void scan(icl_buffer *data, icl_buffer *flag, UINT n) {
// use size for actual n, not overapproximation as in allocation
UINT numWorkGroups = ((n + wx - 1) / wx);
UINT sizeScanBuff = ((numWorkGroups + wx -1) / wx) * wx;
size_t gx = numWorkGroups * wx;
INT init = -1;
#if TIMING
clFinish(dev->queue);
icl_start_timer(timer);
#endif
icl_run_kernel(perBlockScanByKey, 1, &gx, &wx, NULL, perBlockScanEvent, 8,
(size_t)0, (void *)flag,
(size_t)0, (void *)data,
sizeof(INT), &init,
sizeof(UINT), &n,
sizeof(UINT) * wx, NULL,
sizeof(INT) * wx, NULL,
(size_t)0, (void *)keySumArray,
(size_t)0, (void *)preSumArray);
/*
V_OPENCL( kernels[0].setArg( 0, firstKey.getBuffer()), "Error setArg kernels[ 0 ]" ); // Input keys
V_OPENCL( kernels[0].setArg( 1, firstValue.getBuffer()),"Error setArg kernels[ 0 ]" ); // Input buffer
V_OPENCL( kernels[0].setArg( 2, result.getBuffer( ) ), "Error setArg kernels[ 0 ]" ); // Output buffer
V_OPENCL( kernels[0].setArg( 3, init ), "Error setArg kernels[ 0 ]" ); // Initial value exclusive
V_OPENCL( kernels[0].setArg( 4, numElements ), "Error setArg kernels[ 0 ]" ); // Size of scratch buffer
V_OPENCL( kernels[0].setArg( 5, ldsKeySize, NULL ), "Error setArg kernels[ 0 ]" ); // Scratch buffer
V_OPENCL( kernels[0].setArg( 6, ldsValueSize, NULL ), "Error setArg kernels[ 0 ]" ); // Scratch buffer
V_OPENCL( kernels[0].setArg( 7, *binaryPredicateBuffer),"Error setArg kernels[ 0 ]" ); // User provided functor
V_OPENCL( kernels[0].setArg( 8, *binaryFunctionBuffer ),"Error setArg kernels[ 0 ]" ); // User provided functor
V_OPENCL( kernels[0].setArg( 9, *keySumArray ), "Error setArg kernels[ 0 ]" ); // Output per block sum
V_OPENCL( kernels[0].setArg(10, *preSumArray ), "Error setArg kernels[ 0 ]" ); // Output per block sum
V_OPENCL( kernels[0].setArg(11, doExclusiveScan ), "Error setArg kernels[ 0 ]" ); // Exclusive scan?
*/
UINT workPerThread = sizeScanBuff / wx;
icl_run_kernel(intraBlockInclusiveScanByKey, 1, &wx, &wx, NULL, intraBlockEvent, 7,
(size_t)0, (void *)keySumArray,
(size_t)0, (void *)preSumArray,
(size_t)0, (void *)postSumArray,
sizeof(UINT), &numWorkGroups,
sizeof(UINT) * wx, NULL,
sizeof(INT) * wx, NULL,
sizeof(UINT), &workPerThread);
/*
INT* output = (UINT*)malloc(n * sizeof(INT));
icl_read_buffer(keySumArray, CL_TRUE, n * sizeof(INT), output, NULL, NULL);
for(int i = 0; i < n; ++i)
printf("%d ", output[i]);
printf("\n"); */
/*
V_OPENCL( kernels[1].setArg( 0, *keySumArray ), "Error setArg kernels[ 1 ]" ); // Input keys
V_OPENCL( kernels[1].setArg( 1, *preSumArray ), "Error setArg kernels[ 1 ]" ); // Input buffer
V_OPENCL( kernels[1].setArg( 2, *postSumArray ), "Error setArg kernels[ 1 ]" ); // Output buffer
V_OPENCL( kernels[1].setArg( 3, numWorkGroupsK0 ), "Error setArg kernels[ 1 ]" ); // Size of scratch buffer
V_OPENCL( kernels[1].setArg( 4, ldsKeySize, NULL ), "Error setArg kernels[ 1 ]" ); // Scratch buffer
V_OPENCL( kernels[1].setArg( 5, ldsValueSize, NULL ), "Error setArg kernels[ 1 ]" ); // Scratch buffer
V_OPENCL( kernels[1].setArg( 6, workPerThread ), "Error setArg kernels[ 1 ]" ); // User provided functor
V_OPENCL( kernels[1].setArg( 7, *binaryPredicateBuffer ),"Error setArg kernels[ 1 ]" ); // User provided functor
V_OPENCL( kernels[1].setArg( 8, *binaryFunctionBuffer ),"Error setArg kernels[ 1 ]" ); // User provided functor
*/
icl_run_kernel(perBlockAdditionByKey, 1, &gx, &wx, NULL, perBlockAdditionEvent, 5,
(size_t)0, (void *)keySumArray,
(size_t)0, (void *)postSumArray,
(size_t)0, (void *)flag,
(size_t)0, (void *)data,
sizeof(UINT), &n);
/*
V_OPENCL( kernels[2].setArg( 0, *keySumArray ), "Error setArg kernels[ 2 ]" ); // Input buffer
V_OPENCL( kernels[2].setArg( 1, *postSumArray ), "Error setArg kernels[ 2 ]" ); // Input buffer
