hw: fix MDA multicast Xbar integration

Author: Lore0599

hw/snitch/src/snitch.sv

@@ -281,7 +281,7 @@ module snitch import snitch_pkg::*; import riscv_instr::*; #(
   dm::dcsr_t dcsr_d, dcsr_q;
   logic [31:0] dpc_d, dpc_q;
   logic [31:0] dscratch_d, dscratch_q;
-  
+
   logic debug_d, debug_q;
   // Multicast mask
   logic [31:0] csr_mcast_d, csr_mcast_q;

hw/snitch_cluster/src/snitch_cluster.sv

@@ -329,7 +329,7 @@ module snitch_cluster
     UniqueIds: 1'b0,
     AxiAddrWidth: PhysicalAddrWidth,
     AxiDataWidth: WideDataWidth,
-    NoAddrRules: 2,
+    NoAddrRules: NrWideRules,
     NoMulticastRules: 1,
     NoMulticastPorts: 2,
     default: '0

sw/snRuntime/src/dma.h

@@ -60,7 +60,8 @@ inline uint32_t snrt_dma_start_1d_wideptr(uint64_t dst, uint64_t src,
 }
 
 /**
- * @brief Start an asynchronous multicast 1D DMA transfer with 64-bit wide pointers.
+ * @brief Start an asynchronous multicast 1D DMA transfer with 64-bit wide
+ * pointers.
  * @param dst The destination address.
  * @param src The source address.
  * @param size The size of the transfer in bytes.
@@ -119,15 +120,17 @@ inline snrt_dma_txid_t snrt_dma_start_1d(void *dst, const void *src,
 }
 
 /**
- * @brief Start an asynchronous multicast 1D DMA transfer with native-size pointers.
+ * @brief Start an asynchronous multicast 1D DMA transfer with native-size
+ * pointers.
  * @param dst The destination pointer.
  * @param src The source pointer.
  * @param size The size of the transfer in bytes.
  * @return The DMA transfer ID.
  */
 inline snrt_dma_txid_t snrt_dma_start_1d_mcast(void *dst, const void *src,
                                                size_t size, uint32_t mcast) {
-    return snrt_dma_start_1d_mcast_wideptr((size_t)dst, (size_t)src, size, mcast);
+    return snrt_dma_start_1d_mcast_wideptr((size_t)dst, (size_t)src, size,
+                                           mcast);
 }
 
 /**
@@ -487,10 +490,13 @@ inline snrt_dma_txid_t snrt_dma_load_1d_tile(void *dst, void *src,
  * @param mcast Multicast mask applied on the destination address.
  */
 inline snrt_dma_txid_t snrt_dma_mcast_load_1d_tile(void *dst, void *src,
-                                             size_t tile_idx, size_t tile_size,
-                                             uint32_t prec, uint32_t mcast) {
+                                                   size_t tile_idx,
+                                                   size_t tile_size,
+                                                   uint32_t prec,
+                                                   uint32_t mcast) {
     size_t tile_nbytes = tile_size * prec;
-    return snrt_dma_start_1d_mcast(dst, src + tile_idx * tile_nbytes, tile_nbytes, mcast);
+    return snrt_dma_start_1d_mcast(dst, src + tile_idx * tile_nbytes,
+                                   tile_nbytes, mcast);
 }
 
 /**

sw/snRuntime/src/start.c

@@ -57,7 +57,7 @@ static inline void snrt_init_bss() {
         if (snrt_is_dm_core()) {
             size_t size = (size_t)(&__bss_end) - (size_t)(&__bss_start);
             snrt_dma_start_1d_wideptr((uint64_t)(&__bss_start),
-                                  (uint64_t)(snrt_zero_memory_ptr()), size);
+                                      (uint64_t)(snrt_zero_memory_ptr()), size);
             snrt_dma_wait_all();
         }
         snrt_cluster_hw_barrier();
@@ -66,9 +66,9 @@ static inline void snrt_init_bss() {
 #endif
 
 #ifdef SNRT_WAKE_UP
-static inline void snrt_wake_up(){
-    if(snrt_cluster_idx() == 0 && snrt_cluster_core_idx() == 0){
-        snrt_wake_all((1 << snrt_cluster_core_num()) - 1 );
+static inline void snrt_wake_up() {
+    if (snrt_cluster_idx() == 0 && snrt_cluster_core_idx() == 0) {
+        snrt_wake_all((1 << snrt_cluster_core_num()) - 1);
     } else {
         snrt_int_clr_mcip();
     }
@@ -117,7 +117,7 @@ extern void snrt_exit(int exit_code);
 
 void snrt_main() {
     int exit_code = 0;
-    if (snrt_cluster_core_idx() == 0 && snrt_cluster_idx() == 0){
+    if (snrt_cluster_core_idx() == 0 && snrt_cluster_idx() == 0) {
         snrt_int_clr_mcip();
     }
 

sw/snRuntime/src/sync.h

@@ -75,22 +75,24 @@ inline void snrt_mutex_release(volatile uint32_t *pmtx) {
 // Barrier functions
 //================================================================================
 
-inline void snrt_wake_all(uint32_t core_mask){
+inline void snrt_wake_all(uint32_t core_mask) {
 #ifdef SUPPORTS_MULTICAST
     // Multicast cluster interrupt to every other cluster's core
     // Note: we need to address another cluster's address space
     //       because the cluster XBAR has not been extended to support
     //       multicast yet. We address the second cluster, if we are the
     //       first cluster, and the second otherwise.
-    uintptr_t addr = (uintptr_t)snrt_cluster_clint_set_ptr() - SNRT_CLUSTER_OFFSET * snrt_cluster_idx();
+    uintptr_t addr = (uintptr_t)snrt_cluster_clint_set_ptr() -
+                     SNRT_CLUSTER_OFFSET * snrt_cluster_idx();
     if (snrt_cluster_idx() == 0) addr += SNRT_CLUSTER_OFFSET;
     snrt_enable_multicast(BCAST_MASK_ALL);
     *((uint32_t *)addr) = core_mask;
     snrt_disable_multicast();
 #else
-    for (int i = 0; i < snrt_cluster_num(); i++){
-        if (snrt_cluster_idx() != i){
-            void* ptr = snrt_remote_l1_ptr(snrt_cluster_clint_set_ptr(), snrt_cluster_idx(), i);
+    for (int i = 0; i < snrt_cluster_num(); i++) {
+        if (snrt_cluster_idx() != i) {
+            void *ptr = snrt_remote_l1_ptr(snrt_cluster_clint_set_ptr(),
+                                           snrt_cluster_idx(), i);
             *((uint32_t *)ptr) = core_mask;
         }
     }
@@ -147,7 +149,6 @@ inline void snrt_global_barrier() {
     // Synchronize all DM cores in software
     if (snrt_is_dm_core()) {
         snrt_inter_cluster_barrier();
-
     }
     // Synchronize cores in a cluster with the HW barrier
     snrt_cluster_hw_barrier();

target/snitch_cluster/sw/runtime/rtl/src/snitch_cluster_start.h

@@ -4,8 +4,9 @@
 
 #pragma once
 
-#define SNRT_INIT_TLS
 #define SNRT_INIT_BSS
+#define SNRT_WAKE_UP
+#define SNRT_INIT_TLS
 #define SNRT_INIT_CLS
 #define SNRT_INIT_LIBS
 #define SNRT_CRT0_PRE_BARRIER