diff options
Diffstat (limited to 'TS')
| -rw-r--r-- | TS/Makefile | 27 | ||||
| -rw-r--r-- | TS/baselines/cpu/mprofile.h | 6 | ||||
| -rwxr-xr-x | TS/benchmark-scripts/ccmcc25-sim.sh | 25 | ||||
| -rwxr-xr-x | TS/benchmark-scripts/ccmcc25.sh | 30 | ||||
| -rw-r--r-- | TS/dpu/task.c | 138 | ||||
| -rw-r--r-- | TS/host/app.c | 339 | ||||
| -rw-r--r--[-rwxr-xr-x] | TS/include/common.h (renamed from TS/support/common.h) | 38 | ||||
| -rw-r--r-- | TS/include/dfatool_host.ah | 31 | ||||
| -rw-r--r-- | TS/include/params.h | 67 | ||||
| -rw-r--r-- | TS/include/timer.h | 5 | ||||
| -rwxr-xr-x | TS/run-paper-strong-full.sh | 29 | ||||
| -rwxr-xr-x | TS/run-paper-strong-rank.sh | 28 | ||||
| -rwxr-xr-x | TS/run-paper-weak.sh | 29 | ||||
| -rw-r--r-- | TS/support/params.h | 58 | ||||
| -rwxr-xr-x | TS/support/timer.h | 66 |
15 files changed, 491 insertions, 425 deletions
diff --git a/TS/Makefile b/TS/Makefile index ac081bd..2fce611 100644 --- a/TS/Makefile +++ b/TS/Makefile @@ -5,14 +5,31 @@ WITH_ALLOC_OVERHEAD ?= 0 WITH_LOAD_OVERHEAD ?= 0 WITH_FREE_OVERHEAD ?= 0 -COMMON_INCLUDES := support HOST_SOURCES := $(wildcard host/*.c) DPU_SOURCES := $(wildcard dpu/*.c) -COMMON_FLAGS := -Wall -Wextra -g -I${COMMON_INCLUDES} -DBL=${BL} -HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD} -DWITH_LOAD_OVERHEAD=${WITH_LOAD_OVERHEAD} -DWITH_FREE_OVERHEAD=${WITH_FREE_OVERHEAD} -lm +aspectc ?= 0 +aspectc_timing ?= 0 +dfatool_timing ?= 1 + +HOST_CC := ${CC} + +COMMON_FLAGS := -Wall -Wextra -g -Iinclude -DBL=${BL} +HOST_FLAGS := ${COMMON_FLAGS} -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD} -DWITH_LOAD_OVERHEAD=${WITH_LOAD_OVERHEAD} -DWITH_FREE_OVERHEAD=${WITH_FREE_OVERHEAD} -DDFATOOL_TIMING=${dfatool_timing} -DASPECTC=${aspectc} -lm DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} +ifeq (${aspectc_timing}, 1) + ASPECTC_HOST_FLAGS += -ainclude/dfatool_host_dpu.ah -ainclude/dfatool_host.ah +endif + +ASPECTC_HOST_FLAGS ?= -a0 + +ifeq (${aspectc}, 1) + HOST_CC = ag++ -r repo.acp -v 0 ${ASPECTC_HOST_FLAGS} --c_compiler ${UPMEM_HOME}/bin/clang++ -p . --Xcompiler +else + HOST_FLAGS += -std=c11 +endif + QUIET = @ ifdef verbose @@ -25,7 +42,9 @@ bin: ${QUIET}mkdir -p bin bin/ts_host: ${HOST_SOURCES} ${COMMON_INCLUDES} bin - ${QUIET}${CC} -o $@ ${HOST_SOURCES} ${HOST_FLAGS} + ${QUIET}cp ../include/dfatool_host_dpu.ah include + ${QUIET}${HOST_CC} -o $@ ${HOST_SOURCES} ${HOST_FLAGS} + ${QUIET}rm -f include/dfatool_host_dpu.ah bin/ts_dpu: ${DPU_SOURCES} ${COMMON_INCLUDES} bin ${QUIET}dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES} diff --git a/TS/baselines/cpu/mprofile.h b/TS/baselines/cpu/mprofile.h index 120c225..bfaf052 100644 --- a/TS/baselines/cpu/mprofile.h +++ b/TS/baselines/cpu/mprofile.h @@ -10,5 +10,7 @@ //#define HBM_ALOC //#define RANDOM_DIAGS -int loadTimeSeriesFromFile (std::string infilename, std::vector<DTYPE> &A, int &timeSeriesLength); -int saveProfileToFile(std::string outfilename, DTYPE * profile, int * profileIndex, int timeSeriesLength, int windowSize); +int loadTimeSeriesFromFile(std::string infilename, std::vector < DTYPE > &A, + int &timeSeriesLength); +int saveProfileToFile(std::string outfilename, DTYPE * profile, + int *profileIndex, int timeSeriesLength, int windowSize); diff --git a/TS/benchmark-scripts/ccmcc25-sim.sh b/TS/benchmark-scripts/ccmcc25-sim.sh new file mode 100755 index 0000000..0df03d9 --- /dev/null +++ b/TS/benchmark-scripts/ccmcc25-sim.sh @@ -0,0 +1,25 @@ +#!/bin/bash + +mkdir -p log/$(hostname) + +run_benchmark_nmc() { + local "$@" + set -e + make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=8 \ + aspectc=1 aspectc_timing=1 dfatool_timing=0 + bin/ts_host -w 0 -e 5 -n ${ts_size} 2>&1 +} + +export -f run_benchmark_nmc + +fn=log/$(hostname)/ccmcc25-sim + +source ~/lib/local/upmem/upmem-2025.1.0-Linux-x86_64/upmem_env.sh simulator + +echo "prim-benchmarks TS $(git describe --all --long) $(git rev-parse HEAD) $(date -R)" >> ${fn}.txt + +parallel -j1 --eta --joblog ${fn}.joblog --resume --header : \ + run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=8 ts_size={ts_size} \ + ::: nr_dpus 1 2 4 8 16 32 48 64 \ + ::: ts_size 2048 4096 8192 16384 32768 \ +>> ${fn}.txt diff --git a/TS/benchmark-scripts/ccmcc25.sh b/TS/benchmark-scripts/ccmcc25.sh new file mode 100755 index 0000000..74c8371 --- /dev/null +++ b/TS/benchmark-scripts/ccmcc25.sh @@ -0,0 +1,30 @@ +#!