", sizeNames[vectorSize] ); return MakeKernels(kern, (cl_uint) kernSize, testName, kernel_count, k, p); } static int BuildKernelDouble( const char *name, int vectorSize, cl_uint kernel_count, cl_kernel *k, cl_program *p ) { const char *c[] = { "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n", "__kernel void math_kernel", sizeNames[vectorSize], "( __global double", sizeNames[vectorSize], "* out, __global double", sizeNames[vectorSize], "* in)\n" "{\n" " int i = get_global_id(0);\n" " out[i] = ", name, "( in[i] );\n" "}\n" }; const char *c3[] = { "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n", "__kernel void math_kernel", sizeNames[vectorSize], "( __global double* out, __global double* in)\n" "{\n" " size_t i = get_global_id(0);\n" " if( i + 1 < get_global_size(0) )\n" " {\n" " double3 f0 = vload3( 0, in + 3 * i );\n" " f0 = ", name, "( f0 );\n" " vstore3( f0, 0, out + 3*i );\n" " }\n" " else\n" " {\n" " size_t parity = i & 1; // Figure out how many elements are left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two buffer size \n" " double3 f0;\n" " switch( parity )\n" " {\n" " case 1:\n" " f0 = (double3)( in[3*i], NAN, NAN ); \n" " break;\n" " case 0:\n" " f0 = (double3)( in[3*i], in[3*i+1], NAN ); \n" " break;\n" " }\n" " f0 = ", name, "( f0 );\n" " switch( parity )\n" " {\n" " case 0:\n" " out[3*i+1] = f0.y; \n" " // fall through\n" " case 1:\n" " out[3*i] = f0.x; \n" " break;\n" " }\n" " }\n" "}\n" }; const char **kern = c; size_t kernSize = sizeof(c)/sizeof(c[0]); if( sizeValues[vectorSize] == 3 ) { kern = c3; kernSize = sizeof(c3)/sizeof(c3[0]); } char testName[32]; snprintf( testName, sizeof( testName ) -1, "math_kernel%s", sizeNames[vectorSize] ); return MakeKernels(kern, (cl_uint) kernSize, testName, kernel_count, k, p); } typedef struct BuildKernelInfo { cl_uint offset; // the first vector size to build cl_uint kernel_count; cl_kernel **kernels; cl_program *programs; const char *nameInCode; }BuildKernelInfo; static cl_int BuildKernel_FloatFn( cl_uint job_id, cl_uint thread_id UNUSED, void *p ); static cl_int BuildKernel_FloatFn( cl_uint job_id, cl_uint thread_id UNUSED, void *p ) { BuildKernelInfo *info = (BuildKernelInfo*) p; cl_uint i = info->offset + job_id; return BuildKernel( info->nameInCode, i, info->kernel_count, info->kernels[i], info->programs + i ); } static cl_int BuildKernel_DoubleFn( cl_uint job_id, cl_uint thread_id UNUSED, void *p ); static cl_int BuildKernel_DoubleFn( cl_uint job_id, cl_uint thread_id UNUSED, void *p ) { BuildKernelInfo *info = (BuildKernelInfo*) p; cl_uint i = info->offset + job_id; return BuildKernelDouble( info->nameInCode, i, info->kernel_count, info->kernels[i], info->programs + i ); } //Thread specific data for a worker thread typedef struct ThreadInfo { cl_mem inBuf; // input buffer for the thread cl_mem outBuf[ VECTOR_SIZE_COUNT ]; // output buffers for the thread float maxError; // max error value. Init to 0. double maxErrorValue; // position of the max error value. Init to 0. cl_command_queue tQueue; // per thread command queue to improve performance }ThreadInfo; typedef struct TestInfo { size_t subBufferSize; // Size of the sub-buffer in elements const Func *f; // A pointer to the function info cl_program programs[ VECTOR_SIZE_COUNT ]; // programs for various vector sizes cl_kernel *k[VECTOR_SIZE_COUNT ]; // arrays of thread-specific kernels for each worker thread: k[vector_size][thread_id] ThreadInfo *tinfo; // An array of thread specific information for each worker thread cl_uint threadCount; // Number of worker threads cl_uint jobCount; // Number of jobs cl_uint step; // step between each chunk and the next. cl_uint scale; // stride between individual test values float ulps; // max_allowed ulps int ftz; // non-zero if running in flush to zero mode int isRangeLimited; // 1 if the function is only to be evaluated over a range float half_sin_cos_tan_limit; }TestInfo; static cl_int TestFloat( cl_uint job_id, cl_uint thread_id, void *p ); int TestFunc_Float_Float(const