Multi-Scalar Configuration

.gitlab-ci.yml

@@ -16,7 +16,9 @@ variables:
   python3: 'python3'
   # Config to build and test
   CI_CONFIG: 'cachepool_fpu_4g'
-  SW_PREFIX: 'test-cachepool-'
+  CI_CONFIG_DUAL: 'cachepool_dual_fpu_4g'
+  CI_CONFIG_16G: 'cachepool_fpu_16g'
+  SW_PREFIX: 'test-'
 
 default:
   tags: [shared]
@@ -60,6 +62,63 @@ build:
       - hardware/deps/dram_rtl_sim/dramsys_lib/DRAMSys/build/lib/
       - hardware/deps/dram_rtl_sim/dramsys_lib/DRAMSys/configs/
 
+# ---------------------------------------------------------------------------
+# Build stage: same as `build`, for the dual-scalar-per-CC config. Only the
+# kernels adapted for dual-scalar (see `test-dual`) are exercised against it.
+# ---------------------------------------------------------------------------
+build-dual:
+  stage: build
+  script:
+    - echo "Using CC=$CC"
+    - echo "Using CXX=$CXX"
+    - test -x "$CC"
+    - test -x "$CXX"
+    - source iis-env.sh
+    - make init
+    - make dram-build
+    - make clean generate update-floonoc bootrom vsim sw config=$CI_CONFIG_DUAL DEBUG=0 noc_profiling=0
+  artifacts:
+    when: always
+    expire_in: 2h
+    paths:
+      - sim/work/
+      - sim/bin/cachepool_cluster.vsim
+      - sim/bin/cachepool_cluster.vsim.gui
+      - sim/work-dpi/
+      - software/build/CachePoolTests/
+      - hardware/deps/dram_rtl_sim/dramsys_lib/DRAMSys/build/lib/
+      - hardware/deps/dram_rtl_sim/dramsys_lib/DRAMSys/configs/
+
+# ---------------------------------------------------------------------------
+# Build stage: same as `build`, for the largest available config (16 groups).
+# Manual (see test-16g): slow, so not run automatically on every push.
+# ---------------------------------------------------------------------------
+build-16g:
+  stage: build
+  when: manual
+  allow_failure: true
+  timeout: 4h
+  script:
+    - echo "Using CC=$CC"
+    - echo "Using CXX=$CXX"
+    - test -x "$CC"
+    - test -x "$CXX"
+    - source iis-env.sh
+    - make init
+    - make dram-build
+    - make clean all config=$CI_CONFIG_16G DEBUG=0 noc_profiling=0
+  artifacts:
+    when: always
+    expire_in: 2h
+    paths:
+      - sim/work/
+      - sim/bin/cachepool_cluster.vsim
+      - sim/bin/cachepool_cluster.vsim.gui
+      - sim/work-dpi/
+      - software/build/CachePoolTests/
+      - hardware/deps/dram_rtl_sim/dramsys_lib/DRAMSys/build/lib/
+      - hardware/deps/dram_rtl_sim/dramsys_lib/DRAMSys/configs/
+
 # ---------------------------------------------------------------------------
 # Test stage: run each kernel in parallel on a separate runner.
 # Each job downloads the build artifacts, runs one simulation, and checks
@@ -90,9 +149,69 @@ test:
     - python3 util/auto-benchmark/check-ci.py test_${KERNEL}.log
   artifacts:
     when: always
-    expire_in: 1 week
+    expire_in: 3 days
     paths:
       # Full simulation log
       - test_*.log
       # Performance-monitor trace files written by the simulator
       - sim/bin/logs/
+
+# ---------------------------------------------------------------------------
+# Test stage: dual-scalar-per-CC config. Limited to the kernels adapted and
+# confirmed working under this config (host0-only Spatz access via
+# snrt_cluster_vpu_*/snrt_cluster_is_primary); other tests are not yet
+# verified for dual-scalar and are left to the default-config matrix above.
+# ---------------------------------------------------------------------------
+test-dual:
+  stage: test
+  needs: [build-dual]
+  parallel:
+    matrix:
+      - KERNEL:
+          - fdotp-32b_M32768
+          - gemv_M512_N128_K32
+          - fmatmul-32b_M64_N64_K64
+          - fft-32b_M1024_N16
+  script:
+    - mkdir -p sim/bin/logs sim/bin/logs/core sim/bin/logs/noc sim/bin/logs/others
+    - chmod +x sim/bin/cachepool_cluster.vsim
+    - BIN="${SW_PREFIX}${KERNEL}"
+    - sim/bin/cachepool_cluster.vsim software/build/CachePoolTests/$BIN 2>&1 | tee test_dual_${KERNEL}.log
+    - python3 util/auto-benchmark/check-ci.py test_dual_${KERNEL}.log
+  artifacts:
+    when: always
+    expire_in: 3 days
+    paths:
+      - test_dual_*.log
+      - sim/bin/logs/
+
+# ---------------------------------------------------------------------------
+# Test stage: largest available config (16 groups), largest available size
+# per kernel. Runs once build-16g (manual) is triggered and succeeds -- see
+# build-16g.
+# ---------------------------------------------------------------------------
+test-16g:
+  stage: test
+  needs: [build-16g]
+  timeout: 4h
+  parallel:
+    matrix:
+      - KERNEL:
+          - fdotp-32b_M65536
+          - gemv_M1024_N128_K32
+          - fmatmul-32b_M128_N128_K128
+          - fft-32b_M1024_N16
+          - byte-enable
+          - 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}"
+    - 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:
+    when: always
+    expire_in: 3 days
+    paths:
+      - test_16g_*.log
+      - sim/bin/logs/

