Skip to content

Commit

Permalink
#0: Enable metal on galaxy.
Browse files Browse the repository at this point in the history 
#8305: add Galaxy cluster apis
#8305: cleanup, add print
#8450: Establish tunnels originating from an mmio device. Determine the remote chips as well as their order on the tunnel.
#8452: add tests for tg pipeline
#0:    patch for tg workflows.
#8450: Add tables for tunnel dispatch workers with build settings.
       Populate build settings for tunnel kernels.
       Launch FD2 kernels based on information in tunnel device dispatch worker map.
       Enable 4 devices per hugepage/channel
#0:    disable hanging/failing tests for Galaxy
#0:    skip using channel 3, 7 which use huge page channel 3. This (4th) huepage is not a full 1GB in size. 256 MB is taken up by syseng tools 4th huge page.
#0:    re-enable Galaxy sharded tests, reduce one test runtime for Galaxy
#0:    fix cluster init for Galaxy
#8953: Fix hardcoding of queue sizes in tests.
#8450: Fix compute grid selection for N150. N150 can be standalone system or part of a TG system. On TG compute grid for N150 is different than standalone N150.
#0:    Reduce prefetch q entries to account for Galaxy CQ size.
  • Loading branch information
@ubcheema
ubcheema committed Jun 3, 2024
1 parent 354370a commit d79350e
Show file tree
Hide file tree
Showing 13 changed files with 1,419 additions and 698 deletions.
87 changes: 41 additions & 46 deletions .../unit_tests_fast_dispatch/command_queue/test_EnqueueWriteBuffer_and_EnqueueReadBuffer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -94,11 +94,11 @@ void test_EnqueueWriteBuffer_and_EnqueueReadBuffer(Device *device, CommandQueue
// Clear out command queue
uint16_t channel = tt::Cluster::instance().get_assigned_channel_for_device(device->id());
chip_id_t mmio_device_id = tt::Cluster::instance().get_associated_mmio_device(device->id());
uint32_t cq_size = tt::Cluster::instance().get_host_channel_size(mmio_device_id, channel) / device->num_hw_cqs();
uint32_t cq_size = device->sysmem_manager().get_cq_size();

std::vector<uint32_t> cq_zeros((cq_size - CQ_START) / sizeof(uint32_t), 0);

tt::Cluster::instance().write_sysmem(cq_zeros.data(), (cq_size - CQ_START), CQ_START, mmio_device_id, channel);
tt::Cluster::instance().write_sysmem(cq_zeros.data(), (cq_size - CQ_START), get_absolute_cq_offset(channel, 0, cq_size) + CQ_START, mmio_device_id, channel);

for (const bool cq_write : {true, false}) {
for (const bool cq_read : {true, false}) {
Expand Down Expand Up @@ -327,6 +327,7 @@ namespace dram_tests {
TEST_F(CommandQueueSingleCardFixture, WriteOneTileToDramBank0) {
TestBufferConfig config = {.num_pages = 1, .page_size = 2048, .buftype = BufferType::DRAM};
for (Device *device : devices_) {
tt::log_info("Running On Device {}", device->id());
local_test_functions::test_EnqueueWriteBuffer_and_EnqueueReadBuffer(device, device->command_queue(), config);
}
}
Expand Down Expand Up @@ -428,11 +429,9 @@ TEST_F(CommandQueueFixture, TestPageSizeTooLarge) {
// Requires enqueue write buffer
TEST_F(CommandQueueSingleCardFixture, TestWrapHostHugepageOnEnqueueReadBuffer) {
for (Device *device : this->devices_) {
tt::log_info("Running On Device {}", device->id());
uint32_t page_size = 2048;
uint16_t channel = tt::Cluster::instance().get_assigned_channel_for_device(device->id());
chip_id_t mmio_device_id = tt::Cluster::instance().get_associated_mmio_device(device->id());
uint32_t command_queue_size = tt::Cluster::instance().get_host_channel_size(mmio_device_id, channel);
uint32_t command_issue_region_size = 805310400;
uint32_t command_issue_region_size = device->sysmem_manager().get_issue_queue_size(0);

uint32_t max_command_size = command_issue_region_size - CQ_START;
uint32_t buffer = 14240;
Expand All @@ -446,10 +445,9 @@ TEST_F(CommandQueueSingleCardFixture, TestWrapHostHugepageOnEnqueueReadBuffer) {

