#11512: Add sweeps for eltwise sharded ops

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

@@ -39,10 +39,12 @@ on:
           - eltwise.unary.tril.tril_pytorch2
           - eltwise.unary.clamp.clamp
           - eltwise.unary.clamp.clamp_forge
+          - eltwise.unary.clamp.clamp_sharded
           - eltwise.unary.clamp.clamp_pytorch2
           - eltwise.unary.clamp.clamp_min_pytorch2
           - eltwise.unary.clip.clip
           - eltwise.unary.cbrt.cbrt
+          - eltwise.unary.cbrt.cbrt_sharded
           - eltwise.unary.rsub.rsub
           - eltwise.unary.rsub.rsub_pytorch2
           - eltwise.unary.rsqrt.rsqrt_pytorch2
@@ -56,6 +58,7 @@ on:
           - eltwise.unary.trunc.trunc_sharded
           - eltwise.unary.floor.floor
           - eltwise.unary.floor.floor_forge
+          - eltwise.unary.floor.floor_sharded
           - eltwise.unary.floor.floor_pytorch2
           - eltwise.unary.clone.clone
           - eltwise.unary.elu.elu
@@ -216,6 +219,7 @@ on:
           - eltwise.unary.isneginf.isneginf_sharded
           - eltwise.unary.isposinf.isposinf
           - eltwise.unary.isposinf.isposinf_sharded
+          - eltwise.binary.add.add_sharded
           - eltwise.binary.add.add_all_pytorch2
           - eltwise.binary.add.add_set2_pytorch2
           - eltwise.binary.add.add_different_memory_configs
@@ -235,14 +239,17 @@ on:
           - eltwise.unary_complex.angle_bw.angle_bw
           - eltwise.unary_complex.conj_bw
           - eltwise.binary.subtract.subtract
+          - eltwise.binary.subtract.subtract_sharded
           - eltwise.binary.subtract.subtract_tensor_pytorch2
           - eltwise.binary.multiply.multiply
           - eltwise.binary.multiply.mul_tensor_pytorch2
           - eltwise.binary.multiply.multiply_scalar_pytorch2
           - eltwise.binary.div.div
+          - eltwise.binary.div.div_sharded
           - eltwise.binary.div.div_tensor_pytorch2
           - eltwise.binary.div.div_forge
           - eltwise.binary.div_no_nan.div_no_nan
+          - eltwise.binary.div_no_nan.div_no_nan_sharded
           - eltwise.binary.logical_or.logical_or_
           - eltwise.binary.logical_or.logical_or
           - eltwise.binary.logical_or.logical_or_output

tests/sweep_framework/sweeps/eltwise/binary/add/add_sharded.py

@@ -0,0 +1,121 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional, Tuple
+from functools import partial
+
+import json
+import torch
+import random
+import ttnn
+import math
+from tests.sweep_framework.sweep_utils.utils import gen_shapes, sanitize_shape_rm
+from tests.sweep_framework.sweep_utils.sharding_utils import gen_sharded_spec_unary, parse_sharding_spec
+from tests.tt_eager.python_api_testing.sweep_tests.generation_funcs import gen_func_with_cast_tt
+from tests.tt_eager.python_api_testing.sweep_tests.generation_funcs import gen_rand_inf
+
+from tests.sweep_framework.sweep_utils.sharding_utils import (
+    gen_sharded_spec_unary,
+    parse_sharding_spec,
+    invalidate_vector_sharding,
+)
+from tests.ttnn.utils_for_testing import check_with_pcc, start_measuring_time, stop_measuring_time
+from models.utility_functions import torch_random
+
+# Override the default timeout in seconds for hang detection.
+TIMEOUT = 120
+
+random.seed(0)
+
+
+# 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_spec": gen_sharded_spec_unary(16, layouts=["TILE_LAYOUT"]),  # add op only supports tiled layout
+        "input_a_dtype": [ttnn.bfloat16, ttnn.bfloat8_b],
+        "input_b_dtype": [ttnn.bfloat16, ttnn.bfloat8_b],
+    },
+}
+
+
+# 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]]:
+    input_layout = test_vector["input_spec"]["input_layout"]
+    sharding_invalidated, output_str = invalidate_vector_sharding(test_vector["input_spec"])
+
+    if input_layout == "ROW_MAJOR_LAYOUT":
+        return True, "Inputs to eltwise binary must be tilized"
+    if sharding_invalidated:
+        return sharding_invalidated, output_str
+    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_spec,
+    input_a_dtype,
+    input_b_dtype,
+    *,
+    device,
+) -> list:
+    data_seed = random.randint(0, 20000000)
+    torch.manual_seed(data_seed)
+
+    (
+        input_shape,
+        core_grid,
+        sharding_strategy,
+        shard_orientation,
+        tensor_hw_as_shard_shape,
+        input_layout,
+        shard_height_mul_of_32,
+    ) = parse_sharding_spec(input_spec)
+    torch_input_tensor_a = gen_func_with_cast_tt(
+        partial(torch_random, low=-100, high=100, dtype=torch.float32), input_a_dtype
+    )(input_shape)
+    torch_input_tensor_b = gen_func_with_cast_tt(
+        partial(torch_random, low=-100, high=100, dtype=torch.float32), input_b_dtype
+    )(input_shape)
+    golden_function = ttnn.get_golden_function(ttnn.add)
+    torch_output_tensor = golden_function(torch_input_tensor_a, torch_input_tensor_b)
+
+    sharded_config = ttnn.create_sharded_memory_config_(
+        shape=input_shape,
+        core_grid=core_grid,
+        strategy=sharding_strategy,
+        orientation=shard_orientation,
+        use_height_and_width_as_shard_shape=tensor_hw_as_shard_shape,
+        tile_layout=shard_height_mul_of_32,
+    )
+    input_tensor_a = ttnn.from_torch(
+        torch_input_tensor_a,
+        dtype=input_a_dtype,
+        layout=input_layout,
+        device=device,
+        memory_config=sharded_config,
+    )
+
+    input_tensor_b = ttnn.from_torch(
+        torch_input_tensor_b,
+        dtype=input_b_dtype,
+        layout=input_layout,
+        device=device,
+        memory_config=sharded_config,
+    )
+
+    start_time = start_measuring_time()
+    output_tensor = ttnn.add(input_tensor_a, input_tensor_b, memory_config=sharded_config)
+    e2e_perf = stop_measuring_time(start_time)
+    output_tensor = ttnn.to_torch(output_tensor)
+
+    pcc = check_with_pcc(torch_output_tensor, output_tensor, 0.999)
+    return [pcc, e2e_perf]

