diff options
Diffstat (limited to 'TRNS/baselines')
| -rw-r--r-- | TRNS/baselines/cpu/Makefile | 49 | ||||
| -rw-r--r-- | TRNS/baselines/cpu/README | 16 | ||||
| -rw-r--r-- | TRNS/baselines/cpu/kernel.cpp | 131 | ||||
| -rw-r--r-- | TRNS/baselines/cpu/kernel.h | 41 | ||||
| -rw-r--r-- | TRNS/baselines/cpu/main.cpp | 219 | ||||
| -rw-r--r-- | TRNS/baselines/cpu/support/common.h | 53 | ||||
| -rw-r--r-- | TRNS/baselines/cpu/support/setup.h | 50 | ||||
| -rw-r--r-- | TRNS/baselines/cpu/support/timer.h | 63 | ||||
| -rw-r--r-- | TRNS/baselines/cpu/support/verify.h | 72 | ||||
| -rw-r--r-- | TRNS/baselines/gpu/Makefile | 51 | ||||
| -rw-r--r-- | TRNS/baselines/gpu/README | 16 | ||||
| -rw-r--r-- | TRNS/baselines/gpu/kernel.cu | 170 | ||||
| -rw-r--r-- | TRNS/baselines/gpu/kernel.h | 44 | ||||
| -rw-r--r-- | TRNS/baselines/gpu/main.cpp | 298 | ||||
| -rw-r--r-- | TRNS/baselines/gpu/support/common.h | 53 | ||||
| -rw-r--r-- | TRNS/baselines/gpu/support/cuda-setup.h | 78 | ||||
| -rw-r--r-- | TRNS/baselines/gpu/support/timer.h | 73 | ||||
| -rw-r--r-- | TRNS/baselines/gpu/support/verify.h | 71 |
18 files changed, 1548 insertions, 0 deletions
diff --git a/TRNS/baselines/cpu/Makefile b/TRNS/baselines/cpu/Makefile new file mode 100644 index 0000000..cb2e264 --- /dev/null +++ b/TRNS/baselines/cpu/Makefile @@ -0,0 +1,49 @@ +# +# Copyright (c) 2016 University of Cordoba and University of Illinois +# All rights reserved. +# +# Developed by: IMPACT Research Group +# University of Cordoba and University of Illinois +# http://impact.crhc.illinois.edu/ +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to deal +# with the Software without restriction, including without limitation the +# rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +# sell copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# > Redistributions of source code must retain the above copyright notice, +# this list of conditions and the following disclaimers. +# > Redistributions in binary form must reproduce the above copyright +# notice, this list of conditions and the following disclaimers in the +# documentation and/or other materials provided with the distribution. +# > Neither the names of IMPACT Research Group, University of Cordoba, +# University of Illinois nor the names of its contributors may be used +# to endorse or promote products derived from this Software without +# specific prior written permission. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +# CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH +# THE SOFTWARE. +# + +CXX=g++ +CXX_FLAGS=-std=c++11 + +LIB=-L/usr/lib/ -lm -pthread + +DEP=kernel.cpp kernel.h main.cpp support/common.h support/setup.h support/timer.h +SRC=main.cpp kernel.cpp +EXE=trns + +all: + $(CXX) $(CXX_FLAGS) $(SRC) $(LIB) -o $(EXE) + +clean: + rm -f $(EXE) + diff --git a/TRNS/baselines/cpu/README b/TRNS/baselines/cpu/README new file mode 100644 index 0000000..eec9460 --- /dev/null +++ b/TRNS/baselines/cpu/README @@ -0,0 +1,16 @@ +In-place matrix transposition (TRNS) + +Compilation instructions + + make + +Execution instructions + + ./trns -w 0 -r 1 -m 16 -n 8 -o 4096 -p 2556 + +For more options + + ./trns -h + +Read more +J. Gomez-Luna et al., “In-place Matrix Transposition on GPUs,” IEEE TPDS, 2016. diff --git a/TRNS/baselines/cpu/kernel.cpp b/TRNS/baselines/cpu/kernel.cpp new file mode 100644 index 0000000..cd51953 --- /dev/null +++ b/TRNS/baselines/cpu/kernel.cpp @@ -0,0 +1,131 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#include "kernel.h" +#include <math.h> +#include <thread> +#include <vector> +#include <algorithm> + +// CPU threads----------------------------------------------------------------- +void run_cpu_threads_100(T *input, std::atomic_int *finished, std::atomic_int *head, int A, int B, int b, int threads) { +///////////////// Run CPU worker threads ///////////////////////////////// +#if PRINT + printf("Starting %d CPU threads\n", threads); +#endif + + std::vector<std::thread> cpu_threads; + for(int i = 0; i < threads; i++) { + + cpu_threads.push_back(std::thread([=]() { + + T data[b]; + T backup[b]; + int done; + int m = A * B - 1; + // Dynamic fetch + int gid = (head)->fetch_add(1); + + while(gid < m) { + int next_in_cycle = (gid * A) - m * (gid / B); + if(next_in_cycle == gid) { + // Dynamic fetch + gid = (head)->fetch_add(1); + continue; + } + for(int i = 0; i < b; i++) { + data[i] = input[gid * b + i]; + } + //make sure the read is not cached + done = (finished + gid)->load(); + for(; done == 0; next_in_cycle = (next_in_cycle * A) - m * (next_in_cycle / B)) { + for(int i = 0; i < b; i++) { + backup[i] = input[next_in_cycle * b + i]; + } + done = (finished + next_in_cycle)->exchange(1); + if(!done) { + for(int i = 0; i < b; i++) { + input[next_in_cycle * b + i] = data[i]; + } + } + for(int i = 0; i < b; i++) { + data[i] = backup[i]; + } + } + // Dynamic fetch + gid = (head)->fetch_add(1); + } + })); + } + + std::for_each(cpu_threads.begin(), cpu_threads.end(), [](std::thread &t) { t.join(); }); +} + + +// CPU threads----------------------------------------------------------------- +void run_cpu_threads_010(T *input, std::atomic_int* head, int a, int b, int tiles, int threads) { +///////////////// Run CPU worker threads ///////////////////////////////// +#if PRINT + printf("Starting %d CPU threads\n", threads); +#endif + + std::vector<std::thread> cpu_threads; + for(int i = 0; i < threads; i++) { + + cpu_threads.push_back(std::thread([=]() { + + T tile[a * b]; + int m = a * b - 1; + + // Dynamic fetch + int gid = (head)->fetch_add(1); + + while(gid < tiles) { + T* input_array = input + a * b * gid; + for (int j = 0; j < a * b; j++) { + int next = (j * a)-m*(j/b); + tile[next] = input_array[j]; + } + for (int j = 0; j < a * b; j++) { + input_array[j] = tile[j]; + } + // Dynamic fetch + gid = (head)->fetch_add(1); + } + })); + } + + std::for_each(cpu_threads.begin(), cpu_threads.end(), [](std::thread &t) { t.join(); }); +} diff --git a/TRNS/baselines/cpu/kernel.h b/TRNS/baselines/cpu/kernel.h new file mode 100644 index 0000000..b0eec1f --- /dev/null +++ b/TRNS/baselines/cpu/kernel.h @@ -0,0 +1,41 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#include <stdlib.h> +#include <atomic> +#include "support/common.h" + +void run_cpu_threads_100(T *input, std::atomic_int *finished, std::atomic_int *head, int A, int B, int b, int threads); +void run_cpu_threads_010(T *input, std::atomic_int *head, int m, int n, int tiles, int threads); diff --git a/TRNS/baselines/cpu/main.cpp b/TRNS/baselines/cpu/main.cpp new file mode 100644 index 0000000..2c3bdc3 --- /dev/null +++ b/TRNS/baselines/cpu/main.cpp @@ -0,0 +1,219 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#include "support/setup.h" +#include "kernel.h" +#include "support/common.h" +#include "support/timer.h" +#include "support/verify.h" + +#include <unistd.h> +#include <thread> +#include <string.h> +#include <assert.h> + +// Params --------------------------------------------------------------------- +struct Params { + + int n_threads; + int n_warmup; + int n_reps; + int M_; + int m; + int N_; + int n; + + Params(int argc, char **argv) { + n_threads = 4; + n_warmup = 5; + n_reps = 50; + M_ = 128; + m = 16; + N_ = 128; + n = 8; + int opt; + while((opt = getopt(argc, argv, "ht:w:r:m:n:o:p:")) >= 0) { + switch(opt) { + case 'h': + usage(); + exit(0); + break; + case 't': n_threads = atoi(optarg); break; + case 'w': n_warmup = atoi(optarg); break; + case 'r': n_reps = atoi(optarg); break; + case 'm': m = atoi(optarg); break; + case 'n': n = atoi(optarg); break; + case 'o': M_ = atoi(optarg); break; + case 'p': N_ = atoi(optarg); break; + default: + fprintf(stderr, "\nUnrecognized option!\n"); + usage(); + exit(0); + } + } + } + + void usage() { + fprintf(stderr, + "\nUsage: ./trns [options]" + "\n" + "\nGeneral options:" + "\n -h help" + "\n -t <T> # of host threads (default=4)" + "\n -w <W> # of untimed warmup iterations (default=5)" + "\n -r <R> # of timed repetition iterations (default=50)" + "\n" + "\nData-partitioning-specific options:" + "\n TRNS only supports CPU-only or GPU-only execution" + "\n" + "\nBenchmark-specific options:" + "\n -m <I> m (default=16 elements)" + "\n -n <I> n (default=8 elements)" + "\n -o <I> M_ (default=128 elements)" + "\n -p <I> N_ (default=128 elements)" + "\n"); + } +}; + +// Input Data ----------------------------------------------------------------- +void read_input(T *x_vector, const Params &p) { + int in_size = p.M_ * p.m * p.N_ * p.n; + srand(5432); + for(int i = 0; i < in_size; i++) { + x_vector[i] = ((T)(rand() % 100) / 100); + } +} + +// Main ------------------------------------------------------------------------------------------ +int main(int argc, char **argv) { + + const Params p(argc, argv); + Timer timer; + + // Allocate + timer.start("Allocation"); + int M_ = p.M_; + int m = p.m; + int N_ = p.N_; + int n = p.n; + int in_size = M_ * m * N_ * n; + int finished_size = M_ * m * N_; + T * h_in_out = (T *)malloc(in_size * sizeof(T)); + std::atomic_int *h_finished = + (std::atomic_int *)malloc(sizeof(std::atomic_int) * finished_size); + std::atomic_int *h_head = (std::atomic_int *)malloc(N_ * sizeof(std::atomic_int)); + ALLOC_ERR(h_in_out, h_finished, h_head); + T *h_in_backup = (T *)malloc(in_size * sizeof(T)); + ALLOC_ERR(h_in_backup); + timer.stop("Allocation"); + timer.print("Allocation", 1); + + // Initialize + timer.start("Initialization"); + read_input(h_in_out, p); + memset((void *)h_finished, 0, sizeof(std::atomic_int) * finished_size); + for(int i = 0; i < N_; i++) + h_head[i].store(0); + timer.stop("Initialization"); + timer.print("Initialization", 1); + memcpy(h_in_backup, h_in_out, in_size * sizeof(T)); // Backup for reuse across iterations + + // Loop over main kernel + for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) { + + // Reset + memcpy(h_in_out, h_in_backup, in_size * sizeof(T)); + memset((void *)h_finished, 0, sizeof(std::atomic_int) * finished_size); + for(int i = 0; i < N_; i++) + h_head[i].store(0); + + // start timer + if(rep >= p.n_warmup) + timer.start("Step 1"); + // Launch CPU threads + std::thread main_thread_1(run_cpu_threads_100, h_in_out, h_finished, h_head, M_ * m, N_, n, p.n_threads); //M_ * m * N_); + main_thread_1.join(); + // end timer + if(rep >= p.n_warmup) + timer.stop("Step 1"); + + for(int i = 0; i < N_; i++) + h_head[i].store(0); + + // start timer + if(rep >= p.n_warmup) + timer.start("Step 2"); + // Launch CPU threads + std::thread main_thread_2(run_cpu_threads_010, h_in_out, h_head, m, n, M_ * N_, p.n_threads); + main_thread_2.join(); + // end timer + if(rep >= p.n_warmup) + timer.stop("Step 2"); + + memset((void *)h_finished, 0, sizeof(std::atomic_int) * finished_size); + for(int i = 0; i < N_; i++) + h_head[i].store(0); + + // start timer + if(rep >= p.n_warmup) + timer.start("Step 3"); + // Launch CPU threads + for(int i = 0; i < N_; i++){ + std::thread main_thread_3(run_cpu_threads_100, h_in_out + i * M_ * n * m, h_finished + i * M_ * n, h_head + i, M_, n, m, p.n_threads); //M_ * n); + main_thread_3.join(); + } + // end timer + if(rep >= p.n_warmup) + timer.stop("Step 3"); + } + timer.print("Step 1", p.n_reps); + timer.print("Step 2", p.n_reps); + timer.print("Step 3", p.n_reps); + + // Verify answer + //verify(h_in_out, h_in_backup, M_ * m, N_ * n, 1); + + // Free memory + timer.start("Deallocation"); + free(h_in_out); + free(h_finished); + free(h_head); + free(h_in_backup); + timer.stop("Deallocation"); + timer.print("Deallocation", 1); + + printf("Test Passed\n"); + return 0; +} diff --git a/TRNS/baselines/cpu/support/common.h b/TRNS/baselines/cpu/support/common.h new file mode 100644 index 0000000..f03900a --- /dev/null +++ b/TRNS/baselines/cpu/support/common.h @@ -0,0 +1,53 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#ifndef _COMMON_H_ +#define _COMMON_H_ + +#ifndef DOUBLE_PRECISION +#define DOUBLE_PRECISION 1 +#endif + +#if DOUBLE_PRECISION +#define T double +#else +#define T float +#endif + +#define PRINT 0 + +#define divceil(n, m) (((n)-1) / (m) + 1) + +#endif diff --git a/TRNS/baselines/cpu/support/setup.h b/TRNS/baselines/cpu/support/setup.h new file mode 100644 index 0000000..a978152 --- /dev/null +++ b/TRNS/baselines/cpu/support/setup.h @@ -0,0 +1,50 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#include <fstream> + +// Allocation error checking +#define ERR_1(v1) \ + if(v1 == NULL) { \ + fprintf(stderr, "Allocation error at %s, %d\n", __FILE__, __LINE__); \ + exit(-1); \ + } +#define ERR_2(v1,v2) ERR_1(v1) ERR_1(v2) +#define ERR_3(v1,v2,v3) ERR_2(v1,v2) ERR_1(v3) +#define ERR_4(v1,v2,v3,v4) ERR_3(v1,v2,v3) ERR_1(v4) +#define ERR_5(v1,v2,v3,v4,v5) ERR_4(v1,v2,v3,v4) ERR_1(v5) +#define ERR_6(v1,v2,v3,v4,v5,v6) ERR_5(v1,v2,v3,v4,v5) ERR_1(v6) +#define GET_ERR_MACRO(_1,_2,_3,_4,_5,_6,NAME,...) NAME +#define ALLOC_ERR(...) GET_ERR_MACRO(__VA_ARGS__,ERR_6,ERR_5,ERR_4,ERR_3,ERR_2,ERR_1)(__VA_ARGS__) diff --git a/TRNS/baselines/cpu/support/timer.h b/TRNS/baselines/cpu/support/timer.h new file mode 100644 index 0000000..70ee386 --- /dev/null +++ b/TRNS/baselines/cpu/support/timer.h @@ -0,0 +1,63 @@ +/*
+ * Copyright (c) 2016 University of Cordoba and University of Illinois
+ * All rights reserved.
+ *
+ * Developed by: IMPACT Research Group
+ * University of Cordoba and University of Illinois
+ * http://impact.crhc.illinois.edu/
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * with the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * > Redistributions of source code must retain the above copyright notice,
+ * this list of conditions and the following disclaimers.
+ * > Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimers in the
+ * documentation and/or other materials provided with the distribution.
+ * > Neither the names of IMPACT Research Group, University of Cordoba,
+ * University of Illinois nor the names of its contributors may be used
+ * to endorse or promote products derived from this Software without
+ * specific prior written permission.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+ * THE SOFTWARE.
+ *
+ */
+
+#include <sys/time.h>
+#include <iostream>
+#include <map>
+#include <string>
+
+using namespace std;
+
+struct Timer {
+
+ map<string, struct timeval> startTime;
+ map<string, struct timeval> stopTime;
+ map<string, double> time;
+
+ void start(string name) {
+ if(!time.count(name)) {
+ time[name] = 0.0;
+ }
+ gettimeofday(&startTime[name], NULL);
+ }
+
+ void stop(string name) {
+ gettimeofday(&stopTime[name], NULL);
+ time[name] += (stopTime[name].tv_sec - startTime[name].tv_sec) * 1000000.0 +
+ (stopTime[name].tv_usec - startTime[name].tv_usec);
+ }
+
+ void print(string name, int REP) { printf("%s Time (ms): %f\n", name.c_str(), time[name] / (1000 * REP)); }
+};
diff --git a/TRNS/baselines/cpu/support/verify.h b/TRNS/baselines/cpu/support/verify.h new file mode 100644 index 0000000..ea02f0c --- /dev/null +++ b/TRNS/baselines/cpu/support/verify.h @@ -0,0 +1,72 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#include "common.h" +#include <math.h> + +inline int compare_output(T *output, T *ref, int dim) { + int i; + for(i = 0; i < dim; i++) { + T diff = fabs(ref[i] - output[i]); + if((diff - 0.0f) > 0.00001f && diff > 0.01 * fabs(ref[i])) { + printf("line: %d ref: %f actual: %f diff: %f\n", i, ref[i], output[i], diff); + exit(EXIT_FAILURE); + } + } + return 0; +} + +// Sequential transposition for comparison purposes +//[w][h/t][t] to [h/t][w][t] +static void trns_host(T* input, unsigned int A, unsigned int B, unsigned int b){ + T* output = (T*) malloc(sizeof(T) * A * B * b); + unsigned int next; + for (unsigned int j = 0; j < b; j++){ + for (unsigned int i = 0; i < A * B; i++){ + next = (i * A) - (A * B - 1) * (i / B); + output[next * b + j] = input[i*b+j]; + } + } + for (unsigned int k = 0; k < A * B * b; k++){ + input[k] = output[k]; + } + free(output); +} + +inline void verify(T *input2, T *input, int height, int width, int tile_size) { + trns_host(input, height, width, tile_size); + compare_output(input2, input, height * width); +} + diff --git a/TRNS/baselines/gpu/Makefile b/TRNS/baselines/gpu/Makefile new file mode 100644 index 0000000..12395b7 --- /dev/null +++ b/TRNS/baselines/gpu/Makefile @@ -0,0 +1,51 @@ +# +# Copyright (c) 2016 University of Cordoba and University of Illinois +# All rights reserved. +# +# Developed by: IMPACT Research Group +# University of Cordoba and University of Illinois +# http://impact.crhc.illinois.edu/ +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to deal +# with the Software without restriction, including without limitation the +# rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +# sell copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# > Redistributions of source code must retain the above copyright notice, +# this list of conditions and the following disclaimers. +# > Redistributions in binary form must reproduce the above copyright +# notice, this list of conditions and the following disclaimers in the +# documentation and/or other materials provided with the distribution. +# > Neither the names of IMPACT Research Group, University of Cordoba, +# University of Illinois nor the names of its contributors may be used +# to endorse or promote products derived from this Software without +# specific prior written permission. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +# CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH +# THE SOFTWARE. +# + +CXX=/usr/local/cuda/bin/nvcc +CXX_FLAGS=-std=c++11 + +LIB=-L/usr/lib/ -L$/usr/local/cuda/lib64/ -lm + +INC=-I/usr/local/cuda/include/ + +DEP=kernel.h main.cpp kernel.cu support/common.h support/cuda-setup.h support/timer.h support/verify.h +SRC=main.cpp kernel.cu +EXE=trns + +all: + $(CXX) $(CXX_FLAGS) $(SRC) $(LIB) $(INC) -o $(EXE) + +clean: + rm -f $(EXE) + diff --git a/TRNS/baselines/gpu/README b/TRNS/baselines/gpu/README new file mode 100644 index 0000000..c36ee77 --- /dev/null +++ b/TRNS/baselines/gpu/README @@ -0,0 +1,16 @@ +In-place matrix transposition (TRNS) + +Compilation instructions + + make + +Execution instructions + + ./trns -w 0 -r 1 -m 16 -n 8 -o 4096 -p 2556 -i 64 + +For more options + + ./trns -h + +Read more +J. Gomez-Luna et al., “In-place Matrix Transposition on GPUs,” IEEE TPDS, 2016. diff --git a/TRNS/baselines/gpu/kernel.cu