TEST_F(CommandQueueSingleCardFixture, TestIssueMultipleReadWriteCommandsForOneBuffer) {
for (Device *device : this->devices_) {
tt::log_info("Running On Device {}", device->id());
uint32_t page_size = 2048;
uint16_t channel = tt::Cluster::instance().get_assigned_channel_for_device(device->id());
chip_id_t mmio_device_id = tt::Cluster::instance().get_associated_mmio_device(device->id());
uint32_t command_queue_size = tt::Cluster::instance().get_host_channel_size(mmio_device_id, channel);
uint32_t command_queue_size = device->sysmem_manager().get_cq_size();
uint32_t num_pages = command_queue_size / page_size;

TestBufferConfig config = {.num_pages = num_pages, .page_size = page_size, .buftype = BufferType::DRAM};
Expand All @@ -465,10 +463,8 @@ TEST_F(CommandQueueSingleCardFixture, TestWrapCompletionQOnInsufficientSpace) {
uint32_t small_page_size = 2048; // page size for second read

for (Device *device : devices_) {
uint16_t channel = tt::Cluster::instance().get_assigned_channel_for_device(device->id());
chip_id_t mmio_device_id = tt::Cluster::instance().get_associated_mmio_device(device->id());
uint32_t command_queue_size = tt::Cluster::instance().get_host_channel_size(mmio_device_id, channel);
uint32_t command_completion_region_size = 268431360;
tt::log_info("Running On Device {}", device->id());
uint32_t command_completion_region_size = device->sysmem_manager().get_completion_queue_size(0);

uint32_t first_buffer_size = tt::round_up(command_completion_region_size * 0.95, large_page_size);

Expand Down Expand Up @@ -504,10 +500,8 @@ TEST_F(CommandQueueSingleCardFixture, TestWrapCompletionQOnInsufficientSpace) {
TEST_F(CommandQueueSingleCardFixture, TestWrapCompletionQOnInsufficientSpace2) {
// Using default 75-25 issue and completion queue split
for (Device *device : devices_) {
uint16_t channel = tt::Cluster::instance().get_assigned_channel_for_device(device->id());
chip_id_t mmio_device_id = tt::Cluster::instance().get_associated_mmio_device(device->id());
uint32_t command_queue_size = tt::Cluster::instance().get_host_channel_size(mmio_device_id, channel);
uint32_t command_completion_region_size = 268431360;
tt::log_info("Running On Device {}", device->id());
uint32_t command_completion_region_size = device->sysmem_manager().get_completion_queue_size(0);

uint32_t num_pages_buff_1 = 9;
uint32_t page_size_buff_1 = 2048;
Expand Down Expand Up @@ -653,6 +647,7 @@ TEST_F(CommandQueueSingleCardFixture, WritesToRandomBufferTypeAndThenReadsBlocki
.seed = 0, .num_pages_total = 50000, .page_size = 2048, .max_num_pages_per_buffer = 16};

for (Device *device : devices_) {
tt::log_info("Running on Device {}", device->id());
EXPECT_TRUE(local_test_functions::stress_test_EnqueueWriteBuffer_and_EnqueueReadBuffer<true>(
device, device->command_queue(), config));
}
Expand All @@ -672,34 +667,34 @@ TEST_F(CommandQueueSingleCardFixture, WritesToRandomBufferTypeAndThenReadsNonblo

