Skip to content

Commit

Permalink
#14466: cleanup unary composite (#14682)
Browse files Browse the repository at this point in the history 
### Ticket
Link to Github Issue #14466

### Problem description
Use binary tensor-scalar support in composite ops instead of creation
ops

### What's changed
wherever binary tensor-scalar support is possible, removed usage of
creation ops

on main
op,count,python min dispatch time (ms),python mean dispatch
time(ms),python mean dispatch + sync time (ms),C++ mean dispatch time
(ms)
ttnn.cosh,800,0.178,0.184,0.71,0.061
ttnn.hardsigmoid,800,0.185,0.212,0.561,0.056
ttnn.lgamma,120,1.279,1.403,5.029,0.493
ttnn.sinh,800,0.179,0.191,0.71,0.062

vs 
on branch
ttnn.cosh,800,0.137,0.144,0.576,0.051
ttnn.hardsigmoid,800,0.091,0.097,0.29,0.031
ttnn.lgamma,120,1.208,1.225,4.953,0.464
ttnn.sinh,800,0.136,0.141,0.576,0.05

### Checklist
- [x] Post commit CI passes
https://github.com/tenstorrent/tt-metal/actions/runs/11741081069
https://github.com/tenstorrent/tt-metal/actions/runs/11766921873
- [x] Nightly FD
https://github.com/tenstorrent/tt-metal/actions/runs/11741696706
https://github.com/tenstorrent/tt-metal/actions/runs/11766922914
- [ ] Blackhole Post commit (if applicable)
- [ ] Model regression CI testing passes (if applicable)
- [ ] Device performance regression CI testing passes (if applicable)
- [x] New/Existing tests provide coverage for changes
  • Loading branch information
@KalaivaniMCW
KalaivaniMCW authored Nov 10, 2024
1 parent b75f637 commit 486862f
Show file tree
Hide file tree
Showing 3 changed files with 15 additions and 28 deletions.
14 changes: 6 additions & 8 deletions tests/ttnn/unit_tests/operations/eltwise/test_ternary.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -11,10 +11,6 @@
from tests.ttnn.utils_for_testing import assert_with_pcc


def torch_mac(input, tensor1, tensor2):
return torch.add(torch.mul(input, tensor1), tensor2)


@pytest.mark.parametrize("h", [64])
@pytest.mark.parametrize("w", [128])
def test_mac_all_tensors(device, h, w):
Expand All @@ -23,7 +19,9 @@ def test_mac_all_tensors(device, h, w):
torch_input_tensor = torch.rand((h, w), dtype=torch.bfloat16)
torch_input_tensor1 = torch.rand((h, w), dtype=torch.bfloat16)
torch_input_tensor2 = torch.rand((h, w), dtype=torch.bfloat16)
torch_output_tensor = torch_mac(torch_input_tensor, torch_input_tensor1, torch_input_tensor2)

golden_fn = ttnn.get_golden_function(ttnn.mac)
torch_output_tensor = golden_fn(torch_input_tensor, torch_input_tensor1, torch_input_tensor2)

input_tensor = ttnn.from_torch(torch_input_tensor, layout=ttnn.TILE_LAYOUT, device=device)
input_tensor = ttnn.to_device(input_tensor, device)
Expand All @@ -49,9 +47,9 @@ def test_mac_tensor_with_2_scalaras(device, h, w, scalar1, scalar2):
torch_input_tensor = torch.rand((h, w), dtype=torch.bfloat16)
torch_input_tensor1 = scalar1
torch_input_tensor2 = scalar2
torch_output_tensor = torch.unsqueeze(
torch.unsqueeze(torch_mac(torch_input_tensor, torch_input_tensor1, torch_input_tensor2), 0), 0
)

golden_fn = ttnn.get_golden_function(ttnn.mac)
torch_output_tensor = golden_fn(torch_input_tensor, torch_input_tensor1, torch_input_tensor2)

input_tensor = ttnn.from_torch(torch_input_tensor, layout=ttnn.TILE_LAYOUT, device=device)
input_tensor = ttnn.to_device(input_tensor, device)
Expand Down
8 changes: 2 additions & 6 deletions ttnn/cpp/ttnn/operations/eltwise/ternary/ternary_composite_op.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -106,13 +106,9 @@ Tensor _mac(const Tensor& a, const Tensor& b, const Tensor& c, const std::option
return ttnn::add(ttnn::multiply(a, b), c);
}

// y = a * b + c
Tensor _mac_overload(const Tensor& a, float b, float c, const std::optional<MemoryConfig>& output_mem_config) {
Tensor t_b = ttnn::operations::creation::create_scalar(b, a.get_dtype(), Layout::TILE, a.device());
Tensor t_c = ttnn::operations::creation::create_scalar(c, a.get_dtype(), Layout::TILE, a.device());
Tensor return_tensor = _mac(a, t_b, t_c, output_mem_config);
t_b.deallocate();
t_c.deallocate();
return return_tensor;
return ttnn::add(ttnn::multiply(a, b, std::nullopt, output_mem_config), c, std::nullopt, output_mem_config);
}

