eml-eda · moyne · Jun 3, 2025 · Jun 3, 2025 · Jun 3, 2025 · Jun 3, 2025
diff --git a/examples/targets/carfield/config/Makefile b/examples/targets/carfield/config/Makefile
@@ -47,7 +47,7 @@ export PULP_SDK_HOME    = $(PULPRT_HOME)
 PULP_APPS = pulp_cluster_runtime
 PULP_APP_SRCS = $(MATCH_COMMON_SRCS) $(MATCH_pulp_cluster_OFFLOAD_SRC) $(PULPRT_HOME)/lib/libc/minimal/io.c $(PULPRT_HOME)/lib/libc/minimal/prf.c 
 
-PULP_CFLAGS = -O3 $(MATCH_INCLUDES) -DCLUSTER_COMPILATION -DARCHI_CLUSTER_NB_PE=8 -I$(PULPRT_HOME)/lib/libc/minimal/include -D__pulp_cluster__
+PULP_CFLAGS = -O3 $(MATCH_INCLUDES) -DCLUSTER_COMPILATION -DARCHI_CLUSTER_NB_PE=8 -I$(PULPRT_HOME)/lib/libc/minimal/include -D__pulp_cluster__ -Dhalf=float16 -D_Float16=float16
 PULPD_ELF_REMOVE_SECTIONS := --remove-section .l1cluster_g --remove-section .bss_l1
 
 -include $(PULP_SDK_HOME)/install/rules/pulp.mk
@@ -61,7 +61,7 @@ build-offload:
 	$(PULPD_RISCV)-objcopy $(PULPD_ELF_REMOVE_SECTIONS) $(BUILD_DIR)/build/pulp_cluster_runtime/pulp_cluster_runtime;
 
 	@echo "Generating objdump..."
-	$(PULPD_RISCV)-objdump -d -S $(BUILD_DIR)/build/pulp_cluster_runtime/pulp_cluster_runtime > $(BUILD_DIR)/build/pulp_cluster_runtime/pulp_cluster_runtime.dump;
+	$(PULPD_RISCV)-objdump -drwCS $(BUILD_DIR)/build/pulp_cluster_runtime/pulp_cluster_runtime > $(BUILD_DIR)/build/pulp_cluster_runtime/pulp_cluster_runtime.dump;
 
 	@echo "Runtime offload build done."
 
@@ -74,7 +74,7 @@ build-offload:
 CAR_SW_DIR  := $(CAR_ROOT)/sw
 CHS_ROOT 	?= $(shell $(BENDER) path cheshire)
 
-CHS_SW_GCC_BINROOT  ?= /usr/pack/riscv-1.0-kgf/riscv64-gcc-11.2.0/bin
+CHS_SW_GCC_BINROOT  ?= /usr/pack/riscv-1.0-kgf/riscv64-gcc-14.2.0/bin
 
 -include $(CHS_ROOT)/cheshire.mk
 CHS_BOOTMODE ?= 0 # default passive bootmode
@@ -105,7 +105,7 @@ $(HOST_LIB): $(HOST_LIB_SRCS_O)
 $(CAR_SW_DIR)/%.car.o: $(CAR_SW_DIR)/%.c
 	$(CHS_SW_CC) $(CAR_SW_INCLUDES) $(CHS_SW_CCFLAGS) -c $< -o $@
 
-HOST_FLAGS := -T$(HOST_LD_SCRIPT) -Wno-pointer-to-int-cast -DIntClustNumCores=8 -g
+HOST_FLAGS := -T$(HOST_LD_SCRIPT) -Wno-pointer-to-int-cast -DIntClustNumCores=8 -Dhalf=_Float16 -g -march=rv64gc_zifencei
 
 
 
