sw: Align collective support with hardware

Author: Lore0599

sw/runtime/api/sync_decls.h

@@ -23,7 +23,8 @@ typedef struct {
 typedef snrt_comm_info_t *snrt_comm_t;
 
 typedef enum {
-    SNRT_REDUCTION_NONE = 0,
+    SNRT_COLLECTIVE_UNICAST = 0,
+    SNRT_COLLECTIVE_MULTICAST = 1,
     SNRT_REDUCTION_BARRIER = 2,
     SNRT_REDUCTION_FADD = 4,
     SNRT_REDUCTION_FMUL = 5,
@@ -35,21 +36,13 @@ typedef enum {
     SNRT_REDUCTION_MAX = 11,
     SNRT_REDUCTION_MINU = 14,
     SNRT_REDUCTION_MAXU = 15
-} snrt_reduction_opcode_t;
-
-typedef enum {
-    SNRT_COLLECTIVE_UNICAST = 0,
-    SNRT_COLLECTIVE_MULTICAST = 1,
-    SNRT_COLLECTIVE_PARALLEL_REDUCTION = 2,
-    SNRT_COLLECTIVE_OFFLOAD_REDUCTION = 3
 } snrt_collective_opcode_t;
 
 typedef union {
     struct __attribute__((__packed__)) {
-        // snrt_reduction_opcode_t reduction_opcode : SNRT_REDUCTION_OPCODE_WIDTH;
-        // snrt_collective_opcode_t collective_opcode
-        //     : SNRT_COLLECTIVE_OPCODE_WIDTH;
-        uint64_t mask : SNRT_COLLECTIVE_MASK_WIDTH;
+        snrt_collective_opcode_t collective_op
+            : SNRT_COLLECTIVE_OPCODE_WIDTH;
+        uint64_t mask : (64 - SNRT_COLLECTIVE_OPCODE_WIDTH);
     } f;
     uint64_t w;
 } snrt_collective_op_t;
@@ -79,7 +72,7 @@ inline void snrt_enable_multicast(uint64_t mask);
 inline void snrt_disable_multicast();
 
 inline void snrt_enable_reduction(uint64_t mask,
-                                  snrt_reduction_opcode_t reduction);
+                                  snrt_collective_opcode_t reduction);
 
 inline void snrt_disable_reduction();
 

sw/runtime/impl/snitch_cluster_cfg.h.tpl

@@ -34,8 +34,7 @@
 #define SNRT_SUPPORTS_NARROW_REDUCTION
 % endif
 
-#define SNRT_REDUCTION_OPCODE_WIDTH ${cfg['cluster']['reduction_opcode_width']}
-#define SNRT_COLLECTIVE_OPCODE_WIDTH ${cfg['cluster']['collective_width'] - cfg['cluster']['reduction_opcode_width']}
+#define SNRT_COLLECTIVE_OPCODE_WIDTH ${cfg['cluster']['collective_width']}
 
 // Software configuration
 #define SNRT_LOG2_STACK_SIZE 10

sw/runtime/src/dma.c

@@ -7,7 +7,7 @@ extern void snrt_dma_set_awuser(uint64_t field);
 extern void snrt_dma_enable_multicast(uint64_t mask);
 
 extern void snrt_dma_enable_reduction(uint64_t mask,
-                                      snrt_reduction_opcode_t opcode);
+                                      snrt_collective_opcode_t opcode);
 
 extern void snrt_dma_disable_multicast();
 
@@ -32,7 +32,7 @@ extern snrt_dma_txid_t snrt_dma_load_1d_tile_mcast(void *dst, void *src,
 
 extern snrt_dma_txid_t snrt_dma_reduction_load_1d_tile(
     void *dst, void *src, size_t tile_idx, size_t tile_size, uint32_t prec,
-    uint64_t mask, snrt_reduction_opcode_t opcode);
+    uint64_t mask, snrt_collective_opcode_t opcode);
 
 extern snrt_dma_txid_t snrt_dma_1d_to_2d(volatile void *dst, volatile void *src,
                                          size_t size, size_t row_size,

