Skip to content

Commit b678d6d

Browse files
MellichMellich
committed
Merge branch 'cl2hpp-default' into 'master'
cl2.hpp compatibility for all benchmarks See merge request pc2/HPCC_FPGA!58
2 parents 48e0386 + c993b1a commit b678d6d

13 files changed

Lines changed: 43 additions & 51 deletions

File tree

‎.gitlab-ci.yml‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -9,7 +9,7 @@ default:
99
tags:
1010
- jacamar
1111
before_script:
12-
- module load intelFPGA_pro/20.4.0 bittware_520n/20.4.0_max intel devel/CMake/3.15.3-GCCcore-8.3.0
12+
- module load intelFPGA_pro/21.1.0 bittware_520n/20.4.0_max intel devel/CMake/3.15.3-GCCcore-8.3.0
1313

1414
###
1515
#

‎FFT/src/host/execution.h‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -28,7 +28,6 @@ SOFTWARE.
2828
#include <vector>
2929

3030
/* External library headers */
31-
#include "CL/cl.hpp"
3231
#include "parameters.h"
3332
#include "fft_benchmark.hpp"
3433

‎FFT/src/host/execution_default.cpp‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -176,10 +176,10 @@ namespace bm_execution {
176176
for (uint r =0; r < config.programSettings->numRepetitions; r++) {
177177
auto startCalculation = std::chrono::high_resolution_clock::now();
178178
for (int r=0; r < config.programSettings->kernelReplications; r++) {
179-
fetchQueues[r].enqueueTask(fetchKernels[r]);
180-
fftQueues[r].enqueueTask(fftKernels[r]);
179+
fetchQueues[r].enqueueNDRangeKernel(fetchKernels[r], cl::NullRange, cl::NDRange(1), cl::NDRange(1));
180+
fftQueues[r].enqueueNDRangeKernel(fftKernels[r], cl::NullRange, cl::NDRange(1), cl::NDRange(1));
181181
#ifdef XILINX_FPGA
182-
storeQueues[r].enqueueTask(storeKernels[r]);
182+
storeQueues[r].enqueueNDRangeKernel(storeKernels[r], cl::NullRange, cl::NDRange(1), cl::NDRange(1));
183183
#endif
184184
}
185185
for (int r=0; r < config.programSettings->kernelReplications; r++) {

‎LINPACK/src/host/execution_types/execution_iec.hpp‎

Lines changed: 20 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -185,7 +185,7 @@ calculate(const hpcc_base::ExecutionSettings<linpack::LinpackProgramSettings>&co
185185
err =kernels.back().back().setArg(3, config.programSettings->matrixSize / config.programSettings->blockSize);
186186
ASSERT_CL(err)
187187
all_events.back().emplace_back();
188-
err = lu_queues.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))), &all_events.back().back());
188+
err = lu_queues.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))), &all_events.back().back());
189189
ASSERT_CL(err)
190190

191191

@@ -216,11 +216,11 @@ calculate(const hpcc_base::ExecutionSettings<linpack::LinpackProgramSettings>&co
216216

217217
if (tops + 1 == (config.programSettings->matrixSize / config.programSettings->blockSize)) {
218218
all_events.back().emplace_back();
219-
err = top_queues.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
219+
err = top_queues.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
220220
ASSERT_CL(err)
221221
}
222222
else {
223-
err = top_queues.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))));
223+
err = top_queues.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))));
224224
ASSERT_CL(err)
225225
}
226226

@@ -253,11 +253,11 @@ calculate(const hpcc_base::ExecutionSettings<linpack::LinpackProgramSettings>&co
253253

254254
if (tops + 1 == (config.programSettings->matrixSize / config.programSettings->blockSize)) {
255255
all_events.back().emplace_back();
256-
err = left_queues.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
256+
err = left_queues.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
257257
ASSERT_CL(err)
258258
}
259259
else {
260-
err = left_queues.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))));
260+
err = left_queues.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))));
261261
ASSERT_CL(err)
262262
}
263263
network_layer_op_flags[0] |= LEFT_BLOCK;
@@ -325,7 +325,7 @@ calculate(const hpcc_base::ExecutionSettings<linpack::LinpackProgramSettings>&co
325325
err = kernels.back().back().setArg(1, network_forward_flags);
326326
ASSERT_CL(err)
327327

328-
err = network_queues_bottomright.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))));
328+
err = network_queues_bottomright.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))));
329329
ASSERT_CL(err)
330330
}
331331
// Create the network kernel for down -> top direction
@@ -344,11 +344,11 @@ calculate(const hpcc_base::ExecutionSettings<linpack::LinpackProgramSettings>&co
344344

