Merge branch 'main' into brosko/logical_mutex

tenstorrent · Dec 18, 2024 · 86d9432 · 86d9432
2 parents 2fb24e0 + e454c69
commit 86d9432
Show file tree

Hide file tree

Showing 18 changed files with 114 additions and 93 deletions.
diff --git a/.github/workflows/build-and-run-all-tests.yml b/.github/workflows/build-and-run-all-tests.yml
@@ -19,9 +19,14 @@ jobs:
           {arch: wormhole_b0},
           {arch: blackhole},
         ]
+        ubuntu-version: [
+          'ubuntu-22.04',
+          'ubuntu-20.04',
+        ]
     uses: ./.github/workflows/build-tests.yml
     with:
       arch: ${{ matrix.test-group.arch}}
+      ubuntu-version: ${{ matrix.ubuntu-version}}
       timeout: 15
 
   test-all:
@@ -39,8 +44,14 @@ jobs:
           # Enable once we have functional cards.
           # {arch: blackhole},
         ]
+        ubuntu-version: [
+          # Running tests on ubuntu-20.04 should be sufficient. Reduce load on CI.
+          # 'ubuntu-22.04',
+          'ubuntu-20.04',
+        ]
     uses: ./.github/workflows/run-tests.yml
     with:
       arch: ${{ matrix.test-group.arch}}
+      ubuntu-version: ${{ matrix.ubuntu-version}}
       card: ${{ matrix.test-group.card}}
       timeout: ${{ matrix.test-group.timeout}}
diff --git a/.github/workflows/build-tests.yml b/.github/workflows/build-tests.yml
@@ -8,6 +8,9 @@ on:
       arch:
         required: true
         type: string
+      ubuntu-version:
+        required: true
+        type: string
       timeout:
         required: true
         type: number
@@ -21,6 +24,13 @@ on:
           - grayskull
           - wormhole_b0
           - blackhole
+      ubuntu-version:
+        required: true
+        description: 'The version of Ubuntu to build on'
+        type: choice
+        options:
+          - ubuntu-22.04
+          - ubuntu-20.04
       timeout:
         required: true
         description: 'The timeout for the build job in minutes'
@@ -40,18 +50,11 @@ jobs:
   build:
     # Due to parsing bug, fromJSON is used to convert string to number
     timeout-minutes: ${{ fromJSON(inputs.timeout) }}
-    strategy:
-      fail-fast: false
-      matrix:
-        build: [
-          {runs-on: ubuntu-22.04, docker-image: tt-umd-ci-ubuntu-22.04},
-          {runs-on: ubuntu-20.04, docker-image: tt-umd-ci-ubuntu-20.04},
-        ]
 
-    name: Build umd_tests for ${{ inputs.arch }} on ${{ matrix.build.runs-on }}
-    runs-on: ${{ matrix.build.runs-on }}
+    name: Build umd_tests for ${{ inputs.arch }} on ${{ inputs.ubuntu-version }}
+    runs-on: ${{ inputs.ubuntu-version }}
     container:
-      image: ghcr.io/${{ github.repository }}/${{ matrix.build.docker-image }}:latest
+      image: ghcr.io/${{ github.repository }}/tt-umd-ci-${{ inputs.ubuntu-version }}:latest
       options: --user root
 
