Update on "[ET-VK] Using push constants for unary op."

This diff transitions the unary op to utilize push constants, replacing the previous ubo implementation.

Differential Revision: [D77706459](https://our.internmc.facebook.com/intern/diff/D77706459/)

[ghstack-poisoned]

Author: trivedivivek

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-7cda4017ddda554752e89069ae205be5e8388f59
+9b498d3bb28b8e3411ce464dd2755c5b96d92c8f

.ci/scripts/check_c10_sync.sh

@@ -12,4 +12,4 @@ pushd pytorch
 git checkout "$pytorch_pin"
 popd
 "$(dirname "${BASH_SOURCE[0]}")"/compare_dirs.sh runtime/core/portable_type/c10/c10 pytorch/c10
-"$(dirname "${BASH_SOURCE[0]}")"/compare_dirs.sh runtime/core/portable_type/c10/torch/headeronly pytorch/torch/headeronly
+"$(dirname "${BASH_SOURCE[0]}")"/compare_dirs.sh runtime/core/portable_type/c10/torch/standalone pytorch/torch/standalone

backends/arm/_passes/arm_pass_manager.py

@@ -6,6 +6,8 @@
 # LICENSE file in the root directory of this source tree.
 
 # pyre-unsafe
+
+import executorch.backends.arm.tosa.dialect  # noqa: unused
 from executorch.backends.arm._passes import (
     AddBiasPass,
     AnnotateChannelsLastDimOrder,

backends/arm/test/passes/test_decorate_fp32_to_int32_casting_pass.py

@@ -6,7 +6,7 @@
 from typing import Tuple
 
 import torch
-from executorch.backends.arm.test import common
+from executorch.backends.arm.test import common, conftest
 
 from executorch.backends.arm.test.tester.test_pipeline import (
     OpNotSupportedPipeline,
@@ -55,6 +55,7 @@ def test_decorate_fp32_to_int32_casting_tosa_MI(test_data: Tuple):
         (test_tensor,),
         aten_op=[],
         exir_op=[],
+        run_on_tosa_ref_model=conftest.is_option_enabled("tosa_ref_model"),
     )
     pipeline.run()
 

backends/cadence/aot/functions_hifi.yaml

@@ -75,7 +75,7 @@
 - op: clamp.Tensor_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::clamp_tensor_out
+      kernel_name: cadence::impl::HiFi::clamp_Tensor_out
 
 - op: clone.out
   kernels:
@@ -100,7 +100,7 @@
 - op: eq.Tensor_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::eq_tensor_out
+      kernel_name: cadence::impl::HiFi::eq_Tensor_out
 
 - op: fmod.Tensor_out
   kernels:
@@ -120,12 +120,12 @@
 - op: ge.Scalar_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::ge_scalar_out
+      kernel_name: cadence::impl::HiFi::ge_Scalar_out
 
 - op: ge.Tensor_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::ge_tensor_out
+      kernel_name: cadence::impl::HiFi::ge_Tensor_out
 
 - op: gelu.out
   kernels:
@@ -135,12 +135,12 @@
 - op: gt.Scalar_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::gt_scalar_out
+      kernel_name: cadence::impl::HiFi::gt_Scalar_out
 
 - op: gt.Tensor_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::gt_tensor_out
+      kernel_name: cadence::impl::HiFi::gt_Tensor_out
 
 - op: hardtanh.out
   kernels:
@@ -150,27 +150,27 @@
 - op: le.Scalar_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::le_scalar_out
+      kernel_name: cadence::impl::HiFi::le_Scalar_out
 
 - op: le.Tensor_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::le_tensor_out
+      kernel_name: cadence::impl::HiFi::le_Tensor_out
 
 - op: lt.Scalar_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::lt_scalar_out
+      kernel_name: cadence::impl::HiFi::lt_Scalar_out
 
 - op: lt.Tensor_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::lt_tensor_out
+      kernel_name: cadence::impl::HiFi::lt_Tensor_out
 
 - op: masked_fill.Scalar_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::masked_fill_scalar_out
+      kernel_name: cadence::impl::HiFi::masked_fill_Scalar_out
 
