Thu Mar 7 09:22:15 PM EST 2024

diff --git a/examples/slow.txt b/examples/slow.txt
index f4126d6591edb7805ece613083561c0d244610fa..ecdb1f1badc822258e486e248920be443814e013 100644 (file)
--- a/examples/slow.txt
+++ b/examples/slow.txt
@@ -1,5 +1,5 @@
 //designed to be slow
-qreg q[14];
+qreg q[11];
 x q[0];
 x q[1];
 x q[2];
@@ -11,8 +11,8 @@ x q[7];
 x q[8];
 x q[9];
 x q[10];
-x q[11];
-x q[12];
-x q[13];
+//x q[11];
+//x q[12];
+//x q[13];
 
-born;
\ No newline at end of file
+//born;
\ No newline at end of file

diff --git a/src/QAnsel.c b/src/QAnsel.c
index 851ec99d3a7c3da637e23a46ae6367a5dabebbb1..a7ae070f5a88c4ae4489325987b4298ea3d1caef 100644 (file)
--- a/src/QAnsel.c
+++ b/src/QAnsel.c
@@ -10,13 +10,17 @@
 unsigned char HIDDEN_VARIABLE = 0;
 FILE* RANDOM_FILE;
 #define GPU_ENABLED
-unsigned char USE_GPU = 1;
 unsigned char USE_THREADS = 1;
-#define SPEED_TEST
+#define MODE_BARE 1
+#define MODE_THREADED 2
+#define MODE_METAL 3
+#define MODE_METAL_THREADED 4
+unsigned char MODE = MODE_BARE;
+//#define SPEED_TEST
 
 typedef struct
 {
-       char n[128];//2082378
+       char n[128];
        unsigned char q0, q1, q2;
        float arg0, arg1, arg2;
 } QInstr;
@@ -223,22 +227,6 @@ void qansel_instruction(cpx_mtx_t* stateVector, unsigned char qubitCount, QInstr
                cpx_mtx_knk_metal_2x2(tmp.ptr, filter.ptr, gate.ptr, filter.rows, filter.cols, gate.rows, gate.cols);
                us2 = get_time();
                printf("\tMetal2x2: %lu\n", us2 - us1);
-
-               us1 = get_time();
-               for (int i = 0; i < filter.rows; i++)
-               {
-                       for (int j = 0; j < filter.cols; j++)
-                       {
-                               int x = (j * 2) + (i * filter.cols * 2);
-                               int y = (i * 2) + (j * filter.cols * 2);
-                               filter.ptr[x] = filter.ptr[y];
-                               filter.ptr[x + 1] = filter.ptr[y + 1];
-                       }
-               }
-               us2 = get_time();
-               printf("\tTranspose: %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();
@@ -261,11 +249,11 @@ void qansel_instruction(cpx_mtx_t* stateVector, unsigned char qubitCount, QInstr
                //us2 = get_time();
                //printf("\tTranspose: %lu\n", us2 - us1);
                #else
-               if (USE_GPU && tmp.rows >= 512)
+               if (MODE == MODE_METAL && tmp.cols >= 64)
                {
                        cpx_mtx_knk_metal_2x2(tmp.ptr, filter.ptr, gate.ptr, filter.rows, filter.cols, gate.rows, gate.cols);
                }
-               else if (USE_THREADS && tmp.rows >= 512)
+               else if ((MODE == MODE_THREADED || MODE == MODE_METAL_THREADED) && tmp.cols >= 64)
                {
                        cpx_mtx_knk_threads_2x2(tmp.ptr, filter.ptr, gate.ptr, filter.rows, filter.cols, gate.rows, gate.cols);
                }
@@ -299,11 +287,11 @@ void qansel_instruction(cpx_mtx_t* stateVector, unsigned char qubitCount, QInstr
        us2 = get_time();
        printf("\tBare: %lu\n", us2 - us1);
        #else
-       if (USE_GPU && tmp.cols >= 64)
+       if ((MODE == MODE_METAL || MODE == MODE_METAL_THREADED) && tmp.cols >= 64)
        {
                cpx_mtx_dot_metal(tmp.ptr, stateVector->ptr, filter.ptr, stateVector->rows, stateVector->cols, filter.rows, filter.cols);
        }
-       else if (USE_THREADS && tmp.cols >= 512)
+       else if (MODE == MODE_THREADED && tmp.cols >= 64)
        {
                cpx_mtx_dot_threads(tmp.ptr, stateVector->ptr, filter.ptr, stateVector->rows, stateVector->cols, filter.rows, filter.cols);
        }
@@ -1434,12 +1422,15 @@ void process(int argc, char** argv)
 
