From d88590f4022bfb6eda3bad4fa599727bab723667 Mon Sep 17 00:00:00 2001 From: George Wort Date: Wed, 12 Dec 2018 17:39:58 +0000 Subject: COMPMID-1753: NEON: Implement Less, Greater, GreaterEqual, Equal, Not Equal Change-Id: I6fa95badcdecb826ac5bd9113f118603d5ac2e82 Reviewed-on: https://review.mlplatform.org/393 Tested-by: Arm Jenkins Reviewed-by: Georgios Pinitas --- .../NEON/kernels/NEElementwiseOperationKernel.cpp | 618 +++++++++++++++++---- 1 file changed, 508 insertions(+), 110 deletions(-) (limited to 'src/core') diff --git a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp index ee9c10014f..88fd730554 100644 --- a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp +++ b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -61,6 +61,20 @@ float32x4x4_t load_quantized(const uint8_t *input1_ptr, const int32x4_t &offset, return out; } +void store_quantized(uint8_t *output_ptr, const uint32x4x4_t &out) +{ + const uint8x8_t pa = vqmovn_u16(vcombine_u16(vqmovn_u32(out.val[0]), vqmovn_u32(out.val[1]))); + const uint8x8_t pb = vqmovn_u16(vcombine_u16(vqmovn_u32(out.val[2]), vqmovn_u32(out.val[3]))); + vst1q_u8(output_ptr, vcombine_u8(pa, pb)); +} + +void store_quantized(uint8_t *output_ptr, const int32x4x4_t &out) +{ + const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(out.val[0]), vqmovn_s32(out.val[1]))); + const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(out.val[2]), vqmovn_s32(out.val[3]))); + vst1q_u8(output_ptr, vcombine_u8(pa, pb)); +} + void store_quantized(uint8_t *output_ptr, const float32x4x4_t &rf, const float32x4_t &offset, const float32x4_t &invscale) { int32x4x4_t out = @@ -70,10 +84,7 @@ void store_quantized(uint8_t *output_ptr, const float32x4x4_t &rf, const float32 vcvtq_s32_f32(vmlaq_f32(offset, rf.val[2], invscale)), vcvtq_s32_f32(vmlaq_f32(offset, rf.val[3], invscale)), }; - - const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(out.val[0]), vqmovn_s32(out.val[1]))); - const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(out.val[2]), vqmovn_s32(out.val[3]))); - vst1q_u8(output_ptr, vcombine_u8(pa, pb)); + store_quantized(output_ptr, out); } float32x4x4_t dup_quantized(qasymm8_t broadcast_value, int offset, float scale) @@ -95,7 +106,7 @@ float32x4x4_t dup_quantized(qasymm8_t broadcast_value, int offset, float scale) } template -inline ScalarType elementwise_op_scalar(const ScalarType &a, const ScalarType &b) +inline ScalarType elementwise_arithm_op_scalar(const ScalarType &a, const ScalarType &b) { auto res = ScalarType(0); @@ -118,8 +129,14 @@ inline ScalarType elementwise_op_scalar(const ScalarType &a, const ScalarType &b return res; } +template +inline uint8_t elementwise_arithm_op_quantized_scalar(const float &a, const float &b, QuantizationInfo qinfo) +{ + return qinfo.quantize(elementwise_arithm_op_scalar(a, b), RoundingPolicy::TO_NEAREST_UP); +} + template -inline VectorType elementwise_op(const VectorType &a, const VectorType &b) +inline VectorType elementwise_arithm_op(const VectorType &a, const VectorType &b) { VectorType res = { 0, 0, 0, 0 }; @@ -145,28 +162,297 @@ inline VectorType elementwise_op(const VectorType &a, const VectorType &b) return res; } -template -inline VectorType elementwise_op_broadcast(const VectorType &a, const ScalarType &broadcast_value) +template +inline float32x4x4_t elementwise_arithm_op(const float32x4x4_t &a, const float32x4x4_t &b) +{ + float32x4x4_t out = + { + elementwise_arithm_op(a.val[0], b.val[0]), + elementwise_arithm_op(a.val[1], b.val[1]), + elementwise_arithm_op(a.val[2], b.val[2]), + elementwise_arithm_op(a.val[3], b.val[3]), + }; + return out; +} + +template +inline VectorType elementwise_arithm_op_broadcast(const VectorType &a, const ScalarType &broadcast_value, const bool reorder) { VectorType broadcast_vector = wrapper::vdup_n(broadcast_value, wrapper::traits::vector_128_tag()); - return elementwise_op(a, broadcast_vector); + return elementwise_arithm_op(reorder ? broadcast_vector : a, reorder ? a : broadcast_vector); } -template -float32x4x4_t elementwise_op(const float32x4x4_t &a, const float32x4x4_t &b) +template +inline uint8_t elementwise_comp_op_scalar(const InputScalarType &a, const InputScalarType &b) { - float32x4x4_t out = + bool