Hierarchical peripheral configuration and a narrow peripheral mesh

.gitlab-ci.yml

@@ -19,6 +19,7 @@ variables:
   CI_CONFIG_DUAL: 'cachepool_dual_fpu_4g'
   CI_CONFIG_16G: 'cachepool_fpu_16g'
   CI_CONFIG_DUAL_16G: 'cachepool_dual_fpu_16g'
+  # Binary prefix of the tests whose KERNEL has no group prefix (rlc-, ctrl-)
   SW_PREFIX: 'test-'
 
 default:
@@ -169,16 +170,16 @@ test:
           - fmatmul-32b_M64_N64_K64
           - fft-32b_M1024_N16
           - byte-enable
-          - multi_producer_single_consumer_double_linked_list_M1_N1350_K10
-          - multi_producer_single_consumer_double_linked_list_M1_N1350_K4_P2_C2_am_G1024_tbchk
-          - multi_producer_single_consumer_double_linked_list_P1_C3_ul_ci
+          - rlc-multi_producer_single_consumer_double_linked_list_M1_N1350_K10
+          - rlc-multi_producer_single_consumer_double_linked_list_M1_N1350_K4_P2_C2_am_G1024_tbchk
+          - rlc-multi_producer_single_consumer_double_linked_list_P1_C3_ul_ci
   script:
     # The vsim script writes a .rtlbinary marker here; monitor logs need the
     # core/noc/others subfolders. Build artifacts exclude sim/bin/logs/, so
     # these must be recreated in the test job too.
     - mkdir -p sim/bin/logs sim/bin/logs/core sim/bin/logs/noc sim/bin/logs/others
     - chmod +x sim/bin/cachepool_cluster.vsim
-    - BIN="${SW_PREFIX}${KERNEL}"
+    - case "$KERNEL" in rlc-*|ctrl-*) BIN="$KERNEL" ;; *) BIN="${SW_PREFIX}${KERNEL}" ;; esac
     - sim/bin/cachepool_cluster.vsim software/build/CachePoolTests/$BIN 2>&1 | tee test_${KERNEL}.log
     - python3 util/auto-benchmark/check-ci.py test_${KERNEL}.log
   artifacts:
@@ -209,7 +210,7 @@ test-dual:
   script:
     - mkdir -p sim/bin/logs sim/bin/logs/core sim/bin/logs/noc sim/bin/logs/others
     - chmod +x sim/bin/cachepool_cluster.vsim
-    - BIN="${SW_PREFIX}${KERNEL}"
+    - case "$KERNEL" in rlc-*|ctrl-*) BIN="$KERNEL" ;; *) BIN="${SW_PREFIX}${KERNEL}" ;; esac
     - sim/bin/cachepool_cluster.vsim software/build/CachePoolTests/$BIN 2>&1 | tee test_dual_${KERNEL}.log
     - python3 util/auto-benchmark/check-ci.py test_dual_${KERNEL}.log
   artifacts:
@@ -235,11 +236,11 @@ test-16g:
           - fmatmul-32b_M128_N128_K128
           - fft-32b_M1024_N16
           - byte-enable
-          - multi_producer_single_consumer_double_linked_list_M1_N1350_K10
+          - rlc-multi_producer_single_consumer_double_linked_list_M1_N1350_K10
   script:
     - mkdir -p sim/bin/logs sim/bin/logs/core sim/bin/logs/noc sim/bin/logs/others
     - chmod +x sim/bin/cachepool_cluster.vsim
-    - BIN="${SW_PREFIX}${KERNEL}"
+    - case "$KERNEL" in rlc-*|ctrl-*) BIN="$KERNEL" ;; *) BIN="${SW_PREFIX}${KERNEL}" ;; esac
     - sim/bin/cachepool_cluster.vsim software/build/CachePoolTests/$BIN 2>&1 | tee test_16g_${KERNEL}.log
     - python3 util/auto-benchmark/check-ci.py test_16g_${KERNEL}.log
   artifacts:
@@ -268,7 +269,7 @@ test-dual-16g:
   script:
     - mkdir -p sim/bin/logs sim/bin/logs/core sim/bin/logs/noc sim/bin/logs/others
     - chmod +x sim/bin/cachepool_cluster.vsim
-    - BIN="${SW_PREFIX}${KERNEL}"
+    - case "$KERNEL" in rlc-*|ctrl-*) BIN="$KERNEL" ;; *) BIN="${SW_PREFIX}${KERNEL}" ;; esac
     - sim/bin/cachepool_cluster.vsim software/build/CachePoolTests/$BIN 2>&1 | tee test_dual_16g_${KERNEL}.log
     - python3 util/auto-benchmark/check-ci.py test_dual_16g_${KERNEL}.log
   artifacts:

