Reuse clccubes context. Makes things go zooom!

src/CCubes.c

@@ -45,7 +45,7 @@ SEXP CCubes(SEXP tt) {
     double elapsed_time;
 
     config_set_int("log", LOG_LEVEL_WARN);
-    config_set_int("log:clccubes", LOG_LEVEL_WARN);
+    config_set_int("log:clccubes", LOG_LEVEL_DEBUG);
     config_set_int("log:ccubes", LOG_LEVEL_DEBUG);
     config_set_int("log:cl", LOG_LEVEL_DEBUG);
 
@@ -53,6 +53,10 @@ SEXP CCubes(SEXP tt) {
     sprintf(log, "log-ccubes");
     config_set_string("out", log);
 
+    struct ccubes_context *ctx = ccubes_create();
+    if (ctx == NULL) {
+	    log_error("ccubes", "ccubes_do_tasks failed");
+    }
 
 
     if (PRINT_INFO) {
@@ -195,16 +199,21 @@ SEXP CCubes(SEXP tt) {
             Rprintf("---k: %d\n", k);
         }
 
-	int n_tasks = 512;
-	struct ccubes_context *ctx = NULL;
-        for (int task = 0; task < nchoosek(ninputs, k); task+=n_tasks) {
+	int n_tasks = nchoosek(ninputs, k);
+	int n_tasks_batch = 512;
+        for (int task = 0; task < n_tasks; task+=n_tasks_batch) {
+		/* adjust if batch size is larger than total job size */
+		int current_batch = n_tasks < n_tasks_batch ? n_tasks : n_tasks_batch;
+
 		log_debug("ccubes", "Tasks %d - %d out of %d",
-		    task, task + n_tasks, nchoosek(ninputs, k));
+		    task, task + current_batch - 1, n_tasks);
+
 		bool *coverage;
 		unsigned int *fixed_bits;
 		unsigned int *value_bits;
 		unsigned int *pichart_values;
-		ctx = ccubes_do_tasks(n_tasks,
+		ccubes_do_tasks(ctx,
+		    current_batch,
 		    task,
 		    k,
 		    ninputs,
@@ -227,11 +236,8 @@ SEXP CCubes(SEXP tt) {
 		    value_bits,
 		    pichart_values
 			);
-		if (ctx == NULL) {
-			log_error("ccubes", "ccubes_do_tasks failed");
-		}
 
-		for (int i = 0; i < n_tasks; i++) {
+		for (int i = 0; i < current_batch; i++) {
 			log_debug("ccubes", "Task %d", i);
 
 			log_debug_raw("ccubes", "coverage[%d]:", i);
@@ -262,7 +268,9 @@ SEXP CCubes(SEXP tt) {
 			}
 			log_debug_raw("ccubes", "\n");
 		}
-		break;
+
+		/* change to something less aggresive for reuse */
+		ccubes_clean_up(ctx);
 	}
 
         #ifdef _OPENMP
@@ -657,6 +665,8 @@ SEXP CCubes(SEXP tt) {
         }
     }
 
+    ccubes_destroy(ctx);
+
     R_Free(p_pichart);
     R_Free(p_implicants_val);
     R_Free(p_implicants_pos);

src/cl_setup.c

@@ -225,6 +225,7 @@ cl_init(struct cl_uctx *puctx)
 	/*
 	 * CPU
 	 */
+	puctx->cpu_queue = NULL;
 #if 0
 	puctx->cpu_queue = clCreateCommandQueue(ctx, devices[1], 0,
 	    &result);

src/clccubes.c

@@ -27,20 +27,10 @@
 
 
 struct ccubes_context *
-ccubes_init(int n_tasks,
-    int n_tasks_off,
-    int k,
-    int ninputs,
-    int posrows,
-    int negrows,
-    int implicant_words,
-    int value_bit_width,
-    int pichart_words,
-    int estimPI
-    )
+ccubes_create()
 {
-	int rc = 0;
 	struct ccubes_context *ctx = NULL;
+	int rc = 0;
 
 	ctx = malloc(sizeof *ctx);
 	if (ctx == NULL) {
@@ -55,6 +45,37 @@ ccubes_init(int n_tasks,
 		goto err;
 	}
 