Func *f, MTdata d) { TestInfo test_info; cl_int error; size_t i, j; float maxError = 0.0f; double maxErrorVal = 0.0; int skipTestingRelaxed = ( gTestFastRelaxed && strcmp(f->name,"tan") == 0 ); logFunctionInfo(f->name,sizeof(cl_float),gTestFastRelaxed); // Init test_info memset( &test_info, 0, sizeof( test_info ) ); test_info.threadCount = GetThreadCount(); test_info.subBufferSize = BUFFER_SIZE / (sizeof( cl_float) * RoundUpToNextPowerOfTwo(test_info.threadCount)); test_info.scale = 1; if (gWimpyMode) { test_info.subBufferSize = gWimpyBufferSize / (sizeof( cl_float) * RoundUpToNextPowerOfTwo(test_info.threadCount)); test_info.scale = (cl_uint) sizeof(cl_float) * 2 * gWimpyReductionFactor; } test_info.step = (cl_uint) test_info.subBufferSize * test_info.scale; if (test_info.step / test_info.subBufferSize != test_info.scale) { //there was overflow test_info.jobCount = 1; } else { test_info.jobCount = (cl_uint)((1ULL << 32) / test_info.step); } test_info.f = f; test_info.ulps = gIsEmbedded ? f->float_embedded_ulps : f->float_ulps; test_info.ftz = f->ftz || gForceFTZ || 0 == (CL_FP_DENORM & gFloatCapabilities); // cl_kernels aren't thread safe, so we make one for each vector size for every thread for( i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++ ) { size_t array_size = test_info.threadCount * sizeof( cl_kernel ); test_info.k[i] = (cl_kernel*)malloc( array_size ); if( NULL == test_info.k[i] ) { vlog_error( "Error: Unable to allocate storage for kernels!\n" ); error = CL_OUT_OF_HOST_MEMORY; goto exit; } memset( test_info.k[i], 0, array_size ); } test_info.tinfo = (ThreadInfo*)malloc( test_info.threadCount * sizeof(*test_info.tinfo) ); if( NULL == test_info.tinfo ) { vlog_error( "Error: Unable to allocate storage for thread specific data.\n" ); error = CL_OUT_OF_HOST_MEMORY; goto exit; } memset( test_info.tinfo, 0, test_info.threadCount * sizeof(*test_info.tinfo) ); for( i = 0; i < test_info.threadCount; i++ ) { cl_buffer_region region = { i * test_info.subBufferSize * sizeof( cl_float), test_info.subBufferSize * sizeof( cl_float) }; test_info.tinfo[i].inBuf = clCreateSubBuffer( gInBuffer, CL_MEM_READ_ONLY, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &error); if( error || NULL == test_info.tinfo[i].inBuf) { vlog_error( "Error: Unable to create sub-buffer of gInBuffer for region {%zd, %zd}\n", region.origin, region.size ); goto exit; } for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) { test_info.tinfo[i].outBuf[j] = clCreateSubBuffer( gOutBuffer[j], CL_MEM_WRITE_ONLY, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &error); if( error || NULL == test_info.tinfo[i].outBuf[j] ) { vlog_error( "Error: Unable to create sub-buffer of gInBuffer for region {%zd, %zd}\n", region.origin, region.size ); goto exit; } } test_info.tinfo[i].tQueue = clCreateCommandQueueWithProperties(gContext, gDevice, 0, &error); if( NULL == test_info.tinfo[i].tQueue || error ) { vlog_error( "clCreateCommandQueue failed. (%d)\n", error ); goto exit; } } // Check for special cases for unary float test_info.isRangeLimited = 0; test_info.half_sin_cos_tan_limit = 0; if( 0 == strcmp( f->name, "half_sin") || 0 == strcmp( f->name, "half_cos") ) { test_info.isRangeLimited = 1; test_info.half_sin_cos_tan_limit = 1.0f + test_info.ulps * (FLT_EPSILON/2.0f); // out of range results from finite inputs must be in [-1,1] } else if( 0 == strcmp( f->name, "half_tan")) { test_info.isRangeLimited = 1; test_info.half_sin_cos_tan_limit = INFINITY; // out of range resut from finite inputs must be numeric } // Init the kernels { BuildKernelInfo build_info = { gMinVectorSizeIndex, test_info.threadCount, test_info.k, test_info.programs, f->nameInCode }; if( (error = ThreadPool_Do( BuildKernel_FloatFn, gMaxVectorSizeIndex - gMinVectorSizeIndex, &build_info ) )) goto exit; } if( !gSkipCorrectnessTesting || skipTestingRelaxed) { error = ThreadPool_Do( TestFloat, test_info.jobCount, &test_info ); // Accumulate the arithmetic errors for( i = 0; i < test_info.threadCount; i++ ) { if( test_info.tinfo[i].maxError > maxError ) { maxError = test_info.tinfo[i].maxError; maxErrorVal = test_info.tinfo[i].maxErrorValue; } } if( error ) goto exit; if( gWimpyMode ) vlog( "Wimp pass" ); else vlog( "passed" ); if( skipTestingRelaxed ) { vlog(" (rlx skip correctness testing)\n"); goto exit; } } if( gMeasureTimes ) { //Init input array uint32_t *p = (uint32_t *)gIn; if( strstr( f->name, "exp" ) || strstr( f->name, "sin" ) || strstr( f->name, "cos" ) || strstr( f->name, "tan" ) ) for( j = 0; j < BUFFER_SIZE / sizeof( float ); j++ ) ((float*)p)[j] = (float) genrand_real1(d); else if( strstr( f->name, "log" ) ) for( j = 0; j < BUFFER_SIZE / sizeof( float ); j++ ) p[j] = genrand_int32(d) & 0x7fffffff; else for( j = 0; j < BUFFER_SIZE / sizeof( float ); j++ ) p[j] = genrand_int32(d); if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0, BUFFER_SIZE, gIn, 0, NULL, NULL) )) { vlog_error( "\n*** Error %d in clEnqueueWriteBuffer ***\n", error ); return error; } // Run the kernels for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) { size_t vectorSize = sizeValues[j] * sizeof(cl_float); size_t localCount = (BUFFER_SIZE + vectorSize - 1) / vectorSize; if( ( error = clSetKernelArg( test_info.k[j][0], 0, sizeof( gOutBuffer[j] ), &gOutBuffer[j] ) )) { LogBuildError( test_info.programs[j]); goto exit; } if( ( error = clSetKernelArg( test_info.k[j][0], 1, sizeof( gInBuffer ), &gInBuffer ) )) { LogBuildError(test_info.programs[j]); goto exit; } double sum = 0.0; double bestTime = INFINITY; for( i = 0; i < PERF_LOOP_COUNT; i++ ) { uint64_t startTime = GetTime(); if( (error = clEnqueueNDRangeKernel(gQueue, test_info.k[j][0], 1, NULL, &localCount, NULL, 0, NULL, NULL)) ) { vlog_error( "FAILED -- could not execute kernel\n" ); goto exit; } // Make sure OpenCL is done if( (error = clFinish(gQueue) ) ) { vlog_error( "Error %d at clFinish\n", error ); goto exit; } uint64_t endTime = GetTime(); double current_time = SubtractTime( endTime, startTime ); sum += current_time; if( current_time < bestTime ) bestTime = current_time; } if( gReportAverageTimes ) bestTime = sum / PERF_LOOP_COUNT; double clocksPerOp = bestTime * (double) gDeviceFrequency * gComputeDevices * gSimdSize * 1e6 / (BUFFER_SIZE / sizeof( float ) ); vlog_perf( clocksPerOp, LOWER_IS_BETTER, "clocks / element", "%sf%s", f->name, sizeNames[j] ); } } if( ! gSkipCorrectnessTesting ) vlog( "\t%8.2f @ %a", maxError, maxErrorVal ); vlog( "\n" ); exit: for( i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++ ) { clReleaseProgram(test_info.programs[i]); if( test_info.k[i] ) { for( j = 0; j < test_info.threadCount; j++ ) clReleaseKernel(test_info.k[i][j]); free( test_info.k[i] ); } } if( test_info.tinfo ) { for( i = 0; i < test_info.threadCount; i++ ) { clReleaseMemObject(test_info.tinfo[i].inBuf); for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) clReleaseMemObject(test_info.tinfo[i].outBuf[j]); clReleaseCommandQueue(test_info.tinfo[i].tQueue); } free( test_info.tinfo ); } return error; } static cl_int TestFloat( cl_uint job_id, cl_uint thread_id, void *data ) { const TestInfo *job = (const TestInfo *) data; size_t buffer_elements = job->subBufferSize; size_t buffer_size = buffer_elements * sizeof( cl_float ); cl_uint scale = job->scale; cl_uint base = job_id * (cl_uint) job->step; ThreadInfo *tinfo = job->tinfo + thread_id; float ulps = job->ulps; fptr func = job->f->func; const char * fname = job->f->name; if ( gTestFastRelaxed ) { ulps = job->f->relaxed_error; func = job->f->rfunc; } cl_uint j, k; cl_int error; int isRangeLimited = job->isRangeLimited; float half_sin_cos_tan_limit = job->half_sin_cos_tan_limit; int ftz = job->ftz; // start the map of the output arrays cl_event e[ VECTOR_SIZE_COUNT ]; cl_uint *out[ VECTOR_SIZE_COUNT ]; for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) { out[j] = (uint32_t*) clEnqueueMapBuffer( tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_WRITE, 0, buffer_size, 0, NULL, e + j, &error); if( error || NULL == out[j]) { vlog_error( "Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error ); return error; } } // Get that moving if( (error = clFlush(tinfo->tQueue) )) vlog( "clFlush failed\n" ); // Write the new values to the input array cl_uint *p = (cl_uint*) gIn + thread_id * buffer_elements; for( j = 0; j < buffer_elements; j++ ) { p[j] = base + j * scale; if( gTestFastRelaxed ) { float p_j = *(float *) &p[j]; if ( strcmp(fname,"sin")==0 || strcmp(fname,"cos")==0 ) //the domain of the function is [-pi,pi] { if( fabs(p_j) > M_PI ) p[j] = NAN; } if ( strcmp( fname, "reciprocal" ) == 0 ) { if( fabs(p_j) > 0x7E800000 ) //the domain of the function is [2^-126,2^126] p[j] = NAN; } } } if( (error = clEnqueueWriteBuffer( tinfo->tQueue, tinfo->inBuf, CL_FALSE, 0, buffer_size, p, 0, NULL, NULL) )) { vlog_error( "Error: clEnqueueWriteBuffer failed! err: %d\n", error ); return error; } for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) { //Wait for the map to finish if( (error = clWaitForEvents(1, e + j) )) { vlog_error( "Error: clWaitForEvents failed! err: %d\n", error ); return error; } if( (error = clReleaseEvent( e[j] ) )) { vlog_error( "Error: clReleaseEvent failed! err: %d\n", error ); return error; } // Fill the result buffer with garbage, so that old results don't carry over uint32_t pattern = 0xffffdead; memset_pattern4(out[j], &pattern, buffer_size); if( (error = clEnqueueUnmapMemObject( tinfo->tQueue, tinfo->outBuf[j], out[j], 0, NULL, NULL) )) { vlog_error( "Error: clEnqueueMapBuffer failed! err: %d\n", error ); return error; } // run the kernel size_t vectorCount = (buffer_elements + sizeValues[j] - 1) / sizeValues[j]; cl_kernel kernel = job->k[j][thread_id]; //each worker thread has its own copy of the cl_kernel cl_program program = job->programs[j]; if( ( error = clSetKernelArg( kernel, 0, sizeof( tinfo->outBuf[j] ), &tinfo->outBuf[j] ))){ LogBuildError(program); return error; } if( ( error = clSetKernelArg( kernel, 1, sizeof( tinfo->inBuf ), &tinfo->inBuf ) )) { LogBuildError(program); return error; } if( (error = clEnqueueNDRangeKernel(tinfo->tQueue, kernel, 1, NULL, &vectorCount, NULL, 0, NULL, NULL))) { vlog_error( "FAILED -- could not execute kernel\n" ); return error; } } // Get that moving if( (error = clFlush(tinfo->tQueue) )) vlog( "clFlush 2 failed\n" ); if( gSkipCorrectnessTesting ) return CL_SUCCESS; //Calculate the correctly rounded reference result float *r = (float *)gOut_Ref + thread_id * buffer_elements; float *s = (float *)p; for( j = 0; j < buffer_elements; j++ ) r[j] = (float) func.f_f( s[j] ); // Read the data back -- no need to wait for the first N-1 buffers. This is an in order queue. for( j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++ ) { out[j] = (uint32_t*) clEnqueueMapBuffer( tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0, buffer_size, 0, NULL, NULL, &error); if( error || NULL == out[j] ) { vlog_error( "Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error ); return error; } } // Wait for the last buffer out[j] = (uint32_t*) clEnqueueMapBuffer( tinfo->tQueue, tinfo->outBuf[j], CL_TRUE, CL_MAP_READ, 0, buffer_size, 0, NULL, NULL, &error); if( error || NULL == out[j] ) { vlog_error( "Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error ); return error; } //Verify data uint32_t *t = (uint32_t *)r; for( j = 0; j < buffer_elements; j++ ) { for( k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++ ) { uint32_t *q = out[k]; // If we aren't getting the correctly rounded result if( t[j] != q[j] ) { float test = ((float*) q)[j]; double correct = func.f_f( s[j] ); float err = Ulp_Error( test, correct ); float abs_error = Abs_Error( test, correct ); int fail = 0; int use_abs_error = 0; // it is possible for the output to not match the reference result but for Ulp_Error // to be zero, for example -1.#QNAN vs. 1.