Bender.lock

@@ -150,7 +150,7 @@ packages:
     dependencies:
     - tech_cells_generic
   spatz:
-    revision: f6ea19fb26c8bf992ae51a3e4acd3f2a36c6659f
+    revision: f4741b0a5bcc49103bd8a3db8d51fdf76d240df7
     version: null
     source:
       Git: https://github.com/pulp-platform/spatz.git

Bender.yml

@@ -42,6 +42,9 @@ sources:
   - hardware/generated/floo_cachepool_noc_pkg.sv
   - hardware/src/cachepool_pkg.sv
   - hardware/src/cachepool_cc.sv
+  - hardware/src/cachepool_spatz_lock.sv
+  - hardware/src/acc_mux.sv
+  - hardware/src/cachepool_cc_dual.sv
   # Barrier
   - hardware/src/cachepool_tile_barrier.sv
   - hardware/src/cachepool_cluster_barrier.sv

Makefile

@@ -190,12 +190,12 @@ $(info FLOO_DIR: $(FLOO_DIR))
 # Generates the sources for FlooNoC
 .PHONY: update-floonoc install-floogen clean-floonoc
 install-floogen:
-	pip install -e $(FLOO_DIR) --quiet
+	$(PYTHON) -m pip install -e $(FLOO_DIR) --quiet
 
 update-floonoc: $(FLOO_NOC)
 $(FLOO_NOC): $(FLOO_CFG)
 	mkdir -p $(FLOO_GEN_OUTDIR)
-	PATH="$(HOME)/.local/bin:$(PATH)" floogen pkg -c $(FLOO_CFG) -o $(FLOO_GEN_OUTDIR) --no-format
+	floogen pkg -c $(FLOO_CFG) -o $(FLOO_GEN_OUTDIR) --no-format
 
 clean-floonoc:
 	rm -f $(FLOO_NOC)
@@ -213,7 +213,7 @@ dram-build: $(DRAMSYS_PATH)/README.md dram-config
 	if [ ! -d "build" ]; then \
 		mkdir build && cd build; \
 		$(CMAKE) -DCMAKE_CXX_FLAGS=-fPIC -DCMAKE_C_FLAGS=-fPIC -D DRAMSYS_WITH_DRAMPOWER=ON .. ; \
-		make -j; \
+		make -j4; \
 	fi
 
 $(DRAMSYS_PATH)/README.md: dram-init
@@ -242,7 +242,7 @@ dram-clean:
 	fi
 
 dram-init:
-	make -C ${DRAMSYS_DIR} -j8 dramsys CXX=$(CXX) CC=$(CC)
+	make -C ${DRAMSYS_DIR} -j4 dramsys CXX=$(CXX) CC=$(CC)
 