b/TRNS/baselines/gpu/kernel.cu new file mode 100644 index 0000000..f2251cd --- /dev/null +++ b/TRNS/baselines/gpu/kernel.cu @@ -0,0 +1,170 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#include "support/common.h" + +extern __shared__ int l_mem[]; + +// GPU kernel ------------------------------------------------------------------------------------------ +__global__ void PTTWAC_soa_asta(int A, int B, int b, T *input, int *finished, int *head) { + + int* done = l_mem; + int* gid_ = &done[1]; + + const int tid = threadIdx.x; + int m = A * B - 1; + + if(tid == 0) // Dynamic fetch + gid_[0] = atomicAdd(&head[0], 1); + __syncthreads(); + + while(gid_[0] < m) { + int next_in_cycle = (gid_[0] * A) - m * (gid_[0] / B); + if(next_in_cycle == gid_[0]) { + if(tid == 0) // Dynamic fetch + gid_[0] = atomicAdd(&head[0], 1); + __syncthreads(); + continue; + } + T data1, data2, data3, data4; + int i = tid; + if(i < b) + data1 = input[gid_[0] * b + i]; + i += blockDim.x; + if(i < b) + data2 = input[gid_[0] * b + i]; + i += blockDim.x; + if(i < b) + data3 = input[gid_[0] * b + i]; + i += blockDim.x; + if(i < b) + data4 = input[gid_[0] * b + i]; + + if(tid == 0) { + //make sure the read is not cached + done[0] = atomicAdd(&finished[gid_[0]], 0); + } + __syncthreads(); + + for(; done[0] == 0; next_in_cycle = (next_in_cycle * A) - m * (next_in_cycle / B)) { + T backup1, backup2, backup3, backup4; + i = tid; + if(i < b) + backup1 = input[next_in_cycle * b + i]; + i += blockDim.x; + if(i < b) + backup2 = input[next_in_cycle * b + i]; + i += blockDim.x; + if(i < b) + backup3 = input[next_in_cycle * b + i]; + i += blockDim.x; + if(i < b) + backup4 = input[next_in_cycle * b + i]; + + if(tid == 0) { + done[0] = atomicExch(&finished[next_in_cycle], (int)1); + } + __syncthreads(); + + if(!done[0]) { + i = tid; + if(i < b) + input[next_in_cycle * b + i] = data1; + i += blockDim.x; + if(i < b) + input[next_in_cycle * b + i] = data2; + i += blockDim.x; + if(i < b) + input[next_in_cycle * b + i] = data3; + i += blockDim.x; + if(i < b) + input[next_in_cycle * b + i] = data4; + } + i = tid; + if(i < b) + data1 = backup1; + i += blockDim.x; + if(i < b) + data2 = backup2; + i += blockDim.x; + if(i < b) + data3 = backup3; + i += blockDim.x; + if(i < b) + data4 = backup4; + } + + if(tid == 0) // Dynamic fetch + gid_[0] = atomicAdd(&head[0], 1); + __syncthreads(); + } +} + +cudaError_t call_PTTWAC_soa_asta(int blocks, int threads, int A, int B, int b, T *input, + int *finished, int *head, int l_mem_size){ + dim3 dimGrid(blocks); + dim3 dimBlock(threads); + PTTWAC_soa_asta<<<dimGrid, dimBlock, l_mem_size>>>(A, B, b, input, + finished, head); + cudaError_t err = cudaGetLastError(); + return err; +} + +__global__ void BS_marshal(T *input, int tile_size, int width) { + + T* tile = (T*)l_mem; + + int tidx = threadIdx.x; + int m = width*tile_size-1; + int bid = blockIdx.x; + + input += tile_size*width*bid; + for (int i = tidx; i < tile_size*width; i+=blockDim.x) { + int next = (i * tile_size)-m*(i/width); + tile[next] = input[i]; + } + __syncthreads(); + for (int i = tidx; i < tile_size*width; i+=blockDim.x) { + input[i] = tile[i]; + } +} + +cudaError_t call_BS_marshal(int blocks, int threads, int m, int n, T *input, int l_mem_size){ + dim3 dimGrid(blocks); + dim3 dimBlock(threads); + BS_marshal<<<dimGrid, dimBlock, l_mem_size>>>(input, m, n); + cudaError_t err = cudaGetLastError(); + return err; +} diff --git a/TRNS/baselines/gpu/kernel.h b/TRNS/baselines/gpu/kernel.h new file mode 100644 index 0000000..47ecc16 --- /dev/null +++ b/TRNS/baselines/gpu/kernel.h @@ -0,0 +1,44 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#include "cuda_runtime.h" +#include <stdlib.h> +#include <atomic> +#include "support/common.h" + +cudaError_t call_PTTWAC_soa_asta(int blocks, int threads, int A, int B, int b, T *input, + int *finished, int *head, int l_mem_size); + +cudaError_t call_BS_marshal(int blocks, int threads, int m, int n, T *input, int l_mem_size); diff --git a/TRNS/baselines/gpu/main.cpp b/TRNS/baselines/gpu/main.cpp new file mode 100644 index 0000000..fa4c1e5 --- /dev/null +++ b/TRNS/baselines/gpu/main.cpp @@ -0,0 +1,298 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#include "support/cuda-setup.h" +#include "kernel.h" +#include "support/common.h" +#include "support/timer.h" +#include "support/verify.h" + +#include <unistd.h> +#include <thread> +#include <string.h> +#include <assert.h> + +// Params --------------------------------------------------------------------- +struct Params { + + int device; + int n_gpu_threads; + int n_gpu_blocks; + int n_threads; + int n_warmup; + int n_reps; + int M_; + int m; + int N_; + int n; + + Params(int argc, char **argv) { + device = 