port TRNS to dfatool

9 files changed, 172 insertions, 64 deletions

diff --git a/TRNS/Makefile b/TRNS/Makefile
index 8a6fd64..f58eb0c 100644
--- a/TRNS/Makefile
+++ b/TRNS/Makefile
@@ -1,44 +1,36 @@
-DPU_DIR := dpu
-HOST_DIR := host
-BUILDDIR ?= bin
 NR_DPUS ?= 1
 NR_TASKLETS ?= 16
 ENERGY ?= 0
 
-define conf_filename
-	${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2).conf
-endef
-CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS})
-
-HOST_TARGET := ${BUILDDIR}/host_code
-DPU_TARGET := ${BUILDDIR}/dpu_code
-
 COMMON_INCLUDES := support
-HOST_SOURCES := $(wildcard ${HOST_DIR}/*.c)
-DPU_SOURCES := $(wildcard ${DPU_DIR}/*.c)
-
-.PHONY: all clean test
-
-__dirs := $(shell mkdir -p ${BUILDDIR})
+HOST_SOURCES := $(wildcard host/*.c)
+DPU_SOURCES := $(wildcard dpu/*.c)
 
 COMMON_FLAGS := -Wall -Wextra -g -I${COMMON_INCLUDES}
 HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DENERGY=${ENERGY} 
 DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} 
 
-all: ${HOST_TARGET} ${DPU_TARGET}
+QUIET = @
 
-${CONF}:
-	$(RM) $(call conf_filename,*,*)
-	touch ${CONF}
+ifdef verbose
+	QUIET =
+endif
 
-${HOST_TARGET}: ${HOST_SOURCES} ${COMMON_INCLUDES} ${CONF}
-	$(CC) -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
+all: bin/host_code bin/dpu_code
 
-${DPU_TARGET}: ${DPU_SOURCES} ${COMMON_INCLUDES} ${CONF}
-	dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
+bin:
+	${QUIET}mkdir -p bin
+
+bin/host_code: ${HOST_SOURCES} ${COMMON_INCLUDES} bin
+	${QUIET}${CC} -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
+
+bin/dpu_code: ${DPU_SOURCES} ${COMMON_INCLUDES} bin
+	${QUIET}dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
 
 clean:
-	$(RM) -r $(BUILDDIR)
+	${QUIET}rm -rf bin
 
 test: all
-	./${HOST_TARGET}
+	${QUIET}bin/host_code
+
+.PHONY: all clean test
diff --git a/TRNS/baselines/cpu/Makefile b/TRNS/baselines/cpu/Makefile
index cb2e264..781e2be 100644
--- a/TRNS/baselines/cpu/Makefile
+++ b/TRNS/baselines/cpu/Makefile
@@ -41,9 +41,29 @@ DEP=kernel.cpp kernel.h main.cpp support/common.h support/setup.h support/timer.
 SRC=main.cpp kernel.cpp
 EXE=trns
 
-all:
-	$(CXX) $(CXX_FLAGS) $(SRC) $(LIB) -o $(EXE)
+all: trns
+
+trns: ${SRC}
+	$(CXX) -O2 $(CXX_FLAGS) $(SRC) $(LIB) -o $(EXE)
+
+trns_O0: ${SRC}
+	$(CXX) $(CXX_FLAGS) $(SRC) $(LIB) -o $(EXE)_O0
+
+trns_O2: ${SRC}
+	$(CXX) -O2 $(CXX_FLAGS) $(SRC) $(LIB) -o $(EXE)_O2
+
+run: trns
+	./trns -w 0 -r 1 -m 16 -n 8 -o 4096 -p 2556
+
+# upstream uses -r 1
+
+run_O0: trns_O0
+	./trns_O0 -w 0 -r 50 -m 16 -n 8 -o 4096 -p 2556
+
+run_O2: trns_O2
+	./trns_O2 -w 0 -r 50 -m 16 -n 8 -o 4096 -p 2556
 
 clean:
-	rm -f $(EXE)
+	rm -f $(EXE) ${EXE}_O0 ${EXE}_O2
 
+.PHONY: all run run_O0 run_O2 clean
diff --git a/TRNS/baselines/cpu/main.cpp b/TRNS/baselines/cpu/main.cpp
index 2c3bdc3..2c481df 100644
--- a/TRNS/baselines/cpu/main.cpp
+++ b/TRNS/baselines/cpu/main.cpp
@@ -44,6 +44,9 @@
 #include <string.h>
 #include <assert.h>
 
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
 // Params ---------------------------------------------------------------------
 struct Params {
 
@@ -138,7 +141,7 @@ int main(int argc, char **argv) {
     T *h_in_backup = (T *)malloc(in_size * sizeof(T));
     ALLOC_ERR(h_in_backup);
     timer.stop("Allocation");
-    timer.print("Allocation", 1);
+    //timer.print("Allocation", 1);
 