 ############
 # Modelsim #
@@ -298,6 +298,10 @@ VLOG_DEFS += -DSPATZ_NUM_FPU=$(spatz_num_fpu)
 VLOG_DEFS += -DSPATZ_NUM_IPU=$(spatz_num_ipu)
 VLOG_DEFS += -DSPATZ_MAX_TRANS=$(spatz_max_trans)
 VLOG_DEFS += -DSNITCH_MAX_TRANS=$(snitch_max_trans)
+VLOG_DEFS += -DNUM_SCALAR_PER_CC=$(num_scalar_per_core)
+ifeq ($(num_scalar_per_core),2)
+VLOG_DEFS += -DCACHEPOOL_DUAL_CC
+endif
 VLOG_DEFS += -DLG_PORT_PER_CORE=$(num_lg_ports_per_core)
 VLOG_DEFS += -DRG_PORT_PER_CORE=$(num_rg_ports_per_core)
 VLOG_DEFS += -DNOC_PORT_PER_TILE=$(num_noc_ports_per_tile)
@@ -351,24 +355,28 @@ 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
+
 # Auto-discover every test that has a script/gen_data.py.
 # Convention: script/data_<params>.json → data/data_<params>.h
 # Adding a new test or a new data variant needs no Makefile changes:
 #   - new test:    drop gen_data.py + data_<params>.json into its script/ dir
 #   - new variant: add data_<params>.json to an existing test's script/ dir
-DATA_TESTS := $(patsubst $(TESTS_DIR)/%/script/gen_data.py,%, \
-                $(wildcard $(TESTS_DIR)/*/script/gen_data.py))
+DATA_TESTS := $(patsubst $(SOFTWARE_DIR)/%/script/gen_data.py,%, \
+                $(foreach root,$(TEST_ROOTS), \
+                  $(wildcard $(SOFTWARE_DIR)/$(root)/*/script/gen_data.py)))
 
 define gen_data_rules
-$(1)_DATA    := $$(patsubst $(TESTS_DIR)/$(1)/script/data_%.json, \
-                             $(TESTS_DIR)/$(1)/data/data_%.h, \
-                             $$(wildcard $(TESTS_DIR)/$(1)/script/data_*.json))
+$(1)_DATA    := $$(patsubst $(SOFTWARE_DIR)/$(1)/script/data_%.json, \
+                             $(SOFTWARE_DIR)/$(1)/data/data_%.h, \
+                             $$(wildcard $(SOFTWARE_DIR)/$(1)/script/data_*.json))
 ALL_GEN_DATA += $$($(1)_DATA)
 
-$(TESTS_DIR)/$(1)/data/data_%.h: \
-    $(TESTS_DIR)/$(1)/script/data_%.json \
-    $(TESTS_DIR)/$(1)/script/gen_data.py
-	$$(PYTHON) $(TESTS_DIR)/$(1)/script/gen_data.py -c $$<
+$(SOFTWARE_DIR)/$(1)/data/data_%.h: \
+    $(SOFTWARE_DIR)/$(1)/script/data_%.json \
+    $(SOFTWARE_DIR)/$(1)/script/gen_data.py
+	$$(PYTHON) $(SOFTWARE_DIR)/$(1)/script/gen_data.py -c $$<
 endef
 
 $(foreach test,$(DATA_TESTS),$(eval $(call gen_data_rules,$(test))))
@@ -384,23 +392,32 @@ $(BANDWIDTH_DATA): $(TESTS_DIR)/bandwidth/script/data.json \
 .PHONY: gen-data
 gen-data: $(ALL_GEN_DATA)
 
+# Clean targets, used to delete generated files
 .PHONY: clean.data
 clean.data:
 	rm -f $(ALL_GEN_DATA)
 
