diff options
Diffstat (limited to 'src/core/NEON/kernels/NEElementwiseOperationKernel.cpp')
| -rw-r--r-- | src/core/NEON/kernels/NEElementwiseOperationKernel.cpp | 46 |
1 files changed, 22 insertions, 24 deletions
diff --git a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp index 33457e1fca..0fe05d2044 100644 --- a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp +++ b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp @@ -142,9 +142,9 @@ inline ScalarType elementwise_arithm_op_scalar(const ScalarType &a, const Scalar } template <ArithmeticOperation op> -inline uint8_t elementwise_arithm_op_quantized_scalar(const float &a, const float &b, QuantizationInfo qinfo) +inline uint8_t elementwise_arithm_op_quantized_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo) { - return qinfo.quantize(elementwise_arithm_op_scalar<op>(a, b), RoundingPolicy::TO_NEAREST_UP); + return quantize_qasymm8(elementwise_arithm_op_scalar<op>(a, b), qinfo); } template <ArithmeticOperation op, typename VectorType> @@ -253,7 +253,7 @@ inline uint8_t elementwise_comp_op_scalar(const InputScalarType &a, const InputS } template <ComparisonOperation op> -inline uint8_t elementwise_comp_op_quantized_scalar(const float &a, const float &b, QuantizationInfo qinfo) +inline uint8_t elementwise_comp_op_quantized_scalar(const float &a, const float &b, UniformQuantizationInfo qinfo) { ARM_COMPUTE_UNUSED(qinfo); return elementwise_comp_op_scalar<op>(a, b); @@ -567,7 +567,7 @@ void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const } void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, - uint8_t (*scalar_func)(const float &, const float &, QuantizationInfo), + uint8_t (*scalar_func)(const float &, const float &, UniformQuantizationInfo), int (*broadcast_func)(int, int, int, const uint8_t *, float32x4x4_t, uint8_t *, int32x4_t, float32x4_t, float32x4_t, float32x4_t, const bool), int (*neon_func)(int, int, int, const uint8_t *, const uint8_t *, uint8_t *, @@ -587,12 +587,11 @@ void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *o const auto window_end_x = static_cast<int>(window.x().end()); const bool is_broadcast_across_x = (input1_win.x().step() == 0) || (input2_win.x().step() == 0); - const float output_scale = out->info()->quantization_info().scale; - const int output_offset = out->info()->quantization_info().offset; + const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform(); // Output quantization info (add 0.5 to round toward the nearest integer - 0.5 rounds away from zero) - const float32x4_t voffseto = vdupq_n_f32(output_offset + 0.5f); - const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_scale); + const float32x4_t voffseto = vdupq_n_f32(output_qinfo.offset + 0.5f); + const float32x4_t invvscaleo = vdupq_n_f32(1.f / output_qinfo.scale); if(is_broadcast_across_x) { @@ -603,8 +602,8 @@ void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *o const ITensor *broadcast_tensor = is_broadcast_input_2 ? in2 : in1; const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1; - const QuantizationInfo broadcast_qinfo = broadcast_tensor->info()->quantization_info(); - const QuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info(); + const UniformQuantizationInfo broadcast_qinfo = broadcast_tensor->info()->quantization_info().uniform(); + const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform(); const int32x4_t voffset_non_broadcast = vdupq_n_s32(non_broadcast_qinfo.offset); const float32x4_t vscale_non_broadcast = vdupq_n_f32(non_broadcast_qinfo.scale); @@ -628,31 +627,30 @@ void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *o voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2); for(; x < window_end_x; ++x) { - const float afs = scvt_f32_qasymm8(*(non_broadcast_input_ptr + x), non_broadcast_qinfo.scale, non_broadcast_qinfo.offset); - const float bfs = scvt_f32_qasymm8(broadcast_value, broadcast_qinfo.scale, broadcast_qinfo.offset); - *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? bfs : afs, !is_broadcast_input_2 ? afs : bfs, - out->info()->quantization_info()); + const float afs = dequantize_qasymm8(*(non_broadcast_input_ptr + x), non_broadcast_qinfo); + const float bfs = dequantize_qasymm8(broadcast_value, broadcast_qinfo); + *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? bfs : afs, !is_broadcast_input_2 ? afs : bfs, output_qinfo); } }, broadcast_input, non_broadcast_input, output); } else { + const UniformQuantizationInfo input1_qinfo = in1->info()->quantization_info().uniform(); + const UniformQuantizationInfo input2_qinfo = in2->info()->quantization_info().uniform(); + // Input1 quantization info - const int32x4_t voffset1 = vdupq_n_s32(in1->info()->quantization_info().offset); - const float32x4_t vscale1 = vdupq_n_f32(in1->info()->quantization_info().scale); + const int32x4_t voffset1 = vdupq_n_s32(input1_qinfo.offset); + const float32x4_t vscale1 = vdupq_n_f32(input1_qinfo.scale); // Input2 quantization info - const int32x4_t voffset2 = vdupq_n_s32(in2->info()->quantization_info().offset); - const float32x4_t vscale2 = vdupq_n_f32(in2->info()->quantization_info().scale); + const int32x4_t voffset2 = vdupq_n_s32(input2_qinfo.offset); + const float32x4_t vscale2 = vdupq_n_f32(input2_qinfo.scale); // Clear X Dimension on execution window as we handle manually input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); - const QuantizationInfo input1_qinfo = in1->info()->quantization_info(); - const QuantizationInfo input2_qinfo = in2->info()->quantization_info(); - Iterator input1(in1, input1_win); Iterator input2(in2, input2_win); Iterator output(out, win); @@ -667,9 +665,9 @@ void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *o vscale1, vscale2, voffseto, invvscaleo); for(; x < window_end_x; ++x) { - const float afs = scvt_f32_qasymm8(*(input1_ptr + x), input1_qinfo.scale, input1_qinfo.offset); - const float bfs = scvt_f32_qasymm8(*(input2_ptr + x), input2_qinfo.scale, input2_qinfo.offset); - *(output_ptr + x) = (*scalar_func)(afs, bfs, out->info()->quantization_info()); + const float afs = dequantize_qasymm8(*(input1_ptr + x), input1_qinfo); + const float bfs = dequantize_qasymm8(*(input2_ptr + x), input2_qinfo); + *(output_ptr + x) = (*scalar_func)(afs, bfs, output_qinfo); } }, input1, input2, output); |