#QNAN. In such cases there is no failure if (err == 0.0f) { fail = 0; } else if( gTestFastRelaxed ) { if ( strcmp(fname,"sin")==0 || strcmp(fname,"cos")==0 ) { fail = ! (fabsf(abs_error) <= ulps); use_abs_error = 1; } if ( strcmp(fname, "reciprocal") == 0 ) { fail = ! (fabsf(err) <= ulps); } if ( strcmp(fname, "exp") == 0 || strcmp(fname, "exp2") == 0 ) { float exp_error = 3+floor(fabs(2*s[j])); fail = ! (fabsf(err) <= exp_error); ulps = exp_error; } if (strcmp(fname, "tan") == 0) { if( !gFastRelaxedDerived ) { fail = ! (fabsf(err) <= ulps); } // Else fast math derived implementation does not require ULP verification } if (strcmp(fname, "exp10") == 0) { if( !gFastRelaxedDerived ) { fail = ! (fabsf(err) <= ulps); } // Else fast math derived implementation does not require ULP verification } if ( strcmp(fname,"log") == 0 || strcmp(fname,"log2") == 0 ) { if( s[j] >= 0.5 && s[j] <= 2 ) { fail = ! (fabsf(abs_error) <= ulps ); } else { ulps = gIsEmbedded ? job->f->float_embedded_ulps : job->f->float_ulps; fail = ! (fabsf(err) <= ulps); } } // fast-relaxed implies finite-only if( IsFloatInfinity(correct) || IsFloatNaN(correct) || IsFloatInfinity(s[j]) || IsFloatNaN(s[j]) ) { fail = 0; err = 0; } } else { fail = ! (fabsf(err) <= ulps); } // half_sin/cos/tan are only valid between +-2**16, Inf, NaN if( isRangeLimited && fabsf(s[j]) > MAKE_HEX_FLOAT(0x1.0p16f, 0x1L, 16) && fabsf(s[j]) < INFINITY ) { if( fabsf( test ) <= half_sin_cos_tan_limit ) { err = 0; fail = 0; } } if( fail ) { if( ftz ) { typedef int (*CheckForSubnormal) (double,float); // If we are in fast relaxed math, we have a different calculation for the subnormal threshold. CheckForSubnormal isFloatResultSubnormalPtr; if ( gTestFastRelaxed ) { isFloatResultSubnormalPtr = &IsFloatResultSubnormalAbsError; } else { isFloatResultSubnormalPtr = &IsFloatResultSubnormal; } // retry per section 6.5.3.2 if( (*isFloatResultSubnormalPtr)(correct, ulps) ) { fail = fail && ( test != 0.0f ); if( ! fail ) err = 0.0f; } // retry per section 6.5.3.3 if( IsFloatSubnormal( s[j] ) ) { double correct2 = func.f_f( 0.0 ); double correct3 = func.f_f( -0.0 ); float err2; float err3; if( use_abs_error ) { err2 = Abs_Error( test, correct2 ); err3 = Abs_Error( test, correct3 ); } else { err2 = Ulp_Error( test, correct2 ); err3 = Ulp_Error( test, correct3 ); } fail = fail && ((!(fabsf(err2) <= ulps)) && (!(fabsf(err3) <= ulps))); if( fabsf( err2 ) < fabsf(err ) ) err = err2; if( fabsf( err3 ) < fabsf(err ) ) err = err3; // retry per section 6.5.3.4 if( (*isFloatResultSubnormalPtr)(correct2, ulps ) || (*isFloatResultSubnormalPtr)(correct3, ulps ) ) { fail = fail && ( test != 0.0f); if( ! fail ) err = 0.0f; } } } } if( fabsf(err ) > tinfo->maxError ) { tinfo->maxError = fabsf(err); tinfo->maxErrorValue = s[j]; } if( fail ) { vlog_error( "\nERROR: %s%s: %f ulp error at %a (0x%8.8x): *%a vs. %a\n", job->f->name, sizeNames[k], err, ((float*) s)[j], ((uint32_t*) s)[j], ((float*) t)[j], test); return -1; } } } } for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) { if( (error = clEnqueueUnmapMemObject( tinfo->tQueue, tinfo->outBuf[j], out[j], 0, NULL, NULL)) ) { vlog_error( "Error: clEnqueueUnmapMemObject %d failed 2! err: %d\n", j, error ); return error; } } if( (error = clFlush(tinfo->tQueue) )) vlog( "clFlush 3 failed\n" ); if( 0 == ( base & 0x0fffffff) ) { if (gVerboseBruteForce) { vlog("base:%14u step:%10u scale:%10u buf_elements:%10zd ulps:%5.3f ThreadCount:%2u\n", base, job->step, job->scale, buffer_elements, job->ulps, job->threadCount); } else { vlog("." ); } fflush(stdout); } return CL_SUCCESS; } static cl_int TestDouble( cl_uint job_id, cl_uint thread_id, void *data ) { const TestInfo *job = (const TestInfo *) data; size_t buffer_elements = job->subBufferSize; size_t buffer_size = buffer_elements * sizeof( cl_double ); cl_uint scale = job->scale; cl_uint base = job_id * (cl_uint) job->step; ThreadInfo *tinfo = job->tinfo + thread_id; float ulps = job->ulps; dptr func = job->f->dfunc; cl_uint j, k; cl_int error; int ftz = job->ftz; Force64BitFPUPrecision(); // start the map of the output arrays cl_event e[ VECTOR_SIZE_COUNT ]; cl_ulong *out[ VECTOR_SIZE_COUNT ]; for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) { out[j] = (cl_ulong*) clEnqueueMapBuffer( tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_WRITE, 0, buffer_size, 0, NULL, e + j, &error); if( error || NULL == out[j]) { vlog_error( "Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error ); return error; } } // Get that moving if( (error = clFlush(tinfo->tQueue) )) vlog( "clFlush failed\n" ); // Write the new values to the input array cl_double *p = (cl_double*) gIn + thread_id * buffer_elements; for( j = 0; j < buffer_elements; j++ ) p[j] = DoubleFromUInt32( base + j * scale); if( (error = clEnqueueWriteBuffer( tinfo->tQueue, tinfo->inBuf, CL_FALSE, 0, buffer_size, p, 0, NULL, NULL) )) { vlog_error( "Error: clEnqueueWriteBuffer failed! err: %d\n", error ); return error; } for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) { //Wait for the map to finish if( (error = clWaitForEvents(1, e + j) )) { vlog_error( "Error: clWaitForEvents failed! err: %d\n", error ); return error; } if( (error = clReleaseEvent( e[j] ) )) { vlog_error( "Error: clReleaseEvent failed! err: %d\n", error ); return error; } // Fill the result buffer with garbage, so that old results don't carry over uint32_t pattern = 0xffffdead; memset_pattern4(out[j], &pattern, buffer_size); if( (error = clEnqueueUnmapMemObject( tinfo->tQueue, tinfo->outBuf[j], out[j], 0, NULL, NULL) )) { vlog_error( "Error: clEnqueueMapBuffer failed! err: %d\n", error ); return error; } // run the kernel size_t vectorCount = (buffer_elements + sizeValues[j] - 1) / sizeValues[j]; cl_kernel kernel = job->k[j][thread_id]; //each worker thread has its own copy of the cl_kernel cl_program program = job->programs[j]; if( ( error = clSetKernelArg( kernel, 0, sizeof( tinfo->outBuf[j] ), &tinfo->outBuf[j] ))){ LogBuildError(program); return error; } if( ( error = clSetKernelArg( kernel, 1, sizeof( tinfo->inBuf ), &tinfo->inBuf ) )) { LogBuildError(program); return error; } if( (error = clEnqueueNDRangeKernel(tinfo->tQueue, kernel, 1, NULL, &vectorCount, NULL, 0, NULL, NULL))) { vlog_error( "FAILED -- could not execute kernel\n" ); return error; } } // Get that moving if( (error = clFlush(tinfo->tQueue) )) vlog( "clFlush 2 failed\n" ); if( gSkipCorrectnessTesting ) return CL_SUCCESS; //Calculate the correctly rounded reference result cl_double *r = (cl_double *)gOut_Ref + thread_id * buffer_elements; cl_double *s = (cl_double *)p; for( j = 0; j < buffer_elements; j++ ) r[j] = (cl_double) func.f_f( s[j] ); // Read the data back -- no need to wait for the first N-1 buffers. This is an in order queue. for( j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++ ) { out[j] = (cl_ulong*) clEnqueueMapBuffer( tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_READ, 0, buffer_size, 0, NULL, NULL, &error); if( error || NULL == out[j] ) { vlog_error( "Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error ); return error; } } // Wait for the last buffer out[j] = (cl_ulong*) clEnqueueMapBuffer( tinfo->tQueue, tinfo->outBuf[j], CL_TRUE, CL_MAP_READ, 0, buffer_size, 0, NULL, NULL, &error); if( error || NULL == out[j] ) { vlog_error( "Error: clEnqueueMapBuffer %d failed! err: %d\n", j, error ); return error; } //Verify data cl_ulong *t = (cl_ulong *)r; for( j = 0; j < buffer_elements; j++ ) { for( k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++ ) { cl_ulong *q = out[k]; // If we aren't getting the correctly rounded result if( t[j] != q[j] ) { cl_double test = ((cl_double*) q)[j]; long double correct = func.f_f( s[j] ); float err = Ulp_Error_Double( test, correct ); int fail = ! (fabsf(err) <= ulps); if( fail ) { if( ftz ) { // retry per section 6.5.3.2 if( IsDoubleResultSubnormal(correct, ulps) ) { fail = fail && ( test != 0.0f ); if( ! fail ) err = 0.0f; } // retry per section 6.5.3.3 if( IsDoubleSubnormal( s[j] ) ) { long double correct2 = func.f_f( 0.0L ); long double correct3 = func.f_f( -0.0L ); float err2 = Ulp_Error_Double( test, correct2 ); float err3 = Ulp_Error_Double( test, correct3 ); fail = fail && ((!