NNlib update to latest version for HiFi1 build (pytorch#18112)

Differential Revision: D96245585

Pull Request resolved: pytorch#18112

Author: zonglinpeng

backends/cadence/aot/functions_hifi.yaml

@@ -286,12 +286,12 @@
 - func: cadence::im2row.out(Tensor input, int[2] kernel_size, int[2] dilation, int[2] padding, int[2] stride, Tensor in_zero_point, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
     - arg_meta: null
-      kernel_name: impl::HiFi::native::im2row_out
+      kernel_name: impl::HiFi::im2row_out
 
 - func: cadence::im2row.per_tensor_out(Tensor input, int[2] kernel_size, int[2] dilation, int[2] padding, int[2] stride, int in_zero_point, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
     - arg_meta: null
-      kernel_name: impl::HiFi::native::im2row_per_tensor_out
+      kernel_name: impl::HiFi::im2row_per_tensor_out
 
 - func: cadence::quantize_per_tensor.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
   variants: function

backends/cadence/hifi/kernels/CMakeLists.txt

@@ -8,25 +8,15 @@
 add_library(
   cadence_kernels
   kernels.cpp
-  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/matmul_asym8uxasym8u_asym8u.cpp
-  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_broadcast_32.c
-  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_concat_32.c
-  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_add_f32_broadcast.c
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_atan2_f32.c
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_clamp_f32_broadcast.c
-  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_div_f32_broadcast.c
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_div_mode_f32_broadcast.c
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_fmod_broadcast_f32.c
-  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_greater_lesser_equal_f32.c
-  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_im2row.c
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_logicalxor_bool_bool.c
-  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_minimum_maximum_f32.c
-  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_mul_f32_broadcast.c
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_pow_f32.c
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_remainder_broadcast_f32.c
-  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_elm_where_f32xf32_f32.c
+  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_im2row.c
   ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_reduce_32_32.c
-  ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/xa_nn_transpose_32.c
 )
 # Let files say "include <executorch/path/to/header.h>".
 set(_common_include_directories

backends/cadence/hifi/kernels/kernels.h

@@ -18,24 +18,9 @@ using executorch::runtime::Result;
 
 /* Potential NNLIB function/APIs */
 
-extern "C" WORD32 xa_nn_elm_add_broadcast_4D_f32xf32_f32(
-    FLOAT32* __restrict__ p_out,
-    const WORD32* const p_out_shape,
-    const FLOAT32* __restrict__ p_inp1,
-    const WORD32* const p_inp1_shape,
-    const FLOAT32* __restrict__ p_inp2,
-    const WORD32* const p_inp2_shape);
-
 extern "C" void
 xa_nn_elm_atan2_f32(FLOAT32* z, const FLOAT32* y, const FLOAT32* x, WORD32 N);
 
-extern "C" WORD32 xa_nn_elm_clamp_f32xf32xf32_f32(
-    FLOAT32* __restrict__ p_out,
-    const FLOAT32* __restrict__ p_inp,
-    const FLOAT32* __restrict__ p_min,
-    const FLOAT32* __restrict__ p_max,
-    WORD32 num_elm);
-
 extern "C" WORD32 xa_nn_elm_clamp_broadcast_4D_f32Xf32xf32_f32(
     FLOAT32* __restrict__ p_out,
     const WORD32* const p_out_shape,
@@ -46,14 +31,6 @@ extern "C" WORD32 xa_nn_elm_clamp_broadcast_4D_f32Xf32xf32_f32(
     const FLOAT32* __restrict__ p_max,
     const WORD32* const p_max_shape);
 
-extern "C" WORD32 xa_nn_elm_div_broadcast_4D_f32xf32_f32(
-    FLOAT32* __restrict__ p_out,
-    const WORD32* const p_out_shape,
-    const FLOAT32* __restrict__ p_inp1,
-    const WORD32* const p_inp1_shape,
-    const FLOAT32* __restrict__ p_inp2,
-    const WORD32* const p_inp2_shape);
-
 extern "C" WORD32 xa_nn_elm_div_mode_f32xf32_f32(
     FLOAT32* __restrict__ p_out,
     const FLOAT32* __restrict__ p_inp1,
@@ -70,22 +47,6 @@ extern "C" WORD32 xa_nn_elm_div_mode_broadcast_4D_f32xf32_f32(
     const WORD32* const p_inp2_shape,
     WORD32 mode);
 
