implementing gpu library

src/riscv_gpgpu/lib/lib.c

@@ -0,0 +1,108 @@
+#include "lib.h"
+
+// namespace Sphinx
+// {
+
+#define FUNC void (func)(unsigned)
+
+
+
+void set_wid(unsigned i)
+{
+	SET_WID(i);
+}
+
+void set_func(FUNC)
+{
+	SET_FUNC(func);
+}
+
+unsigned get_func()
+{
+	unsigned ret;
+	GET_FUNC(ret);
+	return ret;
+}
+
+unsigned get_wid()
+{
+	unsigned ret;
+	GET_WID(ret);
+	return ret;
+}
+
+void createThreads(unsigned num_threads, unsigned wid, unsigned func_addr)
+{
+
+	asm __volatile__("addi t5, sp, 0");
+	for (unsigned i = 1; i < num_threads; i++)
+	{
+
+		register unsigned cur_tid asm("t1") = i;
+		asm __volatile__("addi sp, sp, -256");
+		CLONE;
+	}
+	asm __volatile__("addi sp, t5, 0");
+
+
+	register unsigned cur_tid asm("t1") = 0;
+
+
+	// jalis TO FUNC
+	register unsigned num_lanes asm("t6")  = func_addr;
+	register unsigned link      asm("s11") = num_threads;
+
+	register unsigned n_threads asm("a0")  = wid;
+	JALRS;
+	ECALL;
+
+}
+
+void wspawn(unsigned num_threads, unsigned wid, FUNC)
+{
+
+	// set_wid(wid);
+	// set_func(func);
+
+	
+
+	register unsigned n_threads asm("a0") = num_threads;
+	register unsigned wwid   asm("a1") = wid;
+	register unsigned ffunc  asm("a2") = (unsigned) func;
+
+	register unsigned func_add  asm("t1") = (unsigned) &createThreads;
+	WSPAWN; // THIS SHOULD COPY THE CSR REGISTERS TO THE NEW WARP
+
+}
+
+
+
+void createWarps(unsigned num_Warps, unsigned num_threads, FUNC)
+{
+	// asm __volatile__("addi t5, sp, 0");
+
+	// for (unsigned i = 1; i < num_Warps; i++)
+	// {
+	// 	asm __volatile__("addi sp, sp, -2048");
+	// 	wspawn(num_threads, i, func);
+	// }
+
+	// asm __volatile__("addi sp, t5, 0");
+
+	createThreads(num_threads, 0, (unsigned) func);
+
+
+	ECALL;
+
+}
+
+
+unsigned get_tid()
+{
+	register unsigned tid asm("t1");
+}
+
+void initiate_stack()
+{
+	asm __volatile__("lui  sp,0x7ffff");
+}

src/riscv_gpgpu/lib/lib.h

@@ -0,0 +1,30 @@
+
+#ifndef __RISCV_GP_
+#define __RISCV_GP_
+
+
+#define WID_CSR  0x00E
+#define FUNC_CSR 0x00F
+
+#define SET_WID(val)    asm __volatile__("csrw 0x00e,%0"::"r"(val));
+#define GET_WID(ret)   asm __volatile__("csrr %0,0x00e":"=r"(ret));
+
+#define SET_FUNC(val)   asm __volatile__("csrw 0x00f,%0"::"r"(val));
+#define GET_FUNC(ret)   asm __volatile__("csrr %0,0x00f":"=r"(ret));
+
+
+#define WSPAWN          asm __volatile__(".word 0x3006b"::);
+#define CLONE           asm __volatile__(".word 0x3506b":::"t1");
+#define JALRS           asm __volatile__(".word 0x1bfe0eb":::"s10")
+#define ECALL           asm __volatile__(".word 0x00000073")
+
+
+#define FUNC void (func)(unsigned)
+void createWarps(unsigned num_Warps, unsigned num_threads, FUNC);
+unsigned get_tid(void);
+void initiate_stack(void);
+
+
+
+#endif
+