(fabsf(err2) <= ulps)) && (!(fabsf(err3) <= ulps))); if( fabsf( err2 ) < fabsf(err ) ) err = err2; if( fabsf( err3 ) < fabsf(err ) ) err = err3; // retry per section 6.5.3.4 if( IsDoubleResultSubnormal(correct2, ulps ) || IsDoubleResultSubnormal(correct3, ulps ) ) { fail = fail && ( test != 0.0f); if( ! fail ) err = 0.0f; } } } } if( fabsf(err ) > tinfo->maxError ) { tinfo->maxError = fabsf(err); tinfo->maxErrorValue = s[j]; } if( fail ) { vlog_error( "\nERROR: %s%s: %f ulp error at %.13la (0x%16.16llx): *%.13la vs. %.13la\n", job->f->name, sizeNames[k], err, ((cl_double*) gIn)[j], ((cl_ulong*) gIn)[j], ((cl_double*) gOut_Ref)[j], test ); return -1; } } } } for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) { if( (error = clEnqueueUnmapMemObject( tinfo->tQueue, tinfo->outBuf[j], out[j], 0, NULL, NULL)) ) { vlog_error( "Error: clEnqueueUnmapMemObject %d failed 2! err: %d\n", j, error ); return error; } } if( (error = clFlush(tinfo->tQueue) )) vlog( "clFlush 3 failed\n" ); if( 0 == ( base & 0x0fffffff) ) { if (gVerboseBruteForce) { vlog("base:%14u step:%10u scale:%10zd buf_elements:%10u ulps:%5.3f ThreadCount:%2u\n", base, job->step, buffer_elements, job->scale, job->ulps, job->threadCount); } else { vlog("." ); } fflush(stdout); } return CL_SUCCESS; } int TestFunc_Double_Double(const Func *f, MTdata d) { TestInfo test_info; cl_int error; size_t i, j; float maxError = 0.0f; double maxErrorVal = 0.0; #if defined( __APPLE__ ) struct timeval time_val; gettimeofday( &time_val, NULL ); double start_time = time_val.tv_sec + 1e-6 * time_val.tv_usec; double end_time; #endif logFunctionInfo(f->name,sizeof(cl_double),gTestFastRelaxed); // Init test_info memset( &test_info, 0, sizeof( test_info ) ); test_info.threadCount = GetThreadCount(); test_info.subBufferSize = BUFFER_SIZE / (sizeof( cl_double) * RoundUpToNextPowerOfTwo(test_info.threadCount)); test_info.scale = 1; if (gWimpyMode) { test_info.subBufferSize = gWimpyBufferSize / (sizeof( cl_double) * RoundUpToNextPowerOfTwo(test_info.threadCount)); test_info.scale = (cl_uint) sizeof(cl_double) * 2 * gWimpyReductionFactor; } test_info.step = (cl_uint) test_info.subBufferSize * test_info.scale; if (test_info.step / test_info.subBufferSize != test_info.scale) { //there was overflow test_info.jobCount = 1; } else { test_info.jobCount = (cl_uint)((1ULL << 32) / test_info.step); } test_info.f = f; test_info.ulps = f->double_ulps; test_info.ftz = f->ftz || gForceFTZ; // cl_kernels aren't thread safe, so we make one for each vector size for every thread for( i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++ ) { size_t array_size = test_info.threadCount * sizeof( cl_kernel ); test_info.k[i] = (cl_kernel*)malloc( array_size ); if( NULL == test_info.k[i] ) { vlog_error( "Error: Unable to allocate storage for kernels!\n" ); error = CL_OUT_OF_HOST_MEMORY; goto exit; } memset( test_info.k[i], 0, array_size ); } test_info.tinfo = (ThreadInfo*)malloc( test_info.threadCount * sizeof(*test_info.tinfo) ); if( NULL == test_info.tinfo ) { vlog_error( "Error: Unable to allocate storage for thread specific data.\n" ); error = CL_OUT_OF_HOST_MEMORY; goto exit; } memset( test_info.tinfo, 0, test_info.threadCount * sizeof(*test_info.tinfo) ); for( i = 0; i < test_info.threadCount; i++ ) { cl_buffer_region region = { i * test_info.subBufferSize * sizeof( cl_double), test_info.subBufferSize * sizeof( cl_double) }; test_info.tinfo[i].inBuf = clCreateSubBuffer( gInBuffer, CL_MEM_READ_ONLY, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &error); if( error || NULL == test_info.tinfo[i].inBuf) { vlog_error( "Error: Unable to create sub-buffer of gInBuffer for region {%zd, %zd}\n", region.origin, region.size ); goto exit; } for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) { /* Qualcomm fix: 9461 read-write flags must be compatible with parent buffer */ test_info.tinfo[i].outBuf[j] = clCreateSubBuffer( gOutBuffer[j], CL_MEM_WRITE_ONLY, CL_BUFFER_CREATE_TYPE_REGION, ®ion, &error); /* Qualcomm