#7957: update doc- zeros, empty, zeros_like (#14104)

.github/workflows/ttnn-run-sweeps.yaml

@@ -14,6 +14,9 @@ on:
           - ccl.line_all_gather
           - ccl.all_gather_n300
           - ccl.all_gather_n300_focused
+          - creation.zeros.zeros
+          - creation.empty.empty
+          - creation.zeros_like.zeros_like
           - eltwise.unary.abs.abs_pytorch2
           - eltwise.unary.relu.relu
           - eltwise.unary.relu.relu_pytorch2

tests/sweep_framework/sweeps/creation/empty/empty.py

@@ -0,0 +1,72 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional, Tuple
+
+import torch
+import ttnn
+
+from tests.ttnn.utils_for_testing import check_with_pcc, start_measuring_time, stop_measuring_time
+
+
+# Parameters provided to the test vector generator are defined here.
+# They are defined as dict-type suites that contain the arguments to the run function as keys, and lists of possible inputs as values.
+# Each suite has a key name (in this case "suite_1") which will associate the test vectors to this specific suite of inputs.
+# Developers can create their own generator functions and pass them to the parameters as inputs.
+parameters = {
+    "nightly": {
+        "batch_sizes": [(1, 2), (3, 6)],
+        "height": [384, 1024],
+        "width": [1024, 4096],
+        "input_dtype": [ttnn.bfloat16, ttnn.bfloat8_b, ttnn.float32],
+        "output_memory_config": [ttnn.DRAM_MEMORY_CONFIG],
+        "layout": [ttnn.TILE_LAYOUT, ttnn.ROW_MAJOR_LAYOUT],
+    },
+}
+
+
+# Invalidate vector is called during the generation phase where each vector will be passed in.
+# If invalidated, the vector will still be stored but will be skipped.
+# Returns False, None if the vector is valid, and True, str with a reason for invalidation if it is invalid.
+def invalidate_vector(test_vector) -> Tuple[bool, Optional[str]]:
+    if test_vector["layout"] == ttnn.ROW_MAJOR_LAYOUT and test_vector["input_dtype"] == ttnn.bfloat8_b:
+        return True, "Skipped as ROW_MAJOR_LAYOUT and ttnn.bfloat8_b not supported"
+    return False, None
+
+
+def check_output(torch_output_tensor, output_tensor):
+    status = list(torch_output_tensor.shape) == list(output_tensor.shape)
+    msg = ""
+    msg = "pass" if status else "fail"
+
+    return status, msg
+
+
+# This is the run instructions for the test, defined by the developer.
+# The run function must take the above-defined parameters as inputs.
+# The runner will call this run function with each test vector, and the returned results from this function will be stored.
+# If you defined a device_mesh_fixture above, the object you yielded will be passed into this function as 'device'. Otherwise, it will be the default ttnn device opened by the infra.
+def run(
+    batch_sizes,
+    height,
+    width,
+    input_dtype,
+    output_memory_config,
+    layout,
+    *,
+    device,
+) -> list:
+    torch.manual_seed(0)
+
+    input_shape = (*batch_sizes, height, width)
+
+    torch_output_tensor = torch.empty(input_shape)
+
+    start_time = start_measuring_time()
+
+    output_tensor = ttnn.empty(input_shape, input_dtype, layout, device=device, memory_config=output_memory_config)
+    output_tensor = ttnn.to_torch(output_tensor)
+    e2e_perf = stop_measuring_time(start_time)
+
+    return [check_output(torch_output_tensor, output_tensor), e2e_perf]