345345
if (std::distance(it,network_layer_op_flags.end()) == 1) {
346346
all_events.back().emplace_back();
347-
err = network_queues_top.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
347+
err = network_queues_top.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
348348
ASSERT_CL(err)
349349
}
350350
else {
351-
err = network_queues_top.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))));
351+
err = network_queues_top.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))));
352352
ASSERT_CL(err)
353353
}
354354

@@ -368,11 +368,11 @@ calculate(const hpcc_base::ExecutionSettings<linpack::LinpackProgramSettings>&co
368368

369369
if (std::distance(it,network_layer_op_flags.end()) == 1) {
370370
all_events.back().emplace_back();
371-
err = network_queues_left.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
371+
err = network_queues_left.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
372372
ASSERT_CL(err)
373373
}
374374
else {
375-
err = network_queues_left.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))));
375+
err = network_queues_left.back().enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))));
376376
ASSERT_CL(err)
377377
}
378378

@@ -418,16 +418,16 @@ calculate(const hpcc_base::ExecutionSettings<linpack::LinpackProgramSettings>&co
418418
// this is the last taks that will be enqueued in this queue, so create an event
419419
all_events.back().emplace_back();
420420
// Distribute the workload over all available matrix multiplication kernels
421-
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
422-
//err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &communication_events, &(all_events.back().back()));
421+
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
422+
//err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &communication_events, &(all_events.back().back()));
423423
}
424424
else {
425425
#ifndef NDEBUG
426426
std::cout << "Torus " << config.programSettings->torus_row << "," << config.programSettings->torus_col << " Inner L " << block_row << "," << block_col << std::endl;
427427
#endif
428428
// Distribute the workload over all available matrix multiplication kernels
429-
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))));
430-
//err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &communication_events);
429+
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))));
430+
//err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &communication_events);
431431
}
432432
current_update++;
433433
current_replication = (current_replication + 1) % config.programSettings->kernelReplications;
@@ -463,14 +463,14 @@ calculate(const hpcc_base::ExecutionSettings<linpack::LinpackProgramSettings>&co
463463
// this is the last taks that will be enqueued in this queue, so create an event
464464
all_events.back().emplace_back();
465465
// Distribute the workload over all available matrix multiplication kernels
466-
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
466+
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
467467
}
468468
else {
469469
#ifndef NDEBUG
470470
std::cout << "Torus " << config.programSettings->torus_row << "," << config.programSettings->torus_col << " Inner T " << block_row << "," << block_col << std::endl;
471471
#endif
472472
// Distribute the workload over all available matrix multiplication kernels
473-
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))));
473+
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))));
474474
}
475475
ASSERT_CL(err)
476476
current_update++;
@@ -518,14 +518,14 @@ calculate(const hpcc_base::ExecutionSettings<linpack::LinpackProgramSettings>&co
518518
// this is the last taks that will be enqueued in this queue, so create an event
519519
all_events.back().emplace_back();
520520
// Distribute the workload over all available matrix multiplication kernels
521-
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(std::prev(all_events.end())))), &(all_events.back().back()));
521+
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(std::prev(all_events.end())))), &(all_events.back().back()));
522522
}
523523
else {
524524
#ifndef NDEBUG
525525
std::cout << "Torus " << config.programSettings->torus_row << "," << config.programSettings->torus_col << " Inner " << block_row << "," << block_col << std::endl;
526526
#endif
527527
// Distribute the workload over all available matrix multiplication kernels
528-
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(std::prev(all_events.end())))));
528+
err = inner_queues.back()[(current_replication)].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(std::prev(all_events.end())))));
529529
}
530530

531531
ASSERT_CL(err)
@@ -563,13 +563,13 @@ calculate(const hpcc_base::ExecutionSettings<linpack::LinpackProgramSettings>&co
563563
// // this is the last taks that will be enqueued in this queue, so create an event
564564
// all_events.back().emplace_back();
565565
// // Distribute the workload over all available matrix multiplication kernels
566-
// err = inner_queues.back()[(current_replication) + 1].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
566+
// err = inner_queues.back()[(current_replication) + 1].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))), &(all_events.back().back()));
567567
// current_update = 0;
568568
// current_replication++;
569569
// }
570570
// else {
571571
// // Distribute the workload over all available matrix multiplication kernels
572-
// err = inner_queues.back()[(current_replication) + 1].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NullRange, &(*std::prev(std::prev(all_events.end()))));
572+
// err = inner_queues.back()[(current_replication) + 1].enqueueNDRangeKernel(kernels.back().back(), cl::NullRange, cl::NDRange(1), cl::NDRange(1), &(*std::prev(std::prev(all_events.end()))));
573573
// current_update++;
574574
// }
575575
}

0 commit comments

Comments
 (0)