fix: end */ if( error || NULL == test_info.tinfo[i].outBuf[j] ) { vlog_error( "Error: Unable to create sub-buffer of gInBuffer for region {%zd, %zd}\n", region.origin, region.size ); goto exit; } } test_info.tinfo[i].tQueue = clCreateCommandQueueWithProperties(gContext, gDevice, 0, &error); if( NULL == test_info.tinfo[i].tQueue || error ) { vlog_error( "clCreateCommandQueue failed. (%d)\n", error ); goto exit; } } // Init the kernels { BuildKernelInfo build_info = { gMinVectorSizeIndex, test_info.threadCount, test_info.k, test_info.programs, f->nameInCode }; if( (error = ThreadPool_Do( BuildKernel_DoubleFn, gMaxVectorSizeIndex - gMinVectorSizeIndex, &build_info ) )) goto exit; } if( !gSkipCorrectnessTesting ) { error = ThreadPool_Do( TestDouble, test_info.jobCount, &test_info ); // Accumulate the arithmetic errors for( i = 0; i < test_info.threadCount; i++ ) { if( test_info.tinfo[i].maxError > maxError ) { maxError = test_info.tinfo[i].maxError; maxErrorVal = test_info.tinfo[i].maxErrorValue; } } if( error ) goto exit; if( gWimpyMode ) vlog( "Wimp pass" ); else vlog( "passed" ); } #if defined( __APPLE__ ) gettimeofday( &time_val, NULL); end_time = time_val.tv_sec + 1e-6 * time_val.tv_usec; #endif if( gMeasureTimes ) { //Init input array double *p = (double *)gIn; if( strstr( f->name, "exp" ) ) for( j = 0; j < BUFFER_SIZE / sizeof( double ); j++ ) p[j] = (double)genrand_real1(d); else if( strstr( f->name, "log" ) ) for( j = 0; j < BUFFER_SIZE / sizeof( double ); j++ ) p[j] = fabs(DoubleFromUInt32( genrand_int32(d))); else for( j = 0; j < BUFFER_SIZE / sizeof( double ); j++ ) p[j] = DoubleFromUInt32( genrand_int32(d) ); if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0, BUFFER_SIZE, gIn, 0, NULL, NULL) )) { vlog_error( "\n*** Error %d in clEnqueueWriteBuffer ***\n", error ); return error; } // Run the kernels for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) { size_t vectorSize = sizeValues[j] * sizeof(cl_double); size_t localCount = (BUFFER_SIZE + vectorSize - 1) / vectorSize; if( ( error = clSetKernelArg( test_info.k[j][0], 0, sizeof( gOutBuffer[j] ), &gOutBuffer[j] ) )) { LogBuildError(test_info.programs[j]); goto exit; } if( ( error = clSetKernelArg( test_info.k[j][0], 1, sizeof( gInBuffer ), &gInBuffer ) )) { LogBuildError(test_info.programs[j]); goto exit; } double sum = 0.0; double bestTime = INFINITY; for( i = 0; i < PERF_LOOP_COUNT; i++ ) { uint64_t startTime = GetTime(); if( (error = clEnqueueNDRangeKernel(gQueue, test_info.k[j][0], 1, NULL, &localCount, NULL, 0, NULL, NULL)) ) { vlog_error( "FAILED -- could not execute kernel\n" ); goto exit; } // Make sure OpenCL is done if( (error = clFinish(gQueue) ) ) { vlog_error( "Error %d at clFinish\n", error ); goto exit; } uint64_t endTime = GetTime(); double current_time = SubtractTime( endTime, startTime ); sum += current_time; if( current_time < bestTime ) bestTime = current_time; } if( gReportAverageTimes ) bestTime = sum / PERF_LOOP_COUNT; double clocksPerOp = bestTime * (double) gDeviceFrequency * gComputeDevices * gSimdSize * 1e6 / (BUFFER_SIZE / sizeof( double ) ); vlog_perf( clocksPerOp, LOWER_IS_BETTER, "clocks / element", "%sD%s", f->name, sizeNames[j] ); } for( ; j < gMaxVectorSizeIndex; j++ ) vlog( "\t -- " ); } if( ! gSkipCorrectnessTesting ) vlog( "\t%8.2f @ %a", maxError, maxErrorVal ); #if defined( __APPLE__ ) vlog( "\t(%2.2f seconds)", end_time - start_time ); #endif vlog( "\n" ); exit: for( i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++ ) { clReleaseProgram(test_info.programs[i]); if( test_info.k[i] ) { for( j = 0; j < test_info.threadCount; j++ ) clReleaseKernel(test_info.k[i][j]); free( test_info.k[i] ); } } if( test_info.tinfo ) { for( i = 0; i < test_info.threadCount; i++ ) { clReleaseMemObject(test_info.tinfo[i].inBuf); for( j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++ ) clReleaseMemObject(test_info.tinfo[i].outBuf[j]); clReleaseCommandQueue(test_info.tinfo[i].tQueue); } free( test_info.tinfo ); } return error; }