// TODO: Split this into separate tests
TEST_F(CommandQueueSingleCardFixture, ShardedBufferL1ReadWrites) {
std::map<std::string, std::vector<std::array<uint32_t, 2>>> test_params;

for (Device *device : devices_) {
for (const std::array<uint32_t, 2> cores :
{std::array<uint32_t, 2>{1, 1},
std::array<uint32_t, 2>{5, 1},
std::array<uint32_t, 2>{1, 5},
std::array<uint32_t, 2>{5, 3},
std::array<uint32_t, 2>{3, 5},
std::array<uint32_t, 2>{5, 5},
std::array<uint32_t, 2>{
static_cast<uint32_t>(device->compute_with_storage_grid_size().x),
static_cast<uint32_t>(device->compute_with_storage_grid_size().y)}}) {
for (const std::array<uint32_t, 2> num_pages : {
std::array<uint32_t, 2>{1, 1},
std::array<uint32_t, 2>{2, 1},
std::array<uint32_t, 2>{1, 2},
std::array<uint32_t, 2>{2, 2},
std::array<uint32_t, 2>{7, 11},
std::array<uint32_t, 2>{3, 65},
std::array<uint32_t, 2>{67, 4},
std::array<uint32_t, 2>{3, 137},
}) {
for (const std::array<uint32_t, 2> page_shape : {
std::array<uint32_t, 2>{32, 32},
std::array<uint32_t, 2>{1, 4},
std::array<uint32_t, 2>{1, 120},
std::array<uint32_t, 2>{1, 1024},
std::array<uint32_t, 2>{1, 2048},
}) {
if (tt::Cluster::instance().is_galaxy_cluster()) {
test_params = {
{"cores",
{{1, 1},
{static_cast<uint32_t>(device->compute_with_storage_grid_size().x),
static_cast<uint32_t>(device->compute_with_storage_grid_size().y)}}},
{"num_pages", {{3, 65}}},
{"page_shape", {{32, 32}}}};
} else {
test_params = {
{"cores",
{{1, 1},
{5, 1},
{1, 5},
{5, 3},
{3, 5},
{5, 5},
{static_cast<uint32_t>(device->compute_with_storage_grid_size().x),
static_cast<uint32_t>(device->compute_with_storage_grid_size().y)}}},
{"num_pages", {{1, 1}, {2, 1}, {1, 2}, {2, 2}, {7, 11}, {3, 65}, {67, 4}, {3, 137}}},
{"page_shape", {{32, 32}, {1, 4}, {1, 120}, {1, 1024}, {1, 2048}}}};
}
for (const std::array<uint32_t, 2> cores : test_params.at("cores")) {
for (const std::array<uint32_t, 2> num_pages : test_params.at("num_pages")) {
for (const std::array<uint32_t, 2> page_shape : test_params.at("page_shape")) {
for (const TensorMemoryLayout shard_strategy :
{TensorMemoryLayout::HEIGHT_SHARDED,
TensorMemoryLayout::WIDTH_SHARDED,
Expand All @@ -712,7 +707,7 @@ TEST_F(CommandQueueSingleCardFixture, ShardedBufferL1ReadWrites) {
config.num_iterations = num_iterations;
config.mem_config = shard_strategy;
config.page_shape = page_shape;
tt::log_info(tt::LogTest, fmt::format("cores: [{},{}] num_pages: [{},{}] page_shape: [{},{}], shard_strategy: {}, num_iterations: {}", cores[0],cores[1], num_pages[0],num_pages[1], page_shape[0],page_shape[1], magic_enum::enum_name(shard_strategy).data(), num_iterations).c_str());
tt::log_info(tt::LogTest, fmt::format("Device: {} cores: [{},{}] num_pages: [{},{}] page_shape: [{},{}], shard_strategy: {}, num_iterations: {}", device->id(), cores[0],cores[1], num_pages[0],num_pages[1], page_shape[0],page_shape[1], magic_enum::enum_name(shard_strategy).data(), num_iterations).c_str());
local_test_functions::stress_test_EnqueueWriteBuffer_and_EnqueueReadBuffer_sharded(
device, device->command_queue(), config, BufferType::L1, false);
}
Expand Down Expand Up @@ -800,7 +795,7 @@ TEST_F(CommandQueueSingleCardFixture, ShardedBufferLargeL1ReadWrites) {
config.num_iterations = num_iterations;
config.mem_config = shard_strategy;
config.page_shape = page_shape;
tt::log_info(tt::LogTest, fmt::format("cores: [{},{}] num_pages: [{},{}] page_shape: [{},{}], shard_strategy: {}, num_iterations: {}", cores[0],cores[1], num_pages[0],num_pages[1], page_shape[0],page_shape[1], magic_enum::enum_name(shard_strategy).data(), num_iterations).c_str());
tt::log_info(tt::LogTest, fmt::format("Device: {} cores: [{},{}] num_pages: [{},{}] page_shape: [{},{}], shard_strategy: {}, num_iterations: {}", device->id(), cores[0],cores[1], num_pages[0],num_pages[1], page_shape[0],page_shape[1], magic_enum::enum_name(shard_strategy).data(), num_iterations).c_str());
local_test_functions::stress_test_EnqueueWriteBuffer_and_EnqueueReadBuffer_sharded(
device, device->command_queue(), config, BufferType::L1, true);
}
Expand Down
10 changes: 9 additions & 1 deletion tt_metal/common/core_descriptor.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -119,6 +119,10 @@ inline const core_descriptor_t &get_core_descriptor_config(chip_id_t device_id,

