From d23a251df7b248067e06d5559e985ae1c523be27 Mon Sep 17 00:00:00 2001 From: Sang-Hoon Park Date: Mon, 11 Jan 2021 22:19:49 +0000 Subject: Add SVE2 kernels for quantized elementwise operations Partially implements: COMPMID-3872 Change-Id: I76d81f2b8aa343f9d830298bc931e410c7c901bc Signed-off-by: Sang-Hoon Park Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4842 Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins Reviewed-by: Michele Di Giorgio --- .../NEON/kernels/NEElementwiseOperationKernel.cpp | 17 +- .../elementwise/impl/elementwise_quantized_list.h | 369 +++++++++++++++++++++ 2 files changed, 384 insertions(+), 2 deletions(-) create mode 100644 src/core/SVE/kernels/elementwise/impl/elementwise_quantized_list.h diff --git a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp index 29ae9037af..4d67ec3986 100644 --- a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp +++ b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp @@ -30,6 +30,7 @@ #include "src/core/NEON/NEFixedPoint.h" #include "src/core/NEON/wrapper/wrapper.h" #include "src/core/SVE/kernels/elementwise/impl/elementwise_list.h" +#include "src/core/SVE/kernels/elementwise/impl/elementwise_quantized_list.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" @@ -717,6 +718,7 @@ void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const } } +#if !defined(__ARM_FEATURE_SVE2) void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, 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, @@ -1038,6 +1040,7 @@ void elementwise_op_quantized_signed(const ITensor *in1, const ITensor *in2, ITe input1, input2, output); } } +#endif /* !defined(__ARM_FEATURE_SVE2) */ template void elementwise_comp_op_8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) @@ -1143,9 +1146,14 @@ configure_arithm_func(const ITensorInfo *input1, const ITensorInfo *input2, ITen { "op_F32_F32_F32", &elementwise_arithm_op> }, { "op_S32_S32_S32", &elementwise_arithm_op> }, #endif /* defined(__ARM_FEATURE_SVE) */ - { "op_S16_S16_S16", &elementwise_arithm_op> }, +#if defined(__ARM_FEATURE_SVE2) + { "op_QASYMM8_QASYMM8_QASYMM8", &arm_compute::cpu::sve::elementwise_arithmetic_quantized_op }, + { "op_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED", &arm_compute::cpu::sve::elementwise_arithmetic_quantized_op }, +#else /* defined(__ARM_FEATURE_SVE2) */ { "op_QASYMM8_QASYMM8_QASYMM8", &elementwise_arithm_op_quantized }, - { "op_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED", &elementwise_arithm_op_quantized_signed } + { "op_QASYMM8_SIGNED_QASYMM8_SIGNED_QASYMM8_SIGNED", &elementwise_arithm_op_quantized_signed }, +#endif /* defined(__ARM_FEATURE_SVE2) */ + { "op_S16_S16_S16", &elementwise_arithm_op> }, }; #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC #if defined(__ARM_FEATURE_SVE) @@ -1175,8 +1183,13 @@ configure_comp_func(const ITensorInfo *input1, const ITensorInfo *input2, ITenso { "op_S16_S16_U8", &elementwise_comp_op_16 }, { "op_S32_S32_U8", &elementwise_comp_op_32 }, #endif /* defined(__ARM_FEATURE_SVE) */ +#if defined(__ARM_FEATURE_SVE2) + { "op_QASYMM8_SIGNED_QASYMM8_SIGNED_U8", &arm_compute::cpu::sve::elementwise_comparison_quantized_op }, + { "op_QASYMM8_QASYMM8_U8", &arm_compute::cpu::sve::elementwise_comparison_quantized_op } +#else /* defined(__ARM_FEATURE_SVE2) */ { "op_QASYMM8_SIGNED_QASYMM8_SIGNED_U8", &elementwise_comp_op_quantized_signed }, { "op_QASYMM8_QASYMM8_U8", &elementwise_comp_op_quantized } +#endif /* defined(__ARM_FEATURE_SVE2) */ }; #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC #if defined(__ARM_FEATURE_SVE) diff --git a/src/core/SVE/kernels/elementwise/impl/elementwise_quantized_list.h b/src/core/SVE/kernels/elementwise/impl/elementwise_quantized_list.h new file mode 100644 index 0000000000..e85b0891f5 --- /dev/null +++ b/src/core/SVE/kernels/elementwise/impl/elementwise_quantized_list.h @@ -0,0 +1,369 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef SRC_CORE_SVE_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H +#define SRC_CORE_SVE_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H + +#if defined(__ARM_FEATURE_SVE2) + +#include "src/core/SVE/kernels/elementwise/impl/elementwise_list.h" + +namespace