+	ctx->clctx->device_type = CL_DEVICE_TYPE_GPU;
+	strcpy(ctx->clctx->platform_name, "NVIDIA Corporation\0");
+	/* strcpy(ctx->clctx->platform_name, "Advanced Micro Devices, Inc.\0"); */
+	/* strcpy(ctx->clctx->platform_name, "Intel(R) Corporation\0"); */
+	rc = cl_init(ctx->clctx);
+	if (rc != CL_SUCCESS) {
+		log_error("clccubes",
+		    "[%d] Failed to initialize the OpenCL framework", rc);
+		goto err;
+	}
+
+err:
+	return ctx;
+}
+
+int
+ccubes_init(struct ccubes_context *ctx,
+    int n_tasks,
+    int n_tasks_off,
+    int k,
+    int ninputs,
+    int posrows,
+    int negrows,
+    int implicant_words,
+    int value_bit_width,
+    int pichart_words,
+    int estimPI
+    )
+{
+	int rc = 0;
+
 	ctx->k = k;
 	ctx->ninputs = ninputs;
 	ctx->posrows = posrows;
@@ -68,7 +89,7 @@ ccubes_init(int n_tasks,
 	ctx->goff = n_tasks_off;
 
 err:
-	return ctx;
+	return rc;
 }
 
 int
@@ -108,17 +129,6 @@ ccubes_build(struct ccubes_context *ctx)
 {
 	int rc = 0;
 
-	ctx->clctx->device_type = CL_DEVICE_TYPE_GPU;
-	strcpy(ctx->clctx->platform_name, "NVIDIA Corporation\0");
-	/* strcpy(ctx->clctx->platform_name, "Advanced Micro Devices, Inc.\0"); */
-	/*strcpy(ctx->clctx->platform_name, "Intel(R) Corporation\0");*/
-	rc = cl_init(ctx->clctx);
-	if (rc != CL_SUCCESS) {
-		printf("[%d] Failed to initialize the OpenCL framework\n",
-		    rc);
-		goto err;
-	}
-
 
 	rc = cl_build(*ctx->clctx, CL_DEVICE_TYPE_GPU,
 	    "ccubes.cl", &ctx->ccubes_program);
@@ -392,8 +402,9 @@ err:
 	return rc;
 }
 
-struct ccubes_context *
-ccubes_do_tasks(int n_tasks,
+int
+ccubes_do_tasks(struct ccubes_context *ccubesctx,
+    int n_tasks,
     int n_tasks_off,
     int k,
     int ninputs,
@@ -418,9 +429,9 @@ ccubes_do_tasks(int n_tasks,
     )
 {
 	int rc = 0;
-	struct ccubes_context *ccubesctx = NULL;
 
-	ccubesctx = ccubes_init(n_tasks,
+	rc = ccubes_init(ccubesctx,
+	    n_tasks,
 	    n_tasks_off,
 	    k,
 	    ninputs,
@@ -477,15 +488,17 @@ ccubes_do_tasks(int n_tasks,
 		goto err;
 	}
 err:
-	return ccubesctx;
+	return rc;
 }
 
 void
 ccubes_clean_up(struct ccubes_context *ctx)
 {
 	clReleaseProgram(ctx->ccubes_program);
+	log_debug("clccubes", "clReleaseProgram ccubes_program");
 
 	ccubes_unmap(ctx);
+	log_debug("clccubes", "ccubes_unmap");
 
 	/* INPUTS */
 	clReleaseMemObject(ctx->nofvalues);
@@ -496,15 +509,25 @@ ccubes_clean_up(struct ccubes_context *ctx)
 	clReleaseMemObject(ctx->last_index);
 	clReleaseMemObject(ctx->p_covered);
 	clReleaseMemObject(ctx->p_pichart_pos);
+	log_debug("clccubes", "clReleaseMemObject INPUTS");
 
 	/* OUTPUTS */
 	clReleaseMemObject(ctx->coverage);
 	clReleaseMemObject(ctx->fixed_bits);
 	clReleaseMemObject(ctx->value_bits);
 	clReleaseMemObject(ctx->pichart_values);
-
-	cl_clean_up(*ctx->clctx);
+	log_debug("clccubes", "clReleaseMemObject OUTPUTS");
 
 	return;
 }
 
+void
+ccubes_destroy(struct ccubes_context *ctx)
+{
+	cl_clean_up(*ctx->clctx);
+	log_debug("clccubes", "cl_clean_up");
+
+	free(ctx);
+
+	return;
+}

src/clccubes.h

@@ -74,7 +74,11 @@ struct ccubes_context {
 };
 
 struct ccubes_context *
-ccubes_do_tasks(int n_tasks,
+ccubes_create();
+
+int
+ccubes_do_tasks(struct ccubes_context *ccubesctx,
+    int n_tasks,
     int n_tasks_off,
     int k,
     int ninputs,
@@ -98,4 +102,10 @@ ccubes_do_tasks(int n_tasks,
     unsigned int *pichart_values	/* OUT: RW */
 );
 
+void
+ccubes_clean_up(struct ccubes_context *ctx);
+
+void
+ccubes_destroy(struct ccubes_context *ctx);
+
 #endif