#0: Preload kernel before starting

tenstorrent · Nov 13, 2024 · 87bd9a2 · 87bd9a2
1 parent a28571f
commit 87bd9a2
Show file tree

Hide file tree

Showing 8 changed files with 36 additions and 6 deletions.
diff --git a/tt_metal/hostdevcommon/dprint_common.h b/tt_metal/hostdevcommon/dprint_common.h
@@ -18,7 +18,8 @@
 #if !defined(KERNEL_BUILD) && !defined(FW_BUILD) // SW
 #include "common/tt_backend_api_types.hpp"
 typedef tt::DataFormat CommonDataFormat;
-#else // HW already includes tensix_types.h
+#else
+#include "tensix_types.h"
 typedef DataFormat CommonDataFormat;
 #endif
 

diff --git a/tt_metal/hw/firmware/src/brisc.cc b/tt_metal/hw/firmware/src/brisc.cc
@@ -369,7 +369,8 @@ int main() {
 
         WAYPOINT("GW");
         uint8_t go_message_signal = RUN_MSG_DONE;
-        while ((go_message_signal = mailboxes->go_message.signal) != RUN_MSG_GO) {
+        while (((go_message_signal = mailboxes->go_message.signal) != RUN_MSG_GO) &&
+               !(mailboxes->launch[mailboxes->launch_msg_rd_ptr].kernel_config.enables & DISPATCH_ENABLE_FLAG_PRELOAD)) {
             // While the go signal for kernel execution is not sent, check if the worker was signalled
             // to reset its launch message read pointer.
             if (go_message_signal == RUN_MSG_RESET_READ_PTR) {

diff --git a/tt_metal/hw/firmware/src/brisck.cc b/tt_metal/hw/firmware/src/brisck.cc
@@ -22,6 +22,11 @@ extern uint32_t __kernel_init_local_l1_base[];
 extern uint32_t __fw_export_end_text[];
 
 void kernel_launch(uint32_t kernel_base_addr) {
+    uint8_t go_message_signal;
+    tt_l1_ptr mailboxes_t *const mailboxes = (tt_l1_ptr mailboxes_t *)(MEM_MAILBOX_BASE);
+
+    while ((go_message_signal = mailboxes->go_message.signal) != RUN_MSG_GO) {
+    }
 
 #if defined(DEBUG_NULL_KERNELS) && !defined(DISPATCH_KERNEL)
 #ifdef KERNEL_RUN_TIME

diff --git a/tt_metal/hw/firmware/src/ncrisck.cc b/tt_metal/hw/firmware/src/ncrisck.cc
@@ -2,13 +2,16 @@
 //
 // SPDX-License-Identifier: Apache-2.0
 
-#include "risc_common.h"
-#include "tensix.h"
-#include "tensix_types.h"
+#include <cstdint>
+
 #include "noc.h"
-#include "noc_overlay_parameters.h"
 #include "noc_nonblocking_api.h"
+#include "noc_overlay_parameters.h"
+#include "risc_attribs.h"
+#include "risc_common.h"
 #include "stream_io_map.h"
+#include "tensix.h"
+#include "tensix_types.h"
 #ifdef PERF_DUMP
 #include "risc_perf.h"
 #endif
@@ -30,7 +33,11 @@ extern uint32_t __kernel_init_local_l1_base[];
 extern uint32_t __fw_export_end_text[];
 
 void kernel_launch(uint32_t kernel_base_addr) {
+    uint8_t go_message_signal;
+    tt_l1_ptr mailboxes_t *const mailboxes = (tt_l1_ptr mailboxes_t *)(MEM_MAILBOX_BASE);
 
+    while ((go_message_signal = mailboxes->go_message.signal) != RUN_MSG_GO) {
+    }
   DeviceZoneScopedMainChildN("NCRISC-KERNEL");
 #if defined(DEBUG_NULL_KERNELS) && !defined(DISPATCH_KERNEL)
 #ifdef KERNEL_RUN_TIME

diff --git a/tt_metal/hw/firmware/src/trisck.cc b/tt_metal/hw/firmware/src/trisck.cc
@@ -38,6 +38,11 @@ extern uint32_t __fw_export_end_text[];
 
 void kernel_launch(uint32_t kernel_base_addr)
 {
+    uint8_t go_message_signal;
+    tt_l1_ptr mailboxes_t *const mailboxes = (tt_l1_ptr mailboxes_t *)(MEM_MAILBOX_BASE);
+
+    while ((go_message_signal = mailboxes->go_message.signal) != RUN_MSG_GO) {
+    }
   DeviceZoneScopedMainChildN("TRISC-KERNEL");
 #if defined(DEBUG_NULL_KERNELS) && !defined(DISPATCH_KERNEL)
 #ifdef KERNEL_RUN_TIME

diff --git a/tt_metal/hw/inc/dev_msgs.h b/tt_metal/hw/inc/dev_msgs.h
@@ -89,6 +89,10 @@ struct rta_offset_t {
     volatile uint16_t crta_offset;
 };
 