res = false; + + switch(op) + { + case ComparisonOperation::Equal: + res = (a == b); + break; + case ComparisonOperation::NotEqual: + res = (a != b); + break; + case ComparisonOperation::Greater: + res = (a > b); + break; + case ComparisonOperation::GreaterEqual: + res = (a >= b); + break; + case ComparisonOperation::Less: + res = (a < b); + break; + case ComparisonOperation::LessEqual: + res = (a <= b); + break; + default: + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); + } + return res ? ~static_cast(0) : static_cast(0); +} + +template +inline uint8_t elementwise_comp_op_quantized_scalar(const float &a, const float &b, QuantizationInfo qinfo) +{ + ARM_COMPUTE_UNUSED(qinfo); + return elementwise_comp_op_scalar(a, b); +} + +template +inline OutputVectorType elementwise_comp_op(const InputVectorType &a, const InputVectorType &b) +{ + OutputVectorType res = { 0, 0, 0, 0 }; + + switch(op) + { + case ComparisonOperation::Equal: + res = wrapper::vceq(a, b); + break; + case ComparisonOperation::NotEqual: + res = wrapper::vnot(wrapper::vceq(a, b)); + break; + case ComparisonOperation::Greater: + res = wrapper::vcgt(a, b); + break; + case ComparisonOperation::GreaterEqual: + res = wrapper::vcge(a, b); + break; + case ComparisonOperation::Less: + res = wrapper::vcgt(b, a); + break; + case ComparisonOperation::LessEqual: + res = wrapper::vcge(b, a); + break; + default: + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); + } + + return res; +} + +template +inline uint32x4x4_t elementwise_comp_op(const float32x4x4_t &a, const float32x4x4_t &b) +{ + uint32x4x4_t out = { - elementwise_op(a.val[0], b.val[0]), - elementwise_op(a.val[1], b.val[1]), - elementwise_op(a.val[2], b.val[2]), - elementwise_op(a.val[3], b.val[3]), + elementwise_comp_op(a.val[0], b.val[0]), + elementwise_comp_op(a.val[1], b.val[1]), + elementwise_comp_op(a.val[2], b.val[2]), + elementwise_comp_op(a.val[3], b.val[3]) }; return out; } -template -void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +template +inline OutputVectorType elementwise_comp_op_broadcast(const InputVectorType &a, const InputScalarType &broadcast_value, const bool reorder) +{ + InputVectorType broadcast_vector = wrapper::vdup_n(broadcast_value, wrapper::traits::vector_128_tag()); + return elementwise_comp_op(reorder ? broadcast_vector : a, reorder ? a : broadcast_vector); +} + +template +inline int elementwise_arithm_op_loop(int window_start_x, int window_end_x, int window_step_x, + const ScalarType *input1_ptr, const ScalarType *input2_ptr, ScalarType *output_ptr) +{ + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto a = wrapper::vloadq(input1_ptr + x); + const auto b = wrapper::vloadq(input2_ptr + x); + wrapper::vstore(output_ptr + x, elementwise_arithm_op(a, b)); + } + return x; +} + +template +inline int elementwise_arithm_op_quantized_loop(int window_start_x, int window_end_x, int window_step_x, + const uint8_t *input1_ptr, const uint8_t *input2_ptr, uint8_t *output_ptr, + int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2, + float32x4_t voffseto, float32x4_t invvscaleo) +{ + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + // Get inputs and compute output + const float32x4x4_t af = load_quantized(input1_ptr + x, voffset1, vscale1); + const float32x4x4_t bf = load_quantized(input2_ptr + x, voffset2, vscale2); + const float32x4x4_t rf = elementwise_arithm_op(af, bf); + store_quantized(output_ptr + x, rf, voffseto, invvscaleo); + } + return x; +} + +template +inline int elementwise_arithm_op_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, + const ScalarType *non_broadcast_input_ptr, const ScalarType &broadcast_value, ScalarType *output_ptr, const bool reorder) +{ + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto a = wrapper::vloadq((non_broadcast_input_ptr + x)); + wrapper::vstore(output_ptr + x, elementwise_arithm_op_broadcast(a, broadcast_value, reorder)); + } + return x; +} + +template +inline int elementwise_arithm_op_quantized_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, + const uint8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr, + int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast, + float32x4_t voffseto, float32x4_t invvscaleo, bool