/bin/bash + +mkdir -p log/$(hostname) + +run_benchmark_nmc() { + local "$@" + set -e + sudo limit_ranks_to_numa_node ${numa_rank} + make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=8 \ + aspectc=1 aspectc_timing=1 dfatool_timing=0 + bin/ts_host -w 0 -e 50 -n ${ts_size} 2>&1 +} + +export -f run_benchmark_nmc + +for sdk in 2023.2.0 2024.1.0 2024.2.0 2025.1.0; do + + fn=log/$(hostname)/ccmcc25-sdk${sdk} + + source /opt/upmem/upmem-${sdk}-Linux-x86_64/upmem_env.sh + + echo "prim-benchmarks TS $(git describe --all --long) $(git rev-parse HEAD) $(date -R)" >> ${fn}.txt + + parallel -j1 --eta --joblog ${fn}.joblog --resume --header : \ + run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=8 numa_rank=any ts_size={ts_size} \ + ::: nr_dpus 64 128 256 512 768 1024 1536 2048 2304 \ + ::: ts_size 8388608 16777216 33554432 67108864 \ + >> ${fn}.txt + +done diff --git a/TS/dpu/task.c b/TS/dpu/task.c index d704160..5a756aa 100644 --- a/TS/dpu/task.c +++ b/TS/dpu/task.c @@ -18,18 +18,18 @@ __host dpu_arguments_t DPU_INPUT_ARGUMENTS; __host dpu_result_t DPU_RESULTS[NR_TASKLETS]; // Dot product -static void dot_product(DTYPE *vectorA, DTYPE *vectorA_aux, DTYPE *vectorB, DTYPE * result) { - - for(uint32_t i = 0; i < BLOCK_SIZE / sizeof(DTYPE); i++) - { - for(uint32_t j = 0; j < DOTPIP; j++) - { - if((j + i) > BLOCK_SIZE / sizeof(DTYPE) - 1) - { - result[j] += vectorA_aux[(j + i) - BLOCK_SIZE / sizeof(DTYPE)] * vectorB[i]; - } - else - { +static void dot_product(DTYPE *vectorA, DTYPE *vectorA_aux, DTYPE *vectorB, + DTYPE *result) +{ + + for (uint32_t i = 0; i < BLOCK_SIZE / sizeof(DTYPE); i++) { + for (uint32_t j = 0; j < DOTPIP; j++) { + if ((j + i) > BLOCK_SIZE / sizeof(DTYPE) - 1) { + result[j] += + vectorA_aux[(j + i) - + BLOCK_SIZE / sizeof(DTYPE)] * + vectorB[i]; + } else { result[j] += vectorA[j + i] * vectorB[i]; } } @@ -40,43 +40,46 @@ BARRIER_INIT(my_barrier, NR_TASKLETS); extern int main_kernel1(void); -int(*kernels[nr_kernels])(void) = {main_kernel1}; +int (*kernels[nr_kernels])(void) = { main_kernel1 }; -int main(void){ +int main(void) +{ // Kernel - return kernels[DPU_INPUT_ARGUMENTS.kernel](); + return kernels[DPU_INPUT_ARGUMENTS.kernel] (); } // main_kernel1 -int main_kernel1() { +int main_kernel1() +{ unsigned int tasklet_id = me(); #if PRINT printf("tasklet_id = %u\n", tasklet_id); #endif - if(tasklet_id == 0){ - mem_reset(); // Reset the heap + if (tasklet_id == 0) { + mem_reset(); // Reset the heap } // Barrier barrier_wait(&my_barrier); // Input arguments - uint32_t query_length = DPU_INPUT_ARGUMENTS.query_length; - DTYPE query_mean = DPU_INPUT_ARGUMENTS.query_mean; - DTYPE query_std = DPU_INPUT_ARGUMENTS.query_std; + uint32_t query_length = DPU_INPUT_ARGUMENTS.query_length; + DTYPE query_mean = DPU_INPUT_ARGUMENTS.query_mean; + DTYPE query_std = DPU_INPUT_ARGUMENTS.query_std; uint32_t slice_per_dpu = DPU_INPUT_ARGUMENTS.slice_per_dpu; // Boundaries for current tasklet - uint32_t myStartElem = tasklet_id * (slice_per_dpu / (NR_TASKLETS)); - uint32_t myEndElem = myStartElem + (slice_per_dpu / (NR_TASKLETS)) - 1; + uint32_t myStartElem = tasklet_id * (slice_per_dpu / (NR_TASKLETS)); + uint32_t myEndElem = myStartElem + (slice_per_dpu / (NR_TASKLETS)) - 1; // Check time series limit - if(myEndElem > slice_per_dpu - query_length) myEndElem = slice_per_dpu - query_length; + if (myEndElem > slice_per_dpu - query_length) + myEndElem = slice_per_dpu - query_length; // Starting address of the current processing block in MRAM uint32_t mem_offset = (uint32_t) DPU_MRAM_HEAP_POINTER; // Starting address of the query subsequence - uint32_t current_mram_block_addr_query = (uint32_t)(mem_offset); + uint32_t current_mram_block_addr_query = (uint32_t) (mem_offset); mem_offset += query_length * sizeof(DTYPE); // Starting address of the time series slice @@ -86,18 +89,18 @@ int main_kernel1() { // Starting address of the time series means mem_offset += (slice_per_dpu + query_length) * sizeof(DTYPE); - uint32_t current_mram_block_addr_TSMean = (uint32_t)(mem_offset); + uint32_t current_mram_block_addr_TSMean = (uint32_t) (mem_offset); // Starting address of the time series standard deviations mem_offset += (slice_per_dpu + query_length) * sizeof(DTYPE); - uint32_t current_mram_block_addr_TSSigma = (uint32_t)(mem_offset); + uint32_t current_mram_block_addr_TSSigma = (uint32_t) (mem_offset); // Initialize local caches to store the MRAM blocks - DTYPE *cache_TS = (DTYPE *) mem_alloc(BLOCK_SIZE); - DTYPE *cache_TS_aux = (DTYPE *) mem_alloc(BLOCK_SIZE); - DTYPE *cache_query = (DTYPE *) mem_alloc(BLOCK_SIZE); - DTYPE *cache_TSMean = (DTYPE *) mem_alloc(BLOCK_SIZE); - DTYPE *cache_TSSigma = (DTYPE *) mem_alloc(BLOCK_SIZE); + DTYPE *cache_TS = (DTYPE *) mem_alloc(BLOCK_SIZE); + DTYPE *cache_TS_aux = (DTYPE *) mem_alloc(BLOCK_SIZE); + DTYPE *cache_query = (DTYPE *) mem_alloc(BLOCK_SIZE); + DTYPE *cache_TSMean = (DTYPE *) mem_alloc(BLOCK_SIZE); + DTYPE *cache_TSSigma = (DTYPE *) mem_alloc(BLOCK_SIZE); DTYPE *cache_dotprods = (DTYPE *) mem_alloc(BLOCK_SIZE); // Create result structure pointer @@ -108,41 +111,56 @@ int main_kernel1() { DTYPE min_distance = DTYPE_MAX; uint32_t min_index = 0; - - for(uint32_t i = myStartElem; i < myEndElem; i+= (BLOCK_SIZE / sizeof(DTYPE))) - { - for(uint32_t d = 0; d < DOTPIP; d++) + for (uint32_t i = myStartElem; i < myEndElem; + i += (BLOCK_SIZE / sizeof(DTYPE))) { + for (uint32_t d = 0; d < DOTPIP; d++) cache_dotprods[d] = 0; - current_mram_block_addr_TS = (uint32_t) starting_offset_ts + (i - myStartElem) * sizeof(DTYPE); - current_mram_block_addr_query = (uint32_t) DPU_MRAM_HEAP_POINTER; - - for(uint32_t j = 0; j < (query_length) / (BLOCK_SIZE / sizeof(DTYPE)); j++) - { - mram_read((__mram_ptr void const *) current_mram_block_addr_TS, cache_TS, BLOCK_SIZE); - mram_read((__mram_ptr void const *) current_mram_block_addr_TS + BLOCK_SIZE, cache_TS_aux, BLOCK_SIZE); - mram_read((__mram_ptr void const *) current_mram_block_addr_query, cache_query, BLOCK_SIZE); - - current_mram_block_addr_TS += BLOCK_SIZE; + current_mram_block_addr_TS = + (uint32_t) starting_offset_ts + (i - + myStartElem) * + sizeof(DTYPE); + current_mram_block_addr_query = + (uint32_t) DPU_MRAM_HEAP_POINTER; + + for (uint32_t j = 0; + j < (query_length) / (BLOCK_SIZE / sizeof(DTYPE)); j++) { + mram_read((__mram_ptr void const *) + current_mram_block_addr_TS, cache_TS, + BLOCK_SIZE); + mram_read((__mram_ptr void const *) + current_mram_block_addr_TS + BLOCK_SIZE, + cache_TS_aux, BLOCK_SIZE); + mram_read((__mram_ptr void const *) + current_mram_block_addr_query, cache_query, + BLOCK_SIZE); + + current_mram_block_addr_TS += BLOCK_SIZE; current_mram_block_addr_query += BLOCK_SIZE; - dot_product(cache_TS, cache_TS_aux, cache_query, cache_dotprods); + dot_product(cache_TS, cache_TS_aux, cache_query, + cache_dotprods); } - - mram_read((__mram_ptr void const *) current_mram_block_addr_TSMean, cache_TSMean, BLOCK_SIZE); - mram_read((__mram_ptr void const *) current_mram_block_addr_TSSigma, cache_TSSigma, BLOCK_SIZE); - current_mram_block_addr_TSMean += BLOCK_SIZE; + mram_read((__mram_ptr void const *) + current_mram_block_addr_TSMean, cache_TSMean, + BLOCK_SIZE); + mram_read((__mram_ptr void const *) + current_mram_block_addr_TSSigma, cache_TSSigma, + BLOCK_SIZE); + current_mram_block_addr_TSMean += BLOCK_SIZE; current_mram_block_addr_TSSigma += BLOCK_SIZE; - for (uint32_t k = 0; k < (BLOCK_SIZE / sizeof(DTYPE)); k++) - { - distance = 2 * ((DTYPE) query_length - (cache_dotprods[k] - (DTYPE) query_length * cache_TSMean[k] - * query_mean) / (cache_TSSigma[k] * query_std)); - - if(distance < min_distance) - { - min_distance = distance; - min_index = i + k; + for (uint32_t k = 0; k < (BLOCK_SIZE / sizeof(DTYPE)); k++) { + distance = + 2 * ((DTYPE) query_length - + (cache_dotprods[k] - + (DTYPE) query_length * cache_TSMean[k] + * query_mean) / (cache_TSSigma[k] * + query_std)); + + if (distance < min_distance) { + min_distance = distance; + min_index = i + k; } } } diff --git a/TS/host/app.c b/TS/host/app.c index b9faa9c..bfa14df 100644 --- a/TS/host/app.c +++ b/TS/host/app.c @@ -7,8 +7,18 @@ #include <stdlib.h> #include <stdbool.h> #include <string.h> + +#if ASPECTC +extern "C" { +#endif + #include <dpu.h> #include <dpu_log.h> + +#if ASPECTC +} +#endif + #include <unistd.h> #include <getopt.h> #include <assert.h> @@ -31,23 +41,23 @@ #define MAX_DATA_VAL 127 static DTYPE tSeries[1 << 26]; -static DTYPE query [1 << 15]; -static DTYPE AMean [1 << 26]; -static DTYPE ASigma [1 << 26]; +static DTYPE query[1 << 15]; +static DTYPE AMean[1 << 26]; +static DTYPE ASigma[1 << 26]; static DTYPE minHost; static DTYPE minHostIdx; // Create input arrays -static DTYPE *create_test_file(unsigned int ts_elements, unsigned int query_elements) { +static DTYPE *create_test_file(unsigned int ts_elements, + unsigned int query_elements) +{ srand(0); - for (uint64_t i = 0; i < ts_elements; i++) - { + for (uint64_t i = 0; i < ts_elements; i++) { tSeries[i] = i % MAX_DATA_VAL; } - for (uint64_t i = 0; i < query_elements; i++) - { + for (uint64_t i = 0; i < query_elements; i++) { query[i] = i % MAX_DATA_VAL; } @@ -55,61 +65,62 @@ static DTYPE *create_test_file(unsigned int ts_elements, unsigned int query_elem } // Compute output in the host -static void streamp(DTYPE* tSeries, DTYPE* AMean, DTYPE* ASigma, int ProfileLength, - DTYPE* query, int queryLength, DTYPE queryMean, DTYPE queryStdDeviation) +static void streamp(DTYPE *tSeries, DTYPE *AMean, DTYPE *ASigma, + int ProfileLength, DTYPE *query, int queryLength, + DTYPE queryMean, DTYPE queryStdDeviation) { DTYPE distance; DTYPE dotprod; - minHost = INT32_MAX; + minHost = INT32_MAX; minHostIdx = 0; - for (int subseq = 0; subseq < ProfileLength; subseq++) - { + for (int subseq = 0; subseq < ProfileLength; subseq++) { dotprod = 0; - for(int j = 0; j < queryLength; j++) - { + for (int j = 0; j < queryLength; j++) { dotprod += tSeries[j + subseq] * query[j]; } - distance = 2 * (queryLength - (dotprod - queryLength * AMean[subseq] - * queryMean) / (ASigma[subseq] * queryStdDeviation)); + distance = + 2 * (queryLength - (dotprod - queryLength * AMean[subseq] + * queryMean) / (ASigma[subseq] * + queryStdDeviation)); - if(distance < minHost) - { + if (distance < minHost) { minHost = distance; minHostIdx = subseq; } } } -static void compute_ts_statistics(unsigned int timeSeriesLength, unsigned int ProfileLength, unsigned int queryLength) +static void compute_ts_statistics(unsigned int timeSeriesLength, + unsigned int ProfileLength, + unsigned int queryLength) { - double* ACumSum = malloc(sizeof(double) * timeSeriesLength); + double *ACumSum = (double*)malloc(sizeof(double) * timeSeriesLength); ACumSum[0] = tSeries[0]; for (uint64_t i = 1; i < timeSeriesLength; i++) ACumSum[i] = tSeries[i] + ACumSum[i - 1]; - double* ASqCumSum = malloc(sizeof(double) * timeSeriesLength); + double *ASqCumSum = (double*)malloc(sizeof(double) * timeSeriesLength); ASqCumSum[0] = tSeries[0] * tSeries[0]; for (uint64_t i = 1; i < timeSeriesLength; i++) ASqCumSum[i] = tSeries[i] * tSeries[i] + ASqCumSum[i - 1]; - double* ASum = malloc(sizeof(double) * ProfileLength); + double *ASum = (double*)malloc(sizeof(double) * ProfileLength); ASum[0] = ACumSum[queryLength - 1]; for (uint64_t i = 0; i < timeSeriesLength - queryLength; i++) ASum[i + 1] = ACumSum[queryLength + i] - ACumSum[i]; - double* ASumSq = malloc(sizeof(double) * ProfileLength); + double *ASumSq = (double*)malloc(sizeof(double) * ProfileLength); ASumSq[0] = ASqCumSum[queryLength - 1]; for (uint64_t i = 0; i < timeSeriesLength - queryLength; i++) ASumSq[i + 1] = ASqCumSum[queryLength + i] - ASqCumSum[i]; - double * AMean_tmp = malloc(sizeof(double) * ProfileLength); + double *AMean_tmp = (double*)malloc(sizeof(double) * ProfileLength); for (uint64_t i = 0; i < ProfileLength; i++) AMean_tmp[i] = ASum[i] / queryLength; - double* ASigmaSq = malloc(sizeof(double) * ProfileLength); + double *ASigmaSq = (double*)malloc(sizeof(double) * ProfileLength); for (uint64_t i = 0; i < ProfileLength; i++) ASigmaSq[i] = ASumSq[i] / queryLength - AMean[i] * AMean[i]; - for (uint64_t i = 0; i < ProfileLength; i++) - { + for (uint64_t i = 0; i < ProfileLength; i++) { ASigma[i] = sqrt(ASigmaSq[i]); - AMean[i] = (DTYPE) AMean_tmp[i]; + AMean[i] = (DTYPE) AMean_tmp[i]; } free(ACumSum); @@ -121,7 +132,8 @@ static void compute_ts_statistics(unsigned int timeSeriesLength, unsigned int Pr } // Main of the Host Application -int main(int argc, char **argv) { +int main(int argc, char **argv) +{ // Timer declaration Timer timer; @@ -129,22 +141,28 @@ int main(int argc, char **argv) { struct Params p = input_params(argc, argv); struct dpu_set_t dpu_set, dpu; uint32_t nr_of_dpus; - uint32_t nr_of_ranks; + uint32_t nr_of_ranks; - // Allocate DPUs and load binary + // Allocate DPUs and load binary #if !WITH_ALLOC_OVERHEAD - DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set)); - timer.time[0] = 0; // alloc + DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set)); +#if DFATOOL_TIMING + timer.time[0] = 0; // alloc +#endif #endif #if !WITH_LOAD_OVERHEAD - DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL)); - DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus)); - DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks)); - assert(nr_of_dpus == NR_DPUS); - timer.time[1] = 0; // load + DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL)); + DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus)); + DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks)); + assert(nr_of_dpus == NR_DPUS); +#if DFATOOL_TIMING + timer.time[1] = 0; // load +#endif #endif #if !WITH_FREE_OVERHEAD - timer.time[6] = 0; // free +#if DFATOOL_TIMING + timer.time[6] = 0; // free +#endif #endif #if ENERGY @@ -152,12 +170,15 @@ int main(int argc, char **argv) { DPU_ASSERT(dpu_probe_init("energy_probe", &probe)); #endif - unsigned long int ts_size = p.input_size_n; + unsigned long int ts_size = p.input_size_n; const unsigned int query_length = p.input_size_m; // Size adjustment - if(ts_size % (NR_DPUS * NR_TASKLETS*query_length)) - ts_size = ts_size + (NR_DPUS * NR_TASKLETS * query_length - ts_size % (NR_DPUS * NR_TASKLETS*query_length)); + if (ts_size % (NR_DPUS * NR_TASKLETS * query_length)) + ts_size = + ts_size + (NR_DPUS * NR_TASKLETS * query_length - + ts_size % (NR_DPUS * NR_TASKLETS * + query_length)); // Create an input file with arbitrary data create_test_file(ts_size, query_length); @@ -165,30 +186,34 @@ int main(int argc, char **argv) { DTYPE query_mean; double queryMean = 0; - for(unsigned i = 0; i < query_length; i++) queryMean += query[i]; - queryMean /= (double) query_length; + for (unsigned i = 0; i < query_length; i++) + queryMean += query[i]; + queryMean /= (double)query_length; query_mean = (DTYPE) queryMean; DTYPE query_std; double queryStdDeviation; double queryVariance = 0; - for(unsigned i = 0; i < query_length; i++) - { - queryVariance += (query[i] - queryMean) * (query[i] - queryMean); + for (unsigned i = 0; i < query_length; i++) { + queryVariance += + (query[i] - queryMean) * (query[i] - queryMean); } - queryVariance /= (double) query_length; + queryVariance /= (double)query_length; queryStdDeviation = sqrt(queryVariance); query_std = (DTYPE) queryStdDeviation; - DTYPE *bufferTS = tSeries; - DTYPE *bufferQ = query; - DTYPE *bufferAMean = AMean; + DTYPE *bufferTS = tSeries; + DTYPE *bufferQ = query; + DTYPE *bufferAMean = AMean; DTYPE *bufferASigma = ASigma; uint32_t slice_per_dpu = ts_size / NR_DPUS; unsigned int kernel = 0; - dpu_arguments_t input_arguments = {ts_size, query_length, query_mean, query_std, slice_per_dpu, 0, kernel}; + dpu_arguments_t input_arguments = + { (uint32_t)ts_size, query_length, query_mean, query_std, slice_per_dpu, 0, + (enum kernels) kernel + }; uint32_t mem_offset; dpu_result_t result; @@ -199,21 +224,30 @@ int main(int argc, char **argv) { for (int rep = 0; rep < p.n_warmup + p.n_reps; rep++) { + if (rep >= p.n_warmup) { + start(&timer, 6, 0); + } + streamp(tSeries, AMean, ASigma, ts_size - query_length - 1, + query, query_length, query_mean, query_std); + if (rep >= p.n_warmup) { + stop(&timer, 6); + } + #if WITH_ALLOC_OVERHEAD - if(rep >= p.n_warmup) { + if (rep >= p.n_warmup) { start(&timer, 0, 0); } DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set)); - if(rep >= p.n_warmup) { + if (rep >= p.n_warmup) { stop(&timer, 0); } #endif #if WITH_LOAD_OVERHEAD - if(rep >= p.n_warmup) { + if (rep >= p.n_warmup) { start(&timer, 1, 0); } DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL)); - if(rep >= p.n_warmup) { + if (rep >= p.n_warmup) { stop(&timer, 1); } DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus)); @@ -225,58 +259,70 @@ int main(int argc, char **argv) { start(&timer, 2, 0); } uint32_t i = 0; - - DPU_FOREACH(dpu_set, dpu) { - input_arguments.exclusion_zone = 0; - - DPU_ASSERT(dpu_copy_to(dpu, "DPU_INPUT_ARGUMENTS", 0, (const void *) &input_arguments, sizeof(input_arguments))); - i++; + DPU_FOREACH(dpu_set, dpu, i) { + DPU_ASSERT(dpu_prepare_xfer(dpu, &input_arguments)); } + DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, "DPU_INPUT_ARGUMENTS", 0, sizeof(input_arguments), DPU_XFER_DEFAULT)); i = 0; mem_offset = 0; - DPU_FOREACH(dpu_set, dpu, i) - { + DPU_FOREACH(dpu_set, dpu, i) { DPU_ASSERT(dpu_prepare_xfer(dpu, bufferQ)); } - DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, DPU_MRAM_HEAP_POINTER_NAME, 0, query_length * sizeof(DTYPE), DPU_XFER_DEFAULT)); + DPU_ASSERT(dpu_push_xfer + (dpu_set, DPU_XFER_TO_DPU, + DPU_MRAM_HEAP_POINTER_NAME, 0, + query_length * sizeof(DTYPE), DPU_XFER_DEFAULT)); i = 0; mem_offset += query_length * sizeof(DTYPE); DPU_FOREACH(dpu_set, dpu, i) { - DPU_ASSERT(dpu_prepare_xfer(dpu, bufferTS + slice_per_dpu * i)); + DPU_ASSERT(dpu_prepare_xfer + (dpu, bufferTS + slice_per_dpu * i)); } - DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, DPU_MRAM_HEAP_POINTER_NAME, mem_offset,(slice_per_dpu + query_length)*sizeof(DTYPE), DPU_XFER_DEFAULT)); + DPU_ASSERT(dpu_push_xfer + (dpu_set, DPU_XFER_TO_DPU, + DPU_MRAM_HEAP_POINTER_NAME, mem_offset, + (slice_per_dpu + query_length) * sizeof(DTYPE), + DPU_XFER_DEFAULT)); mem_offset += ((slice_per_dpu + query_length) * sizeof(DTYPE)); i = 0; DPU_FOREACH(dpu_set, dpu, i) { - DPU_ASSERT(dpu_prepare_xfer(dpu, bufferAMean + slice_per_dpu * i)); + DPU_ASSERT(dpu_prepare_xfer + (dpu, bufferAMean + slice_per_dpu * i)); } - DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, DPU_MRAM_HEAP_POINTER_NAME, mem_offset, (slice_per_dpu + query_length)*sizeof(DTYPE), DPU_XFER_DEFAULT)); + DPU_ASSERT(dpu_push_xfer + (dpu_set, DPU_XFER_TO_DPU, + DPU_MRAM_HEAP_POINTER_NAME, mem_offset, + (slice_per_dpu + query_length) * sizeof(DTYPE), + DPU_XFER_DEFAULT)); i = 0; mem_offset += ((slice_per_dpu + query_length) * sizeof(DTYPE)); DPU_FOREACH(dpu_set, dpu, i) { - DPU_ASSERT(dpu_prepare_xfer(dpu, bufferASigma + slice_per_dpu * i)); + DPU_ASSERT(dpu_prepare_xfer + (dpu, bufferASigma + slice_per_dpu * i)); } - DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, DPU_MRAM_HEAP_POINTER_NAME, mem_offset, (slice_per_dpu + query_length)*sizeof(DTYPE), DPU_XFER_DEFAULT)); + DPU_ASSERT(dpu_push_xfer + (dpu_set, DPU_XFER_TO_DPU, + DPU_MRAM_HEAP_POINTER_NAME, mem_offset, + (slice_per_dpu + query_length) * sizeof(DTYPE), + DPU_XFER_DEFAULT)); if (rep >= p.n_warmup) { stop(&timer, 2); } - // Run kernel on DPUs - if (rep >= p.n_warmup) - { + if (rep >= p.n_warmup) { start(&timer, 3, 0); #if ENERGY DPU_ASSERT(dpu_probe_start(&probe)); @@ -285,37 +331,49 @@ int main(int argc, char **argv) { DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS)); - if (rep >= p.n_warmup) - { + if (rep >= p.n_warmup) { stop(&timer, 3); #if ENERGY DPU_ASSERT(dpu_probe_stop(&probe)); #endif } - dpu_result_t* results_retrieve[NR_DPUS]; + dpu_result_t *results_retrieve[NR_DPUS]; if (rep >= p.n_warmup) { start(&timer, 4, 0); } DPU_FOREACH(dpu_set, dpu, i) { - results_retrieve[i] = (dpu_result_t*)malloc(NR_TASKLETS * sizeof(dpu_result_t)); + results_retrieve[i] = + (dpu_result_t *) malloc(NR_TASKLETS * + sizeof(dpu_result_t)); } - DPU_FOREACH(dpu_set, dpu, i) { DPU_ASSERT(dpu_prepare_xfer(dpu, results_retrieve[i])); } - DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_FROM_DPU, "DPU_RESULTS", 0, NR_TASKLETS * sizeof(dpu_result_t), DPU_XFER_DEFAULT)); + DPU_ASSERT(dpu_push_xfer + (dpu_set, DPU_XFER_FROM_DPU, "DPU_RESULTS", 0, + NR_TASKLETS * sizeof(dpu_result_t), + DPU_XFER_DEFAULT)); i = 0; DPU_FOREACH(dpu_set, dpu, i) { - for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) { - if(results_retrieve[i][each_tasklet].minValue < result.minValue && results_retrieve[i][each_tasklet].minValue > 0) - { - result.minValue = results_retrieve[i][each_tasklet].minValue; - result.minIndex = (DTYPE)results_retrieve[i][each_tasklet].minIndex + (i * slice_per_dpu); + for (unsigned int each_tasklet = 0; + each_tasklet < NR_TASKLETS; each_tasklet++) { + if (results_retrieve[i][each_tasklet].minValue < + result.minValue + && + results_retrieve[i][each_tasklet].minValue > + 0) { + result.minValue = + results_retrieve[i] + [each_tasklet].minValue; + result.minIndex = (DTYPE) + results_retrieve[i] + [each_tasklet].minIndex + + (i * slice_per_dpu); } } @@ -323,11 +381,9 @@ int main(int argc, char **argv) { i++; } - if(rep >= p.n_warmup) { + if (rep >= p.n_warmup) { stop(&timer, 4); } - - #if PRINT printf("LOGS\n"); DPU_FOREACH(dpu_set, dpu) { @@ -337,67 +393,89 @@ int main(int argc, char **argv) { #if WITH_ALLOC_OVERHEAD #if WITH_FREE_OVERHEAD - if(rep >= p.n_warmup) { + if (rep >= p.n_warmup) { start(&timer, 5, 0); } #endif DPU_ASSERT(dpu_free(dpu_set)); #if WITH_FREE_OVERHEAD - if(rep >= p.n_warmup) { + if (rep >= p.n_warmup) { stop(&timer, 5); } #endif #endif - if (rep >= p.n_warmup) { - start(&timer, 6, 0); - } - streamp(tSeries, AMean, ASigma, ts_size - query_length - 1, query, query_length, query_mean, query_std); - if(rep >= p.n_warmup) { - stop(&timer, 6); - } - int status = (minHost == result.minValue); if (status) { - printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] results are equal\n"); + printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET + "] results are equal\n"); if (rep >= p.n_warmup) { - printf("[::] TS UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%lu", - NR_DPUS, nr_of_ranks, NR_TASKLETS, XSTR(DTYPE), BLOCK_SIZE, ts_size); - printf(" b_with_alloc_overhead=%d b_with_load_overhead=%d b_with_free_overhead=%d ", - WITH_ALLOC_OVERHEAD, WITH_LOAD_OVERHEAD, WITH_FREE_OVERHEAD); - printf("| latency_alloc_us=%f latency_load_us=%f latency_write_us=%f latency_kernel_us=%f latency_read_us=%f latency_free_us=%f latency_cpu_us=%f ", - timer.time[0], // alloc - timer.time[1], // load - timer.time[2], // write - timer.time[3], // kernel - timer.time[4], // read - timer.time[5], // free - timer.time[6]); // CPU - printf(" throughput_cpu_MBps=%f throughput_upmem_kernel_MBps=%f throughput_upmem_total_MBps=%f", - ts_size * sizeof(DTYPE) / timer.time[6], - ts_size * sizeof(DTYPE) / (timer.time[3]), - ts_size * sizeof(DTYPE) / (timer.time[0] + timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5])); - printf(" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f", - ts_size * sizeof(DTYPE) / (timer.time[2] + timer.time[3] + timer.time[4]), - ts_size * sizeof(DTYPE) / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4]), - ts_size * sizeof(DTYPE) / (timer.time[0] + timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4])); - printf(" throughput_cpu_MOpps=%f throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f", - ts_size / timer.time[6], - ts_size / (timer.time[3]), - ts_size / (timer.time[0] + timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4] + timer.time[5])); - printf(" throughput_upmem_wxr_MOpps=%f throughput_upmem_lwxr_MOpps=%f throughput_upmem_alwxr_MOpps=%f\n", - ts_size / (timer.time[2] + timer.time[3] + timer.time[4]), - ts_size / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4]), - ts_size / (timer.time[0] + timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4])); + dfatool_printf + ("[::] TS UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%lu", + NR_DPUS, nr_of_ranks, NR_TASKLETS, + XSTR(DTYPE), BLOCK_SIZE, ts_size); + dfatool_printf + (" b_with_alloc_overhead=%d b_with_load_overhead=%d b_with_free_overhead=%d ", + WITH_ALLOC_OVERHEAD, WITH_LOAD_OVERHEAD, + WITH_FREE_OVERHEAD); + dfatool_printf("| latency_alloc_us=%f latency_load_us=%f latency_write_us=%f latency_kernel_us=%f latency_read_us=%f latency_free_us=%f latency_cpu_us=%f ", timer.time[0], // alloc + timer.time[1], // load + timer.time[2], // write + timer.time[3], // kernel + timer.time[4], // read + timer.time[5], // free + timer.time[6]); // CPU + dfatool_printf + (" throughput_cpu_MBps=%f throughput_upmem_kernel_MBps=%f throughput_upmem_total_MBps=%f", + ts_size * sizeof(DTYPE) / timer.time[6], + ts_size * sizeof(DTYPE) / (timer.time[3]), + ts_size * sizeof(DTYPE) / (timer.time[0] + + timer.time[1] + + timer.time[2] + + timer.time[3] + + timer.time[4] + + timer.time[5])); + dfatool_printf + (" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f", + ts_size * sizeof(DTYPE) / (timer.time[2] + + timer.time[3] + + timer.time[4]), + ts_size * sizeof(DTYPE) / (timer.time[1] + + timer.time[2] + + timer.time[3] + + timer.time[4]), + ts_size * sizeof(DTYPE) / (timer.time[0] + + timer.time[1] + + timer.time[2] + + timer.time[3] + + timer.time[4])); + dfatool_printf + (" throughput_cpu_MOpps=%f throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f", + ts_size / timer.time[6], + ts_size / (timer.time[3]), + ts_size / (timer.time[0] + timer.time[1] + + timer.time[2] + timer.time[3] + + timer.time[4] + timer.time[5])); + dfatool_printf + (" throughput_upmem_wxr_MOpps=%f throughput_upmem_lwxr_MOpps=%f throughput_upmem_alwxr_MOpps=%f\n", + ts_size / (timer.time[2] + timer.time[3] + + timer.time[4]), + ts_size / (timer.time[1] + timer.time[2] + + timer.time[3] + timer.time[4]), + ts_size / (timer.time[0] + timer.time[1] + + timer.time[2] + timer.time[3] + + timer.time[4])); } } else { - printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] results differ!