fix grayskull barrier tests

tenstorrent · Dec 13, 2024 · e14c7fa · e14c7fa
1 parent 8760ff9
commit e14c7fa
Show file tree

Hide file tree

Showing 7 changed files with 67 additions and 43 deletions.
diff --git a/tests/blackhole/test_bh_common.h b/tests/blackhole/test_bh_common.h
@@ -15,8 +15,8 @@ constexpr std::uint32_t DRAM_BARRIER_BASE = 0;
 
 namespace tt::umd::test::utils {
 
-static void set_params_for_remote_txn(Cluster& device) {
-    // Populate address map and NOC parameters that the driver needs for remote transactions
+static void set_barrier_params(Cluster& device) {
+    // Populate address map and NOC parameters that the driver needs for memory barriers and remote transactions.
     device.set_barrier_address_params(
         {l1_mem::address_map::L1_BARRIER_BASE, eth_l1_mem::address_map::ERISC_BARRIER_BASE, DRAM_BARRIER_BASE});
 }
@@ -58,7 +58,7 @@ class BlackholeTestFixture : public ::testing::Test {
         assert(device != nullptr);
         assert(device->get_cluster_description()->get_number_of_chips() == get_detected_num_chips());
 
-        set_params_for_remote_txn(*device);
+        set_barrier_params(*device);
 
         tt_device_params default_params;
         device->start_device(default_params);

diff --git a/tests/blackhole/test_silicon_driver_bh.cpp b/tests/blackhole/test_silicon_driver_bh.cpp
@@ -20,8 +20,8 @@ using namespace tt::umd;
 
 constexpr std::uint32_t DRAM_BARRIER_BASE = 0;
 
-static void set_params_for_remote_txn(Cluster& device) {
-    // Populate address map and NOC parameters that the driver needs for remote transactions
+static void set_barrier_params(Cluster& device) {
+    // Populate address map and NOC parameters that the driver needs for memory barriers and remote transactions.
     device.set_barrier_address_params(
         {l1_mem::address_map::L1_BARRIER_BASE, eth_l1_mem::address_map::ERISC_BARRIER_BASE, DRAM_BARRIER_BASE});
 }
@@ -92,7 +92,7 @@ TEST(SiliconDriverBH, CreateDestroy) {
             false,
             true,
             false);
-        set_params_for_remote_txn(device);
+        set_barrier_params(device);
         device.start_device(default_params);
         device.deassert_risc_reset();
         device.close_device();
@@ -190,7 +190,7 @@ TEST(SiliconDriverBH, CreateDestroy) {
 //         true,
 //         true,
 //         simulated_harvesting_masks);
-//     set_params_for_remote_txn(device);
+//     set_barrier_params(device);
 //     auto mmio_devices = device.get_target_mmio_device_ids();
 
 //     for (int i = 0; i < target_devices.size(); i++) {
@@ -280,7 +280,7 @@ TEST(SiliconDriverBH, UnalignedStaticTLB_RW) {
     uint32_t num_host_mem_ch_per_mmio_device = 1;
 
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
     auto mmio_devices = device.get_target_mmio_device_ids();
 
     for (int i = 0; i < target_devices.size(); i++) {
@@ -337,7 +337,7 @@ TEST(SiliconDriverBH, StaticTLB_RW) {
     uint32_t num_host_mem_ch_per_mmio_device = 1;
 
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
     auto mmio_devices = device.get_target_mmio_device_ids();
 
     for (int i = 0; i < target_devices.size(); i++) {
@@ -402,7 +402,7 @@ TEST(SiliconDriverBH, DynamicTLB_RW) {
     uint32_t num_host_mem_ch_per_mmio_device = 1;
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
 
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
 
     tt_device_params default_params;
     device.start_device(default_params);
@@ -492,7 +492,7 @@ TEST(SiliconDriverBH, MultiThreadedDevice) {
     uint32_t num_host_mem_ch_per_mmio_device = 1;
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
 