     env:
@@ -84,5 +87,5 @@ jobs:
       - name: Upload build artifacts archive
         uses: actions/upload-artifact@v4
         with:
-          name: build-artifacts-${{ inputs.arch }}-${{ matrix.build.runs-on }}
+          name: build-artifacts-${{ inputs.arch }}-${{ inputs.ubuntu-version }}
           path: artifact.tar
diff --git a/.github/workflows/run-tests.yml b/.github/workflows/run-tests.yml
@@ -7,6 +7,9 @@ on:
       arch:
         required: true
         type: string
+      ubuntu-version:
+        required: true
+        type: string
       card:
         required: true
         type: string
@@ -23,6 +26,13 @@ on:
           - grayskull
           - wormhole_b0
           - blackhole
+      ubuntu-version:
+        required: true
+        description: 'The version of Ubuntu to build on'
+        type: choice
+        options:
+          - ubuntu-22.04
+          - ubuntu-20.04
       card:
         required: true
         description: 'The card to run tests on'
@@ -47,20 +57,13 @@ jobs:
   test:
     # Due to parsing bug, fromJSON is used to convert string to number
     timeout-minutes: ${{ fromJSON(inputs.timeout) }}
-    strategy:
-      fail-fast: false
-      matrix:
-        build: [
-          {runs-on: ubuntu-22.04, docker-image: tt-umd-ci-ubuntu-22.04},
-          {runs-on: ubuntu-20.04, docker-image: tt-umd-ci-ubuntu-20.04},
-        ]
 
-    name: Run tests for ${{ inputs.arch }} on ${{ inputs.card }} on ${{ matrix.build.runs-on }}
+    name: Run tests for ${{ inputs.arch }} on ${{ inputs.card }} on ${{ inputs.ubuntu-version }}
     runs-on:
       - self-hosted
       - ${{ inputs.card }}
     container:
-      image: ghcr.io/${{ github.repository }}/${{ matrix.build.docker-image }}:latest
+      image: ghcr.io/${{ github.repository }}/tt-umd-ci-${{ inputs.ubuntu-version }}:latest
       options: --user root --device /dev/tenstorrent/0
       volumes:
         - /dev/hugepages:/dev/hugepages
@@ -83,7 +86,7 @@ jobs:
       - name: Use build artifacts
         uses: actions/download-artifact@v4
         with:
-          name: build-artifacts-${{ inputs.arch }}-${{ matrix.build.runs-on }}
+          name: build-artifacts-${{ inputs.arch }}-${{ inputs.ubuntu-version }}
           path: ./
 
       # This is needed to preserve file permissions

diff --git a/device/api/umd/device/architecture_implementation.h b/device/api/umd/device/architecture_implementation.h
@@ -11,9 +11,9 @@
 #include <tuple>
 #include <vector>
 
-#include "umd/device/tlb.h"
 #include "umd/device/tt_xy_pair.h"
 #include "umd/device/types/arch.h"
+#include "umd/device/types/tlb.h"
 #include "umd/device/types/xy_pair.h"
 
 struct tt_device_l1_address_params;

diff --git a/device/api/umd/device/blackhole_implementation.h b/device/api/umd/device/blackhole_implementation.h
@@ -10,7 +10,7 @@
 #include <stdexcept>
 
 #include "umd/device/architecture_implementation.h"
-#include "umd/device/tlb.h"
+#include "umd/device/types/tlb.h"
 
 namespace tt::umd {
 

diff --git a/device/api/umd/device/cluster.h b/device/api/umd/device/cluster.h
@@ -18,12 +18,12 @@
 #include "tt_soc_descriptor.h"
 #include "tt_xy_pair.h"
 #include "umd/device/chip/chip.h"
-#include "umd/device/tlb.h"
 #include "umd/device/tt_device/tt_device.h"
 #include "umd/device/tt_io.hpp"
 #include "umd/device/types/arch.h"
 #include "umd/device/types/cluster_descriptor_types.h"
 #include "umd/device/types/cluster_types.h"
+#include "umd/device/types/tlb.h"
 
 using TLB_DATA = tt::umd::tlb_data;
 
@@ -609,7 +609,6 @@ class Cluster : public tt_device {
     static void harvest_rows_in_soc_descriptor(tt::ARCH arch, tt_SocDescriptor& sdesc, uint32_t harvested_rows);
     static std::unordered_map<tt_xy_pair, tt_xy_pair> create_harvested_coord_translation(
         const tt::ARCH arch, bool identity_map);
-    std::unordered_map<tt_xy_pair, tt_xy_pair> get_harvested_coord_translation_map(chip_id_t logical_device_id);
     virtual std::uint32_t get_num_dram_channels(std::uint32_t device_id);
     virtual std::uint64_t get_dram_channel_size(std::uint32_t device_id, std::uint32_t channel);
     virtual std::uint32_t get_num_host_channels(std::uint32_t device_id);

diff --git a/device/api/umd/device/grayskull_implementation.h b/device/api/umd/device/grayskull_implementation.h
@@ -9,7 +9,7 @@
 #include <array>
 