-extern "C" WORD32 xa_nn_elm_greater_lesser_equal_f32xf32_f32(
-    WORD8* __restrict__ p_out,
-    const FLOAT32* __restrict__ p_inp1,
-    const FLOAT32* __restrict__ p_inp2,
-    WORD32 num_elm,
-    WORD32 kernel_type);
-
-extern "C" WORD32 xa_nn_elm_greater_lesser_equal_broadcast_4D_f32xf32_f32(
-    WORD8* __restrict__ p_out,
-    const WORD32* const p_out_shape,
-    const FLOAT32* __restrict__ p_inp1,
-    const WORD32* const p_inp1_shape,
-    const FLOAT32* __restrict__ p_inp2,
-    const WORD32* const p_inp2_shape,
-    WORD32 kernel_type);
-
 extern "C" WORD32 xa_nn_elm_fmod_f32xf32_f32(
     FLOAT32* __restrict__ p_out,
     const FLOAT32* __restrict__ p_inp1,
@@ -106,42 +67,6 @@ extern "C" WORD32 xa_nn_elm_logicalxor_boolxbool_bool(
     const WORD8* __restrict__ p_inp2,
     WORD32 num_elm);
 
-extern "C" WORD32 xa_nn_elm_maximum_f32xf32_f32(
-    FLOAT32* __restrict__ p_out,
-    const FLOAT32* __restrict__ p_inp1,
-    const FLOAT32* __restrict__ p_inp2,
-    WORD32 num_elm);
-
-extern "C" WORD32 xa_nn_elm_maximum_broadcast_4D_f32xf32_f32(
-    FLOAT32* __restrict__ p_out,
-    const WORD32* const p_out_shape,
-    const FLOAT32* __restrict__ p_inp1,
-    const WORD32* const p_inp1_shape,
-    const FLOAT32* __restrict__ p_inp2,
-    const WORD32* const p_inp2_shape);
-
-extern "C" WORD32 xa_nn_elm_minimum_f32xf32_f32(
-    FLOAT32* __restrict__ p_out,
-    const FLOAT32* __restrict__ p_inp1,
-    const FLOAT32* __restrict__ p_inp2,
-    WORD32 num_elm);
-
-extern "C" WORD32 xa_nn_elm_minimum_broadcast_4D_f32xf32_f32(
-    FLOAT32* __restrict__ p_out,
-    const WORD32* const p_out_shape,
-    const FLOAT32* __restrict__ p_inp1,
-    const WORD32* const p_inp1_shape,
-    const FLOAT32* __restrict__ p_inp2,
-    const WORD32* const p_inp2_shape);
-
-extern "C" WORD32 xa_nn_elm_mul_broadcast_4D_f32xf32_f32(
-    FLOAT32* __restrict__ p_out,
-    const WORD32* const p_out_shape,
-    const FLOAT32* __restrict__ p_inp1,
-    const WORD32* const p_inp1_shape,
-    const FLOAT32* __restrict__ p_inp2,
-    const WORD32* const p_inp2_shape);
-
 extern "C" void xa_nn_elm_pow_f32(
     FLOAT32* __restrict__ z,
     const FLOAT32* __restrict__ x,
@@ -162,23 +87,6 @@ extern "C" WORD32 xa_nn_elm_remainder_broadcast_4D_f32xf32_f32(
     const FLOAT32* __restrict__ p_inp2,
     const WORD32* const p_inp2_shape);
 
-extern "C" WORD32 xa_nn_elm_where_f32xf32_f32(
-    FLOAT32* __restrict__ p_out,
-    const FLOAT32* __restrict__ p_inp1,
-    const FLOAT32* __restrict__ p_inp2,
-    const unsigned char* __restrict__ p_condition,
-    WORD32 num_elm);
-
-extern "C" WORD32 xa_nn_elm_where_broadcast_4D_f32xf32_f32(
-    FLOAT32* __restrict__ p_out,
-    const WORD32* const p_out_shape,
-    const FLOAT32* __restrict__ p_inp1,
-    const WORD32* const p_inp1_shape,
-    const FLOAT32* __restrict__ p_inp2,
-    const WORD32* const p_inp2_shape,
-    const unsigned char* __restrict__ p_condition,
-    const WORD32* const p_condition_shape);
-
 extern "C" WORD32 xa_nn_im2row_quantized(
     const WORD8* __restrict__ data_im,
     const WORD32 in_zero_point,
@@ -212,60 +120,12 @@ extern "C" WORD32 xa_nn_reduce_mean_4D_f32_f32(
     WORD32 num_axis_dims,
     void* __restrict__ p_scratch_in);
 
-extern "C" WORD32 xa_nn_transpose_32_32(
-    WORD32* __restrict__ p_out,
-    const WORD32* const p_out_shape,
-    const WORD32* __restrict__ p_inp,
-    const WORD32* const p_inp_shape,
-    const WORD32* __restrict__ p_permute_vec,
-    WORD32 num_out_dims,
-    WORD32 num_inp_dims);
-
 namespace impl {
 namespace HiFi {
 namespace kernels {
 
 void* allocate_temp_memory(KernelRuntimeContext& ctx, size_t size);
 