Bender.lock

@@ -63,7 +63,7 @@ packages:
     dependencies:
     - axi
   floo_noc:
-    revision: 8413d9b044a51f958a7d4ef261e2b01e73db4544
+    revision: 7e5104cd0875c3ded16d53366caaef2dc465c49d
     version: null
     source:
       Git: https://github.com/pulp-platform/FlooNoC.git
@@ -150,7 +150,7 @@ packages:
     dependencies:
     - tech_cells_generic
   spatz:
-    revision: 6acbd2e0cb9131f870209d03ffb85c46fbd2ea9c
+    revision: f574402d29ed54f2dbede81099419889575c4584
     version: null
     source:
       Git: https://github.com/pulp-platform/spatz.git

Bender.yml

@@ -27,20 +27,22 @@ export_include_dirs:
 sources:
   # Level 0
   - hardware/src/reqrsp_to_reg.sv
-  - hardware/src/reqrsp_w2n_converter.sv
   - hardware/src/reqrsp_xbar.sv
   - hardware/src/tcdm_cache_interco.sv
   - hardware/src/tcdm_id_remapper.sv
   - hardware/src/spatz_cache_amo.sv
   # Memory-mapped register
   - hardware/cachepool_peripheral/cachepool_peripheral_reg_pkg.sv
   - hardware/cachepool_peripheral/cachepool_peripheral_reg_top.sv
-  - hardware/cachepool_peripheral/cachepool_peripheral.sv
   # Bootrom
   - hardware/bootrom/bootrom.sv
   # Level 1
   - hardware/generated/floo_cachepool_noc_pkg.sv
   - hardware/src/cachepool_pkg.sv
+  - hardware/cachepool_peripheral/cachepool_peripheral.sv
+  # Group configuration registers (cluster sender + group-level register)
+  - hardware/src/cachepool_group_cfg_sender.sv
+  - hardware/src/cachepool_group_cfg.sv
   - hardware/src/cachepool_acc_cut.sv
   - hardware/src/cachepool_cc.sv
   - hardware/src/cachepool_spatz_lock.sv

Makefile

@@ -309,6 +309,15 @@ CACHEPOOL_DEFS += -DDATA_WIDTH=$(data_width)
 CACHEPOOL_DEFS += -DADDR_WIDTH=$(addr_width)
 CACHEPOOL_DEFS += -DNUM_BARRIER_SLOTS=$(num_barrier_slots)
 
+# Group-level configuration registers and the reset values shared with the cluster peripheral
+CACHEPOOL_DEFS += -DGROUP_CFG_BASE=$(group_cfg_base)
+CACHEPOOL_DEFS += -DGROUP_CFG_STRIDE=$(group_cfg_stride)
+CACHEPOOL_DEFS += -DL1D_XBAR_OFFSET_RST=$(l1d_xbar_offset_rst)
+CACHEPOOL_DEFS += -DL1D_PRIVATE_RST=$(l1d_private_rst)
+CACHEPOOL_DEFS += -DL1D_PRIVATE_ADDR_RST=$(l1d_private_addr_rst)
+CACHEPOOL_DEFS += -DICACHE_PREFETCH_RST=$(icache_prefetch_rst)
+CACHEPOOL_DEFS += -DBOOTROM_DATA_WIDTH=$(bootrom_data_width)
+
 # Tile configuration
 CACHEPOOL_DEFS += -DREFILL_DATA_WIDTH=$(refill_data_width)
 
@@ -395,7 +404,7 @@ include sim/sim.mk
 TESTS_DIR := $(SOFTWARE_DIR)/tests
 
 # Test-group roots (relative to SOFTWARE_DIR) scanned for data-generated tests.
-TEST_ROOTS := cache sync kernels/fp kernels/int rlc tests
+TEST_ROOTS := cache sync kernels/fp kernels/int rlc control tests
 
 # Auto-discover every test that has a script/gen_data.py.
 # Convention: script/data_<params>.json → data/data_<params>.h
@@ -468,7 +477,11 @@ SW_CMAKE_FLAGS = \
   -DGCC_PATH=${GCC_INSTALL_DIR} \
   -DPYTHON=${PYTHON} \
   -DBUILD_TESTS=ON \
-  -DNUM_SCALAR_PER_CORE=$(num_scalar_per_core)
+  -DNUM_SCALAR_PER_CORE=$(num_scalar_per_core) \
+  -DL2_CHANNEL=$(l2_channel) \
+  -DL2_INTERLEAVE=$(l2_interleave) \
+  -DL2_BANK_WIDTH=$(l2_bank_width) \
+  -DL1D_CACHELINE_WIDTH=$(l1d_cacheline_width)
 
 .PHONY: sw
 sw: generate bootrom gen-data

README.md

@@ -302,7 +302,8 @@ The default boundary is `0xA000_0000`. This means data in `.pdcp_src` (at `0xA00
 