@@ -123,5 +123,5 @@ build-host: $(HOST_LIB) build-payload
 	@echo $(HOST_LIB_SRCS_O)
 	$(CHS_SW_CC) $(HOST_INCLUDES) $(MATCH_INCLUDES) $(CHS_SW_LDFLAGS) $(HOST_FLAGS) -o $(BUILD_DIR)/host.elf $(HOST_LIB) $(MATCH_COMMON_SRCS) $(MATCH_HOST_SRC) $(CHS_SW_LIBS)
 	@echo "Generating objdump"
-	@$(CHS_SW_OBJDUMP) -d -S $(BUILD_DIR)/host.elf > $(BUILD_DIR)/host.dump
+	@$(CHS_SW_OBJDUMP) -drwCS $(BUILD_DIR)/host.elf > $(BUILD_DIR)/host.dump
 	@echo "Host build done"
diff --git a/examples/targets/carfield/config/link.ld b/examples/targets/carfield/config/link.ld
@@ -77,11 +77,9 @@ SECTIONS {
     *(.bulk.*)
   } > l2
 
-  .l2_heap : ALIGN(32) {
+  .l2_heap (NOLOAD) : ALIGN(32) {
     __l2_heap_start = .;
-    *(.l2_heap)
-    *(.l2_heap.*)
-    . = ALIGN(32);
+    . = ORIGIN(l2) + LENGTH(l2) - LENGTH(l2_common);
     __l2_heap_end = .;
   } > l2
 

diff --git a/examples/targets/carfield/libs/carfield_lib/include/carfield.h b/examples/targets/carfield/libs/carfield_lib/include/carfield.h
@@ -39,7 +39,7 @@ void carfield_free_ram(void* ext, size_t size);
 extern volatile uint32_t last_completed_node_id;
 extern volatile uint32_t last_task_error_code;
 
-#define GLOBAL_IRQ_ENABLE   0x00001808
+#define GLOBAL_IRQ_ENABLE   (1UL << 3)
 #define EXTERNAL_IRQ_ENABLE 0x00000800
 #define PLIC_BASE_ADDRESS   0x04000000
 

diff --git a/examples/targets/carfield/libs/carfield_lib/include/cluster.h b/examples/targets/carfield/libs/carfield_lib/include/cluster.h
@@ -14,6 +14,10 @@
 #include "pulp.h"
 #include "bench/bench.h"
 #include "pulp_nn/pulp_nn_kernels.h"
+
+typedef float16 fp16;
+typedef fp16 v2f16 __attribute__((vector_size (4)));
+
 #endif
 
 
@@ -76,6 +80,8 @@ void pulp_nn_hoparallel_conv2d_wrapper(MatchCtx* ctx);
 
 void pulp_nn_add_wrapper(MatchCtx* ctx);
 
+void pulp_nn_dense_fp16_wrapper(MatchCtx* ctx);
+
 void pulp_nn_wrapper(MatchCtx* ctx);
 
 #endif // CAR_LIB_CLUSTER_H
diff --git a/examples/targets/carfield/libs/carfield_lib/src/carfield.c b/examples/targets/carfield/libs/carfield_lib/src/carfield.c
@@ -115,10 +115,19 @@ static dif_rv_plic_t plic0;
 void carfield_init_plic() {
     // Reset PLIC
     dif_rv_plic_reset(&plic0);
+
     // Set global interrupt enable in CVA6 csr
-    asm volatile("csrw  mstatus, %0\n" : : "r"(GLOBAL_IRQ_ENABLE));
+    unsigned long mstatus;
+    asm volatile ("csrr %0, mstatus" : "=r"(mstatus));
+    mstatus |= GLOBAL_IRQ_ENABLE;
+    asm volatile ("csrw mstatus, %0" :: "r"(mstatus));
+
     // Set external interrupt enable in CVA6 csr
-    asm volatile("csrw  mie, %0\n" : : "r"(EXTERNAL_IRQ_ENABLE));
+    unsigned long mie;
+    asm volatile ("csrr %0, mie" : "=r"(mie));
+    mie |= EXTERNAL_IRQ_ENABLE;
+    asm volatile ("csrw mie, %0" :: "r"(mie));
+
     // Setup PLIC
     mmio_region_t plic_base_addr = mmio_region_from_addr(PLIC_BASE_ADDRESS);
     dif_result_t t = dif_rv_plic_init(plic_base_addr, &plic0);

diff --git a/examples/targets/carfield/libs/carfield_lib/src/cluster.c b/examples/targets/carfield/libs/carfield_lib/src/cluster.c
@@ -5,6 +5,9 @@
 #include "carfield_lib/mbox.h"
 #include "carfield_lib/utils.h"
 