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
 
     tt_device_params default_params;
     device.start_device(default_params);
@@ -562,7 +562,7 @@ TEST(SiliconDriverBH, MultiThreadedMemBar) {
     uint32_t num_host_mem_ch_per_mmio_device = 1;
 
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
     for (int i = 0; i < target_devices.size(); i++) {
         // Iterate over devices and only setup static TLBs for functional worker cores
         auto& sdesc = device.get_soc_descriptor(i);
@@ -702,7 +702,7 @@ TEST(SiliconDriverBH, DISABLED_BroadcastWrite) {  // Cannot broadcast to tensix/
     uint32_t num_host_mem_ch_per_mmio_device = 1;
 
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
     auto mmio_devices = device.get_target_mmio_device_ids();
 
     tt_device_params default_params;
@@ -790,7 +790,7 @@ TEST(SiliconDriverBH, DISABLED_VirtualCoordinateBroadcast) {  // same problem as
     uint32_t num_host_mem_ch_per_mmio_device = 1;
 
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
     auto mmio_devices = device.get_target_mmio_device_ids();
 
     tt_device_params default_params;
@@ -891,7 +891,7 @@ TEST(SiliconDriverBH, SysmemTestWithPcie) {
         true,   // clean system resources - yes
         true);  // perform harvesting - yes
 
-    set_params_for_remote_txn(cluster);
+    set_barrier_params(cluster);
     cluster.start_device(tt_device_params{});  // no special parameters
 
     const chip_id_t mmio_chip_id = 0;
@@ -962,7 +962,7 @@ TEST(SiliconDriverBH, RandomSysmemTestWithPcie) {
         true,   // clean system resources - yes
         true);  // perform harvesting - yes
 
-    set_params_for_remote_txn(cluster);
+    set_barrier_params(cluster);
     cluster.start_device(tt_device_params{});  // no special parameters
 
     const chip_id_t mmio_chip_id = 0;

diff --git a/tests/galaxy/test_umd_concurrent_threads.cpp b/tests/galaxy/test_umd_concurrent_threads.cpp
@@ -53,7 +53,7 @@ TEST(GalaxyConcurrentThreads, WriteToAllChipsL1) {
         Cluster(test_utils::GetAbsPath(SOC_DESC_PATH), all_devices, num_host_mem_ch_per_mmio_device, false, true);
     const auto sdesc_per_chip = device.get_virtual_soc_descriptors();
 
-    tt::umd::test::utils::set_params_for_remote_txn(device);
+    tt::umd::test::utils::set_barrier_params(device);
 
     tt_device_params default_params;
     device.start_device(default_params);
@@ -150,7 +150,7 @@ TEST(GalaxyConcurrentThreads, WriteToAllChipsDram) {
         Cluster(test_utils::GetAbsPath(SOC_DESC_PATH), all_devices, num_host_mem_ch_per_mmio_device, false, true);
     const auto sdesc_per_chip = device.get_virtual_soc_descriptors();
 
-    tt::umd::test::utils::set_params_for_remote_txn(device);
+    tt::umd::test::utils::set_barrier_params(device);
 
     tt_device_params default_params;
     device.start_device(default_params);
@@ -236,7 +236,7 @@ TEST(GalaxyConcurrentThreads, PushInputsWhileSignalingCluster) {
         Cluster(test_utils::GetAbsPath(SOC_DESC_PATH), target_devices, num_host_mem_ch_per_mmio_device, false, true);
     const auto sdesc_per_chip = device.get_virtual_soc_descriptors();
 
