Optimize Quantized/Integer Bilinear Scale for Neon™
This patch introduces several performance optimizations regarding the Bilinear Scale operator with REPLICATE Border mode. Changes apply only to NHWC.
This patch
- Reduces the memory footprint by disabling precomputation of indices and weights when they're not used
- Rewrites the kernels for QASYMM8/QASYMM8_SIGNED/U8(Uint8)
- Adds S8(Int8) Bilinear Scale for Border mode REPLICATE
- Removes Bilinear Scale SVE kernels for Quantized and Integer types and adjust the heuristics to choose the Neon™ implementation
- Adds new test cases where the input and output of the Bilinear Scale operator have different quantization scale and offset
Resolves: COMPMID-5453, COMPMID-5454
Change-Id: I3d251e76e0c6978fd5a0a1795ec62ab536bec93c
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8250
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/cpu/kernels/scale/sve/qasymm8_signed.cpp b/src/cpu/kernels/scale/sve/qasymm8_signed.cpp
index 63f5154..67bca65 100644
--- a/src/cpu/kernels/scale/sve/qasymm8_signed.cpp
+++ b/src/cpu/kernels/scale/sve/qasymm8_signed.cpp
@@ -83,108 +83,6 @@
},
out);
}
-
-void qasymm8_signed_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
- BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
- bool align_corners, const Window &window)
-{
- // Data layout is NHWC
- const int idx_width = 1;
- const int idx_height = 2;
-
- // Compute the ratio between source height and destination height
- const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(idx_height), dst->info()->dimension(idx_height), align_corners);
- Window win_off;
- win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
- win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
-
- // Don't increment in X and Y direction for the input tensor
- // A pointer to the start of this plane is needed as base for the precomputed offsets
- Window win_in(window);
- win_in.set(idx_width, Window::Dimension(0, 0, 0));
- win_in.set(idx_height, Window::Dimension(0, 0, 0));
-
- for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
- {
- win_off.set(d, Window::Dimension(0, 0, 0));
- }
-
- Iterator in(src, win_in);
- Iterator out(dst, window);
-
- const int32_t in_dim_w = src->info()->dimension(idx_width);
- const int32_t in_dim_h = src->info()->dimension(idx_height);
- const int32_t stride_w = src->info()->strides_in_bytes()[idx_width];
- const int32_t stride_h = src->info()->strides_in_bytes()[idx_height];
-
- const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
- const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
-
- if(border_mode == BorderMode::CONSTANT)
- {
- const int8_t const_border_value = static_cast<int8_t>(constant_border_value.get<int8_t>());
- execute_window_loop(window, [&](const Coordinates & id)
- {
- const int32_t index_h = std::floor((id[idx_height] + sampling_offset) * hr - sampling_offset);
- const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
- const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
- const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
- const auto pixel_row_ptr = reinterpret_cast<const int8_t *>(in.ptr());
-
- const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ?
- (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
- const_border_value;
- const auto a01 = (-1 <= index_w && index_w < in_dim_w - 1 && 0 <= index_h && index_h < in_dim_h) ?
- (*(pixel_row_ptr + (index_w + 1) * stride_w + index_h * stride_h)) :
- const_border_value;
- const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
- (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
- const_border_value;
- const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
- (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
- const_border_value;
-
- const float inp00 = Qasymm8QuantizationHelper<int8_t>::dequantize(a00, iq_info);
- const float inp01 = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info);
- const float inp10 = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info);
- const float inp11 = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info);
- *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
- },
- in, out);
- }
- else if(border_mode == BorderMode::REPLICATE)
- {
- execute_window_loop(window, [&](const Coordinates & id)
- {
- const int index_h = std::floor((id[idx_height] + sampling_offset) * hr - sampling_offset);
- const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
- const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
- const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
- const auto pixel_row_ptr = reinterpret_cast<const int8_t *>(in.ptr());
-
- auto clamped_w = utility::clamp<int>(index_w, 0, in_dim_w - 1);
- auto clamped_w1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
- auto clamped_h = utility::clamp<int>(index_h, 0, in_dim_h - 1);
- auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
-
- const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
- const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
- const auto a10 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h1 * stride_h);
- const auto a11 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h1 * stride_h);
-
- const float inp00 = Qasymm8QuantizationHelper<int8_t>::dequantize(a00, iq_info);
- const float inp01 = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info);
- const float inp10 = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info);
- const float inp11 = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info);
- *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
- },
- in, out);
- }
- else
- {
- ARM_COMPUTE_ERROR("Not implemented");
- }
-}
}
namespace cpu
{
@@ -192,14 +90,15 @@
InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
bool align_corners, const Window &window)
{
- if(policy == InterpolationPolicy::BILINEAR)
- {
- qasymm8_signed_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
- }
- else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ ARM_COMPUTE_UNUSED(dx, dy, border_mode, constant_border_value);
+ if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
{
qasymm8_signed_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
}
+ else
+ {
+ ARM_COMPUTE_ERROR("Not Implemented");
+ }
}
} // namespace cpu
} // namespace arm_compute