tests/sweep_framework/sweeps/creation/zeros/zeros.py

@@ -0,0 +1,64 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional, Tuple
+
+import torch
+import ttnn
+
+from tests.ttnn.utils_for_testing import check_with_pcc, start_measuring_time, stop_measuring_time
+
+
+# Parameters provided to the test vector generator are defined here.
+# They are defined as dict-type suites that contain the arguments to the run function as keys, and lists of possible inputs as values.
+# Each suite has a key name (in this case "suite_1") which will associate the test vectors to this specific suite of inputs.
+# Developers can create their own generator functions and pass them to the parameters as inputs.
+parameters = {
+    "nightly": {
+        "batch_sizes": [(1, 2), (3, 6)],
+        "height": [384, 1024],
+        "width": [1024, 4096],
+        "input_dtype": [ttnn.bfloat16, ttnn.bfloat8_b, ttnn.float32],
+        "output_memory_config": [ttnn.DRAM_MEMORY_CONFIG],
+        "layout": [ttnn.TILE_LAYOUT, ttnn.ROW_MAJOR_LAYOUT],
+    },
+}
+
+
+# Invalidate vector is called during the generation phase where each vector will be passed in.
+# If invalidated, the vector will still be stored but will be skipped.
+# Returns False, None if the vector is valid, and True, str with a reason for invalidation if it is invalid.
+def invalidate_vector(test_vector) -> Tuple[bool, Optional[str]]:
+    if test_vector["input_dtype"] == ttnn.bfloat8_b:
+        return True, "Skipped as ROW_MAJOR_LAYOUT and ttnn.bfloat8_b not supported"
+    return False, None
+
+
+# This is the run instructions for the test, defined by the developer.
+# The run function must take the above-defined parameters as inputs.
+# The runner will call this run function with each test vector, and the returned results from this function will be stored.
+# If you defined a device_mesh_fixture above, the object you yielded will be passed into this function as 'device'. Otherwise, it will be the default ttnn device opened by the infra.
+def run(
+    batch_sizes,
+    height,
+    width,
+    input_dtype,
+    output_memory_config,
+    layout,
+    *,
+    device,
+) -> list:
+    torch.manual_seed(0)
+
+    input_shape = (*batch_sizes, height, width)
+
+    torch_output_tensor = torch.zeros(input_shape)
+
+    start_time = start_measuring_time()
+
+    output_tensor = ttnn.zeros(input_shape, input_dtype, layout, memory_config=output_memory_config)
+    output_tensor = ttnn.to_torch(output_tensor)
+    e2e_perf = stop_measuring_time(start_time)
+
+    return [check_with_pcc(torch_output_tensor, output_tensor, 0.999), e2e_perf]

tests/sweep_framework/sweeps/creation/zeros_like/zeros_like.py

@@ -0,0 +1,75 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional, Tuple
+from functools import partial
+
+import torch
+import ttnn
+from tests.sweep_framework.sweep_utils.utils import gen_shapes
+from tests.tt_eager.python_api_testing.sweep_tests.generation_funcs import gen_func_with_cast_tt
+
+from tests.ttnn.utils_for_testing import check_with_pcc, start_measuring_time, stop_measuring_time
+from models.utility_functions import torch_random
+
+
+# Parameters provided to the test vector generator are defined here.
+# They are defined as dict-type suites that contain the arguments to the run function as keys, and lists of possible inputs as values.
+# Each suite has a key name (in this case "suite_1" and "suite_2") which will associate the test vectors to this specific suite of inputs.
+# Developers can create their own generator functions and pass them to the parameters as inputs.
+parameters = {
+    "nightly": {
+        "input_shape": gen_shapes([1, 1, 32, 32], [2, 6, 256, 256], [1, 1, 320, 320], 128),
+        "input_dtype": [ttnn.bfloat16, ttnn.float32, ttnn.bfloat8_b],
+        "input_a_layout": [ttnn.TILE_LAYOUT, ttnn.ROW_MAJOR_LAYOUT],
+        "input_a_memory_config": [ttnn.DRAM_MEMORY_CONFIG, ttnn.L1_MEMORY_CONFIG],
+        "output_memory_config": [ttnn.DRAM_MEMORY_CONFIG, ttnn.L1_MEMORY_CONFIG],
+    },
+}
+
+
+# Invalidate vector is called during the generation phase where each vector will be passed in.
+# If invalidated, the vector will still be stored but will be skipped.
+# Returns False, None if the vector is valid, and True, str with a reason for invalidation if it is invalid.
+def invalidate_vector(test_vector) -> Tuple[bool, Optional[str]]:
+    if test_vector["input_dtype"] == ttnn.bfloat8_b:
+        return True, "Skipped as bfloat8_b dtype not supported"
+    return False, None
+
+
+# This is the run instructions for the test, defined by the developer.
+# The run function must take the above-defined parameters as inputs.
+# The runner will call this run function with each test vector, and the returned results from this function will be stored.
+# If you defined a mesh_device_fixture above, the object you yielded will be passed into this function as 'device'. Otherwise, it will be the default ttnn device opened by the infra.
+def run(
+    input_shape,
+    input_dtype,
+    input_a_layout,
+    input_a_memory_config,
+    output_memory_config,
+    *,
+    device,
+) -> list:
+    torch.manual_seed(0)
+
+    torch_input_tensor_a = gen_func_with_cast_tt(
+        partial(torch_random, low=-100, high=100, dtype=torch.float32), input_dtype
+    )(input_shape)
+
+    torch_output_tensor = torch.zeros_like(torch_input_tensor_a)
+
+    input_tensor_a = ttnn.from_torch(
+        torch_input_tensor_a,
+        dtype=input_dtype,
+        layout=input_a_layout,
+        device=device,
+        memory_config=input_a_memory_config,
+    )
+
+    start_time = start_measuring_time()
+    result = ttnn.zeros_like(input_tensor_a, memory_config=output_memory_config)
+    output_tensor = ttnn.to_torch(result)
+    e2e_perf = stop_measuring_time(start_time)
+
+    return [check_with_pcc(torch_output_tensor, output_tensor, 0.999), e2e_perf]