     // Initialize
     timer.start("Initialization");
@@ -147,7 +150,7 @@ int main(int argc, char **argv) {
     for(int i = 0; i < N_; i++)
         h_head[i].store(0);
     timer.stop("Initialization");
-    timer.print("Initialization", 1);
+    //timer.print("Initialization", 1);
     memcpy(h_in_backup, h_in_out, in_size * sizeof(T)); // Backup for reuse across iterations
 
     // Loop over main kernel
@@ -197,10 +200,21 @@ int main(int argc, char **argv) {
         // end timer
         if(rep >= p.n_warmup)
             timer.stop("Step 3");
+
+        if (rep >= p.n_warmup) {
+            printf("[::] TRNS CPU | n_threads=%d e_type=%s n_elements=%d "
+                "| throughput_MBps=%f",
+                p.n_threads, XSTR(T), in_size,
+                in_size * sizeof(T) / (timer.get("Step 1") + timer.get("Step 2") + timer.get("Step 3")));
+            printf(" throughput_MOpps=%f",
+                in_size / (timer.get("Step 1") + timer.get("Step 2") + timer.get("Step 3")));
+            printf(" timer1_us=%f timer2_us=%f timer3_us=%f\n",
+                timer.get("Step 1"), timer.get("Step 2"), timer.get("Step 3"));
+        }
     }
-    timer.print("Step 1", p.n_reps);
-    timer.print("Step 2", p.n_reps);
-    timer.print("Step 3", p.n_reps);
+    //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);
@@ -212,7 +226,7 @@ int main(int argc, char **argv) {
     free(h_head);
     free(h_in_backup);
     timer.stop("Deallocation");
-    timer.print("Deallocation", 1);
+    //timer.print("Deallocation", 1);
 
     printf("Test Passed\n");
     return 0;
diff --git a/TRNS/baselines/cpu/run-opti.sh b/TRNS/baselines/cpu/run-opti.sh
new file mode 100755
index 0000000..0a3a4a3
--- /dev/null
+++ b/TRNS/baselines/cpu/run-opti.sh
@@ -0,0 +1,11 @@
+#!/bin/sh
+
+HOST="$(hostname)"
+
+echo $HOST
+
+make clean
+
+make run_O0 | sed 's/CPU/CPU O0/' | tee "${HOST}-O0.txt"
+
+make run_O2 | sed 's/CPU/CPU O2/' | tee "${HOST}-O2.txt"
diff --git a/TRNS/baselines/cpu/support/timer.h b/TRNS/baselines/cpu/support/timer.h
index 70ee386..d8ef221 100644
--- a/TRNS/baselines/cpu/support/timer.h
+++ b/TRNS/baselines/cpu/support/timer.h
@@ -47,9 +47,7 @@ struct Timer {
     map<string, double>         time;

 

     void start(string name) {

-        if(!time.count(name)) {

-            time[name] = 0.0;

-        }

+        time[name] = 0.0;

         gettimeofday(&startTime[name], NULL);

     }

 

@@ -59,5 +57,6 @@ struct Timer {
                       (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)); }

+    void print(string name, int REP) { printf("%s Time (ms): %f\n", name.c_str(), time[name] / (1000 * REP)); }

+    double get(string name) { return time[name]; }

 };

diff --git a/TRNS/host/app.c b/TRNS/host/app.c
index 1aa6392..77e58f2 100644
--- a/TRNS/host/app.c
+++ b/TRNS/host/app.c
@@ -18,6 +18,9 @@
 #include "../support/timer.h"
 #include "../support/params.h"
 
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
 // Define the DPU Binary path as DPU_BINARY here
 #ifndef DPU_BINARY
 #define DPU_BINARY "./bin/dpu_code"
@@ -97,11 +100,10 @@ int main(int argc, char **argv) {
     // Loop over main kernel
     for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
 
-        int timer_fix = 0;
         // Compute output on CPU (performance comparison and verification purposes)
         memcpy(A_host, A_backup, M_ * m * N_ * n * sizeof(T));
         if(rep >= p.n_warmup)
-            start(&timer, 0, rep - p.n_warmup + timer_fix);
+            start(&timer, 0, 0);
         trns_host(A_host, M_ * m, N_ * n, 1);
         if(rep >= p.n_warmup)
             stop(&timer, 0);
@@ -133,7 +135,7 @@ int main(int argc, char **argv) {
 