+#include "pulp_nn/pulp_nn_kernels.h"
+#include "pulp_nn_fp16/pulp_nn_kernels_fp16.h"
+
 //#define CLUSTER_LIB_DEBUG
 #define DEBUG_CALLOC_L1_SCRATCHPAD  0
 #define DEBUG_BLOCKING_DMA          0
@@ -58,6 +61,11 @@ void cluster_sync_cores(MatchCtx* ctx)
 
 void cluster_lib_init(MatchCtx* ctx)
 {
+    #ifdef CLUSTER_LIB_DEBUG
+    for (int i = 0; i < 20000; i++) {
+        asm volatile("fence rw,rw":::"memory");
+    }
+    #endif
     dma_transfer_ = dma_transfer_create();
     #ifdef CLUSTER_LIB_DEBUG
     mini_printf("[PULP] Yo! Cluster is alive! DMA counter is %d\r\n", dma_transfer_);
@@ -719,6 +727,36 @@ void pulp_nn_add_wrapper(MatchCtx* ctx){
     );
 }
 
+
+void pulp_nn_dense_fp16_wrapper(MatchCtx* ctx) {
+    MatchTensor* tensors = ctx->tensors->tensors;
+    int num_ops = ctx->ops->num_ops;
+    int num_tensors = ctx->tensors->num_tensors;
+    int out_ch = tensors[num_tensors-1].tiles[L1_SCRATCHPAD*2+1].size;
+    int inp_ch = tensors[0].tiles[L1_SCRATCHPAD*2+1].size;
+    #ifdef CLUSTER_LIB_DEBUG
+        if(rt_core_id() == 0) {
+            mini_printf("[PULP][KER] pulp_nn_linear_fp16: ");
+            mini_printf("Out. tile (%d,) | ", out_ch);
+            mini_printf("Inp. tile (%d,)\r\n", inp_ch);
+        }
+    #endif
+    pulp_nn_linear_fp16(
+        // activations pt  
+        (float16*)tensors[0].pt, // acts pt
+        // weights pt
+        (float16*)tensors[1].pt, // weights pt
+        // output pt
+        (float16*)tensors[num_tensors-1].pt, // output pt
+        // bias pt
+        num_tensors>4 ? (float16*)NULL : (float16*)tensors[2].pt, // bias pt
+        // dims
+        inp_ch,
+        out_ch
+    );
+}
+
+
 void pulp_nn_wrapper(MatchCtx* ctx){
 
     switch(ctx->pattern_name){
@@ -728,9 +766,9 @@ void pulp_nn_wrapper(MatchCtx* ctx){
         case conv2d:
             pulp_nn_hoparallel_conv2d_wrapper(ctx);
             break;
-        case dense_out:
-            pulp_nn_dense_out_int_wrapper(ctx);
-            break;
+        //case dense_out:
+        //    pulp_nn_dense_out_int_wrapper(ctx);
+        //    break;
         // case pulp_nn_dw_conv2d_less_4_pattern:
         //     pi_team_offload_preset(pulp_nn_dw_conv2d_less_4_wrapper, ctx);
         //     break;
@@ -743,6 +781,8 @@ void pulp_nn_wrapper(MatchCtx* ctx){
         case add_requant:
             pulp_nn_add_wrapper(ctx);
             break;
+        case dense_fp16:
+            pulp_nn_dense_fp16_wrapper(ctx);
         default:
             break;
     }
@@ -807,4 +847,14 @@ uint32_t cluster_timer_stop() {
 }
 
 
+
+double __attribute__((weak)) __extendhfdf2(float16 val)
+{
+  float res;
+  __asm__ __volatile__ ("fcvt.s.h %0, %1": "=f"(res): "f"(val) :);
+  return (double) res;
+}
+
+
+
 #endif