 - op: max_pool2d_with_indices.out
   kernels:
@@ -185,7 +185,7 @@
 - op: mean.out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::mean_out   
+      kernel_name: cadence::impl::HiFi::mean_out
 
 - op: minimum.out
   kernels:
@@ -205,7 +205,7 @@
 - op: ne.Tensor_out
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::ne_tensor_out
+      kernel_name: cadence::impl::HiFi::ne_Tensor_out
 
 - op: permute_copy.out
   kernels:
@@ -289,11 +289,11 @@
   kernels:
     - arg_meta: null
       kernel_name: cadence::impl::HiFi::dequantize_per_tensor_out
-      
+
 - func: cadence::quantized_conv.out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, Tensor weight_zero_point, Tensor bias_scale, float out_scale, int out_zero_point, Tensor out_multiplier, Tensor out_shift, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
     - arg_meta: null
-      kernel_name: cadence::impl::HiFi::quantized_conv_out      
+      kernel_name: cadence::impl::HiFi::quantized_conv_out
 
 - func: cadence::quantized_layer_norm.out(Tensor input, Tensor in_scale, Tensor in_zero_point, int[] normalized_shape, Tensor weight, Tensor bias, float eps, float output_scale, int output_zero_point, *, Tensor(a!) out) -> Tensor(a!)
   kernels:

backends/cadence/fusion_g3/operators/op_exp.cpp

@@ -60,7 +60,7 @@ Tensor& exp_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
     return out;
   } else {
     return torch::executor::native::internal::
-        unary_ufunc_realhbbf16_to_floathbf16(std::exp, ctx, in, out);
+        unary_ufunc_realhbbf16_to_floathbf16(std::exp, std::exp, ctx, in, out);
   }
 }
 

backends/cadence/fusion_g3/operators/op_rsqrt.cpp

@@ -27,7 +27,8 @@ namespace native {
 
 namespace {
 
-double rsqrt(double x) {
+template <typename T>
+T rsqrt(T x) {
   return 1.0 / std::sqrt(x);
 }
 
@@ -61,11 +62,11 @@ Tensor& rsqrt_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
     return out;
   } else {
     return torch::executor::native::internal::
-        unary_ufunc_realhbbf16_to_floathbf16(rsqrt, ctx, in, out);
+        unary_ufunc_realhbbf16_to_floathbf16(rsqrt, rsqrt, ctx, in, out);
   }
 }
 
 } // namespace native
 } // namespace G3
 } // namespace impl
-} // namespace cadence
+} // namespace cadence

backends/cadence/fusion_g3/operators/op_sqrt.cpp

@@ -55,7 +55,8 @@ Tensor& sqrt_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
     return out;
   } else {
     return torch::executor::native::internal::
-        unary_ufunc_realhbbf16_to_floathbf16(std::sqrt, ctx, in, out);
+        unary_ufunc_realhbbf16_to_floathbf16(
+            std::sqrt, std::sqrt, ctx, in, out);
   }
 }
 

backends/cadence/fusion_g3/operators/op_tanh.cpp

@@ -55,7 +55,8 @@ Tensor& tanh_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
     return out;
   } else {
     return torch::executor::native::internal::
-        unary_ufunc_realhbbf16_to_floathbf16(std::tanh, ctx, in, out);
+        unary_ufunc_realhbbf16_to_floathbf16(
+            std::tanh, std::tanh, ctx, in, out);
   }
 }
 

backends/cadence/hifi/operators/CMakeLists.txt

@@ -88,7 +88,7 @@ target_include_directories(
 
 # Custom ops that are needed to run the test model.
 add_library(
-  custom_ops "op_quantized_linear_out.cpp" "op_quantized_layer_norm.cpp" "quantized_matmul_out.cpp"
+  custom_ops "op_quantized_linear_out.cpp" "op_quantized_layer_norm.cpp" "op_quantized_matmul_out.cpp"
              "op_quantize_per_tensor.cpp" "op_quantized_relu_out.cpp" "op_dequantize_per_tensor.cpp"
              "op_quantized_conv_out.cpp" "op_quantized_fully_connected_out"
 )