Support for different prefetching hints.

4 files changed, 77 insertions, 20 deletions

diff --git a/src/experiment.cpp b/src/experiment.cpp
index 4bf4eff..9b6af7c 100644
--- a/src/experiment.cpp
+++ b/src/experiment.cpp
@@ -53,7 +53,7 @@ Experiment::Experiment() :
     seconds          (DEFAULT_SECONDS),
     iterations       (DEFAULT_ITERATIONS),
     experiments      (DEFAULT_EXPERIMENTS),
-    prefetch         (DEFAULT_PREFETCH),
+    prefetch_hint    (NONE),
     output_mode      (TABLE),
     access_pattern   (RANDOM),
     stride           (1),
@@ -80,7 +80,7 @@ Experiment::~Experiment() {
 // -i or --iters            iterations
 // -e or --experiments      experiments
 // -g or --loop				cycles to execute for each iteration (latency hiding)
-// -f or --prefetch			prefetch data
+// -f or --prefetch			use of prefetching
 // -a or --access           memory access pattern
 //         random           random access pattern
 //         forward <stride> exclusive OR and mask
@@ -214,7 +214,25 @@ int Experiment::parse_args(int argc, char* argv[]) {
 			}
 		} else if (strcasecmp(argv[i], "-f") == 0
 				|| strcasecmp(argv[i], "--prefetch") == 0) {
-			this->prefetch = true;
+			i++;
+			if (i == argc) {
+				error = 1;
+				break;
+			}
+			if (strcasecmp(argv[i], "none") == 0) {
+				this->prefetch_hint = Experiment::NONE;
+			} else if (strcasecmp(argv[i], "nta") == 0) {
+				this->prefetch_hint = Experiment::NTA;
+			} else if (strcasecmp(argv[i], "t0") == 0) {
+				this->prefetch_hint = Experiment::T0;
+			} else if (strcasecmp(argv[i], "t1") == 0) {
+				this->prefetch_hint = Experiment::T1;
+			} else if (strcasecmp(argv[i], "t2") == 0) {
+				this->prefetch_hint = Experiment::T2;
+			} else {
+				error = 1;
+				break;
+			}
 		} else if (strcasecmp(argv[i], "-a") == 0
 				|| strcasecmp(argv[i], "--access") == 0) {
 			i++;
@@ -346,7 +364,7 @@ int Experiment::parse_args(int argc, char* argv[]) {
 		printf("    [-n|--numa]        <placement> # numa placement\n");
 		printf("    [-s|--seconds]     <number>    # run each experiment for <number> seconds\n");
 		printf("    [-g|--loop]        <number>    # cycles to execute for each iteration (latency hiding)\n");
-		printf("    [-f|--prefetch]                # prefetch data\n");
+		printf("    [-f|--prefetch]    <hint>      # use of prefetching\n");
 		printf("    [-x|--strict]                  # fail rather than adjust options to sensible values\n");
 		printf("\n");
 		printf("<pattern> is selected from the following:\n");
@@ -363,6 +381,13 @@ int Experiment::parse_args(int argc, char* argv[]) {
 		printf("    both                           # header and results in csv format\n");
 		printf("    table                          # human-readable table of values\n");
 		printf("\n");
+		printf("<hint> is selected from the following:\n");
+		printf("    none                           # do not use prefetching\n");
+		printf("    nta                            # use the NTA hint (non-temporal, only used once)\n");
+		printf("    t0                             # use the T0 hint (prefetch into all caches)\n");
+		printf("    t1                             # use the T1 hint (prefetch into all caches except L1)\n");
+		printf("    t2                             # use the T2 hint (prefetch into all caches except L1 & L2)\n");
+		printf("\n");
 		printf("<placement> is selected from the following:\n");
 		printf("    local                          # all chains are allocated locally\n");
 		printf("    xor <mask>                     # exclusive OR and mask\n");
@@ -656,7 +681,7 @@ void Experiment::print() {
 	printf("num_threads       = %d\n", num_threads);
 	printf("bytes_per_test    = %d\n", bytes_per_test);
 	printf("loop length       = %d\n", loop_length);
-	printf("prefetch          = %s\n", prefetch?"yes":"no");
+	printf("prefetch hint     = %s\n", prefetch_hint_string(prefetch_hint));
 	printf("iterations        = %d\n", iterations);
 	printf("experiments       = %d\n", experiments);
 	printf("access_pattern    = %d\n", access_pattern);
diff --git a/src/experiment.h b/src/experiment.h
index 5ede451..539e96a 100644
--- a/src/experiment.h
+++ b/src/experiment.h
@@ -54,12 +54,14 @@ public:
     int64 num_threads;		// number of threads in the experiment
     int64 bytes_per_test;	// test working set size (bytes)
     int64 loop_length;		// length of the inner loop (cycles)
-    bool prefetch;			// use of prefetching
 
     float seconds;			// number of seconds per experiment
     int64 iterations;		// number of iterations per experiment
     int64 experiments;		// number of experiments per test
 
+    enum { NONE, T0, T1, T2, NTA }
+    prefetch_hint;			// use of prefetching
+
     enum { CSV, BOTH, HEADER, TABLE }
 	output_mode;			// results output mode
 
@@ -101,9 +103,6 @@ public:
     const static int32 DEFAULT_ITERATIONS        = 0;
     const static int32 DEFAULT_EXPERIMENTS       = 1;
 
-    const static int32 DEFAULT_OUTPUT_MODE       = 1;
-    const static bool DEFAULT_PREFETCH           = false;
-
     void alloc_local();
 	void alloc_xor();
 	void alloc_add();
@@ -114,4 +113,20 @@ public:
 private:
 };
 