-    tt::umd::test::utils::set_params_for_remote_txn(device);
+    tt::umd::test::utils::set_barrier_params(device);
 
     tt_device_params default_params;
     device.start_device(default_params);

diff --git a/tests/galaxy/test_umd_remote_api.cpp b/tests/galaxy/test_umd_remote_api.cpp
@@ -33,7 +33,7 @@ void run_remote_read_write_test(uint32_t vector_size, bool dram_write) {
         Cluster(test_utils::GetAbsPath(SOC_DESC_PATH), target_devices, num_host_mem_ch_per_mmio_device, false, true);
     const auto sdesc_per_chip = device.get_virtual_soc_descriptors();
 
-    tt::umd::test::utils::set_params_for_remote_txn(device);
+    tt::umd::test::utils::set_barrier_params(device);
 
     tt_device_params default_params;
     device.start_device(default_params);
@@ -150,7 +150,7 @@ void run_data_mover_test(
     Cluster device =
         Cluster(test_utils::GetAbsPath(SOC_DESC_PATH), target_devices, num_host_mem_ch_per_mmio_device, false, true);
 
-    tt::umd::test::utils::set_params_for_remote_txn(device);
+    tt::umd::test::utils::set_barrier_params(device);
 
     tt_device_params default_params;
     device.start_device(default_params);
@@ -269,7 +269,7 @@ void run_data_broadcast_test(
     Cluster device =
         Cluster(test_utils::GetAbsPath(SOC_DESC_PATH), target_devices, num_host_mem_ch_per_mmio_device, false, true);
 
-    tt::umd::test::utils::set_params_for_remote_txn(device);
+    tt::umd::test::utils::set_barrier_params(device);
 
     tt_device_params default_params;
     device.start_device(default_params);

diff --git a/tests/grayskull/test_silicon_driver.cpp b/tests/grayskull/test_silicon_driver.cpp
@@ -15,6 +15,14 @@
 
 using namespace tt::umd;
 
+constexpr std::uint32_t DRAM_BARRIER_BASE = 0;
+
+static void set_barrier_params(Cluster& device) {
+    // Populate address map and NOC parameters that the driver needs for memory barriers.
+    // Grayskull doesn't have ETH, so we don't need to populate the ETH barrier address.
+    device.set_barrier_address_params({l1_mem::address_map::L1_BARRIER_BASE, 0u, DRAM_BARRIER_BASE});
+}
+
 TEST(SiliconDriverGS, CreateDestroySequential) {
     std::set<chip_id_t> target_devices = {0};
     uint32_t num_host_mem_ch_per_mmio_device = 1;
@@ -421,20 +429,31 @@ TEST(SiliconDriverGS, MultiThreadedMemBar) {  // this tests takes ~5 mins to run
     }
 
     tt_device_params default_params;
+    set_barrier_params(device);
     device.start_device(default_params);
     device.deassert_risc_reset();
     std::vector<uint32_t> readback_membar_vec = {};
     for (auto& core : device.get_soc_descriptor(0).workers) {
         test_utils::read_data_from_device(
-            device, readback_membar_vec, tt_cxy_pair(0, core), 0, 4, "SMALL_READ_WRITE_TLB");
+            device,
+            readback_membar_vec,
+            tt_cxy_pair(0, core),
+            l1_mem::address_map::L1_BARRIER_BASE,
+            4,
+            "SMALL_READ_WRITE_TLB");
         ASSERT_EQ(
             readback_membar_vec.at(0), 187);  // Ensure that memory barriers were correctly initialized on all workers
         readback_membar_vec = {};
     }
 