\n"); + printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET + "] results differ!\n"); } } #if ENERGY double acc_energy, avg_energy, acc_time, avg_time; - DPU_ASSERT(dpu_probe_get(&probe, DPU_ENERGY, DPU_ACCUMULATE, &acc_energy)); + DPU_ASSERT(dpu_probe_get + (&probe, DPU_ENERGY, DPU_ACCUMULATE, &acc_energy)); DPU_ASSERT(dpu_probe_get(&probe, DPU_ENERGY, DPU_AVERAGE, &avg_energy)); DPU_ASSERT(dpu_probe_get(&probe, DPU_TIME, DPU_ACCUMULATE, &acc_time)); DPU_ASSERT(dpu_probe_get(&probe, DPU_TIME, DPU_AVERAGE, &avg_time)); @@ -407,7 +485,6 @@ int main(int argc, char **argv) { printf("Energy (J): %f J\t", avg_energy); #endif - #if !WITH_ALLOC_OVERHEAD DPU_ASSERT(dpu_free(dpu_set)); #endif diff --git a/TS/support/common.h b/TS/include/common.h index b120bb1..6d37bdc 100755..100644 --- a/TS/support/common.h +++ b/TS/include/common.h @@ -14,30 +14,32 @@ #define DTYPE int32_t #define DTYPE_MAX INT32_MAX -typedef struct { +enum kernels { + kernel1 = 0, + nr_kernels = 1, +} kernel; + +typedef struct { uint32_t ts_length; - uint32_t query_length; - DTYPE query_mean; - DTYPE query_std; - uint32_t slice_per_dpu; - int32_t exclusion_zone; - enum kernels { - kernel1 = 0, - nr_kernels = 1, - } kernel; -}dpu_arguments_t; + uint32_t query_length; + DTYPE query_mean; + DTYPE query_std; + uint32_t slice_per_dpu; + int32_t exclusion_zone; + enum kernels kernel; +} dpu_arguments_t; -typedef struct { - DTYPE minValue; - uint32_t minIndex; - DTYPE maxValue; - uint32_t maxIndex; -}dpu_result_t; +typedef struct { + DTYPE minValue; + uint32_t minIndex; + DTYPE maxValue; + uint32_t maxIndex; +} dpu_result_t; #ifndef ENERGY #define ENERGY 0 #endif -#define PRINT 0 +#define PRINT 0 #define ANSI_COLOR_RED "\x1b[31m" #define ANSI_COLOR_GREEN "\x1b[32m" diff --git a/TS/include/dfatool_host.ah b/TS/include/dfatool_host.ah new file mode 100644 index 0000000..4192c73 --- /dev/null +++ b/TS/include/dfatool_host.ah @@ -0,0 +1,31 @@ +#pragma once + +#include <sys/time.h> +#include "dfatool_host_dpu.ah" + +aspect DfatoolHostTiming : public DfatoolHostDPUTiming { + + unsigned long ts_size, query_length; + unsigned int element_size; + + virtual int getKernel() { return kernel; } + + DfatoolHostTiming() { + element_size = sizeof(DTYPE); + } + + advice call("% input_params(...)") : after() { + Params* p = tjp->result(); + ts_size = p->input_size_n; + query_length = p->input_size_m; + printf("[>>] TS | n_dpus=%u n_elements_ts=%lu n_elements_query=%lu\n", NR_DPUS, ts_size, query_length); + } + + advice call("% streamp(...)") : before() { + printf("[--] TS | n_dpus=%u n_elements_ts=%lu n_elements_query=%lu\n", NR_DPUS, ts_size, query_length); + } + + advice execution("% main(...)") : after() { + printf("[<<] TS | n_dpus=%u n_elements_ts=%lu n_elements_query=%lu\n", NR_DPUS, ts_size, query_length); + } +}; diff --git a/TS/include/params.h b/TS/include/params.h new file mode 100644 index 0000000..b7d9763 --- /dev/null +++ b/TS/include/params.h @@ -0,0 +1,67 @@ +#ifndef _PARAMS_H_ +#define _PARAMS_H_ + +#include "common.h" + +// Params --------------------------------------------------------------------- +typedef struct Params { + unsigned long input_size_n; + unsigned long input_size_m; + int n_warmup; + int n_reps; +} Params; + +void usage() +{ + fprintf(stderr, + "\nUsage: ./program [options]" + "\n" + "\nGeneral options:" + "\n -h help" + "\n -w <W> # of untimed warmup iterations (default=1)" + "\n -e <E> # of timed repetition iterations (default=3)" + "\n" + "\nBenchmark-specific options:" + "\n -n <n> n (TS length. Default=64K elements)" + "\n -m <m> m (Query length. Default=256 elements)" "\n"); +} + +struct Params input_params(int argc, char **argv) +{ + struct Params p; + p.input_size_n = 1 << 16; + p.input_size_m = 1 << 8; + + p.n_warmup = 1; + p.n_reps = 3; + + int opt; + while ((opt = getopt(argc, argv, "hw:e:n:m:")) >= 0) { + switch (opt) { + case 'h': + usage(); + exit(0); + break; + case 'w': + p.n_warmup = atoi(optarg); + break; + case 'e': + p.n_reps = atoi(optarg); + break; + case 'n': + p.input_size_n = atol(optarg); + break; + case 'm': + p.input_size_m = atol(optarg); + break; + default: + fprintf(stderr, "\nUnrecognized option!\n"); + usage(); + exit(0); + } + } + assert(NR_DPUS > 0 && "Invalid # of dpus!"); + + return p; +} +#endif diff --git a/TS/include/timer.h b/TS/include/timer.h new file mode 100644 index 0000000..7b80823 --- /dev/null +++ b/TS/include/timer.h @@ -0,0 +1,5 @@ +#pragma once + +#define N_TIMERS 7 +#include "../../include/timer_base.h" +#undef N_TIMERS diff --git a/TS/run-paper-strong-full.sh b/TS/run-paper-strong-full.sh deleted file mode 100755 index 5b7656d..0000000 --- a/TS/run-paper-strong-full.sh +++ /dev/null @@ -1,29 +0,0 @@ -#!