+
+inline const char* prefetch_hint_string(int32 prefetch_hint) {
+	switch (prefetch_hint) {
+	case Experiment::NONE:
+		return "none";
+	case Experiment::T0:
+		return "t0";
+	case Experiment::T1:
+		return "t1";
+	case Experiment::T2:
+		return "t2";
+	case Experiment::NTA:
+		return "nta";
+	}
+}
+
 #endif
diff --git a/src/output.cpp b/src/output.cpp
index 4efb415..ad4aa98 100644
--- a/src/output.cpp
+++ b/src/output.cpp
@@ -50,7 +50,7 @@ void Output::header(Experiment &e, int64 ops, double secs, double ck_res) {
     printf("number of threads,");
     printf("iterations,");
     printf("loop length,");
-    printf("prefetch,");
+    printf("prefetch hint,");
     printf("experiments,");
     printf("access pattern,");
     printf("stride,");
@@ -80,7 +80,7 @@ void Output::csv(Experiment &e, int64 ops, double secs, double ck_res) {
     printf("%ld,", e.num_threads);
     printf("%ld,", e.iterations);
     printf("%ld,", e.loop_length);
-    printf("%s,", e.prefetch?"yes":"no");
+    printf("%s,", prefetch_hint_string(e.prefetch_hint));
     printf("%ld,", e.experiments);
     printf("%s,", e.access());
     printf("%ld,", e.stride);
@@ -123,7 +123,7 @@ void Output::table(Experiment &e, int64 ops, double secs, double ck_res) {
     printf("number of threads    = %ld\n", e.num_threads);
     printf("iterations           = %ld\n", e.iterations);
     printf("loop length          = %ld\n", e.loop_length);
-    printf("prefetch             = %s\n", e.prefetch?"yes":"no");
+    printf("prefetch hint        = %s\n", prefetch_hint_string(e.prefetch_hint));
     printf("experiments          = %ld\n", e.experiments);
     printf("access pattern       = %s\n", e.access());
     printf("stride               = %ld\n", e.stride);
diff --git a/src/run.cpp b/src/run.cpp
index bc9a533..b990a4c 100644
--- a/src/run.cpp
+++ b/src/run.cpp
@@ -40,13 +40,13 @@ static double min(double v1, double v2);
 typedef void (*benchmark)(const Chain**);
 typedef benchmark (*generator)(int64 chains_per_thread,
 		int64 bytes_per_line, int64 bytes_per_chain,
-		int64 stride, int64 loop_length, bool prefetch);
+		int64 stride, int64 loop_length, int32 prefetch_hint);
 static benchmark chase_pointers(int64 chains_per_thread,
 		int64 bytes_per_line, int64 bytes_per_chain,
-		int64 stride, int64 loop_length, bool prefetch);
+		int64 stride, int64 loop_length, int32 prefetch_hint);
 static benchmark follow_streams(int64 chains_per_thread,
 		int64 bytes_per_line, int64 bytes_per_chain,
-		int64 stride, int64 loop_length, bool prefetch);
+		int64 stride, int64 loop_length, int32 prefetch_hint);
 
 Lock Run::global_mutex;
 int64 Run::_ops_per_chain = 0;
@@ -121,7 +121,7 @@ int Run::run() {
 		benchmark bench = gen(this->exp->chains_per_thread,
 				this->exp->bytes_per_line, this->exp->bytes_per_chain,
 				this->exp->stride, this->exp->loop_length,
-				this->exp->prefetch);
+				this->exp->prefetch_hint);
 
 		volatile static double istart = 0;
 		volatile static double istop = 0;
@@ -171,7 +171,7 @@ int Run::run() {
 	benchmark bench = gen(this->exp->chains_per_thread,
 			this->exp->bytes_per_line, this->exp->bytes_per_chain,
 			this->exp->stride, this->exp->loop_length,
-			this->exp->prefetch);
+			this->exp->prefetch_hint);
 
 	for (int e = 0; e < this->exp->experiments; e++) {
 		// barrier
@@ -374,7 +374,7 @@ static benchmark chase_pointers(int64 chains_per_thread, // memory loading per t
 		int64 bytes_per_chain, // ignored
 		int64 stride, // ignored
 		int64 loop_length, // length of the inner loop
-		bool prefetch // prefetch
+		int32 prefetch_hint // use of prefetching
 		) {
 	// Create Compiler.
 	AsmJit::Compiler c;
@@ -417,8 +417,25 @@ static benchmark chase_pointers(int64 chains_per_thread, // memory loading per t
 		c.mov(positions[i], ptr(positions[i], offsetof(Chain, next)));
 
 		// Prefetch next
-		if (prefetch)
+		switch (prefetch_hint)
+		{
+		case Experiment::T0:
 			c.prefetch(ptr(positions[i]), AsmJit::PREFETCH_T0);
+			break;
+		case Experiment::T1:
+			c.prefetch(ptr(positions[i]), AsmJit::PREFETCH_T1);
+			break;
+		case Experiment::T2:
+			c.prefetch(ptr(positions[i]), AsmJit::PREFETCH_T2);
+			break;
+		case Experiment::NTA:
+			c.prefetch(ptr(positions[i]), AsmJit::PREFETCH_NTA);
+			break;
+		case Experiment::NONE:
+		default:
+			break;
+
+		}
 	}
 
 	// Wait
@@ -485,7 +502,7 @@ static benchmark follow_streams(int64 chains_per_thread, // memory loading per t
 		int64 bytes_per_chain, // ignored
 		int64 stride, // ignored
 		int64 loop_length, // ignored
-		bool prefetch // ignored
+		int32 prefetch_hint // ignored
 		) {
 	return 0;
 	/*