tests/sweep_framework/sweeps/eltwise/binary/div/div_sharded.py

@@ -0,0 +1,141 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional, Tuple
+from functools import partial
+
+import json
+import torch
+import random
+import ttnn
+import math
+
+import ttnn.device
+from tests.sweep_framework.sweep_utils.utils import gen_shapes, sanitize_shape_rm
+from tests.sweep_framework.sweep_utils.sharding_utils import gen_sharded_spec_unary, parse_sharding_spec
+from tests.tt_eager.python_api_testing.sweep_tests.generation_funcs import gen_func_with_cast_tt
+from tests.tt_eager.python_api_testing.sweep_tests.generation_funcs import gen_rand_inf
+
+from tests.sweep_framework.sweep_utils.sharding_utils import (
+    gen_sharded_spec_unary,
+    parse_sharding_spec,
+    invalidate_vector_sharding,
+)
+from tests.ttnn.utils_for_testing import check_with_pcc, start_measuring_time, stop_measuring_time
+from models.utility_functions import torch_random
+
+# Override the default timeout in seconds for hang detection.
+TIMEOUT = 120
+
+random.seed(0)
+
+
+# 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_spec": gen_sharded_spec_unary(2, layouts=["TILE_LAYOUT"]),  # div op only supports tiled layout
+        "input_a_dtype": [ttnn.bfloat16],
+        "input_b_dtype": [ttnn.bfloat16],
+        "accurate_mode": [True, False],
+        "round_mode": [None, "floor", "trunc"],
+    },
+}
+
+
+# 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]]:
+    input_layout = test_vector["input_spec"]["input_layout"]
+    sharding_invalidated, output_str = invalidate_vector_sharding(test_vector["input_spec"])
+
+    if input_layout == "ROW_MAJOR_LAYOUT":
+        return True, "Inputs to eltwise binary must be tilized"
+    if sharding_invalidated:
+        return sharding_invalidated, output_str
+    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_spec,
+    input_a_dtype,
+    input_b_dtype,
+    accurate_mode,
+    round_mode,
+    *,
+    device,
+) -> list:
+    data_seed = random.randint(0, 20000000)
+    torch.manual_seed(data_seed)
+
+    (
+        input_shape,
+        core_grid,
+        sharding_strategy,
+        shard_orientation,
+        tensor_hw_as_shard_shape,
+        input_layout,
+        shard_height_mul_of_32,
+    ) = parse_sharding_spec(input_spec)
+
+    torch_input_tensor_a = gen_func_with_cast_tt(
+        partial(torch_random, low=-100, high=100, dtype=torch.float32), input_a_dtype
+    )(input_shape)
+
+    if accurate_mode == False:
+        torch_input_tensor_b = gen_func_with_cast_tt(
+            partial(torch_random, low=0.1, high=100, dtype=torch.float32), input_b_dtype
+        )(input_shape)
+        signs_b = torch.randint(0, 2, input_shape) * 2 - 1
+        torch_input_tensor_b *= signs_b
+    else:
+        torch_input_tensor_b = gen_func_with_cast_tt(
+            partial(torch_random, low=-100, high=100, dtype=torch.float32), input_b_dtype
+        )(input_shape)
+
+    golden_function = ttnn.get_golden_function(ttnn.div)
+    torch_output_tensor = golden_function(torch_input_tensor_a, torch_input_tensor_b)
+
+    sharded_config = ttnn.create_sharded_memory_config_(
+        shape=input_shape,
+        core_grid=core_grid,
+        strategy=sharding_strategy,
+        orientation=shard_orientation,
+        use_height_and_width_as_shard_shape=tensor_hw_as_shard_shape,
+        tile_layout=shard_height_mul_of_32,
+    )
+
+    input_tensor_a = ttnn.from_torch(
+        torch_input_tensor_a,
+        dtype=input_a_dtype,
+        layout=input_layout,
+        device=device,
+        memory_config=sharded_config,
+    )
+
+    input_tensor_b = ttnn.from_torch(
+        torch_input_tensor_b,
+        dtype=input_b_dtype,
+        layout=input_layout,
+        device=device,
+        memory_config=sharded_config,
+    )
+
+    start_time = start_measuring_time()
+    #
+    output_tensor = ttnn.div(
+        input_tensor_a, input_tensor_b, accurate_mode=accurate_mode, round_mode=round_mode, memory_config=sharded_config
+    )
+    e2e_perf = stop_measuring_time(start_time)
+    output_tensor = ttnn.to_torch(output_tensor)
+
+    pcc = check_with_pcc(torch_output_tensor, output_tensor, 0.999)
+    return [pcc, e2e_perf]