sw: Add cluster-to-cluster multicast app

.github/workflows/ci.yml

@@ -53,4 +53,4 @@ jobs:
           submodules: 'recursive'
       - name: Generate single-cluster Occamy RTL sources
         run: |
-          make -C target/sim CFG_OVERRIDE=cfg/single-cluster.hjson rtl
+          make -C target/sim CFG_OVERRIDE=cfg/U-Q1C1.hjson rtl

.gitlab-ci.yml

@@ -48,7 +48,7 @@ docs:
 occamy-single-cluster-vsim:
   script:
     - cd target/sim
-    - make CFG_OVERRIDE=cfg/single-cluster.hjson rtl
+    - make CFG_OVERRIDE=cfg/U-Q1C1.hjson rtl
     - make sw
     - make bin/occamy_top.vsim
     - ./run.py sw/run-single-cluster.yaml --simulator vsim
@@ -61,7 +61,7 @@ occamy-single-cluster-vsim:
 occamy-full-vsim:
   script:
     - cd target/sim
-    - make CFG_OVERRIDE=cfg/full.hjson rtl
+    - make CFG_OVERRIDE=cfg/U-Q6C4.hjson rtl
     - make LENGTH=384 sw
     - make bin/occamy_top.vsim
     - ./run.py sw/run-full-occamy.yaml --simulator vsim
@@ -77,7 +77,7 @@ occamy-mcast-vsim:
     DATA_CFG: "$(PWD)/sw/device/apps/blas/gemm/params.hjson"
   script:
     - cd target/sim
-    - make CFG_OVERRIDE=cfg/M-Q8C4.hjson rtl
+    - make CFG_OVERRIDE=cfg/M-Q4C4.hjson rtl
     - make LENGTH=1024 sw
     - make bin/occamy_top.vsim
-    - ./run.py sw/run-full-occamy.yaml -j
+    - ./run.py sw/run-mcast-occamy.yaml -j

target/sim/Makefile

@@ -67,6 +67,7 @@ CVA6_TRACE            = trace_hart_0.log
 # to be rebuilt. This file is used to express this condition as a
 # prerequisite for other rules.
 CFG = cfg/lru.hjson
+DEFAULT_CFG = cfg/U-Q6C4.hjson
 
 #####################
 # Simulator options #
@@ -147,7 +148,7 @@ VLT_COBJ += $(VLT_BUILDDIR)/vlt/verilated_vcd_c.o
 $(CFG): FORCE
 	@# If the LRU config file doesn't exist, we use the default config.
 	@if [ ! -e $@ ] ; then \
-		DEFAULT_CFG="cfg/full.hjson"; \
+		DEFAULT_CFG=$(DEFAULT_CFG); \
 		echo "Using default config file: $$DEFAULT_CFG"; \
 		cp $$DEFAULT_CFG $@; \
 	fi

target/sim/cfg/M-Q8C4.hjson → target/sim/cfg/M-Q4C4.hjson

@@ -86,7 +86,7 @@
     }
     narrow_xbar_slv_id_width: 4,
     narrow_xbar_user_width: 5, // clog2(total number of clusters)
-    nr_s1_quadrant: 8,
+    nr_s1_quadrant: 4,
     s1_quadrant: {
       nr_clusters: 4,
       // number of pending transactions on the narrow/wide network

target/sim/sw/device/Makefile

@@ -7,6 +7,7 @@
 # Add user applications to APPS variable
 APPS  = blas/axpy
 APPS += blas/gemm
+APPS += mcast
 
 TARGET ?= all
 

target/sim/sw/device/apps/mcast/Makefile

@@ -0,0 +1,10 @@
+# Copyright 2023 ETH Zurich and University of Bologna.
+# Licensed under the Apache License, Version 2.0, see LICENSE for details.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Luca Colagrande <colluca@iis.ee.ethz.ch>
+
+APP = mcast
+SRCS = main.c
+
+include ../common.mk