 void main(int argc, char** argv)
 {
-       USE_GPU = cpx_mtx_begin();
+       MODE = MODE_METAL;
+       unsigned char err = cpx_mtx_begin();
+       if (err == 0 && (MODE == MODE_METAL_THREADED || MODE == MODE_METAL))
+       {
+               MODE = MODE == MODE_METAL_THREADED ? MODE_THREADED : MODE_BARE;
+       }
        RANDOM_FILE = fopen("/dev/TrueRNG0", "r");
        if (!RANDOM_FILE) RANDOM_FILE = fopen("/dev/random", "r");
-       USE_GPU = 0;
-       USE_THREADS = 1;
        process(argc, argv);
        fclose(RANDOM_FILE);
-       if (USE_GPU) cpx_mtx_clean();
+       if (MODE_METAL || MODE_METAL_THREADED) cpx_mtx_clean();
 }
\ No newline at end of file

diff --git a/src/complex.c b/src/complex.c
index 722eef4341e2d9ffe89578cc9f55fb46413a5bea..f9a57b72a31341cc2708890024711d5f0dd59db0 100644 (file)
--- a/src/complex.c
+++ b/src/complex.c
@@ -575,6 +575,57 @@ void cpx_mtx_dot_metal(float* ptrR, float* ptrA, float* ptrB, int rowsA, int col
        err = clReleaseMemObject(memR); gpuerr(err);
 }
 
+/*typedef struct
+{
+       float* ptr;
+       cl_mem* buff;
+       size_t* buff_size;
+       size_t offset;
+       size_t count;
+} cpx_copy_context;
+
+void* cpx_copy_run(void *context)
+{
+       cpx_copy_context* ccc = (cpx_copy_context*)context;
+       cl_int err = clEnqueueWriteBuffer(cpx_mtx_command_queue, *(ccc->buff), CL_FALSE, ccc->offset, ccc->count, ccc->ptr, 0, NULL, NULL);
+       gpuerr(err);
+}
+
+void cpx_copy(float* ptr, cl_mem* buff, size_t* buff_size)
+{
+       int delimeter = (int)(*buff_size);
+       int cores = get_core_count();
+       int threadCount = cores;
+       if (threadCount > delimeter) threadCount = delimeter;
+       int delimetersPerThread = delimeter / threadCount;
+       int leftOvers = delimeter % threadCount;
+
+       cpx_copy_context ctxs[threadCount];
+    pthread_t threads[threadCount];
+
+       for (int i = 0; i < threadCount; i++)
+       {
+               ctxs[i].ptr = ptr;
+               ctxs[i].buff = buff;
+               ctxs[i].buff_size = buff_size;
+               ctxs[i].offset = i * delimetersPerThread;
+               ctxs[i].count = delimetersPerThread + ((i == threadCount - 1) ? leftOvers : 0);
+        if (pthread_create(&(threads[i]), NULL, &cpx_copy_run, (void*)&(ctxs[i])))
+        {
+            fprintf(stderr, "QAnsel: Thread error. (3)\n");
+            exit(1);
+        }
+       }
+    for (uint32_t i = 0; i < threadCount; i++)
+    {
+        if (pthread_join(threads[i], NULL))
+        {
+            fprintf(stderr, "QAnsel: Thread error. (4)\n");
+        }
+    }
+       clFlush(cpx_mtx_command_queue);
+}*/
+
 void cpx_mtx_knk_metal(float* ptrR, float* ptrA, float* ptrB, int rowsA, int colsA, int rowsB, int colsB)
 {
        int rowsR = rowsA * rowsB;
@@ -594,8 +645,7 @@ void cpx_mtx_knk_metal(float* ptrR, float* ptrA, float* ptrB, int rowsA, int col
     gpuerr(err);
        err = clEnqueueWriteBuffer(cpx_mtx_command_queue, memB, CL_TRUE, 0, sizeB, ptrB, 0, NULL, NULL);
     gpuerr(err);
-       printf("%lu!!!!\n", get_time() - q);
-
+       
        //Load and compile program
        cl_program program;
        if (cpx_mtx_cache == NULL)