[WIP] SPSC queue

software/bsg_manycore_lib/bsg_manycore_atomic.h

@@ -1,7 +1,7 @@
 #ifndef _BSG_MANYCORE_ATOMIC_H
 #define _BSG_MANYCORE_ATOMIC_H
 
-inline int bsg_amoswap (int* p, int val)
+inline int bsg_amoswap (volatile int* p, int val)
 {
   int result;
   asm volatile ("amoswap.w %[result], %[val], 0(%[p])" \
@@ -10,7 +10,7 @@ inline int bsg_amoswap (int* p, int val)
   return result;
 }
 
-inline int bsg_amoswap_aq (int* p, int val)
+inline int bsg_amoswap_aq (volatile int* p, int val)
 {
   int result;
   asm volatile ("amoswap.w.aq %[result], %[val], 0(%[p])" \
@@ -19,7 +19,7 @@ inline int bsg_amoswap_aq (int* p, int val)
   return result;
 }
 
-inline int bsg_amoswap_rl(int* p, int val)
+inline int bsg_amoswap_rl(volatile int* p, int val)
 {
   int result;
   asm volatile ("amoswap.w.rl %[result], %[val], 0(%[p])" \
@@ -28,7 +28,7 @@ inline int bsg_amoswap_rl(int* p, int val)
   return result;
 }
 
-inline int bsg_amoswap_aqrl(int* p, int val)
+inline int bsg_amoswap_aqrl(volatile int* p, int val)
 {
   int result;
   asm volatile ("amoswap.w.aqrl %[result], %[val], 0(%[p])" \
@@ -38,7 +38,7 @@ inline int bsg_amoswap_aqrl(int* p, int val)
 }
 
 
-inline int bsg_amoor (int* p, int val)
+inline int bsg_amoor (volatile int* p, int val)
 {
   int result;
   asm volatile ("amoor.w %[result], %[val], 0(%[p])" \
@@ -47,7 +47,7 @@ inline int bsg_amoor (int* p, int val)
   return result;
 }
 
-inline int bsg_amoor_aq (int* p, int val)
+inline int bsg_amoor_aq (volatile int* p, int val)
 {
   int result;
   asm volatile ("amoor.w.aq %[result], %[val], 0(%[p])" \
@@ -56,7 +56,7 @@ inline int bsg_amoor_aq (int* p, int val)
   return result;
 }
 
-inline int bsg_amoor_rl (int* p, int val)
+inline int bsg_amoor_rl (volatile int* p, int val)
 {
   int result;
   asm volatile ("amoor.w.rl %[result], %[val], 0(%[p])" \
@@ -65,7 +65,7 @@ inline int bsg_amoor_rl (int* p, int val)
   return result;
 }
 
-inline int bsg_amoor_aqrl (int* p, int val)
+inline int bsg_amoor_aqrl (volatile int* p, int val)
 {
   int result;
   asm volatile ("amoor.w.aqrl %[result], %[val], 0(%[p])" \
@@ -74,7 +74,7 @@ inline int bsg_amoor_aqrl (int* p, int val)
   return result;
 }
 
-inline int bsg_amoadd (int* p, int val)
+inline int bsg_amoadd (volatile int* p, int val)
 {
   int result;
   asm volatile ("amoadd.w %[result], %[val], 0(%[p])" \
@@ -83,7 +83,7 @@ inline int bsg_amoadd (int* p, int val)
   return result;
 }
 
-inline int bsg_amoadd_aq (int* p, int val)
+inline int bsg_amoadd_aq (volatile int* p, int val)
 {
   int result;
   asm volatile ("amoadd.w.aq %[result], %[val], 0(%[p])" \
@@ -92,7 +92,7 @@ inline int bsg_amoadd_aq (int* p, int val)
   return result;
 }
 
-inline int bsg_amoadd_rl (int* p, int val)
+inline int bsg_amoadd_rl (volatile int* p, int val)
 {
   int result;
   asm volatile ("amoadd.w.rl %[result], %[val], 0(%[p])" \
@@ -101,7 +101,7 @@ inline int bsg_amoadd_rl (int* p, int val)
   return result;
 }
 
-inline int bsg_amoadd_aqrl (int* p, int val)
+inline int bsg_amoadd_aqrl (volatile int* p, int val)
 {
   int result;
   asm volatile ("amoadd.w.aqrl %[result], %[val], 0(%[p])" \

software/bsg_manycore_lib/bsg_manycore_spsc_queue.hpp

@@ -0,0 +1,100 @@
+
+#ifndef _BSG_MANYCORE_SPSC_QUEUE_HPP
+#define _BSG_MANYCORE_SPSC_QUEUE_HPP
+
+#include <bsg_manycore_atomic.h>
+
+template <typename T, int S>
+class bsg_manycore_spsc_queue_recv {
+private:
+    volatile T *buffer;
+    volatile int *count;
+    int rptr;
+
+public:
+    bsg_manycore_spsc_queue_recv(T *buffer, int *count) 
+        : buffer(buffer), count(count), rptr(0) { };
+
+    bool is_empty(void)
+    {
+        return *count == 0;
+    }
+
+    bool try_recv(T *data)
+    {
+        if (is_empty())
+        {
+            return false;
+        }
+
+        *data = buffer[rptr];
+        // Probably faster than modulo, but should see if compiler
+        //   optimizes...
+        if (++rptr == S)
+        {
+            rptr = 0;
+        }
+        bsg_fence();
+        bsg_amoadd(count, -1);
+
+        return true;
+    }
+
+    // TODO: Add timeout?
+    T recv(void)
+    {
+        T data;
+        while (1)
+        {
+            if (try_recv(&data)) break;
+        }
+
+        return data;
+    }
+};
+
+template <typename T, int S>
+class bsg_manycore_spsc_queue_send {
+private:
+    volatile T *buffer;
+    volatile int *count;
+    int wptr;
+
+public:
+    bsg_manycore_spsc_queue_send(T *buffer, int *count) 
+        : buffer(buffer), count(count), wptr(0) { };
+
+    bool is_full(void)
+    {
+        return (*count == S);
+    }
+
+    bool try_send(T data)
+    {
+        if (is_full()) return false;
+
+        buffer[wptr] = data;
+        // Probably faster than modulo, but should see if compiler
+        //   optimizes...
+        if (++wptr == S)
+        {
+            wptr = 0;
+        }
+        bsg_fence();
+        bsg_amoadd(count, 1);
+
+        return true;
+    }
+
+    // TODO: Add timeout?
+    void send(T data)
+    {
+        while (1)
+        {
+            if (try_send(data)) break;
+        }
+    }
+};
+
+#endif
+