target/sim/sw/device/apps/mcast/main.c

@@ -0,0 +1,52 @@
+// Copyright 2023 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Author: Luca Colagrande <colluca@iis.ee.ethz.ch>
+
+#include "snrt.h"
+
+#define LENGTH 32
+#define INITIALIZER 0xAAAAAAAA
+
+// Allocate a buffer in main memory which we will use to copy data around
+// with the DMA.
+uint32_t buffer_src[LENGTH];
+
+int main() {
+    uint32_t is_first_cluster = snrt_cluster_idx() == 0;
+
+    // Allocate destination buffer
+    uint32_t *buffer_dst = snrt_l1_next();
+
+    // First cluster initializes the source buffer in DRAM and multicast-
+    // copies it to the destination buffer in every cluster's TCDM.
+    if (snrt_is_dm_core() && is_first_cluster) {
+        // Initializing the DRAM source buffer using the LSU may cause data
+        // races with the DMA "later" copying the source buffer to the TCDM
+        // destinations. We therefore initialize the DRAM source buffer using
+        // the DMA itself, from a TCDM source buffer initialized with the LSU.
+        uint32_t *buffer_src_tcdm = buffer_dst + LENGTH;
+        for (uint32_t i = 0; i < LENGTH; i++) {
+            buffer_src_tcdm[i] = INITIALIZER;
+        }
+        snrt_dma_start_1d(buffer_src, buffer_src_tcdm, sizeof(buffer_src));
+
+        // Initiate DMA transfer
+        dma_broadcast_to_clusters(buffer_dst, buffer_src, sizeof(buffer_src));
+    }
+
+    // All other clusters wait on a global barrier to signal the transfer
+    // completion.
+    snrt_global_barrier();
+
+    // Every cluster checks that the data in the destination buffer is correct.
+    if (snrt_is_dm_core()) {
+        uint32_t n_errs = LENGTH;
+        for (uint32_t i = 0; i < LENGTH; i++) {
+            if (buffer_dst[i] == INITIALIZER) n_errs--;
+        }
+        return n_errs;
+    } else
+        return 0;
+}

target/sim/sw/device/runtime/src/mcast.c

@@ -0,0 +1,7 @@
+// Copyright 2023 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Author: Luca Colagrande <colluca@iis.ee.ethz.ch>
+
+extern void dma_broadcast_to_clusters();

target/sim/sw/device/runtime/src/mcast.h

@@ -0,0 +1,12 @@
+// Copyright 2023 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Author: Luca Colagrande <colluca@iis.ee.ethz.ch>
+
+#define BCAST_MASK ((snrt_cluster_num() - 1) << 18)
+
+inline void dma_broadcast_to_clusters(void* dst, void* src, size_t size) {
+    snrt_dma_start_1d_mcast(dst, src, BCAST_MASK, size);
+    snrt_dma_wait_all();
+}

target/sim/sw/device/runtime/src/snrt.c

@@ -9,6 +9,7 @@
 #include "cluster_interrupts.c"
 #include "dma.c"
 #include "global_interrupts.c"
+#include "mcast.c"
 #include "occamy_device.c"
 #include "occamy_memory.c"
 #include "occamy_start.c"

target/sim/sw/device/runtime/src/snrt.h

@@ -29,6 +29,7 @@
 #include "dma.h"
 #include "dump.h"
 #include "global_interrupts.h"
+#include "mcast.h"
 #include "occamy_device.h"
 #include "occamy_memory.h"
 #include "printf.h"

target/sim/sw/host/apps/offload/Makefile

@@ -18,6 +18,7 @@ APP  = offload
 SRCS = src/offload.c
 DEVICE_APPS  = blas/axpy
 DEVICE_APPS += blas/gemm
+DEVICE_APPS += mcast
 
 # Compiler toolchain
 RISCV_CC      = riscv64-unknown-elf-gcc

target/sim/sw/run-mcast-occamy.yaml

@@ -0,0 +1,12 @@
+# Copyright 2023 ETH Zurich and University of Bologna.
+# Licensed under the Apache License, Version 2.0, see LICENSE for details.
+# SPDX-License-Identifier: Apache-2.0
+
+runs:
+  - elf: host/apps/offload/build/offload-axpy.elf
+    cmd: [../../../deps/snitch_cluster/sw/blas/axpy/verify.py, --symbols-bin,
+          ./device/apps/blas/axpy/build/axpy.elf, "${sim_bin}", "${elf}"]
+  - elf: host/apps/offload/build/offload-gemm.elf
+    cmd: [../../../deps/snitch_cluster/sw/blas/gemm/verify.py, --symbols-bin,
+          ./device/apps/blas/gemm/build/gemm.elf, "${sim_bin}", "${elf}"]
+  - elf: host/apps/offload/build/offload-mcast.elf