auto compute_with_storage_start = desc_yaml["compute_with_storage_grid_range"]["start"];
auto compute_with_storage_end = desc_yaml["compute_with_storage_grid_range"]["end"];
if (tt::Cluster::instance().is_galaxy_cluster() and product_name == "nebula_x1") {
compute_with_storage_start = desc_yaml["tg_compute_with_storage_grid_range"]["start"];
compute_with_storage_end = desc_yaml["tg_compute_with_storage_grid_range"]["end"];
}
TT_ASSERT(compute_with_storage_start.IsSequence() and compute_with_storage_end.IsSequence());
TT_ASSERT(compute_with_storage_end[0].as<size_t>() >= compute_with_storage_start[0].as<size_t>());
TT_ASSERT(compute_with_storage_end[1].as<size_t>() >= compute_with_storage_start[1].as<size_t>());
Expand All @@ -136,7 +140,11 @@ inline const core_descriptor_t &get_core_descriptor_config(chip_id_t device_id,
}

std::vector<RelativeCoreCoord> dispatch_cores;
for (const auto& core_node : desc_yaml["dispatch_cores"]) {
auto dispatch_cores_string = "dispatch_cores";
if (tt::Cluster::instance().is_galaxy_cluster() and product_name == "nebula_x1") {
dispatch_cores_string = "tg_dispatch_cores";
}
for (const auto& core_node : desc_yaml[dispatch_cores_string]) {
RelativeCoreCoord coord = {};
if (core_node.IsSequence()) {
// Logical coord
Expand Down
9 changes: 9 additions & 0 deletions tt_metal/core_descriptors/wormhole_b0_80_arch.yaml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -49,12 +49,21 @@ nebula_x1:
start: [0, 0]
end: [7, 7]

tg_compute_with_storage_grid_range: # Logical only start and end [x, y]
start: [0, 0]
end: [7, 3]

storage_cores:
[]

dispatch_cores:
[[0, -1], [1, -1], [2, -1], [3, -1], [4, -1], [5, -1], [6, -1], [7, -1]]

tg_dispatch_cores:
[[0, -1], [1, -1], [2, -1], [3, -1], [4, -1], [5, -1], [6, -1], [7, -1],
[0, -2], [1, -2], [2, -2], [3, -2], [4, -2], [5, -2], [6, -2], [7, -2],
[0, -3], [1, -3], [2, -3], [3, -3], [4, -3], [5, -3], [6, -3], [7, -3]]

dispatch_core_type:
"tensix"

Expand Down
Loading

0 comments on commit d79350e

Please sign in to comment.