software/spmd/spsc_queue/Makefile

@@ -0,0 +1,15 @@
+bsg_tiles_X = 4
+bsg_tiles_Y = 2
+
+all: main.run
+
+OBJECT_FILES := main.o
+
+include ../Makefile.include
+
+main.riscv: $(LINK_SCRIPT) $(OBJECT_FILES) $(SPMD_COMMON_OBJECTS) $(BSG_MANYCORE_LIB)
+	$(RISCV_LINK)  $(filter %.o, $^) -L. "-l:$(BSG_MANYCORE_LIB)" -o $@ $(RISCV_LINK_OPTS)
+
+main.o: Makefile
+
+include $(BSG_MANYCORE_DIR)/software/mk/Makefile.tail_rules

software/spmd/spsc_queue/main.cpp

@@ -0,0 +1,68 @@
+//This kernel performs a barrier among all tiles in tile group 
+
+#include "bsg_manycore.h"
+#include "bsg_set_tile_x_y.h"
+#include "bsg_manycore_spsc_queue.hpp"
+#include "bsg_tile_group_barrier.hpp"
+
+bsg_barrier<bsg_tiles_X, bsg_tiles_Y> barrier;
+
+#define BUFFER_ELS 24
+#define CHAIN_LEN   8
+#define NUM_PACKETS 1000
+
+int buffer_chain [CHAIN_LEN*BUFFER_ELS] __attribute__ ((section (".dram"))) = {0};
+int buffer_count [CHAIN_LEN] __attribute__ ((section (".dram"))) = {0};
+
+int main()
+{
+    bsg_set_tile_x_y();
+
+    int *buffer = &buffer_chain[0] + (__bsg_id * BUFFER_ELS);
+    int *count  = &buffer_count[0] + (__bsg_id);
+
+    int *next_buffer = &buffer_chain[0] + ((__bsg_id+1) * BUFFER_ELS);
+    int *next_count = &buffer_count[0] + (__bsg_id+1);
+
+    bsg_manycore_spsc_queue_recv<int, BUFFER_ELS> recv_spsc(buffer, count);
+    bsg_manycore_spsc_queue_send<int, BUFFER_ELS> send_spsc(next_buffer, next_count);
+
+    int packets = 0;
+    int recv_data;
+    int send_data;
+    do
+    {
+        if (__bsg_id == 0)
+        {
+            recv_data = packets;
+        }
+        else
+        {
+            recv_data = recv_spsc.recv();
+        }
+
+        //bsg_printf("[%d] RECV %d\n", __bsg_id, recv_data);
+
+        send_data = recv_data;
+
+        if (__bsg_id == CHAIN_LEN-1)
+        {
+            if (recv_data != packets)
+            {
+                bsg_fail();
+            }
+        }
+        else
+        {
+            send_spsc.send(send_data);
+        }
+
+        //bsg_printf("[%d] SEND %d\n", __bsg_id, send_data);
+    } while (++packets < NUM_PACKETS);
+
+    barrier.sync();
+    bsg_finish();
+    bsg_wait_while(1);
+	return 0;
+}
+

software/spmd/vcache_atomic_mixed/Makefile

@@ -0,0 +1,17 @@
+bsg_tiles_X = 1
+bsg_tiles_Y = 1
+
+
+all: main.run
+
+OBJECT_FILES=main.o
+
+include ../Makefile.include
+
+main.riscv: $(LINK_SCRIPT) $(OBJECT_FILES) $(SPMD_COMMON_OBJECTS) $(BSG_MANYCORE_LIB) crt.o
+	$(RISCV_LINK) $(OBJECT_FILES) $(SPMD_COMMON_OBJECTS) -L. "-l:$(BSG_MANYCORE_LIB)" -o $@ $(RISCV_LINK_OPTS)
+
+
+main.o: Makefile
+
+include ../../mk/Makefile.tail_rules

software/spmd/vcache_atomic_mixed/main.c

@@ -0,0 +1,28 @@
+#include "bsg_manycore.h"
+#include "bsg_set_tile_x_y.h"
+#include "bsg_manycore_atomic.h"
+
+int lock __attribute__ ((section (".dram"))) = {0};
+
+int main()
+{
+
+  bsg_set_tile_x_y();
+
+  if (__bsg_id == 0)
+  {
+    // Using a regular int* here cause an illegal ordering resulting in:
+    //   x = 1, y = 0, z = 1
+    //int* lock_ptr = &lock;
+    volatile int* lock_ptr  = &lock;
+    int x = *lock_ptr;
+    int y = bsg_amoadd(lock_ptr, 1);
+    int z = *lock_ptr;
+
+    if (x != 0 || y != 0 || z != 1) bsg_fail();
+
+    bsg_finish();
+  }
+
+  bsg_wait_while(1);
+}