-void memcpy(void* dst, const void* src, size_t num_bytes);
-
-WORD32 matmul_asym8uxasym8u_asym8u(
-    UWORD8* __restrict__ p_out, // output uint8 matrix
-    const UWORD8* __restrict__ p_mat1, // weight uint8 matrix
-    const UWORD8* __restrict__ p_vec1, // input uint8 matrix
-    const WORD32* __restrict__ p_bias, // bias int32 vec
-    WORD32 rows, // rows of p_mat1
-    WORD32 cols1, // columns of p_mat1
-    WORD32 row_stride1, // row stride of p_mat1
-    WORD32 vec_count, // rows of p_mat2
-    WORD32 vec_offset, // vec_offset of p_mat2.
-    WORD32 out_offset, // out_offset, i.e., offset of next output element
-    WORD32 out_stride, // out_stride, i.e., stride to go to next output row
-    WORD32 mat1_zero_bias, // zero_point of p_mat1
-    WORD32 vec1_zero_bias, // zero_point of p_vec1
-    const WORD32* __restrict__ out_multiplier,
-    const WORD32* __restrict__ out_shift,
-    WORD32 out_zero_bias,
-    bool per_channel_quantized = false); // per-channel quantized weight
-
-WORD32 xa_nn_matmul_asym8uxasym8u_asym8u(
-    UWORD8* __restrict__ p_out,
-    const UWORD8* __restrict__ p_mat1,
-    const UWORD8* __restrict__ p_mat2,
-    const WORD32* __restrict__ p_bias,
-    WORD32 rows,
-    WORD32 cols,
-    WORD32 row_stride,
-    WORD32 vec_count,
-    WORD32 vec_offset,
-    WORD32 out_offset,
-    WORD32 out_stride,
-    WORD32 mat1_zero_bias,
-    WORD32 vec1_zero_bias,
-    WORD32 out_multiplier,
-    WORD32 out_shift,
-    WORD32 out_zero_bias);
-
 template <typename T>
 T quantize(const float x, float scale, int32_t zero_point);
 

backends/cadence/hifi/operators/op_clamp.cpp

@@ -155,13 +155,13 @@ Tensor& clamp_Tensor_out(
           inp_shape[i + off_inp] = in.size(i);
         }
 
-        WORD32 ret_val = xa_nn_elm_minimum_broadcast_4D_f32xf32_f32(
+        WORD32 ret_val = xa_nn_elm_min_4D_Bcast_f32xf32_f32(
             out_data, out_shape, inp_data, inp_shape, max_data, max_shape);
 
         ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
 
       } else {
-        WORD32 ret_val = xa_nn_elm_minimum_f32xf32_f32(
+        WORD32 ret_val = xa_nn_elm_min_f32xf32_f32(
             out_data, inp_data, max_data, out.numel());
 
         ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
@@ -192,13 +192,13 @@ Tensor& clamp_Tensor_out(
           min_shape[i + off_min] = min.size(i);
         for (int i = 0; i < inp_dim; i++)
           inp_shape[i + off_inp] = in.size(i);
-        WORD32 ret_val = xa_nn_elm_maximum_broadcast_4D_f32xf32_f32(
+        WORD32 ret_val = xa_nn_elm_max_4D_Bcast_f32xf32_f32(
             out_data, out_shape, inp_data, inp_shape, min_data, min_shape);
 
         ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
 
       } else {
-        WORD32 ret_val = xa_nn_elm_maximum_f32xf32_f32(
+        WORD32 ret_val = xa_nn_elm_max_f32xf32_f32(
             out_data, inp_data, min_data, out.numel());
 
         ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);

backends/cadence/hifi/operators/op_eq.cpp

@@ -94,15 +94,21 @@ Tensor& eq_Tensor_out(
       for (int i = 0; i < b.dim(); i++)
         inp2_shape[i + off_b] = b.size(i);
 
-      WORD32 ret_val = xa_nn_elm_greater_lesser_equal_broadcast_4D_f32xf32_f32(
-          p_out, out_shape, p_inp1, inp1_shape, p_inp2, inp2_shape, 4);
+      WORD32 ret_val = xa_nn_elm_compare_broadcast_4D_f32xf32_f32(
+          p_out,
+          out_shape,
+          p_inp1,
+          inp1_shape,
+          p_inp2,
+          inp2_shape,
+          COMPARE_EQUAL);
 
       ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
     } else {
       int num_elm = out.numel();
 
-      WORD32 ret_val = xa_nn_elm_greater_lesser_equal_f32xf32_f32(
-          p_out, p_inp1, p_inp2, num_elm, 4);
+      WORD32 ret_val = xa_nn_elm_compare_f32xf32_f32(
+          p_out, p_inp1, p_inp2, num_elm, COMPARE_EQUAL);
 
       ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
     }

backends/cadence/hifi/operators/op_ge.cpp

@@ -94,15 +94,21 @@ Tensor& ge_Tensor_out(
       for (int i = 0; i < b.dim(); i++)
         inp2_shape[i + off_b] = b.size(i);
 
-      WORD32 ret_val = xa_nn_elm_greater_lesser_equal_broadcast_4D_f32xf32_f32(
-          p_out, out_shape, p_inp1, inp1_shape, p_inp2, inp2_shape, 0);
+      WORD32 ret_val = xa_nn_elm_compare_broadcast_4D_f32xf32_f32(
+          p_out,
+          out_shape,
+          p_inp1,
+          inp1_shape,
+          p_inp2,
+          inp2_shape,
+          COMPARE_GREATEREQUAL);
 
       ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
     } else {
       int num_elm = out.numel();
 
-      WORD32 ret_val = xa_nn_elm_greater_lesser_equal_f32xf32_f32(
-          p_out, p_inp1, p_inp2, num_elm, 0);
+      WORD32 ret_val = xa_nn_elm_compare_f32xf32_f32(
+          p_out, p_inp1, p_inp2, num_elm, COMPARE_GREATEREQUAL);
 
       ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
     }

backends/cadence/hifi/operators/op_gt.cpp

@@ -96,15 +96,21 @@ Tensor& gt_Tensor_out(
       for (int i = 0; i < b.dim(); i++)
         inp2_shape[i + off_b] = b.size(i);
 
-      WORD32 ret_val = xa_nn_elm_greater_lesser_equal_broadcast_4D_f32xf32_f32(
-          p_out, out_shape, p_inp1, inp1_shape, p_inp2, inp2_shape, 1);
+      WORD32 ret_val = xa_nn_elm_compare_broadcast_4D_f32xf32_f32(
+          p_out,
+          out_shape,
+          p_inp1,
+          inp1_shape,
+          p_inp2,
+          inp2_shape,
+          COMPARE_GREATER);
 
       ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
     } else {
       int num_elm = out.numel();
 
-      WORD32 ret_val = xa_nn_elm_greater_lesser_equal_f32xf32_f32(
-          p_out, p_inp1, p_inp2, num_elm, 1);
+      WORD32 ret_val = xa_nn_elm_compare_f32xf32_f32(
+          p_out, p_inp1, p_inp2, num_elm, COMPARE_GREATER);
 
       ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
     }

backends/cadence/hifi/operators/op_le.cpp

@@ -95,15 +95,21 @@ Tensor& le_Tensor_out(
       for (int i = 0; i < b.dim(); i++)
         inp2_shape[i + off_b] = b.size(i);
 
-      WORD32 ret_val = xa_nn_elm_greater_lesser_equal_broadcast_4D_f32xf32_f32(
-          p_out, out_shape, p_inp1, inp1_shape, p_inp2, inp2_shape, 2);
+      WORD32 ret_val = xa_nn_elm_compare_broadcast_4D_f32xf32_f32(
+          p_out,
+          out_shape,
+          p_inp1,
+          inp1_shape,
+          p_inp2,
+          inp2_shape,
+          COMPARE_LESSEREQUAL);
 
       ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
     } else {
       int num_elm = out.numel();
 
-      WORD32 ret_val = xa_nn_elm_greater_lesser_equal_f32xf32_f32(
-          p_out, p_inp1, p_inp2, num_elm, 2);
+      WORD32 ret_val = xa_nn_elm_compare_f32xf32_f32(
+          p_out, p_inp1, p_inp2, num_elm, COMPARE_LESSEREQUAL);
 
       ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
     }

backends/cadence/hifi/operators/op_lt.cpp

@@ -93,15 +93,21 @@ Tensor& lt_Tensor_out(
       for (int i = 0; i < b.dim(); i++)
         inp2_shape[i + off_b] = b.size(i);
 
-      WORD32 ret_val = xa_nn_elm_greater_lesser_equal_broadcast_4D_f32xf32_f32(
-          p_out, out_shape, p_inp1, inp1_shape, p_inp2, inp2_shape, 3);
+      WORD32 ret_val = xa_nn_elm_compare_broadcast_4D_f32xf32_f32(
+          p_out,
+          out_shape,
+          p_inp1,
+          inp1_shape,
+          p_inp2,
+          inp2_shape,
+          COMPARE_LESSER);
 
       ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
     } else {
       int num_elm = out.numel();
 
-      WORD32 ret_val = xa_nn_elm_greater_lesser_equal_f32xf32_f32(
-          p_out, p_inp1, p_inp2, num_elm, 3);
+      WORD32 ret_val = xa_nn_elm_compare_f32xf32_f32(
+          p_out, p_inp1, p_inp2, num_elm, COMPARE_LESSER);
 
       ET_KERNEL_CHECK(ctx, ret_val == 0, Internal, out);
     }