reorder) +{ + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const float32x4x4_t af = load_quantized(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast); + const float32x4x4_t rf = elementwise_arithm_op(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector); + store_quantized(output_ptr + x, rf, voffseto, invvscaleo); + } + return x; +} + +template +inline int elementwise_comp_op_16_loop(int window_start_x, int window_end_x, int window_step_x, + const InputScalarType *input1_ptr, const InputScalarType *input2_ptr, uint8_t *output_ptr) +{ + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto a = wrapper::vloadq(input1_ptr + x); + const auto b = wrapper::vloadq(input2_ptr + x); + const auto res = elementwise_comp_op(a, b); + wrapper::vstore(output_ptr + x, wrapper::vmovn(res)); + } + return x; +} + +template +inline int elementwise_comp_op_32_loop(int window_start_x, int window_end_x, int window_step_x, + const InputScalarType *input1_ptr, const InputScalarType *input2_ptr, uint8_t *output_ptr) +{ + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + auto a = wrapper::vloadq(input1_ptr + x); + auto b = wrapper::vloadq(input2_ptr + x); + const auto res = elementwise_comp_op(a, b); + a = wrapper::vloadq(input1_ptr + x + 4); + b = wrapper::vloadq(input2_ptr + x + 4); + const auto res2 = elementwise_comp_op(a, b); + wrapper::vstore(output_ptr + x, wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(res), wrapper::vmovn(res2)))); + } + if(x <= window_end_x - 4) + { + const auto a = wrapper::vloadq(input1_ptr + x); + const auto b = wrapper::vloadq(input2_ptr + x); + const auto res = elementwise_comp_op(a, b); + for(int i = 0; i < 4; i++) + { + *(output_ptr + x + i) = wrapper::vgetlane(res, i); + } + x = +4; + } + return x; +} + +template +inline int elementwise_comp_op_quantized_loop(int window_start_x, int window_end_x, int window_step_x, + const uint8_t *input1_ptr, const uint8_t *input2_ptr, uint8_t *output_ptr, + int32x4_t voffset1, int32x4_t voffset2, float32x4_t vscale1, float32x4_t vscale2, + float32x4_t voffseto, float32x4_t invvscaleo) +{ + ARM_COMPUTE_UNUSED(voffseto, invvscaleo); + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const float32x4x4_t af = load_quantized(input1_ptr + x, voffset1, vscale1); + const float32x4x4_t bf = load_quantized(input2_ptr + x, voffset2, vscale2); + const uint32x4x4_t rf = elementwise_comp_op(af, bf); + store_quantized(output_ptr + x, rf); + } + return x; +} + +template +inline int elementwise_comp_op_broadcast_16_loop(int window_start_x, int window_end_x, int window_step_x, + const InputScalarType *non_broadcast_input_ptr, const InputScalarType &broadcast_value, uint8_t *output_ptr, const bool reorder) +{ + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto a = elementwise_comp_op_broadcast(wrapper::vloadq((non_broadcast_input_ptr + x)), broadcast_value, reorder); + wrapper::vstore(output_ptr + x, wrapper::vmovn(a)); + } + return x; +} + +template +inline int elementwise_comp_op_broadcast_32_loop(int window_start_x, int window_end_x, int window_step_x, + const InputScalarType *non_broadcast_input_ptr, const InputScalarType &broadcast_value, uint8_t *output_ptr, const bool reorder) +{ + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto a = elementwise_comp_op_broadcast(wrapper::vloadq(non_broadcast_input_ptr + x), broadcast_value, reorder); + const auto b = elementwise_comp_op_broadcast(wrapper::vloadq(non_broadcast_input_ptr + x + 4), broadcast_value, reorder); + wrapper::vstore(output_ptr + x, wrapper::vmovn(wrapper::vcombine(wrapper::vmovn(a), wrapper::vmovn(b)))); + } + if(x <= window_end_x - 4) + { + const auto a = elementwise_comp_op_broadcast(wrapper::vloadq((non_broadcast_input_ptr + x)), broadcast_value, reorder); + for(int i = 0; i < 4; i++) + { + *(output_ptr + x + i) = wrapper::vgetlane(a, i); + } + x = +4; + } + return x; +} + +template +inline int elementwise_comp_op_quantized_broadcast_loop(int window_start_x, int window_end_x, int window_step_x, + const uint8_t *non_broadcast_input_ptr, float32x4x4_t broadcast_vector, uint8_t *output_ptr, + int32x4_t voffset_non_broadcast, float32x4_t vscale_non_broadcast, + float32x4_t voffseto, float32x4_t invvscaleo, bool