 #include "architecture_implementation.h"
-#include "umd/device/tlb.h"
+#include "umd/device/types/tlb.h"
 
 namespace tt::umd {
 

diff --git a/device/api/umd/device/pci_device.hpp b/device/api/umd/device/pci_device.hpp
@@ -14,9 +14,9 @@
 
 #include "fmt/format.h"
 #include "umd/device/semver.hpp"
-#include "umd/device/tlb.h"
 #include "umd/device/tt_xy_pair.h"
 #include "umd/device/types/arch.h"
+#include "umd/device/types/tlb.h"
 
 namespace tt::umd {
 class semver_t;

diff --git a/device/api/umd/device/tt_cluster_descriptor.h b/device/api/umd/device/tt_cluster_descriptor.h
@@ -25,12 +25,14 @@ class Node;
 }
 
 enum BoardType : uint32_t {
-    N150 = 0,
-    N300 = 1,
-    E150 = 2,
-    P150A = 3,
-    GALAXY = 4,
-    UNKNOWN = 5,
+    E75 = 0,
+    E150 = 1,
+    E300 = 2,
+    N150 = 3,
+    N300 = 4,
+    P150A = 5,
+    GALAXY = 6,
+    UNKNOWN = 7,
 };
 
 class tt_ClusterDescriptor {

diff --git a/device/api/umd/device/tt_device/tt_device.h b/device/api/umd/device/tt_device/tt_device.h
@@ -69,18 +69,15 @@ class TTDevice {
         tt_xy_pair end,
         std::uint64_t address,
         bool multicast,
-        std::unordered_map<tt_xy_pair, tt_xy_pair> &harvested_coord_translation,
         std::uint64_t ordering);
     dynamic_tlb set_dynamic_tlb(
         unsigned int tlb_index,
         tt_xy_pair target,
         std::uint64_t address,
-        std::unordered_map<tt_xy_pair, tt_xy_pair> &harvested_coord_translation,
         std::uint64_t ordering = tt::umd::tlb_data::Relaxed);
     dynamic_tlb set_dynamic_tlb_broadcast(
         unsigned int tlb_index,
         std::uint64_t address,
-        std::unordered_map<tt_xy_pair, tt_xy_pair> &harvested_coord_translation,
         tt_xy_pair start,
         tt_xy_pair end,
         std::uint64_t ordering = tt::umd::tlb_data::Relaxed);

diff --git a/device/api/umd/device/tlb.h → device/api/umd/device/types/tlb.h b/device/api/umd/device/tlb.h → device/api/umd/device/types/tlb.h
diff --git a/device/api/umd/device/wormhole_implementation.h b/device/api/umd/device/wormhole_implementation.h
@@ -9,7 +9,7 @@
 #include <array>
 
 #include "architecture_implementation.h"
-#include "umd/device/tlb.h"
+#include "umd/device/types/tlb.h"
 
 namespace tt::umd {
 

diff --git a/device/cluster.cpp b/device/cluster.cpp
@@ -43,11 +43,11 @@
 #include "umd/device/chip/remote_chip.h"
 #include "umd/device/driver_atomics.h"
 #include "umd/device/hugepage.h"
-#include "umd/device/tlb.h"
 #include "umd/device/tt_cluster_descriptor.h"
 #include "umd/device/tt_core_coordinates.h"
 #include "umd/device/tt_soc_descriptor.h"
 #include "umd/device/types/arch.h"
+#include "umd/device/types/tlb.h"
 #include "yaml-cpp/yaml.h"
 
 using namespace boost::interprocess;
@@ -299,10 +299,6 @@ void Cluster::create_device(
 
 bool Cluster::using_harvested_soc_descriptors() { return perform_harvesting_on_sdesc && performed_harvesting; }
 