arm_compute +{ +namespace cpu +{ +namespace sve +{ +using namespace arm_compute::wrapper; + +template +struct QuantizedLoopArguments +{ + OperatorType op; + const InputScalarType *input1_ptr; + const InputScalarType *input2_ptr; + OutputScalarType *output_ptr; + + const svint32_t &in1_offset; + const svint32_t &in2_offset; + const svint32_t &out_offset; + const svfloat32_t &in1_scale; + const svfloat32_t &in2_scale; + const svfloat32_t &out_scale; +}; + +template +struct BroadcastQuantizedLoopArguments +{ + OperatorType op; + const InputScalarType *input1_ptr; + float broadcast_value; + OutputScalarType *output_ptr; + bool reorder; + + const svint32_t &in1_offset; + const svint32_t &out_offset; + const svfloat32_t &in1_scale; + const svfloat32_t &out_scale; +}; + +svfloat32x4_t load_quantized(const int8_t *ptr, svbool_t pg, const svint32_t &offset, const svfloat32_t &scale) +{ + auto x = svld1(pg, ptr); + + const auto widened = svcreate4( + svmovlb(svmovlb(x)), + svmovlt(svmovlb(x)), + svmovlb(svmovlt(x)), + svmovlt(svmovlt(x))); + + pg = svptrue_b8(); + + return svcreate4( + svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 0), offset)), scale), + svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 1), offset)), scale), + svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 2), offset)), scale), + svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svget4(widened, 3), offset)), scale)); +} + +svfloat32x4_t load_quantized(const uint8_t *ptr, svbool_t pg, const svint32_t &offset, const svfloat32_t &scale) +{ + auto x = svld1(pg, ptr); + + //vprint(x); + + const auto widened = svcreate4( + svmovlb(svmovlb(x)), + svmovlt(svmovlb(x)), + svmovlb(svmovlt(x)), + svmovlt(svmovlt(x))); + + pg = svptrue_b8(); + + return svcreate4( + svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 0)), offset)), scale), + svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 1)), offset)), scale), + svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 2)), offset)), scale), + svmul_z(pg, svcvt_f32_z(pg, svsub_z(pg, svreinterpret_s32(svget4(widened, 3)), offset)), scale)); +} + +void store_quantized(uint8_t *ptr, svbool_t pg, svfloat32x4_t data, const svint32_t &offset, const svfloat32_t &inv_scale) +{ + const auto quantized = svcreate4( + svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 0), inv_scale))), offset), + svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 1), inv_scale))), offset), + svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 2), inv_scale))), offset), + svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 3), inv_scale))), offset)); + + const auto narrowed_bottom = svqxtunt(svqxtunb(svget4(quantized, 0)), svget4(quantized, 1)); + const auto narrowed_top = svqxtunt(svqxtunb(svget4(quantized, 2)), svget4(quantized, 3)); + const auto narrowed = svqxtnt(svqxtnb(narrowed_bottom), narrowed_top); + svst1(pg, ptr, narrowed); +} + +void store_quantized(int8_t *ptr, svbool_t pg, svfloat32x4_t data, const svint32_t &offset, const svfloat32_t &inv_scale) +{ + const auto quantized = svcreate4( + svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 0), inv_scale))), offset), + svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 1), inv_scale))), offset), + svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 2), inv_scale))), offset), + svadd_z(pg, svcvt_s32_z(pg, svrinta_z(pg, svmul_z(pg, svget4(data, 3), inv_scale))), offset)); + + const auto narrowed_bottom = svqxtnt(svqxtnb(svget4(quantized, 0)), svget4(quantized, 1)); + const auto narrowed_top = svqxtnt(svqxtnb(svget4(quantized, 2)), svget4(quantized, 3)); + const auto narrowed = svqxtnt(svqxtnb(narrowed_bottom), narrowed_top); + + svst1(pg, ptr, narrowed); +} + +template +inline void arithmetic_op_quantized_loop(svbool_t pg, const QuantizedLoopArguments &args) +{ + const auto in1 = load_quantized(args.input1_ptr, pg, args.in1_offset, args.in1_scale); + const auto in2 = load_quantized(args.input2_ptr, pg, args.in2_offset, args.in2_scale); + + const auto result = svcreate4( + elementwise_arithmetic_op(pg, svget4(in1, 0), svget4(in2, 0), args.op), + elementwise_arithmetic_op(pg, svget4(in1, 1), svget4(in2, 