+enum dispatchenable_flags {
+    DISPATCH_ENABLE_FLAG_PRELOAD = 1 << 7,
+};
+
 struct kernel_config_msg_t {
     volatile uint16_t watcher_kernel_ids[DISPATCH_CLASS_MAX];
     volatile uint16_t ncrisc_kernel_size16;  // size in 16 byte units
@@ -107,6 +111,7 @@ struct kernel_config_msg_t {
     volatile uint8_t brisc_noc_mode;
     volatile uint8_t max_cb_index;
     volatile uint8_t exit_erisc_kernel;
+    // Or of enable_flags and dispatch_core_processor_masks.
     volatile uint8_t enables;
     volatile uint8_t pad2[12];
 } __attribute__((packed));

diff --git a/tt_metal/hw/inc/grayskull/noc_nonblocking_api.h b/tt_metal/hw/inc/grayskull/noc_nonblocking_api.h
@@ -8,6 +8,7 @@
 
 #include "noc_parameters.h"
 #include "dev_msgs.h"
+#include "risc_attribs.h"
 
 ////
 

diff --git a/tt_metal/impl/dispatch/command_queue.cpp b/tt_metal/impl/dispatch/command_queue.cpp
@@ -1020,6 +1020,7 @@ void EnqueueProgramCommand::assemble_device_commands(
     uint32_t programmable_core_index = hal.get_programmable_core_type_index(HalProgrammableCoreType::TENSIX);
     for (KernelGroup& kernel_group : program.get_kernel_groups(programmable_core_index)) {
         kernel_group.launch_msg.kernel_config.mode = DISPATCH_MODE_DEV;
+        kernel_group.launch_msg.kernel_config.enables |= DISPATCH_ENABLE_FLAG_PRELOAD;
         for (uint32_t i = 0; i < kernel_config_addrs.size(); i++) {
             kernel_group.launch_msg.kernel_config.kernel_config_base[i] = kernel_config_addrs[i].addr;
         }
@@ -1052,6 +1053,7 @@ void EnqueueProgramCommand::assemble_device_commands(
     if (programmable_core_index != -1) {
         for (KernelGroup& kernel_group : program.get_kernel_groups(programmable_core_index)) {
             kernel_group.launch_msg.kernel_config.mode = DISPATCH_MODE_DEV;
+            kernel_group.launch_msg.kernel_config.enables |= DISPATCH_ENABLE_FLAG_PRELOAD;
             for (uint32_t i = 0; i < kernel_config_addrs.size(); i++) {
                 kernel_group.launch_msg.kernel_config.kernel_config_base[i] = kernel_config_addrs[i].addr;
             }
@@ -1084,6 +1086,8 @@ void EnqueueProgramCommand::assemble_device_commands(
     // if not,  check if the program is active on workers. If active, have dispatch_d issue a write barrier
     cmd_sequence_sizeB += (this->device->dispatch_s_enabled() || program_transfer_info.num_active_cores > 0) * CQ_PREFETCH_CMD_BARE_MIN_SIZE;
 
+    // Wait for eveything to be written to before launch.
+    cmd_sequence_sizeB += CQ_PREFETCH_CMD_BARE_MIN_SIZE;
     // either dispatch_s or dispatch_d will send the go signal (go_signal_mcast command)
     cmd_sequence_sizeB += CQ_PREFETCH_CMD_BARE_MIN_SIZE;
 
@@ -1201,6 +1205,7 @@ void EnqueueProgramCommand::assemble_device_commands(
             kernel_bins_prefetch_subcmds[i].size());
     }
 
+    device_command_sequence.add_dispatch_wait(/*barrier*/ true, 0, 0, false, false, /*do_wait */ false);
     // Go Signals
     program_command_sequence.go_signals.reserve(
         multicast_go_signal_sub_cmds.size() + unicast_go_signal_sub_cmds.size());