-std::unordered_map<tt_xy_pair, tt_xy_pair> Cluster::get_harvested_coord_translation_map(chip_id_t logical_device_id) {
-    return harvested_coord_translation.at(logical_device_id);
-}
-
 std::unordered_map<chip_id_t, uint32_t> Cluster::get_harvesting_masks_for_soc_descriptors() {
     if (using_harvested_soc_descriptors()) {
         return harvested_rows_per_target;
@@ -962,11 +958,10 @@ void Cluster::broadcast_pcie_tensix_risc_reset(chip_id_t chip_id, const TensixSo
     auto [soft_reset_reg, _] = tt_device->set_dynamic_tlb_broadcast(
         architecture_implementation->get_reg_tlb(),
         architecture_implementation->get_tensix_soft_reset_addr(),
-        harvested_coord_translation.at(chip_id),
-        tt_xy_pair(0, 0),
-        tt_xy_pair(
+        harvested_coord_translation.at(chip_id).at(tt_xy_pair(0, 0)),
+        harvested_coord_translation.at(chip_id).at(tt_xy_pair(
             architecture_implementation->get_grid_size_x() - 1,
-            architecture_implementation->get_grid_size_y() - 1 - num_rows_harvested.at(chip_id)),
+            architecture_implementation->get_grid_size_y() - 1 - num_rows_harvested.at(chip_id))),
         TLB_DATA::Posted);
     tt_device->write_regs(soft_reset_reg, 1, &valid);
     tt_driver_atomics::sfence();
@@ -1120,9 +1115,8 @@ void Cluster::write_device_memory(
         while (size_in_bytes > 0) {
             auto [mapped_address, tlb_size] = dev->set_dynamic_tlb(
                 tlb_index,
-                target,
+                harvested_coord_translation.at(target.chip).at(target),
                 address,
-                harvested_coord_translation.at(target.chip),
                 dynamic_tlb_ordering_modes.at(fallback_tlb));
             uint32_t transfer_size = std::min((uint64_t)size_in_bytes, tlb_size);
             dev->write_block(mapped_address, transfer_size, buffer_addr);
@@ -1169,9 +1163,8 @@ void Cluster::read_device_memory(
         while (size_in_bytes > 0) {
             auto [mapped_address, tlb_size] = dev->set_dynamic_tlb(
                 tlb_index,
-                target,
+                harvested_coord_translation.at(target.chip).at(target),
                 address,
-                harvested_coord_translation.at(target.chip),
                 dynamic_tlb_ordering_modes.at(fallback_tlb));
             uint32_t transfer_size = std::min((uint64_t)size_in_bytes, tlb_size);
             dev->read_block(mapped_address, transfer_size, buffer_addr);
@@ -1353,7 +1346,8 @@ void Cluster::configure_tlb(
         tlb_index);
 
     TTDevice* tt_device = get_tt_device(logical_device_id);
-    tt_device->set_dynamic_tlb(tlb_index, core, address, harvested_coord_translation.at(logical_device_id), ordering);
+    tt_device->set_dynamic_tlb(
+        tlb_index, harvested_coord_translation.at(logical_device_id).at(core), address, ordering);
     auto tlb_size = std::get<1>(tt_device->get_architecture_implementation()->describe_tlb(tlb_index).value());
     tlb_config_map.at(logical_device_id).insert({tlb_index, (address / tlb_size) * tlb_size});
     map_core_to_tlb_per_chip.at(logical_device_id).insert({core, tlb_index});
@@ -1447,9 +1441,8 @@ int Cluster::test_setup_interface() {
         uint32_t mapped_reg = tt_device
                                   ->set_dynamic_tlb(
                                       tt_device->get_architecture_implementation()->get_reg_tlb(),
-                                      tt_xy_pair(0, 0),
-                                      0xffb20108,
-                                      harvested_coord_translation.at(chip_id))
+                                      harvested_coord_translation.at(chip_id).at(tt_xy_pair(0, 0)),
+                                      0xffb20108)
                                   .bar_offset;
 
