From: miha-q <>
Date: Mon, 4 Mar 2024 23:25:30 +0000 (-0500)
Subject: Mon Mar  4 06:25:30 PM EST 2024
X-Git-Url: http://www.foleosoft.com/?a=commitdiff_plain;h=1b9174d022a1095a15224c832b72d3d62be421c2;p=QAnsel.git

Mon Mar  4 06:25:30 PM EST 2024
---

diff --git a/src/QAnsel.c b/src/QAnsel.c
index e501f37..90c68aa 100644
--- a/src/QAnsel.c
+++ b/src/QAnsel.c
@@ -212,22 +212,6 @@ void qansel_instruction(cpx_mtx_t* stateVector, unsigned char qubitCount, QInstr
 		tmp.cols = filter.cols * gate.cols;
 		tmp.ptr = malloc(tmp.rows * (tmp.cols * 2) * sizeof(float));
 
-
-		printf("%ix%i (knk)\n", tmp.rows, tmp.cols);
-		unsigned long int us1, us2;
-		us1 = get_time();
-		cpx_mtx_knk_metal(tmp.ptr, filter.ptr, gate.ptr, filter.rows, filter.cols, gate.rows, gate.cols);
-		us2 = get_time();
-		printf("\tMetal: %lu\n", us2 - us1);
-		us1 = get_time();
-		cpx_mtx_knk_threads(tmp.ptr, filter.ptr, gate.ptr, filter.rows, filter.cols, gate.rows, gate.cols);
-		us2 = get_time();
-		printf("\tThreads: %lu\n", us2 - us1);
-		us1 = get_time();
-		cpx_mtx_knk(tmp.ptr, filter.ptr, gate.ptr, filter.rows, filter.cols, gate.rows, gate.cols);
-		us2 = get_time();
-		printf("\tBare: %lu\n", us2 - us1);
-
 		#ifdef SPEED_TEST
 		printf("%ix%i (knk)\n", tmp.rows, tmp.cols);
 		unsigned long int us1, us2;
@@ -244,6 +228,18 @@ void qansel_instruction(cpx_mtx_t* stateVector, unsigned char qubitCount, QInstr
 		us2 = get_time();
 		printf("\tBare: %lu\n", us2 - us1);
 		#else
+		if (USE_GPU && 0)
+		{
+			cpx_mtx_knk_metal(tmp.ptr, filter.ptr, gate.ptr, filter.rows, filter.cols, gate.rows, gate.cols);
+		}
+		else if (USE_THREADS)
+		{
+			cpx_mtx_knk_threads(tmp.ptr, filter.ptr, gate.ptr, filter.rows, filter.cols, gate.rows, gate.cols);
+		}
+		else
+		{
+			cpx_mtx_knk(tmp.ptr, filter.ptr, gate.ptr, filter.rows, filter.cols, gate.rows, gate.cols);
+		}
 		#endif
 
 
@@ -1407,8 +1403,6 @@ void process(int argc, char** argv)
 void main(int argc, char** argv)
 {
 	USE_GPU = cpx_mtx_begin();
-	USE_GPU = 0;
-	USE_THREADS = 0;
 
 	RANDOM_FILE = fopen("/dev/TrueRNG0", "r");
 	if (!RANDOM_FILE) RANDOM_FILE = fopen("/dev/random", "r");
diff --git a/src/complex.c b/src/complex.c
index 3688fae..4aca6ec 100644
--- a/src/complex.c
+++ b/src/complex.c
@@ -474,7 +474,7 @@ void cpx_mtx_knk_metal(float* ptrR, float* ptrA, float* ptrB, int rowsA, int col
 	err = clSetKernelArg(kernel, 6, sizeof(int), &colsB); gpuerr(clSetKernelArg);
 
 	//Run the program
-	err = clEnqueueNDRangeKernel(cpx_mtx_command_queue, kernel, 1, NULL, (size_t[]){rowsR}, NULL, 0, NULL, NULL);
+	err = clEnqueueNDRangeKernel(cpx_mtx_command_queue, kernel, 2, NULL, (size_t[]){rowsR, colsR}, NULL, 0, NULL, NULL);
 	gpuerr(clEnqueueNDRangeKernel);
 
 	//Wait for completion