sw/runtime/src/dma.h

@@ -85,7 +85,7 @@ inline void snrt_dma_set_awuser(uint64_t field) {
  */
 inline void snrt_dma_enable_multicast(uint64_t mask) {
     snrt_collective_op_t op;
-    // op.f.collective_opcode = SNRT_COLLECTIVE_MULTICAST;
+    op.f.collective_op = SNRT_COLLECTIVE_MULTICAST;
     op.f.mask = mask;
     snrt_dma_set_awuser(op.w);
 }
@@ -99,10 +99,9 @@ inline void snrt_dma_enable_multicast(uint64_t mask) {
  * @param opcode Type of reduction operation
  */
 inline void snrt_dma_enable_reduction(uint64_t mask,
-                                      snrt_reduction_opcode_t opcode) {
+                                      snrt_collective_opcode_t opcode) {
     snrt_collective_op_t op;
-    // op.f.reduction_opcode = opcode;
-    // op.f.collective_opcode = SNRT_COLLECTIVE_OFFLOAD_REDUCTION;
+    op.f.collective_op = opcode;
     op.f.mask = mask;
     snrt_dma_set_awuser(op.w);
 }
@@ -129,7 +128,7 @@ inline void snrt_dma_disable_reduction() { snrt_dma_set_awuser(0); }
  */
 static inline uint32_t snrt_dma_start_1d_reduction(
     uint64_t dst, uint64_t src, size_t size, uint64_t mask,
-    snrt_reduction_opcode_t opcode, const uint32_t channel = 0) {
+    snrt_collective_opcode_t opcode, const uint32_t channel = 0) {
     snrt_dma_enable_reduction(mask, opcode);
     uint32_t txid = snrt_dma_start_1d(dst, src, size, channel);
     snrt_dma_disable_reduction();
@@ -162,7 +161,7 @@ static inline uint32_t snrt_dma_start_1d_mcast(uint64_t dst, uint64_t src,
  */
 static inline uint32_t snrt_dma_start_1d_reduction(
     volatile void *dst, volatile void *src, size_t size, uint64_t mask,
-    snrt_reduction_opcode_t opcode, const uint32_t channel = 0) {
+    snrt_collective_opcode_t opcode, const uint32_t channel = 0) {
     return snrt_dma_start_1d_reduction((uint64_t)dst, (uint64_t)src, size, mask,
                                        opcode, channel);
 }
@@ -423,7 +422,7 @@ inline snrt_dma_txid_t snrt_dma_load_1d_tile_mcast(void *dst, void *src,
  */
 inline snrt_dma_txid_t snrt_dma_reduction_load_1d_tile(
     void *dst, void *src, size_t tile_idx, size_t tile_size, uint32_t prec,
-    uint64_t mask, snrt_reduction_opcode_t opcode) {
+    uint64_t mask, snrt_collective_opcode_t opcode) {
     size_t tile_nbytes = tile_size * prec;
     return snrt_dma_start_1d_reduction((uintptr_t)dst,
                                        (uintptr_t)src + tile_idx * tile_nbytes,

sw/runtime/src/sync.h

@@ -432,7 +432,7 @@ inline void snrt_set_awuser(uint64_t field) {
  */
 inline void snrt_enable_multicast(uint64_t mask) {
     snrt_collective_op_t op;
-    // op.f.collective_opcode = SNRT_COLLECTIVE_MULTICAST;
+    op.f.collective_op = SNRT_COLLECTIVE_MULTICAST;
     op.f.mask = mask;
     snrt_set_awuser(op.w);
 }
@@ -454,21 +454,10 @@ inline void snrt_disable_multicast() { snrt_set_awuser(0); }
  * @param opcode Type of reduction operation
  */
 inline void snrt_enable_reduction(uint64_t mask,
-                                  snrt_reduction_opcode_t opcode) {
-    snrt_collective_opcode_t coll_opcode;
-
-    switch (opcode) {
-        case SNRT_REDUCTION_BARRIER:
-            coll_opcode = SNRT_COLLECTIVE_PARALLEL_REDUCTION;
-            break;
-        default:
-            coll_opcode = SNRT_COLLECTIVE_OFFLOAD_REDUCTION;
-            break;
-    }
+                                  snrt_collective_opcode_t opcode) {
 
     snrt_collective_op_t op;
-    // op.f.reduction_opcode = opcode;
-    // op.f.collective_opcode = coll_opcode;
+    op.f.collective_op = opcode;
     op.f.mask = mask;
     snrt_set_awuser(op.w);
 }