re-implement wait_for_interrupt and get rid of multiplication in

check_results for better performance

raven_spn/src/init.h

@@ -12,7 +12,7 @@ typedef void (*function_ptr_t) (void);
 //! Instance data for the PLIC.
 plic_instance_t g_plic;
 std::array<function_ptr_t,PLIC_NUM_INTERRUPTS>  g_ext_interrupt_handlers;
-bool hw_interrupt{false};
+bool hw_interrupt{true};
 
 
 /*! \brief external interrupt handler
@@ -44,14 +44,18 @@ void configure_irq(size_t irq_num, function_ptr_t handler, unsigned char prio=1)
 }
 
 static void msi_interrupt_handler(){
-	hw_interrupt = true;
+	hw_interrupt = false;
 }
 
 void wait_for_interrupt() {
-	while (!hw_interrupt) {
-		delayUS(1);
-	};
-	hw_interrupt = false;
+    // wait until HW is done
+    if(hw_interrupt) {
+        do{
+            asm("wfi");
+            asm("nop");
+        }while(hw_interrupt);
+    }
+	hw_interrupt=true;
 }
 
 /*!\brief initializes platform

raven_spn/src/raven_spn.cpp

@@ -2,14 +2,11 @@
 #include "spn_regs.h"
 #include "init.h"
 
-#include <cmath> // log
-
 using spn =spn_regs<0x90000000>;
 
 // huge arrays of XSPN input and referance data
 extern std::array<uint8_t, 50000> input_data;
 extern std::array<double, 10000> ref_data;
-constexpr auto ln2 = std::log(2);
 
 bool double_equals(double a, double b, double epsilon = 0.001)
 {
@@ -21,12 +18,12 @@ void run_xspn(int in_addr, int out_addr) {
 	spn::input_length_reg() = 500; 		// each sample consists of 5 uint8 values
 	spn::input_addr_reg() = in_addr;
 	spn::output_addr_reg() = out_addr;
-	spn::num_of_in_beats_reg() = 40;		// Number of AXI4 burst beats needed to load all input data
+	spn::num_of_in_beats_reg() = 40;	// Number of AXI4 burst beats needed to load all input data
 	spn::num_of_out_beats_reg() = 64;	// Number of AXI4 burst beats needed to store all result data
 	spn::start_reg() = 1;
 }
 
-bool check_results(int addr, int k) {
+void check_results(int addr, int k) {
 	bool result = 0;
 	int  step = 500;
 	double *res_base = (double*) (addr);
@@ -34,14 +31,13 @@ bool check_results(int addr, int k) {
     printf("Start result comparison %d - %d\n", k, k+step);
 
 	for (int i = 0; i < step; i++) {
-		if (!double_equals(res_base[i] * ln2, ref_data.at(k + i))) {
+		if (!double_equals(res_base[i], ref_data.at(k + i))) {
 			printf("XSPN ref %d comparison FAILED\n", k + i);
 			*error_exit = 0x1;
 			result = 1;
 		}
 	}
     printf("Compared samples %d - %d with the reference\n", k, k+step);
-	return result;
 }
 
 /*! \brief main function
@@ -49,29 +45,29 @@ bool check_results(int addr, int k) {
  */
 int main() {
     platform_init();
-    int ret_val = 0;
 
     spn::mode_reg() = 1;
     spn::start_reg() = 1;
-	wait_for_interrupt();
+    wait_for_interrupt();
     spn::interrupt_reg() = 1;
     printf("READOUT HW:0x%x\n", spn::readout_reg());
 
 	int in_addr  = (int)input_data.data();
 	int out_addr = 0x800B0000;
 
+    run_xspn(in_addr, out_addr);
+
     int step = 500; // number of samples to be process at once
     for (int k = 0; k < 10000; k+=step) {
-
-        run_xspn(in_addr, out_addr);
         printf("XSPN processes samples %d - %d\n", k, k+step);
 
         wait_for_interrupt();
     	spn::interrupt_reg() = 1;
+		check_results(out_addr, k);
 
-		ret_val = check_results(out_addr, k);
 		in_addr += step * 5; // 5 bytes in each sample
+        run_xspn(in_addr, out_addr);
 	}
 
-	return ret_val;
+	return 0;
 }