     for (auto& core : device.get_soc_descriptor(0).workers) {
         test_utils::read_data_from_device(
-            device, readback_membar_vec, tt_cxy_pair(0, core), 0, 4, "SMALL_READ_WRITE_TLB");
+            device,
+            readback_membar_vec,
+            tt_cxy_pair(0, core),
+            l1_mem::address_map::L1_BARRIER_BASE,
+            4,
+            "SMALL_READ_WRITE_TLB");
         ASSERT_EQ(
             readback_membar_vec.at(0), 187);  // Ensure that memory barriers were correctly initialized on all workers
         readback_membar_vec = {};
@@ -443,7 +462,7 @@ TEST(SiliconDriverGS, MultiThreadedMemBar) {  // this tests takes ~5 mins to run
     for (int chan = 0; chan < device.get_soc_descriptor(0).get_num_dram_channels(); chan++) {
         auto core = device.get_soc_descriptor(0).get_core_for_dram_channel(chan, 0);
         test_utils::read_data_from_device(
-            device, readback_membar_vec, tt_cxy_pair(0, core), 0, 4, "SMALL_READ_WRITE_TLB");
+            device, readback_membar_vec, tt_cxy_pair(0, core), DRAM_BARRIER_BASE, 4, "SMALL_READ_WRITE_TLB");
         ASSERT_EQ(
             readback_membar_vec.at(0), 187);  // Ensure that memory barriers were correctly initialized on all DRAM
         readback_membar_vec = {};
@@ -502,7 +521,12 @@ TEST(SiliconDriverGS, MultiThreadedMemBar) {  // this tests takes ~5 mins to run
 
     for (auto& core : device.get_soc_descriptor(0).workers) {
         test_utils::read_data_from_device(
-            device, readback_membar_vec, tt_cxy_pair(0, core), 0, 4, "SMALL_READ_WRITE_TLB");
+            device,
+            readback_membar_vec,
+            tt_cxy_pair(0, core),
+            l1_mem::address_map::L1_BARRIER_BASE,
+            4,
+            "SMALL_READ_WRITE_TLB");
         ASSERT_EQ(readback_membar_vec.at(0), 187);  // Ensure that memory barriers end up in correct sate workers
         readback_membar_vec = {};
     }

diff --git a/tests/wormhole/test_silicon_driver_wh.cpp b/tests/wormhole/test_silicon_driver_wh.cpp
@@ -18,8 +18,8 @@ using namespace tt::umd;
 
 constexpr std::uint32_t DRAM_BARRIER_BASE = 0;
 
-static void set_params_for_remote_txn(Cluster& device) {
-    // Populate address map and NOC parameters that the driver needs for remote transactions
+static void set_barrier_params(Cluster& device) {
+    // Populate address map and NOC parameters that the driver needs for memory barriers and remote transactions.
     device.set_barrier_address_params(
         {l1_mem::address_map::L1_BARRIER_BASE, eth_l1_mem::address_map::ERISC_BARRIER_BASE, DRAM_BARRIER_BASE});
 }
@@ -93,7 +93,7 @@ TEST(SiliconDriverWH, CreateDestroy) {
             false,
             true,
             false);
-        set_params_for_remote_txn(device);
+        set_barrier_params(device);
         device.start_device(default_params);
         device.deassert_risc_reset();
         device.close_device();
@@ -159,7 +159,7 @@ TEST(SiliconDriverWH, HarvestingRuntime) {
     uint32_t num_host_mem_ch_per_mmio_device = 1;
 
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true, simulated_harvesting_masks);
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
     auto mmio_devices = device.get_target_mmio_device_ids();
 
     for(int i = 0; i < target_devices.size(); i++) {
@@ -221,7 +221,7 @@ TEST(SiliconDriverWH, UnalignedStaticTLB_RW) {
 
     uint32_t num_host_mem_ch_per_mmio_device = 1;
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
     auto mmio_devices = device.get_target_mmio_device_ids();
 
     for (int i = 0; i < target_devices.size(); i++) {
@@ -277,7 +277,7 @@ TEST(SiliconDriverWH, StaticTLB_RW) {
 
     uint32_t num_host_mem_ch_per_mmio_device = 1;
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
     auto mmio_devices = device.get_target_mmio_device_ids();
 
     for (int i = 0; i < target_devices.size(); i++) {
@@ -341,7 +341,7 @@ TEST(SiliconDriverWH, DynamicTLB_RW) {
     uint32_t num_host_mem_ch_per_mmio_device = 1;
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
 