ttnn/cpp/pybind11/operations/creation.hpp

@@ -96,7 +96,7 @@ void bind_full_operation(py::module& module, const creation_operation_t& operati
 }
 
 template <typename creation_operation_t>
-void bind_full_operation_with_hard_coded_value(py::module& module, const creation_operation_t& operation, const std::string& value_string) {
+void bind_full_operation_with_hard_coded_value(py::module& module, const creation_operation_t& operation, const std::string& value_string, const std::string& info_doc = "") {
     auto doc = fmt::format(
         R"doc(
         Creates a tensor with the specified shape and fills it with the value of {1}.
@@ -115,6 +115,9 @@ void bind_full_operation_with_hard_coded_value(py::module& module, const creatio
         Returns:
             ttnn.Tensor: A tensor filled with {1}.
 
+        Note:
+            {2}
+
         Example:
             >>> tensor = ttnn.{0}(shape=[1, 2, 2, 2], dtype=ttnn.bfloat16, layout=ttnn.ROW_MAJOR_LAYOUT)
             >>> print(tensor)
@@ -124,7 +127,8 @@ void bind_full_operation_with_hard_coded_value(py::module& module, const creatio
                             [{1}, {1}]]]]], shape=Shape([1, 2, 2, 2]), dtype=DataType::BFLOAT16, layout=Layout::ROW_MAJOR)
         )doc",
         operation.base_name(),
-        value_string);
+        value_string,
+        info_doc);
 
     bind_registered_operation(
         module,
@@ -221,7 +225,7 @@ void bind_full_like_operation(py::module& module, const creation_operation_t& op
 }
 
 template <typename creation_operation_t>
-void bind_full_like_operation_with_hard_coded_value(py::module& module, const creation_operation_t& operation, const std::string& value_string) {
+void bind_full_like_operation_with_hard_coded_value(py::module& module, const creation_operation_t& operation, const std::string& value_string, const std::string& info_doc = "") {
     auto doc = fmt::format(
         R"doc(
         Creates a tensor of the same shape as the input tensor and fills it with the value of {1}. The data type, layout, device, and memory configuration of the resulting tensor can be specified.
@@ -238,6 +242,9 @@ void bind_full_like_operation_with_hard_coded_value(py::module& module, const cr
         Returns:
             ttnn.Tensor: A tensor filled with {1}.
 