0; + n_gpu_threads = 64; + n_gpu_blocks = 16; + n_warmup = 5; + n_reps = 50; + M_ = 128; + m = 16; + N_ = 128; + n = 8; + int opt; + while((opt = getopt(argc, argv, "hd:i:g:t:w:r:m:n:o:p:")) >= 0) { + switch(opt) { + case 'h': + usage(); + exit(0); + break; + case 'd': device = atoi(optarg); break; + case 'i': n_gpu_threads = atoi(optarg); break; + case 'g': n_gpu_blocks = atoi(optarg); break; + case 't': n_threads = atoi(optarg); break; + case 'w': n_warmup = atoi(optarg); break; + case 'r': n_reps = atoi(optarg); break; + case 'm': m = atoi(optarg); break; + case 'n': n = atoi(optarg); break; + case 'o': M_ = atoi(optarg); break; + case 'p': N_ = atoi(optarg); break; + default: + fprintf(stderr, "\nUnrecognized option!\n"); + usage(); + exit(0); + } + } + assert((n_gpu_threads > 0 && n_gpu_blocks > 0) + && "TRNS only runs on CPU-only or GPU-only: './trns -g 0' or './trns -t 0'"); + } + + void usage() { + fprintf(stderr, + "\nUsage: ./trns [options]" + "\n" + "\nGeneral options:" + "\n -h help" + "\n -d <D> CUDA device ID (default=0)" + "\n -i <I> # of device threads per block (default=64)" + "\n -g <G> # of device blocks (default=16)" + "\n -w <W> # of untimed warmup iterations (default=5)" + "\n -r <R> # of timed repetition iterations (default=50)" + "\n" + "\nData-partitioning-specific options:" + "\n TRNS only supports CPU-only or GPU-only execution" + "\n" + "\nBenchmark-specific options:" + "\n -m <I> m (default=16 elements)" + "\n -n <I> n (default=8 elements)" + "\n -o <I> M_ (default=128 elements)" + "\n -p <I> N_ (default=128 elements)" + "\n"); + } +}; + +// Input Data ----------------------------------------------------------------- +void read_input(T *x_vector, const Params &p) { + int in_size = p.M_ * p.m * p.N_ * p.n; + srand(5432); + for(int i = 0; i < in_size; i++) { + x_vector[i] = ((T)(rand() % 100) / 100); + } +} + +// Main ------------------------------------------------------------------------------------------ +int main(int argc, char **argv) { + + const Params p(argc, argv); + CUDASetup setcuda(p.device); + Timer timer; + cudaError_t cudaStatus; + + // Allocate + timer.start("Allocation"); + int M_ = p.M_; + int m = p.m; + int N_ = p.N_; + int n = p.n; + int in_size = M_ * m * N_ * n; + int finished_size = M_ * m * N_; + T * h_in_out = (T *)malloc(in_size * sizeof(T)); + std::atomic_int *h_finished = + (std::atomic_int *)malloc(sizeof(std::atomic_int) * finished_size); + std::atomic_int *h_head = (std::atomic_int *)malloc(N_ * sizeof(std::atomic_int)); + ALLOC_ERR(h_in_out, h_finished, h_head); + T * d_in_out; + int * d_finished; + int * d_head; + if(p.n_gpu_blocks != 0) { + cudaStatus = cudaMalloc((void**)&d_in_out, in_size * sizeof(T)); + cudaStatus = cudaMalloc((void**)&d_finished, (p.n_gpu_blocks != 0) ? sizeof(int) * finished_size : 0); + cudaStatus = cudaMalloc((void**)&d_head, (p.n_gpu_blocks != 0) ? N_ * sizeof(int) : 0); + CUDA_ERR(); + } + T *h_in_backup = (T *)malloc(in_size * sizeof(T)); + ALLOC_ERR(h_in_backup); + cudaDeviceSynchronize(); + timer.stop("Allocation"); + timer.print("Allocation", 1); + + // Initialize + timer.start("Initialization"); + const int max_gpu_threads = setcuda.max_gpu_threads(); + read_input(h_in_out, p); + memset((void *)h_finished, 0, sizeof(std::atomic_int) * finished_size); + for(int i = 0; i < N_; i++) + h_head[i].store(0); + timer.stop("Initialization"); + timer.print("Initialization", 1); + memcpy(h_in_backup, h_in_out, in_size * sizeof(T)); // Backup for reuse across iterations + + // Copy to device + timer.start("Copy To Device"); + if(p.n_gpu_blocks != 0) { + cudaStatus = cudaMemcpy(d_in_out, h_in_backup, in_size * sizeof(T), cudaMemcpyHostToDevice); + cudaStatus = cudaMemcpy(d_finished, h_finished, sizeof(int) * finished_size, cudaMemcpyHostToDevice); + cudaStatus = cudaMemcpy(d_head, h_head, N_ * sizeof(int), cudaMemcpyHostToDevice); + CUDA_ERR(); + } + cudaDeviceSynchronize(); + timer.stop("Copy To Device"); + timer.print("Copy To Device", 1); + + // Loop over main kernel + for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) { + + // Reset + memcpy(h_in_out, h_in_backup, in_size * sizeof(T)); + memset((void *)h_finished, 0, sizeof(std::atomic_int) * finished_size); + for(int i = 0; i < N_; i++) + h_head[i].store(0); + cudaDeviceSynchronize(); + + // Launch GPU threads + if(p.n_gpu_blocks > 0) { + // Kernel launch + assert(p.n_gpu_threads <= max_gpu_threads && + "The thread block size is greater than the maximum thread block size that can be used on this device"); + + cudaStatus = cudaMemcpy(d_in_out, h_in_backup, in_size * sizeof(T), cudaMemcpyHostToDevice); + cudaStatus = cudaMemcpy(d_finished, h_finished, sizeof(int) * finished_size, cudaMemcpyHostToDevice); + cudaStatus = cudaMemcpy(d_head, h_head, N_ * sizeof(int), cudaMemcpyHostToDevice); + CUDA_ERR(); + + // start timer + if(rep >= p.n_warmup) + timer.start("Step 1"); + // Step 1 + cudaStatus = call_PTTWAC_soa_asta(M_ * m * N_, p.n_gpu_threads, M_ * m, N_, n, + d_in_out, (int*)d_finished, (int*)d_head, sizeof(int) + sizeof(int)); + CUDA_ERR(); + // end timer + if(rep >= p.n_warmup) + timer.stop("Step 1"); + + // start timer + if(rep >= p.n_warmup) + timer.start("Step 2"); + // Step 2 + cudaStatus = call_BS_marshal(M_ * N_, p.n_gpu_threads, m, n, d_in_out, m * n * sizeof(T)); + CUDA_ERR(); + // end timer + if(rep >= p.n_warmup) + timer.stop("Step 2"); + + cudaStatus = cudaMemcpy(d_finished, h_finished, sizeof(int) * finished_size, cudaMemcpyHostToDevice); + cudaStatus = cudaMemcpy(d_head, h_head, N_ * sizeof(int), cudaMemcpyHostToDevice); + CUDA_ERR(); + // start timer + if(rep >= p.n_warmup) + timer.start("Step 3"); + // Step 3 + for(int i = 0; i < N_; i++){ + cudaStatus = call_PTTWAC_soa_asta(M_ * n, p.n_gpu_threads, M_, n, m, + d_in_out + i * M_ * n * m, (int*)d_finished + i * M_ * n, (int*)d_head + i, sizeof(int) + sizeof(int)); + CUDA_ERR(); + } + // end timer + if(rep >= p.n_warmup) + timer.stop("Step 3"); + + } + + cudaDeviceSynchronize(); + + } + timer.print("Step 1", p.n_reps); + timer.print("Step 2", p.n_reps); + timer.print("Step 3", p.n_reps); + + // Copy back + timer.start("Copy Back and Merge"); + if(p.n_gpu_blocks != 0) { + cudaStatus = cudaMemcpy(h_in_out, d_in_out, in_size * sizeof(T), cudaMemcpyDeviceToHost); + CUDA_ERR(); + cudaDeviceSynchronize(); + } + timer.stop("Copy Back and Merge"); + timer.print("Copy Back and Merge", 1); + + // Verify answer + verify(h_in_out, h_in_backup, M_ * m, N_ * n, 1); + + // Free memory + timer.start("Deallocation"); + free(h_in_out); + free(h_finished); + free(h_head); + if(p.n_gpu_blocks != 0) { + cudaStatus = cudaFree(d_in_out); + cudaStatus = cudaFree(d_finished); + cudaStatus = cudaFree(d_head); + CUDA_ERR(); + } + free(h_in_backup); + cudaDeviceSynchronize(); + timer.stop("Deallocation"); + timer.print("Deallocation", 1); + + // Release timers + timer.release("Allocation"); + timer.release("Initialization"); + timer.release("Copy To Device"); + timer.release("Step 1"); + timer.release("Step 2"); + timer.release("Step 3"); + timer.release("Copy Back and Merge"); + timer.release("Deallocation"); + + printf("Test Passed\n"); + return 0; +} diff --git a/TRNS/baselines/gpu/support/common.h b/TRNS/baselines/gpu/support/common.h new file mode 100644 index 0000000..8a7f37f --- /dev/null +++ b/TRNS/baselines/gpu/support/common.h @@ -0,0 +1,53 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#ifndef _COMMON_H_ +#define _COMMON_H_ + +#ifndef DOUBLE_PRECISION +#define DOUBLE_PRECISION 1 +#endif + +#if DOUBLE_PRECISION +#define T long int // double +#else +#define T int // float +#endif + +#define PRINT 0 + +#define divceil(n, m) (((n)-1) / (m) + 1) + +#endif diff --git a/TRNS/baselines/gpu/support/cuda-setup.h b/TRNS/baselines/gpu/support/cuda-setup.h new file mode 100644 index 0000000..7b7eefe --- /dev/null +++ b/TRNS/baselines/gpu/support/cuda-setup.h @@ -0,0 +1,78 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#include <cuda.h> +#include <cuda_runtime.h> +#include <fstream> + +// Allocation error checking +#define ERR_1(v1) \ + if(v1 == NULL) { \ + fprintf(stderr, "Allocation error at %s, %d\n", __FILE__, __LINE__); \ + exit(-1); \ + } +#define ERR_2(v1,v2) ERR_1(v1) ERR_1(v2) +#define ERR_3(v1,v2,v3) ERR_2(v1,v2) ERR_1(v3) +#define ERR_4(v1,v2,v3,v4) ERR_3(v1,v2,v3) ERR_1(v4) +#define ERR_5(v1,v2,v3,v4,v5) ERR_4(v1,v2,v3,v4) ERR_1(v5) +#define ERR_6(v1,v2,v3,v4,v5,v6) ERR_5(v1,v2,v3,v4,v5) ERR_1(v6) +#define GET_ERR_MACRO(_1,_2,_3,_4,_5,_6,NAME,...) NAME +#define ALLOC_ERR(...) GET_ERR_MACRO(__VA_ARGS__,ERR_6,ERR_5,ERR_4,ERR_3,ERR_2,ERR_1)(__VA_ARGS__) + +#define CUDA_ERR() \ + if(cudaStatus != cudaSuccess) { \ + fprintf(stderr, "CUDA error: %s\n at %s, %d\n", cudaGetErrorString(cudaStatus), __FILE__, __LINE__); \ + exit(-1); \ + } + +struct CUDASetup { + + cudaDeviceProp device_prop; + + CUDASetup(int device) { + cudaError_t cudaStatus; + cudaStatus = cudaSetDevice(device); + CUDA_ERR(); + + cudaStatus = cudaGetDeviceProperties(&device_prop, device); + CUDA_ERR(); + fprintf(stderr, "%s\t", device_prop.name); + + } + + int max_gpu_threads() { + return device_prop.maxThreadsPerBlock; + } +}; diff --git a/TRNS/baselines/gpu/support/timer.h b/TRNS/baselines/gpu/support/timer.h new file mode 100644 index 0000000..fceab04 --- /dev/null +++ b/TRNS/baselines/gpu/support/timer.h @@ -0,0 +1,73 @@ +/*
+ * Copyright (c) 2016 University of Cordoba and University of Illinois
+ * All rights reserved.