             printf("Load input data (step 1)\n");
             if(rep >= p.n_warmup)
-                start(&timer, 1, rep - p.n_warmup + timer_fix);
+                start(&timer, 1, 0);
             // Load input matrix (step 1)
             for(unsigned int j = 0; j < M_ * m; j++){
                 unsigned int i = 0;
@@ -160,7 +162,7 @@ int main(int argc, char **argv) {
             printf("Run step 2 on DPU(s) \n");
             // Run DPU kernel
             if(rep >= p.n_warmup){
-                start(&timer, 2, rep - p.n_warmup + timer_fix);
+                start(&timer, 2, 0);
 #if ENERGY
                 DPU_ASSERT(dpu_probe_start(&probe));
 #endif
@@ -193,7 +195,7 @@ int main(int argc, char **argv) {
             printf("Run step 3 on DPU(s) \n");
             // Run DPU kernel
             if(rep >= p.n_warmup){
-                start(&timer, 3, rep - p.n_warmup + timer_fix);
+                start(&timer, 3, 0);
 #if ENERGY
                 DPU_ASSERT(dpu_probe_start(&probe));
 #endif
@@ -219,7 +221,7 @@ int main(int argc, char **argv) {
 
             printf("Retrieve results\n");
             if(rep >= p.n_warmup)
-                start(&timer, 4, rep - p.n_warmup + timer_fix);
+                start(&timer, 4, 0);
             DPU_FOREACH(dpu_set, dpu) {
                 DPU_ASSERT(dpu_prepare_xfer(dpu, (T*)(&A_result[curr_dpu * m * n * M_])));
                 curr_dpu++;
@@ -231,10 +233,39 @@ int main(int argc, char **argv) {
             if(first_round){
                 first_round = 0;
             }
-            timer_fix++;
         }
         DPU_ASSERT(dpu_free(dpu_set));
 
+        // Check output
+        bool status = true;
+        for (i = 0; i < M_ * m * N_ * n; i++) {
+            if(A_host[i] != A_result[i]){ 
+                status = false;
+#if PRINT
+                printf("%d: %lu -- %lu\n", i, A_host[i], A_result[i]);
+#endif
+            }
+        }
+        if (status) {
+            printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
+            unsigned long input_size = M_ * m * N_ * n;
+            if (rep >= p.n_warmup) {
+                printf("[::] TRNS NMC | n_dpus=%d n_tasklets=%d e_type=%s n_elements=%lu "
+                    "| throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f",
+                    nr_of_dpus, NR_TASKLETS, XSTR(T), input_size,
+                    input_size * sizeof(T) / timer.time[0],
+                    input_size * sizeof(T) / (timer.time[2] + timer.time[3]),
+                    input_size * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4]));
+                printf(" throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f",
+                    input_size / timer.time[0],
+                    input_size / (timer.time[2] + timer.time[3]),
+                    input_size / (timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4]));
+                printall(&timer, 4);
+            }
+        } else {
+            printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
+        }
+
     }
 
     // Print timing results
@@ -255,27 +286,11 @@ int main(int argc, char **argv) {
     printf("DPU Energy (J): %f\t", energy);
     #endif	
 
-    // Check output
-    bool status = true;
-    for (i = 0; i < M_ * m * N_ * n; i++) {
-        if(A_host[i] != A_result[i]){ 
-            status = false;
-#if PRINT
-            printf("%d: %lu -- %lu\n", i, A_host[i], A_result[i]);
-#endif
-        }
-    }
-    if (status) {
-        printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
-    } else {
-        printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
-    }
-
     // Deallocation
     free(A_host);
     free(A_backup);
     free(A_result);
     free(done_host);
 	
-    return status ? 0 : -1;
+    return 0;
 }
diff --git a/TRNS/run-paper-strong-rank.sh b/TRNS/run-paper-strong-rank.sh
new file mode 100755
index 0000000..205eccb
--- /dev/null
+++ b/TRNS/run-paper-strong-rank.sh
@@ -0,0 +1,27 @@
+#!/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 BS strong-rank (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# 256 and 512 are not part of upstream
+for nr_dpus in 512 256 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets}; then
+			# upstream uses -p 64, but then the number of DPUs is always constant...
+			timeout --foreground -k 1m 60m bin/host_code -w 0 -e 100 -p $nr_dpus -o 12288 -x 1 || true
+		fi
+	done
+done
+)| tee log-paper-strong-rank.txt
diff --git a/TRNS/run-paper-weak.sh b/TRNS/run-paper-weak.sh
new file mode 100755
index 0000000..0840260
--- /dev/null
+++ b/TRNS/run-paper-weak.sh
@@ -0,0 +1,23 @@
+#!/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 BS weak (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# 256 and 512 are not part of upstream
+for nr_dpus in 256 512 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets}; then
+			timeout --foreground -k 1m 30m bin/bs_host -w 0 -e 100 -p 1 -o 12288 -x 0 || true
+		fi
+	done
+done | tee log-paper-weak.txt
diff --git a/TRNS/support/timer.h b/TRNS/support/timer.h
index b53d95f..4d597b9 100755
--- a/TRNS/support/timer.h
+++ b/TRNS/support/timer.h
@@ -57,3 +57,10 @@ void stop(Timer *timer, int i) {
 }

 

 void print(Timer *timer, int i, int REP) { printf("Time (ms): %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");

+}