} // namespace ttnn::operations::ternary
21 changes: 7 additions & 14 deletions ttnn/cpp/ttnn/operations/eltwise/unary/device/unary_composite_op.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -173,8 +173,7 @@ Tensor _cosh(const Tensor& input_a, const std::optional<MemoryConfig>& output_me
Tensor nr_term = ttnn::add(e_pos_x, e_neg_x, std::nullopt, output_mem_config);
e_pos_x.deallocate();
e_neg_x.deallocate();
Tensor scalar = ttnn::full_like(input_a, 0.5f);
return ttnn::multiply(nr_term, scalar, std::nullopt, output_mem_config);
return ttnn::multiply(nr_term, 0.5f, std::nullopt, output_mem_config);
}

// TODO: In future will uplift the op once the floor and tan has supported.
Expand Down Expand Up @@ -294,12 +293,10 @@ Tensor _lgamma(const Tensor& x, const std::optional<MemoryConfig>& output_mem_c
result = ttnn::subtract(result, t, std::nullopt, output_mem_config);
{
{
Tensor t_one = ttnn::full_like(x, 1.0f);
result = ttnn::where(ttnn::eq(x, t_one, std::nullopt, output_mem_config), 0.0f, result);
result = ttnn::where(ttnn::eq(x, 1.0f, std::nullopt, output_mem_config), 0.0f, result);
}
{
Tensor t_two = ttnn::full_like(x, 2.0f);
result = ttnn::where(ttnn::eq(x, t_two, std::nullopt, output_mem_config), 0.0f, result);
result = ttnn::where(ttnn::eq(x, 2.0f, std::nullopt, output_mem_config), 0.0f, result);
}
}
}
Expand All @@ -309,8 +306,7 @@ Tensor _lgamma(const Tensor& x, const std::optional<MemoryConfig>& output_mem_c
// log1p 1
// use transformation y = log(1.0 + x) by broadcast
Tensor _log1p(const Tensor& x, const std::optional<MemoryConfig>& output_mem_config) {
Tensor t_one = ttnn::full_like(x, 1.0f);
Tensor x_1 = ttnn::add(t_one, x, std::nullopt, output_mem_config);
Tensor x_1 = ttnn::add(x, 1.0f, std::nullopt, output_mem_config);
Tensor result_log1p = ttnn::log(x_1, output_mem_config);
return result_log1p;
}
Expand Down Expand Up @@ -350,8 +346,7 @@ Tensor _sinh(const Tensor& input_a, const std::optional<MemoryConfig>& output_me
Tensor nr_term = ttnn::subtract(e_pos_x, e_neg_x, std::nullopt, output_mem_config);
e_pos_x.deallocate();
e_neg_x.deallocate();
Tensor scalar = ttnn::full_like(input_a, 0.5f);
return ttnn::multiply(nr_term, scalar, std::nullopt, output_mem_config);
return ttnn::multiply(nr_term, 0.5f, std::nullopt, output_mem_config);
}

// Function: softsign
Expand All @@ -372,8 +367,8 @@ Tensor _swish(const Tensor& a, const std::optional<MemoryConfig>& output_mem_con

Tensor ExecuteTrunc::invoke(uint8_t queue_id, const Tensor& input, const std::optional<MemoryConfig>& output_mem_config, std::optional<Tensor> output_tensor) {
auto arch = input.device()->arch();
output_tensor = output_tensor.value_or(ttnn::empty_like(input));
TT_FATAL(arch != tt::ARCH::GRAYSKULL, "Op is not supported on Grayskull");
output_tensor = output_tensor.value_or(ttnn::empty_like(input));
Tensor floor_res = ttnn::floor(queue_id, input, output_mem_config);
ttnn::where(queue_id, ttnn::ne(queue_id, input, floor_res), ttnn::add(queue_id, floor_res, 1.0f, std::nullopt, output_mem_config), floor_res, output_mem_config, output_tensor);
ttnn::where(queue_id, ttnn::gtz(queue_id, input, output_mem_config), floor_res, output_tensor.value(), output_mem_config, output_tensor);
Expand Down Expand Up @@ -449,9 +444,7 @@ Tensor _normalize(const Tensor& y, const std::optional<MemoryConfig>& output_mem
// PyTorch version:
// hard sigmoid(x) = { x <= -3: 0, x >= +3: +3, x/6 + 0.5 otherwise}
Tensor _hardsigmoid(const Tensor& a, float value_1, float value_2, const std::optional<MemoryConfig>& output_mem_config) {
Tensor a_t = ttnn::full_like(a,value_1);
Tensor b_t = ttnn::full_like(a,value_2);
Tensor a_mac = ttnn::mac(a, a_t, b_t); // multiply and add.
Tensor a_mac = ttnn::mac(a, value_1, value_2); // multiply and add.
Tensor a_clip = relu_max(a_mac, 1.0f);
return a_clip;
}
Expand Down

0 comments on commit 486862f

Please sign in to comment.