+ *
+ * Developed by: IMPACT Research Group
+ * University of Cordoba and University of Illinois
+ * http://impact.crhc.illinois.edu/
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * with the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * > Redistributions of source code must retain the above copyright notice,
+ * this list of conditions and the following disclaimers.
+ * > Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimers in the
+ * documentation and/or other materials provided with the distribution.
+ * > Neither the names of IMPACT Research Group, University of Cordoba,
+ * University of Illinois nor the names of its contributors may be used
+ * to endorse or promote products derived from this Software without
+ * specific prior written permission.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+ * THE SOFTWARE.
+ *
+ */
+
+#include <cuda_runtime.h>
+#include <sys/time.h>
+#include <iostream>
+#include <map>
+#include <string>
+
+using namespace std;
+
+struct Timer {
+
+ map<string, cudaEvent_t> startTime;
+ map<string, cudaEvent_t> stopTime;
+ map<string, float> time;
+
+ void start(string name) {
+ if(!time.count(name)) {
+ cudaEventCreate(&startTime[name]);
+ cudaEventCreate(&stopTime[name]);
+ time[name] = 0.0;
+ }
+ cudaEventRecord(startTime[name], 0);
+ }
+
+ void stop(string name) {
+ cudaEventRecord(stopTime[name],0);
+ cudaEventSynchronize(stopTime[name]);
+ float part_time = 0.0;
+ cudaEventElapsedTime(&part_time, startTime[name], stopTime[name]);
+ time[name] += part_time;
+ }
+
+ void print(string name, unsigned int REP) { printf("%s Time (ms): %f\n", name.c_str(), time[name] / REP); }
+
+ void release(string name){
+ cudaEventDestroy(startTime[name]);
+ cudaEventDestroy(stopTime[name]);
+ }
+};
diff --git a/TRNS/baselines/gpu/support/verify.h b/TRNS/baselines/gpu/support/verify.h new file mode 100644 index 0000000..c3ba224 --- /dev/null +++ b/TRNS/baselines/gpu/support/verify.h @@ -0,0 +1,71 @@ +/* + * Copyright (c) 2016 University of Cordoba and University of Illinois + * All rights reserved. + * + * Developed by: IMPACT Research Group + * University of Cordoba and University of Illinois + * http://impact.crhc.illinois.edu/ + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * with the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * > Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimers. + * > Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimers in the + * documentation and/or other materials provided with the distribution. + * > Neither the names of IMPACT Research Group, University of Cordoba, + * University of Illinois nor the names of its contributors may be used + * to endorse or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH + * THE SOFTWARE. + * + */ + +#include "common.h" +#include <math.h> + +inline int compare_output(T *output, T *ref, int dim) { + int i; + for(i = 0; i < dim; i++) { + T diff = fabs(ref[i] - output[i]); + if((diff - 0.0f) > 0.00001f && diff > 0.01 * fabs(ref[i])) { + printf("line: %d ref: %f actual: %f diff: %f\n", i, ref[i], output[i], diff); + exit(EXIT_FAILURE); + } + } + return 0; +} + +// Sequential transposition for comparison purposes +//[w][h/t][t] to [h/t][w][t] +static void trns_host(T* input, unsigned int A, unsigned int B, unsigned int b){ + T* output = (T*) malloc(sizeof(T) * A * B * b); + unsigned int next; + for (unsigned int j = 0; j < b; j++){ + for (unsigned int i = 0; i < A * B; i++){ + next = (i * A) - (A * B - 1) * (i / B); + output[next * b + j] = input[i*b+j]; + } + } + for (unsigned int k = 0; k < A * B * b; k++){ + input[k] = output[k]; + } + free(output); +} + +inline void verify(T *input2, T *input, int height, int width, int tile_size) { + trns_host(input, height, width, tile_size); + compare_output(input2, input, height * width); +} |