#17307: Move ncrisc copy into ncrisc firmware

By moving the NCRISC kernel copy into the NCRISC firmware (on wormhole) we can do it in parallel with setting up CBs.

tt_metal/hw/firmware/src/brisc.cc

@@ -42,7 +42,7 @@ constexpr uint32_t RISCV_IC_TRISC_ALL_MASK = RISCV_IC_TRISC0_MASK | RISCV_IC_TRI
 
 #define NCRISC_FIRMWARE_IN_IRAM (defined(ARCH_GRAYSKULL))
 
-#ifdef NCRISC_HAS_IRAM
+#if NCRISC_FIRMWARE_IN_IRAM
 constexpr uint32_t num_cbs_to_early_init = 4;  // safe small number to overlap w/ ncrisc copy
 #else
 constexpr uint32_t num_cbs_to_early_init = 0;
@@ -183,7 +183,7 @@ void set_deassert_addresses() {
 }
 
 void l1_to_ncrisc_iram_copy(uint32_t src_addr, uint16_t size, uint32_t address_offset = 0) {
-#ifdef NCRISC_HAS_IRAM
+#if NCRISC_FIRMWARE_IN_IRAM
     // Always copy ncrisc even if its size is 0 (save branch)...
     // Copy NCRISC firmware from L1 to local IRAM using tensix DMA
     tdma_xmov(
@@ -196,7 +196,7 @@ void l1_to_ncrisc_iram_copy(uint32_t src_addr, uint16_t size, uint32_t address_o
 }
 
 void l1_to_ncrisc_iram_copy_wait() {
-#ifdef NCRISC_HAS_IRAM
+#if NCRISC_FIRMWARE_IN_IRAM
     // Wait for DMA to finish
     wait_tdma_movers_done(RISCV_TDMA_STATUS_FLAG_MOVER0_BUSY_MASK);
 #endif

tt_metal/hw/firmware/src/ncrisc.cc

@@ -14,6 +14,7 @@
 #include "risc_attribs.h"
 #include "circular_buffer.h"
 #include "circular_buffer_init.h"
+#include "tdma_xmov.h"
 
 #include "debug/waypoint.h"
 #include "debug/dprint.h"
@@ -84,6 +85,20 @@ inline __attribute__((always_inline)) void signal_ncrisc_completion() {
 #endif
 }
 
+#if defined(ARCH_WORMHOLE)
+#define MEM_MOVER_VIEW_IRAM_BASE_ADDR (0x4 << 12)
+void l1_to_ncrisc_iram_copy(uint32_t src_addr, uint16_t size, uint32_t address_offset = 0) {
+    // Always copy ncrisc even if its size is 0 (save branch)...
+    // Copy NCRISC firmware from L1 to local IRAM using tensix DMA
+    tdma_xmov(TDMA_MOVER0, src_addr, MEM_MOVER_VIEW_IRAM_BASE_ADDR + address_offset, size, XMOV_L1_TO_L0);
+}
+
+void l1_to_ncrisc_iram_copy_wait() {
+    // Wait for DMA to finish
+    wait_tdma_movers_done(RISCV_TDMA_STATUS_FLAG_MOVER0_BUSY_MASK);
+}
+#endif
+
 int main(int argc, char *argv[]) {
     configure_l1_data_cache();
     DIRTY_STACK_MEMORY();
@@ -109,17 +124,26 @@ int main(int argc, char *argv[]) {
         launch_msg_t* launch_msg = &(mailboxes->launch[launch_msg_rd_ptr]);
 
         uint32_t kernel_config_base = firmware_config_init(mailboxes, ProgrammableCoreType::TENSIX, DISPATCH_CLASS_TENSIX_DM1);
+        int index = static_cast<std::underlying_type<TensixProcessorTypes>::type>(TensixProcessorTypes::DM1);
+
+#if defined(ARCH_WORMHOLE)
+        uint32_t ncrisc_kernel_src_address = kernel_config_base + launch_msg->kernel_config.kernel_text_offset[index];
+        l1_to_ncrisc_iram_copy(ncrisc_kernel_src_address >> 4, launch_msg->kernel_config.ncrisc_kernel_size16, 0);
+#endif
         uint32_t tt_l1_ptr* cb_l1_base =
             (uint32_t tt_l1_ptr*)(kernel_config_base + launch_msg->kernel_config.local_cb_offset);
         uint32_t end_cb_index = launch_msg->kernel_config.max_local_cb_end_index;
         setup_local_cb_read_write_interfaces(cb_l1_base, 0, end_cb_index, true, true, false);
 
+#if defined(ARCH_WORMHOLE)
+        l1_to_ncrisc_iram_copy_wait();
+#endif
+
         cb_l1_base = (uint32_t tt_l1_ptr*)(kernel_config_base + launch_msg->kernel_config.remote_cb_offset);
         end_cb_index = launch_msg->kernel_config.min_remote_cb_start_index;
         experimental::setup_remote_cb_interfaces(cb_l1_base, end_cb_index);
         WAYPOINT("R");
 
-        int index = static_cast<std::underlying_type<TensixProcessorTypes>::type>(TensixProcessorTypes::DM1);
         void (*kernel_address)(uint32_t) = (void (*)(uint32_t))
             (kernel_config_base + launch_msg->kernel_config.kernel_text_offset[index]);
 #ifdef ARCH_BLACKHOLE

tt_metal/jit_build/build.cpp

@@ -219,18 +219,17 @@ void JitBuildState::finish_init() {
     std::string build_dir =
         llrt::RunTimeOptions::get_instance().get_root_dir() + "runtime/hw/lib/" + get_alias(env_.arch_) + "/";
     if (this->is_fw_) {
-        if (this->target_name_ == "brisc") {
+        if (this->target_name_ == "brisc" and this->env_.arch_ == tt::ARCH::GRAYSKULL) {
             this->link_objs_ += build_dir + "tdma_xmov.o ";
         }
         if (this->target_name_ != "erisc") {
             this->link_objs_ += build_dir + "tmu-crt0.o ";
         }
-        if (this->target_name_ == "ncrisc" and
-            ((this->env_.arch_ == tt::ARCH::GRAYSKULL or this->env_.arch_ == tt::ARCH::WORMHOLE_B0))) {
+        if (this->target_name_ == "ncrisc" and this->env_.arch_ == tt::ARCH::GRAYSKULL) {
             this->link_objs_ += build_dir + "ncrisc-halt.o ";
         }
         if (this->target_name_ == "ncrisc" and this->env_.arch_ == tt::ARCH::WORMHOLE_B0) {
-            this->link_objs_ += build_dir + "ncrisc-halt-wormhole.o ";
+            this->link_objs_ += build_dir + "ncrisc-halt-wormhole.o tdma_xmov.o ";
         }
     } else {
         if (this->target_name_ == "ncrisc" and this->env_.arch_ == tt::ARCH::WORMHOLE_B0) {