 Flush instructions are issued through the cluster peripheral and dispatched to cache controllers via a two-level controller:
 
-- **Cluster level** (`cachepool_peripheral`): holds the instruction and a per-tile one-hot tile-select mask. Issues the instruction to all tiles simultaneously and tracks completion at tile granularity.
+- **Cluster level** (`cachepool_peripheral`): holds the instruction and a per-tile one-hot tile-select mask. On `L1D_INSN_COMMIT` it sends the instruction, with each group's slice of the mask, to the group-level configuration register of every group over the L2 mesh (`cachepool_group_cfg_sender`), and stays busy until every group has answered.
+- **Group level** (`cachepool_group_cfg`): starts the instruction in its selected tiles and answers only once all of them have finished.
 - **Tile level** (`cachepool_tile`): receives the instruction, determines which of its local cache controllers to activate based on the partition field, and returns a single ready pulse to the cluster controller when all targeted controllers finish.
 
 ### Flush instruction encoding
@@ -342,7 +343,7 @@ l1d_xbar_config(uint32_t offset);
 
 Lower-level, single-shot primitives (`l1d_flush()`, `l1d_private_flush()`, `l1d_shared_flush()`, `l1d_commit()`, `l1d_wait()`) exist for building custom sequences but issue only the register write — the caller is responsible for cluster-wide fence/barrier synchronization around them.
 
-`l1d_addr()` has no hardware completion status bit (unlike flush, which polls `L1D_FLUSH_STATUS` via `l1d_wait()`); its peripheral register write is posted over the interconnect, so it bridges the gap with a fixed-cycle delay on core 0 before the final barrier.
+Only flush/invalidate instructions are committed. The configuration registers (`L1D_PRIVATE`, `L1D_ADDR`, `XBAR_OFFSET`, `ICACHE_PREFETCH_ENABLE`) need no commit: when a value changes, the peripheral sends it to every group on its own. `L1D_FLUSH_STATUS` stays busy until every group has applied it, so `l1d_part()`, `l1d_addr()` and `l1d_xbar_config()` end with `l1d_wait()`. Writing the value a register already holds sends nothing.
 
 > Changing the partition mode, boundary address, or crossbar offset while the cache contains valid data requires a flush first — `l1d_xbar_config()` and `l1d_part()` flush internally; `l1d_addr()` does not (call a cluster flush before it if needed).
 

config/cachepool.hjson.tmpl

@@ -5,6 +5,7 @@
   cluster: {
     mempool: 0,
     boot_addr: ${boot_addr},               // e.g., 4096 (0x1000)
+    bootrom_data_width: ${bootrom_data_width},
     cluster_base_addr: ${stack_addr},      // 0xBFFF_F800 typically
     cluster_base_offset: 0,
     cluster_base_hartid: 0,

config/config.mk

@@ -74,8 +74,9 @@ data_width ?= 32
 # Core addrwidth
 addr_width ?= 32
 
-# Bootrom datawidth
-bootrom_data_width ?= 32
+# Bootrom datawidth: must equal the L2 refill width, the bootrom is only read by full-line
+# iCache refills on the wide L2 path
+bootrom_data_width ?= $(refill_data_width)
 
 
 ######################
@@ -227,3 +228,21 @@ boot_addr ?= 4096
 
 # UART address (32'hC001_0000)
 uart_addr ?= 3221291008
+
+# Group-level configuration registers, written by the cluster peripheral over the
+# L2 mesh. Group g owns one word at group_cfg_base + g * group_cfg_stride.
+# Base (32'hD000_0000)
+group_cfg_base ?= 3489660928
+# Stride per group (32'h100)
+group_cfg_stride ?= 256
+
+# Reset values of the L1D/xbar configuration, shared by the cluster peripheral
+# registers and the group-level copies so both always come out of reset equal.
+# Cache xbar offset (bit position selecting the cache bank)
+l1d_xbar_offset_rst ?= 14
+# Private partition size and fold (L1D_PRIVATE)
+l1d_private_rst ?= 0
+# Start address of the private L1D partition (32'hA000_0000)
+l1d_private_addr_rst ?= 2684354560
+# Instruction cache prefetching enable
+icache_prefetch_rst ?= 1

hardware/bootrom/bootrom.S

@@ -9,10 +9,13 @@
   .global BOOTDATA
 
 _start:
-  la      t1, exception
+  // lla, not la: la loads the address from the GOT, a 32-bit data read of the bootrom.
+  // The bootrom is only served on the wide L2 path (instruction refills), so it must
+  // never see data loads.
+  lla     t1, exception
   csrw    mtvec, t1
   csrr    a0, mhartid
-  la      a1, BOOTDATA
+  lla     a1, BOOTDATA
 
   // Enable MSI interrupt
   csrw mie, 0xF