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
 
     tt_device_params default_params;
     device.start_device(default_params);
@@ -393,7 +393,7 @@ TEST(SiliconDriverWH, MultiThreadedDevice) {
     uint32_t num_host_mem_ch_per_mmio_device = 1;
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
 
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
 
     tt_device_params default_params;
     device.start_device(default_params);
@@ -463,7 +463,7 @@ TEST(SiliconDriverWH, MultiThreadedMemBar) {
     uint32_t num_host_mem_ch_per_mmio_device = 1;
 
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
     auto mmio_devices = device.get_target_mmio_device_ids();
 
     for (int i = 0; i < target_devices.size(); i++) {
@@ -606,7 +606,7 @@ TEST(SiliconDriverWH, BroadcastWrite) {
     uint32_t num_host_mem_ch_per_mmio_device = 1;
 
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
     auto mmio_devices = device.get_target_mmio_device_ids();
 
     tt_device_params default_params;
@@ -693,7 +693,7 @@ TEST(SiliconDriverWH, VirtualCoordinateBroadcast) {
     uint32_t num_host_mem_ch_per_mmio_device = 1;
 
     Cluster device = Cluster(num_host_mem_ch_per_mmio_device, false, true, true);
-    set_params_for_remote_txn(device);
+    set_barrier_params(device);
     auto mmio_devices = device.get_target_mmio_device_ids();
 
     tt_device_params default_params;
@@ -812,7 +812,7 @@ TEST(SiliconDriverWH, SysmemTestWithPcie) {
         true,   // clean system resources - yes
         true);  // perform harvesting - yes
 
-    set_params_for_remote_txn(cluster);
+    set_barrier_params(cluster);
     cluster.start_device(tt_device_params{});  // no special parameters
 
     const chip_id_t mmio_chip_id = 0;
@@ -878,7 +878,7 @@ TEST(SiliconDriverWH, RandomSysmemTestWithPcie) {
         true,   // clean system resources - yes
         true);  // perform harvesting - yes
 
-    set_params_for_remote_txn(cluster);
+    set_barrier_params(cluster);
     cluster.start_device(tt_device_params{});  // no special parameters
 
     const chip_id_t mmio_chip_id = 0;
@@ -950,7 +950,7 @@ TEST(SiliconDriverWH, LargeAddressTlb) {
     const auto ARC = cluster.get_soc_descriptor(0).arc_cores.at(0);
     const tt_cxy_pair ARC_CORE(0, ARC.x, ARC.y);
 
-    set_params_for_remote_txn(cluster);
+    set_barrier_params(cluster);
     cluster.start_device(tt_device_params{});
 
     auto get_static_tlb_index_callback = [](tt_xy_pair target) { return 0; };

diff --git a/tests/wormhole/test_wh_common.h b/tests/wormhole/test_wh_common.h
@@ -16,8 +16,8 @@ constexpr std::uint32_t DRAM_BARRIER_BASE = 0;
 
 namespace tt::umd::test::utils {
 
-static void set_params_for_remote_txn(Cluster& device) {
-    // Populate address map and NOC parameters that the driver needs for remote transactions
+static void set_barrier_params(Cluster& device) {
+    // Populate address map and NOC parameters that the driver needs for memory barriers and remote transactions.
     device.set_barrier_address_params(
         {l1_mem::address_map::L1_BARRIER_BASE, eth_l1_mem::address_map::ERISC_BARRIER_BASE, DRAM_BARRIER_BASE});
 }
@@ -58,7 +58,7 @@ class WormholeTestFixture : public ::testing::Test {
         assert(device != nullptr);
         assert(device->get_cluster_description()->get_number_of_chips() == get_detected_num_chips());
 
-        set_params_for_remote_txn(*device);
+        set_barrier_params(*device);
 
         tt_device_params default_params;
         device->start_device(default_params);