V_OPENCL( kernels[2].setArg( 2, firstKey.getBuffer()), "Error setArg kernels[ 2 ]" ); // Output buffer
V_OPENCL( kernels[2].setArg( 3, result.getBuffer()), "Error setArg kernels[ 2 ]" ); // Output buffer
V_OPENCL( kernels[2].setArg( 4, numElements ), "Error setArg kernels[ 2 ]" ); // Size of scratch buffer
V_OPENCL( kernels[2].setArg( 5, *binaryPredicateBuffer ),"Error setArg kernels[ 2 ]" ); // User provided functor
V_OPENCL( kernels[2].setArg( 6, *binaryFunctionBuffer ),"Error setArg kernels[ 2 ]" ); // User provided functor
*/
#if TIMING
clFinish(dev->queue);
icl_stop_timer(timer);
perBlockScanTime += icl_profile_event(perBlockScanEvent, MEASURE_START, ICL_FINISHED, ICL_MILLI);
intraBlockTime += icl_profile_event(intraBlockEvent, MEASURE_START, ICL_FINISHED, ICL_MILLI);
perBlockAdditionTime += icl_profile_event(perBlockAdditionEvent, MEASURE_START, ICL_FINISHED, ICL_MILLI);
#endif
}
int main(int argc, char* argv[]) {
int size = 1000;
int* input1 = (int*)malloc(sizeof(int) * size);
int* input2 = (int*) malloc(sizeof(int) * size);
int* output = (int *)malloc(sizeof(int) * size);
for(int i=0; i < size; ++i) {
input1[i] = i;
input2[i] = 1;
}
#ifndef INSIEME
icl_timer* time1 = icl_init_timer(ICL_SEC);
icl_start_timer(time1);
#endif
icl_init_devices(ICL_CPU);
#ifndef INSIEME
printf("TIME for initialization: %f\n", icl_stop_timer(time1));
#endif
if (icl_get_num_devices() != 0) {
icl_device* dev = icl_get_device(0);
icl_print_device_short_info(dev);
icl_kernel* kernel = icl_create_kernel(dev, "vec_mul.cl", "vec_mul", "", ICL_SOURCE);
icl_buffer* buf_input1 = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(int) * size);
icl_buffer* buf_input2 = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(int) * size);
icl_buffer* buf_output = icl_create_buffer(dev, CL_MEM_WRITE_ONLY, sizeof(int) * size);
icl_event* wb1 = icl_create_event();
icl_event* wb2 = icl_create_event();
icl_event* rb = icl_create_event();
icl_write_buffer(buf_input1, CL_FALSE, sizeof(int) * size, &input1[0], NULL, wb1);
icl_write_buffer(buf_input2, CL_FALSE, sizeof(int) * size, &input2[0], NULL, wb2);
size_t szLocalWorkSize = 256;
float multiplier = size/(float)szLocalWorkSize;
if(multiplier > (int)multiplier)
multiplier += 1;
size_t szGlobalWorkSize = (int)multiplier * szLocalWorkSize;
icl_event* rk = icl_create_event();
icl_event* wb_all = icl_create_event_list(2, wb1, wb2);
icl_run_kernel(kernel, 1, &szGlobalWorkSize, &szLocalWorkSize, wb_all, rk, 4,
(size_t)0, (void *)buf_input1,
(size_t)0, (void *)buf_input2,
(size_t)0, (void *)buf_output,
sizeof(cl_int), (void *)&size);
icl_read_buffer(buf_output, CL_TRUE, sizeof(int) * size, &output[0], rk, rb);
printf("Time wb1 %f\n", icl_profile_event(wb1, ICL_STARTED, ICL_FINISHED, ICL_SEC));
printf("Time wb2 %f\n", icl_profile_event(wb2, ICL_STARTED, ICL_FINISHED, ICL_SEC));
printf("Time rk %f\n", icl_profile_event(rk, ICL_STARTED, ICL_FINISHED, ICL_SEC));
printf("Time rb %f\n", icl_profile_event(rb, ICL_STARTED, ICL_FINISHED, ICL_SEC));
icl_release_events(5, wb1, wb2, wb_all, rk, rb);
icl_release_buffers(3, buf_input1, buf_input2, buf_output);
icl_release_kernel(kernel);
}
#ifndef INSIEME
icl_restart_timer(time1);
#endif
icl_release_devices();
#ifndef INSIEME
printf("TIME for releasing the devices: %f\n", icl_stop_timer(time1));
icl_release_timer(time1);
#endif
// CHECK for output
printf("======================\n= Vector Mul Done\n");
unsigned int check = 1;
for(unsigned int i = 0; i < size; ++i) {
if(output[i] != i*size) {
check = 0;
printf("= fail at %d, expected %d / actual %d", i, i*3/2, output[i]);
break;
}
}
printf("= result check: %s\n======================\n", check ? "OK" : "FAIL");
free(input1);
free(input2);
free(output);
}