diff options
| author | Georgios Pinitas <georgios.pinitas@arm.com> | 2020-03-05 15:05:00 +0000 |
|---|---|---|
| committer | Michele Di Giorgio <michele.digiorgio@arm.com> | 2020-03-09 10:16:15 +0000 |
| commit | b6f88b4cce83243068ee8d55a1ebc48567a5fb93 (patch) | |
| tree | c2551aa8e7fb4665771f55a6e2656fd0fe2424e4 | |
| parent | 25d9775b08c59be8d17700f75026a8457aab9838 (diff) | |
| download | ComputeLibrary-b6f88b4cce83243068ee8d55a1ebc48567a5fb93.tar.gz | |
COMPMID-3069: Avoid usage of vld2q in NEDepthConvert
Using vld2q could lead to problems on specific version of gcc.
This is to avoid register allocation issues on gcc 9.2.1
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I87fb9b55c105e0bac5d0be17befcac876c60197c
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2847
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
| -rw-r--r-- | arm_compute/core/NEON/NEColorConvertHelper.inl | 33 |
1 files changed, 28 insertions, 5 deletions
diff --git a/arm_compute/core/NEON/NEColorConvertHelper.inl b/arm_compute/core/NEON/NEColorConvertHelper.inl index 7145d6f206..2cf52e58d2 100644 --- a/arm_compute/core/NEON/NEColorConvertHelper.inl +++ b/arm_compute/core/NEON/NEColorConvertHelper.inl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2019 ARM Limited. + * Copyright (c) 2016-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -539,10 +539,32 @@ void colorconvert_iyuv_to_rgb(const void *__restrict input, void *__restrict out execute_window_loop(win, [&](const Coordinates &) { - const auto ta_y_top = vld2q_u8(in_y.ptr()); - const auto ta_y_bottom = vld2q_u8(in_y.ptr() + input_ptr->plane(0)->info()->strides_in_bytes().y()); - const auto ta_u = vld1q_u8(in_u.ptr()); - const auto ta_v = vld1q_u8(in_v.ptr()); + const auto *y_top_ptr = in_y.ptr(); + const auto *y_bottom_ptr = in_y.ptr() + input_ptr->plane(0)->info()->strides_in_bytes().y(); + const auto *u_ptr = in_u.ptr(); + const auto *v_ptr = in_v.ptr(); + + // Work-around issue in gcc 9(>=) where vld2q might cause issues with register allocation +#if defined(__arch64__) + const auto ta0_y_top = vld1q_u8(y_top_ptr); + const auto ta1_y_top = vld1q_u8(y_top_ptr + 16); + const auto ta0_y_bottom = vld1q_u8(y_bottom_ptr); + const auto ta1_y_bottom = vld1q_u8(y_bottom_ptr + 16); + const auto ta_u = vld1q_u8(u_ptr); + const auto ta_v = vld1q_u8(v_ptr); + + // Convert the uint8x16x4_t to float32x4x4_t + float32x4x4_t yvec_top = arm_compute::convert_uint8x16_to_float32x4x4(vuzp1q_u8(ta0_y_top, ta1_y_top)); + float32x4x4_t yyvec_top = arm_compute::convert_uint8x16_to_float32x4x4(vuzp2q_u8(ta0_y_top, ta1_y_top)); + float32x4x4_t yvec_bottom = arm_compute::convert_uint8x16_to_float32x4x4(vuzp1q_u8(ta0_y_bottom, ta1_y_bottom)); + float32x4x4_t yyvec_bottom = arm_compute::convert_uint8x16_to_float32x4x4(vuzp2q_u8(ta0_y_bottom, ta1_y_bottom)); + float32x4x4_t uvec = arm_compute::convert_uint8x16_to_float32x4x4(ta_u); + float32x4x4_t vvec = arm_compute::convert_uint8x16_to_float32x4x4(ta_v); +#else /* defined(__arch64__) */ + const auto ta_y_top = vld2q_u8(y_top_ptr); + const auto ta_y_bottom = vld2q_u8(y_bottom_ptr); + const auto ta_u = vld1q_u8(u_ptr); + const auto ta_v = vld1q_u8(v_ptr); //ta_y.val[0] = Y0 Y2 Y4 Y6 ... //ta_y.val[1] = Y1 Y3 Y5 Y7 ... //ta_u.val[0] = U0 U2 U4 U6 ... @@ -555,6 +577,7 @@ void colorconvert_iyuv_to_rgb(const void *__restrict input, void *__restrict out float32x4x4_t yyvec_bottom = arm_compute::convert_uint8x16_to_float32x4x4(ta_y_bottom.val[1]); float32x4x4_t uvec = arm_compute::convert_uint8x16_to_float32x4x4(ta_u); float32x4x4_t vvec = arm_compute::convert_uint8x16_to_float32x4x4(ta_v); +#endif /* defined(__arch64__) */ yuyv_to_rgb_calculation(yvec_top.val[0], uvec.val[0], yyvec_top.val[0], vvec.val[0], out.ptr() + 0 * element_size, alpha); yuyv_to_rgb_calculation(yvec_top.val[1], uvec.val[1], yyvec_top.val[1], vvec.val[1], out.ptr() + 1 * element_size, alpha); |