/bin/bash - -set -e - -# BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module -# T: data type -# -w: number of un-timed warmup iterations -# -e: number of timed iterations -# -i; ignored, always uses 262144 elements - -( - -echo "prim-benchmarks TS strong-full (dfatool edition)" -echo "Started at $(date)" -echo "Revision $(git describe --always)" - -# >2048 is not part of upstream -# 12 tasklets are not part of upstream (code does not work with 16…) -for nr_dpus in 2543 2304 256 512 1024 2048; do - for nr_tasklets in 1 2 4 8 12 16; do - echo - # upstream code did not respect $BL in the makefile and used 256B (BL=8) instead. - # This appears to be faster than BL=10. - if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=8; then - timeout --foreground -k 1m 30m bin/ts_host -w 0 -e 100 -n 33554432 || true - fi - done -done -) | tee log-paper-strong-full.txt diff --git a/TS/run-paper-strong-rank.sh b/TS/run-paper-strong-rank.sh deleted file mode 100755 index 58ad641..0000000 --- a/TS/run-paper-strong-rank.sh +++ /dev/null @@ -1,28 +0,0 @@ -#!/bin/bash - -set -e - -# BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module -# T: data type -# -w: number of un-timed warmup iterations -# -e: number of timed iterations -# -i; ignored, always uses 262144 elements - -( - -echo "prim-benchmarks TS strong-rank (dfatool edition)" -echo "Started at $(date)" -echo "Revision $(git describe --always)" - -# >64 are not part of upstream config space -for nr_dpus in 128 1 4 16 64; do - for nr_tasklets in 1 2 4 8 16; do - echo - # upstream code did not respect $BL in the makefile and used 256B (BL=8) instead. - # BL=10 appears to be slightly faster. - if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then - timeout --foreground -k 1m 60m bin/ts_host -w 0 -e 50 -n 524288 || true - fi - done -done -) | tee log-paper-strong-rank.txt diff --git a/TS/run-paper-weak.sh b/TS/run-paper-weak.sh deleted file mode 100755 index 64892f4..0000000 --- a/TS/run-paper-weak.sh +++ /dev/null @@ -1,29 +0,0 @@ -#!/bin/bash - -set -e - -# BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module -# T: data type -# -w: number of un-timed warmup iterations -# -e: number of timed iterations -# -i; ignored, always uses 262144 elements - -( - -echo "prim-benchmarks TS weak (dfatool edition)" -echo "Started at $(date)" -echo "Revision $(git describe --always)" - -# 256 and 512 are not part of upstream -for nr_dpus in 1 4 16 64; do - for nr_tasklets in 1 2 4 8 16; do - echo - # upstream code did not respect $BL in the makefile and used 256B (BL=8) instead. - # BL=10 appears to be slightly faster. - if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=8; then - i=$(( nr_dpus * 524288 )) - timeout --foreground -k 1m 30m bin/ts_host -w 0 -e 100 -n $i || true - fi - done -done -) | tee log-paper-weak.txt diff --git a/TS/support/params.h b/TS/support/params.h deleted file mode 100644 index 4668604..0000000 --- a/TS/support/params.h +++ /dev/null @@ -1,58 +0,0 @@ -#ifndef _PARAMS_H_ -#define _PARAMS_H_ - -#include "common.h" - -// Params --------------------------------------------------------------------- -typedef struct Params { - unsigned long input_size_n; - unsigned long input_size_m; - int n_warmup; - int n_reps; -}Params; - -void usage() { - fprintf(stderr, - "\nUsage: ./program [options]" - "\n" - "\nGeneral options:" - "\n -h help" - "\n -w <W> # of untimed warmup iterations (default=1)" - "\n -e <E> # of timed repetition iterations (default=3)" - "\n" - "\nBenchmark-specific options:" - "\n -n <n> n (TS length. Default=64K elements)" - "\n -m <m> m (Query length. Default=256 elements)" - "\n"); - } - - struct Params input_params(int argc, char **argv) { - struct Params p; - p.input_size_n = 1 << 16; - p.input_size_m = 1 << 8; - - p.n_warmup = 1; - p.n_reps = 3; - - int opt; - while((opt = getopt(argc, argv, "hw:e:n:m:")) >= 0) { - switch(opt) { - case 'h': - usage(); - exit(0); - break; - case 'w': p.n_warmup = atoi(optarg); break; - case 'e': p.n_reps = atoi(optarg); break; - case 'n': p.input_size_n = atol(optarg); break; - case 'm': p.input_size_m = atol(optarg); break; - default: - fprintf(stderr, "\nUnrecognized option!\n"); - usage(); - exit(0); - } - } - assert(NR_DPUS > 0 && "Invalid # of dpus!"); - - return p; - } -#endif diff --git a/TS/support/timer.h b/TS/support/timer.h deleted file mode 100755 index ff1ae1b..0000000 --- a/TS/support/timer.h +++ /dev/null @@ -1,66 +0,0 @@ -/*
- * 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>
-
-typedef struct Timer{
-
- struct timeval startTime[7];
- struct timeval stopTime[7];
- double time[7];
-
-}Timer;
-
-void start(Timer *timer, int i, int rep) {
- if(rep == 0) {
- timer->time[i] = 0.0;
- }
- gettimeofday(&timer->startTime[i], NULL);
-}
-
-void stop(Timer *timer, int i) {
- gettimeofday(&timer->stopTime[i], NULL);
- timer->time[i] += (timer->stopTime[i].tv_sec - timer->startTime[i].tv_sec) * 1000000.0 +
- (timer->stopTime[i].tv_usec - timer->startTime[i].tv_usec);
-}
-
-void print(Timer *timer, int i, int REP) { printf("%f\t", timer->time[i] / (1000 * REP)); }
-
-void printall(Timer *timer, int maxt) {
- for (int i = 0; i <= maxt; i++) {
- printf(" timer%d_us=%f", i, timer->time[i]);
- }
- printf("\n");
-}
|