         uint32_t regval = 0;
@@ -1460,9 +1453,8 @@ int Cluster::test_setup_interface() {
         uint32_t mapped_reg = tt_device
                                   ->set_dynamic_tlb(
                                       tt_device->get_architecture_implementation()->get_reg_tlb(),
-                                      tt_xy_pair(1, 0),
-                                      0xffb20108,
-                                      harvested_coord_translation.at(chip_id))
+                                      harvested_coord_translation.at(chip_id).at(tt_xy_pair(1, 0)),
+                                      0xffb20108)
                                   .bar_offset;
 
         uint32_t regval = 0;
@@ -1475,9 +1467,8 @@ int Cluster::test_setup_interface() {
         // uint32_t mapped_reg = tt_device
         //                           ->set_dynamic_tlb(
         //                               tt_device->get_architecture_implementation()->get_reg_tlb(),
-        //                               tt_xy_pair(1, 0),
-        //                               0xffb20108,
-        //                               harvested_coord_translation.at(logical_device_id))
+        //                               harvested_coord_translation.at(chip_id).at(tt_xy_pair(1, 0)),
+        //                               0xffb20108)
         //                           .bar_offset;
 
         // uint32_t regval = 0;
@@ -2523,9 +2514,8 @@ void Cluster::pcie_broadcast_write(
         auto [mapped_address, tlb_size] = tt_device->set_dynamic_tlb_broadcast(
             tlb_index,
             addr,
-            harvested_coord_translation.at(chip),
-            start,
-            end,
+            harvested_coord_translation.at(chip).at(start),
+            harvested_coord_translation.at(chip).at(end),
             dynamic_tlb_ordering_modes.at(fallback_tlb));
         uint64_t transfer_size = std::min((uint64_t)size_in_bytes, tlb_size);
         tt_device->write_block(mapped_address, transfer_size, buffer_addr);
@@ -3005,8 +2995,8 @@ void Cluster::read_mmio_device_register(
     const scoped_lock<named_mutex> lock(*get_mutex(fallback_tlb, core.chip));
     log_debug(LogSiliconDriver, "  dynamic tlb_index: {}", tlb_index);
 
-    auto [mapped_address, tlb_size] =
-        tt_device->set_dynamic_tlb(tlb_index, core, addr, harvested_coord_translation.at(core.chip), TLB_DATA::Strict);
+    auto [mapped_address, tlb_size] = tt_device->set_dynamic_tlb(
+        tlb_index, harvested_coord_translation.at(core.chip).at(core), addr, TLB_DATA::Strict);
     // Align block to 4bytes if needed.
     auto aligned_buf = tt_4_byte_aligned_buffer(mem_ptr, size);
     tt_device->read_regs(mapped_address, aligned_buf.block_size / sizeof(std::uint32_t), aligned_buf.local_storage);
@@ -3025,8 +3015,8 @@ void Cluster::write_mmio_device_register(
     const scoped_lock<named_mutex> lock(*get_mutex(fallback_tlb, core.chip));
     log_debug(LogSiliconDriver, "  dynamic tlb_index: {}", tlb_index);
 
-    auto [mapped_address, tlb_size] =
-        tt_device->set_dynamic_tlb(tlb_index, core, addr, harvested_coord_translation.at(core.chip), TLB_DATA::Strict);
+    auto [mapped_address, tlb_size] = tt_device->set_dynamic_tlb(
+        tlb_index, harvested_coord_translation.at(core.chip).at(core), addr, TLB_DATA::Strict);
     // Align block to 4bytes if needed.
     auto aligned_buf = tt_4_byte_aligned_buffer(mem_ptr, size);
     if (aligned_buf.input_size != aligned_buf.block_size) {

diff --git a/device/tlb.cpp b/device/tlb.cpp
@@ -2,7 +2,7 @@
 //
 // SPDX-License-Identifier: Apache-2.0
 
-#include "umd/device/tlb.h"
+#include "umd/device/types/tlb.h"
 
 namespace tt::umd {