+        Note:
+            {2}
+
         Example:
             >>> tensor = ttnn.{0}(ttnn.from_torch(torch.randn(1, 2, 2, 2), ttnn.bfloat16, ttnn.TILE_LAYOUT)
             >>> output_tensor = ttnn.{0}(tensor=input_tensor, dtype=ttnn.bfloat16, layout=ttnn.TILE_LAYOUT)
@@ -248,7 +255,8 @@ void bind_full_like_operation_with_hard_coded_value(py::module& module, const cr
                             [{1}, {1}]]]]], shape=Shape([1, 2, 2, 2]), dtype=DataType::BFLOAT16, layout=Layout::TILE_LAYOUT)
         )doc",
         operation.base_name(),
-        value_string);
+        value_string,
+        info_doc);
 
     bind_registered_operation(
         module,
@@ -320,7 +328,7 @@ void bind_arange_operation(py::module& module, const creation_operation_t& opera
             py::arg("memory_config") = ttnn::DRAM_MEMORY_CONFIG});
 }
 
-void bind_empty_operation(py::module& module) {
+void bind_empty_operation(py::module& module, const std::string& info_doc = "") {
     auto doc = fmt::format(
         R"doc(
         Creates a device tensor with uninitialized values of the specified shape, data type, layout, and memory configuration.
@@ -335,12 +343,16 @@ void bind_empty_operation(py::module& module) {
         Returns:
             ttnn.Tensor: The output uninitialized tensor.
 
+        Note:
+            {1}
+
         Example:
             >>> tensor = ttnn.empty(shape=[2, 3], device=device)
             >>> print(tensor)
             ttnn.Tensor([[[[0.9, 0.21, 0.5], [0.67, 0.11, 0.30]]]], shape=Shape([2, 3]), dtype=DataType::BFLOAT16, layout=Layout::TILE)
         )doc",
-        ttnn::empty.base_name());
+        ttnn::empty.base_name(),
+        info_doc);
 
     using EmptyType = decltype(ttnn::empty);
     bind_registered_operation(
@@ -414,16 +426,40 @@ void bind_empty_like_operation(py::module& module) {
 
 void py_module(py::module& module) {
     detail::bind_full_operation(module, ttnn::full);
-    detail::bind_full_operation_with_hard_coded_value(module, ttnn::zeros, "0.0");
+    detail::bind_full_operation_with_hard_coded_value(module, ttnn::zeros, "0.0",
+    R"doc(Supported dtypes, layouts, and ranks:
+
+        +----------------------------+---------------------------------+-------------------+
+        |     Dtypes                 |         Layouts                 |     Ranks         |
+        +----------------------------+---------------------------------+-------------------+
+        |    BFLOAT16, FLOAT32       |       ROW_MAJOR, TILE           |      2, 3, 4      |
+        +----------------------------+---------------------------------+-------------------+)doc");
+
     detail::bind_full_operation_with_hard_coded_value(module, ttnn::ones, "1.0");
 
     detail::bind_full_like_operation(module, ttnn::full_like);
-    detail::bind_full_like_operation_with_hard_coded_value(module, ttnn::zeros_like, "0.0");
+    detail::bind_full_like_operation_with_hard_coded_value(module, ttnn::zeros_like, "0.0",
+    R"doc(Supported dtypes, layouts, and ranks:
+
+        +----------------------------+---------------------------------+-------------------+
+        |     Dtypes                 |         Layouts                 |     Ranks         |
+        +----------------------------+---------------------------------+-------------------+
+        |    BFLOAT16, FLOAT32       |       ROW_MAJOR, TILE           |      2, 3, 4      |
+        +----------------------------+---------------------------------+-------------------+)doc");
     detail::bind_full_like_operation_with_hard_coded_value(module, ttnn::ones_like, "1.0");
 
     detail::bind_arange_operation(module, ttnn::arange);
 
-    detail::bind_empty_operation(module);
+    detail::bind_empty_operation(module,
+    R"doc(Supported dtypes, layouts, and ranks:
+
+        +----------------------------+---------------------------------+-------------------+
+        |     Dtypes                 |         Layouts                 |     Ranks         |
+        +----------------------------+---------------------------------+-------------------+
+        |    BFLOAT16, FLOAT32       |       ROW_MAJOR, TILE           |      2, 3, 4      |
+        +----------------------------+---------------------------------+-------------------+
+        |    BFLOAT_8                |          TILE                   |      2, 3, 4      |
+        +----------------------------+---------------------------------+-------------------+)doc");
     detail::bind_empty_like_operation(module);
 }