 .PHONY: clean.sw
-clean.sw:
+clean.sw: clean.data
 	rm -rf ${SOFTWARE_DIR}/build
 
-.PHONY: clean
-clean: clean.sw clean.vsim clean.data
+.PHONY: clean.generate
+clean.generate:
 	rm -rf $(HJSON_OUT) $(BOOTROM_DIR)/bootdata.cc \
 	                    $(BOOTROM_DIR)/bootdata_bootrom.cc \
 	                    $(BOOTROM_DIR)/bootrom.sv \
 	                    $(BOOTROM_DIR)/bootrom.dump \
 	                    $(BOOTROM_DIR)/bootrom.elf \
+	                    $(CACHEPOOL_DIR)/wlf* \
+	                    $(CACHEPOOL_DIR)/noc_profiling/* \
 	                    $(SNRT_BOOTINFO_H)
 
+.PHONY: clean.hw
+clean.hw: clean.vsim clean.generate clean-floonoc
+
+.PHONY: clean
+clean: clean.hw clean.sw
+
 # Common CMake flags shared by sw and vsim targets.
 # vsim appends -DSNITCH_SIMULATOR to point tests at the compiled binary.
 SW_CMAKE_FLAGS = \
@@ -411,7 +428,8 @@ SW_CMAKE_FLAGS = \
   -DLLVM_PATH=${LLVM_INSTALL_DIR} \
   -DGCC_PATH=${GCC_INSTALL_DIR} \
   -DPYTHON=${PYTHON} \
-  -DBUILD_TESTS=ON
+  -DBUILD_TESTS=ON \
+  -DNUM_SCALAR_PER_CORE=$(num_scalar_per_core)
 
 .PHONY: sw
 sw: generate bootrom gen-data
@@ -428,8 +446,11 @@ vsim: generate bootrom dpi ${SIMBIN_DIR}/cachepool_cluster.vsim
 ############
 
 # Just a shortcut to build everything
+.PHONY: hw
+hw: generate bootrom vsim
+
 .PHONY: all
-all: generate bootrom vsim sw
+all: hw sw
 
 ########
 # Lint #
@@ -463,9 +484,7 @@ avg-log:
 # NoC traffic visualization frontend (fork of https://github.com/ueqri/vis4mesh
 # with CachePool-specific fixes/features). Consumes the Vis4Mesh dataset
 # directories produced by util/scripts/noc_profiling_to_vis4mesh.py from
-# hardware/tb/cachepool_noc_profiling.sv's per-cycle NoC logs. Fetched as a
-# plain pinned clone (same pattern as toolchain.mk's riscv-gnu-toolchain/
-# llvm-project targets), not a git submodule.
+# hardware/tb/cachepool_noc_profiling.sv's per-cycle NoC logs.
 NPM             ?= npm
 VIS4MESH_DIR    ?= ${CACHEPOOL_DIR}/util/vis4mesh
 VIS4MESH_REPO   ?= https://github.com/DiyouS/vis4mesh.git
@@ -531,13 +550,18 @@ help:
 	@echo "SW Build:"
 	@echo ""
 	@echo "*sw*:             build software (generate + bootrom + cmake); overwrites previous build"
-	@echo "*clean.sw*:       remove the software build directory"
+	@echo "*clean.sw*:       remove the software build directory (also runs clean.data)"
 	@echo ""
 	@echo "Simulation:"
 	@echo ""
 	@echo "*vsim*:           build hardware for QuestaSim simulation (use 'sw' to build software separately)"
+	@echo "*hw*:             shortcut for generate + bootrom + vsim"
+	@echo "*all*:            shortcut for hw + sw"
 	@echo "*clean.vsim*:     remove the hardware simulation build [from sim/sim.mk]"
-	@echo "*clean*:          remove SW build, vsim build, and all generated HW files"
+	@echo "*clean.data*:     remove generated data files (gen-data outputs)"
+	@echo "*clean.generate*: remove generated HJSON/bootrom/wlf/noc_profiling files [from 'generate'/'bootrom']"
+	@echo "*clean.hw*:       clean.vsim + clean.generate"
+	@echo "*clean*:          clean.hw + clean.sw (remove SW build, vsim build, and all generated HW files)"
 	@echo ""
 	@echo "Lint:"
 	@echo ""