1), args.op), + elementwise_arithmetic_op(pg, svget4(in1, 2), svget4(in2, 2), args.op), + elementwise_arithmetic_op(pg, svget4(in1, 3), svget4(in2, 3), args.op)); + + store_quantized(args.output_ptr, pg, result, args.out_offset, args.out_scale); +} + +template +inline void arithmetic_op_broadcast_quantized_loop(svbool_t pg, const BroadcastQuantizedLoopArguments &args) +{ + const auto in1 = load_quantized(args.input1_ptr, pg, args.in1_offset, args.in1_scale); + const auto in2 = svcreate4( + svdup_n(args.broadcast_value), svdup_n(args.broadcast_value), svdup_n(args.broadcast_value), svdup_n(args.broadcast_value)); + + const auto &af = args.reorder ? in2 : in1; + const auto &bf = args.reorder ? in1 : in2; + + const auto result = svcreate4( + elementwise_arithmetic_op(pg, svget4(af, 0), svget4(bf, 0), args.op), + elementwise_arithmetic_op(pg, svget4(af, 1), svget4(bf, 1), args.op), + elementwise_arithmetic_op(pg, svget4(af, 2), svget4(bf, 2), args.op), + elementwise_arithmetic_op(pg, svget4(af, 3), svget4(bf, 3), args.op)); + + store_quantized(args.output_ptr, pg, result, args.out_offset, args.out_scale); +} + +template +inline void comparison_op_quantized_loop(svbool_t pg, const QuantizedLoopArguments &args) +{ + const auto in1 = load_quantized(args.input1_ptr, pg, args.in1_offset, args.in1_scale); + const auto in2 = load_quantized(args.input2_ptr, pg, args.in2_offset, args.in2_scale); + + using OutputVectorType = typename sve_vector::type; + + const auto result = svcreate4( + elementwise_comparison_op(pg, svget4(in1, 0), svget4(in2, 0), args.op), + elementwise_comparison_op(pg, svget4(in1, 1), svget4(in2, 1), args.op), + elementwise_comparison_op(pg, svget4(in1, 2), svget4(in2, 2), args.op), + elementwise_comparison_op(pg, svget4(in1, 3), svget4(in2, 3), args.op)); + + const auto zipped_bottom = svzip1(svget4(result, 0), svget4(result, 1)); + const auto zipped_top = svzip1(svget4(result, 2), svget4(result, 3)); + const auto zipped = svzip1(zipped_bottom, zipped_top); + svst1(pg, args.output_ptr, zipped); +} + +template +inline void comparison_op_broadcast_quantized_loop(svbool_t pg, const BroadcastQuantizedLoopArguments &args) +{ + const auto in1 = load_quantized(args.input1_ptr, pg, args.in1_offset, args.in1_scale); + const auto in2 = svcreate4( + svdup_n(args.broadcast_value), svdup_n(args.broadcast_value), svdup_n(args.broadcast_value), svdup_n(args.broadcast_value)); + + const auto &af = args.reorder ? in2 : in1; + const auto &bf = args.reorder ? in1 : in2; + + using OutputVectorType = typename sve_vector::type; + + const auto result = svcreate4( + elementwise_comparison_op(pg, svget4(af, 0), svget4(bf, 0), args.op), + elementwise_comparison_op(pg, svget4(af, 1), svget4(bf, 1), args.op), + elementwise_comparison_op(pg, svget4(af, 2), svget4(bf, 2), args.op), + elementwise_comparison_op(pg, svget4(af, 3), svget4(bf, 3), args.op)); + + const auto zipped_bottom = svzip1(svget4(result, 0), svget4(result, 1)); + const auto zipped_top = svzip1(svget4(result, 2), svget4(result, 3)); + const auto zipped = svzip1(zipped_bottom, zipped_top); + svst1(pg, args.output_ptr, zipped); +} + +template +using LoopQuantizedFuncType = void (*)(svbool_t, const QuantizedLoopArguments &); + +template +using BroadcastQuantizedLoopFuncType = void (*)(svbool_t, const BroadcastQuantizedLoopArguments &); + +template ::type, + typename OutputScalarType = typename sve_scalar::type> +void elementwise_quantized_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, + OperatorType op, + LoopQuantizedFuncType func, + BroadcastQuantizedLoopFuncType broadcast_func) +{ + const auto all_true_pg = wrapper::svptrue(); + + // Create input windows + Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()); + Window input2_win = window.broadcast_if_dimension_le_one(in2->info()->tensor_shape()); + + // Clear X Dimension on execution window as we handle manually + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + const bool is_broadcast_across_x = in1->info()->tensor_shape().x() != in2->info()->tensor_shape().x(); + + const auto output_voffset = svdup_n(out->info()->quantization_info().uniform().offset); + const auto output_vscale = svdup_n(1.f / out->info()->quantization_info().uniform().scale); + + if(is_broadcast_across_x) + { + const bool is_broadcast_input_2 = input2_win.x().step() == 0; + Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win; + Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win; + const ITensor *broadcast_tensor = is_broadcast_input_2 ? in2 : in1; + const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? in2 : in1; + + const auto non_broadcast_qinfo = is_broadcast_input_2 ? in1->info()->quantization_info() : in2->info()->quantization_info(); + const auto broadcast_qinfo = is_broadcast_input_2 ? in2->info()->quantization_info() : in1->info()->quantization_info(); + + const auto non_broadcast_voffset = svdup_n(non_broadcast_qinfo.uniform().offset); + const auto non_broadcast_vscale = svdup_n(non_broadcast_qinfo.uniform().scale); + + // Clear X Dimension on execution window as we handle manually + non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator broadcast_input(broadcast_tensor, broadcast_win); + Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); + Iterator output(out, win); + + execute_window_loop(win, [&](const Coordinates &) + { + auto output_ptr = reinterpret_cast(output.ptr()); + const auto non_broadcast_input_ptr = reinterpret_cast(non_broadcast_input.ptr()); + const InputScalarType broadcast_value = *reinterpret_cast(broadcast_input.ptr()); + + int x = window_start_x; + + svbool_t pg = wrapper::svwhilelt(x, window_end_x); + do + { + const auto args = BroadcastQuantizedLoopArguments + { + op, + non_broadcast_input_ptr + x, + Qasymm8QuantizationHelper::dequantize(broadcast_value, broadcast_qinfo), + output_ptr + x, + !is_broadcast_input_2, + non_broadcast_voffset, output_voffset, + non_broadcast_vscale, output_vscale + }; + broadcast_func(pg, args); + x += wrapper::svcnt(); + pg = wrapper::svwhilelt(x, window_end_x); + } + while(svptest_any(all_true_pg, pg)); + }, + broadcast_input, non_broadcast_input, output); + } + else + { + // 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)); + + Iterator input1(in1, input1_win); + Iterator input2(in2, input2_win); + Iterator output(out, win); + + const auto in1_voffset = svdup_n(in1->info()->quantization_info().uniform().offset); + const auto in1_vscale = svdup_n(in1->info()->quantization_info().uniform().scale); + + const auto in2_voffset = svdup_n(in2->info()->quantization_info().uniform().offset); + const auto in2_vscale = svdup_n(in2->info()->quantization_info().uniform().scale); + + execute_window_loop(win, [&](const Coordinates &) + { + auto output_ptr = reinterpret_cast(output.ptr()); + const auto input1_ptr = reinterpret_cast(input1.ptr()); + const auto input2_ptr = reinterpret_cast(input2.ptr()); + + int x = window_start_x; + + svbool_t pg = wrapper::svwhilelt(x, window_end_x); + do + { + const auto args = QuantizedLoopArguments + { + op, + input1_ptr + x, + input2_ptr + x, + output_ptr + x, + in1_voffset, in2_voffset, output_voffset, + in1_vscale, in2_vscale, output_vscale + }; + func(pg, args); + x += wrapper::svcnt(); + pg = wrapper::svwhilelt(x, window_end_x); + } + while(svptest_any(all_true_pg, pg)); + }, + input1, input2, output); + } +} + +template +void elementwise_arithmetic_quantized_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + using VectorType = typename sve_vector::type; + elementwise_quantized_op(in1, in2, out, window, op, + &arithmetic_op_quantized_loop, + &arithmetic_op_broadcast_quantized_loop); +} + +template +void elementwise_comparison_quantized_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + static_assert(sizeof(InputScalarType) >= sizeof(OutputScalarType), "input data type's width should be equal to or greater than output data type's width"); + using InputVectorType = typename sve_vector::type; + using OutputVectorType = typename sve_vector::type; + elementwise_quantized_op(in1, in2, out, window, op, + &comparison_op_quantized_loop, + &comparison_op_broadcast_quantized_loop); +} + +} // namespace sve +} // namespace cpu +} // namespace arm_compute + +#endif /* defined(__ARM_FEATURE_SVE2) */ +#endif /* SRC_CORE_SVE_KERNELS_ELEMENTWISE_QUANTIZED_LIST_H */ \ No newline at end of file -- cgit v1.2.1