reorder) +{ + ARM_COMPUTE_UNUSED(voffseto, invvscaleo); + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const float32x4x4_t af = load_quantized(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast); + const uint32x4x4_t rf = elementwise_comp_op(reorder ? broadcast_vector : af, reorder ? af : broadcast_vector); + store_quantized(output_ptr + x, rf); + } + return x; +} + +template +void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, + OutputScalarType (*scalar_func)(const InputScalarType &, const InputScalarType &), + int (*broadcast_func)(int, int, int, const InputScalarType *, const InputScalarType &, OutputScalarType *, const bool), + int (*neon_func)(int, int, int, const InputScalarType *, const InputScalarType *, OutputScalarType *)) { // Create input windows Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()); @@ -176,14 +462,13 @@ void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window win = window; win.set(Window::DimX, Window::Dimension(0, 1, 1)); - const int window_step_x = 16 / in1->info()->element_size(); + const int window_step_x = std::min(16 / static_cast(sizeof(OutputScalarType)), 8); 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 = (input1_win.x().step() == 0) || (input2_win.x().step() == 0); if(is_broadcast_across_x) { - // Select the broadcast input on the X axis 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; @@ -199,20 +484,15 @@ void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const execute_window_loop(win, [&](const Coordinates & id) { - auto output_ptr = reinterpret_cast(output.ptr()); - const auto non_broadcast_input_ptr = reinterpret_cast(non_broadcast_input.ptr()); - const ScalarType broadcast_value = *reinterpret_cast(broadcast_input.ptr()); + 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; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto a = wrapper::vloadq((non_broadcast_input_ptr + x)); - wrapper::vstore(output_ptr + x, elementwise_op_broadcast(a, broadcast_value)); - } + int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_value, output_ptr, !is_broadcast_input_2); for(; x < window_end_x; ++x) { const auto a = *(non_broadcast_input_ptr + x); - *(output_ptr + x) = elementwise_op_scalar(a, broadcast_value); + *(output_ptr + x) = (*scalar_func)(!is_broadcast_input_2 ? broadcast_value : a, !is_broadcast_input_2 ? a : broadcast_value); } }, broadcast_input, non_broadcast_input, output); @@ -229,31 +509,29 @@ void elementwise_op(const ITensor *in1, const ITensor *in2, ITensor *out, const execute_window_loop(win, [&](const Coordinates & id) { - auto output_ptr = reinterpret_cast(output.ptr()); - const auto input1_ptr = reinterpret_cast(input1.ptr()); - const auto input2_ptr = reinterpret_cast(input2.ptr()); + 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; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const auto a = wrapper::vloadq(input1_ptr + x); - const auto b = wrapper::vloadq(input2_ptr + x); - wrapper::vstore(output_ptr + x, elementwise_op(a, b)); - } + int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr); for(; x < window_end_x; ++x) { const auto a = *(input1_ptr + x); const auto b = *(input2_ptr + x); - *(output_ptr + x) = elementwise_op_scalar(a, b); + *(output_ptr + x) = (*scalar_func)(a, b); } - }, input1, input2, output); } } -template -void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, + uint8_t (*scalar_func)(const float &, const float &, QuantizationInfo), + 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 *, + int32x4_t, int32x4_t, float32x4_t, float32x4_t, + float32x4_t, float32x4_t)) { // Create input windows Window input1_win = window.broadcast_if_dimension_le_one(in1->info()->tensor_shape()); @@ -305,18 +583,14 @@ void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *o const uint8_t broadcast_value = *reinterpret_cast(broadcast_input.ptr()); const float32x4x4_t broadcast_vector = dup_quantized(broadcast_value, broadcast_qinfo.offset, broadcast_qinfo.scale); - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - const float32x4x4_t af = load_quantized(non_broadcast_input_ptr + x, voffset_non_broadcast, vscale_non_broadcast); - const float32x4x4_t rf = elementwise_op(af, broadcast_vector); - store_quantized(output_ptr + x, rf, voffseto, invvscaleo); - } + int x = (*broadcast_func)(window_start_x, window_end_x, window_step_x, non_broadcast_input_ptr, broadcast_vector, output_ptr, + voffset_non_broadcast, vscale_non_broadcast, voffseto, invvscaleo, !is_broadcast_input_2); for(; x < window_end_x; ++x) { - const float afs = static_cast(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale; - const float bfs = static_cast(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale; - *(output_ptr + x) = out->info()->quantization_info().quantize(elementwise_op_scalar(afs, bfs), RoundingPolicy::TO_NEAREST_UP); + 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()); } }, broadcast_input, non_broadcast_input, output); @@ -348,31 +622,131 @@ void elementwise_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *o const auto input2_ptr = reinterpret_cast(input2.ptr()); const auto output_ptr = reinterpret_cast(output.ptr()); - int x = window_start_x; - for(; x <= (window_end_x - window_step_x); x += window_step_x) - { - // Get inputs and compute output - const float32x4x4_t af = load_quantized(input1_ptr + x, voffset1, vscale1); - const float32x4x4_t bf = load_quantized(input2_ptr + x, voffset2, vscale2); - const float32x4x4_t rf = elementwise_op(af, bf); - store_quantized(output_ptr + x, rf, voffseto, invvscaleo); - } + int x = (*neon_func)(window_start_x, window_end_x, window_step_x, input1_ptr, input2_ptr, output_ptr, voffset1, voffset2, + vscale1, vscale2, voffseto, invvscaleo); for(; x < window_end_x; ++x) { - const float afs = static_cast((*(input1_ptr + x)) - input1_qinfo.offset) * input1_qinfo.scale; - const float bfs = static_cast((*(input2_ptr + x)) - input2_qinfo.offset) * input2_qinfo.scale; - *(output_ptr + x) = out->info()->quantization_info().quantize(elementwise_op_scalar(afs, bfs), RoundingPolicy::TO_NEAREST_UP); + 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()); } }, input1, input2, output); } } -Status validate_arguments_arithmetic(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) +template +void elementwise_comp_op_16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + elementwise_op(in1, in2, out, window, + &elementwise_comp_op_scalar, + &elementwise_comp_op_broadcast_16_loop, + &elementwise_comp_op_16_loop); +} + +template +void elementwise_comp_op_32(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + elementwise_op(in1, in2, out, window, + &elementwise_comp_op_scalar, + &elementwise_comp_op_broadcast_32_loop, + &elementwise_comp_op_32_loop); +} + +template +void elementwise_arithm_op(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + elementwise_op(in1, in2, out, window, + &elementwise_arithm_op_scalar, + &elementwise_arithm_op_broadcast_loop, + &elementwise_arithm_op_loop); +} + +template +void elementwise_arithm_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + elementwise_op_quantized(in1, in2, out, window, &elementwise_arithm_op_quantized_scalar, + &elementwise_arithm_op_quantized_broadcast_loop, + &elementwise_arithm_op_quantized_loop); +} + +template +void elementwise_comp_op_quantized(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + elementwise_op_quantized(in1, in2, out, window, &elementwise_comp_op_quantized_scalar, + &elementwise_comp_op_quantized_broadcast_loop, + &elementwise_comp_op_quantized_loop); +} + +std::function +configure_func(const ITensor *input1, const ITensor *input2, ITensor *output, + std::map map_function) +{ + std::string function_to_call("op_"); + function_to_call += string_from_data_type(input1->info()->data_type()) + "_"; + function_to_call += string_from_data_type(input2->info()->data_type()) + "_"; + function_to_call += string_from_data_type(output->info()->data_type()); + + auto it = map_function.find(function_to_call); + + if(it != map_function.end()) + { + auto func = it->second; + return [func](const ITensor * input1, const ITensor * input2, ITensor * output, const Window & window) + { + func(input1, input2, output, window); + }; + } + return nullptr; +} + +template +std::function +configure_arithm_func(const ITensor *input1, const ITensor *input2, ITensor *output) +{ + static std::map map_function = + { + { "op_F32_F32_F32", &elementwise_arithm_op }, + { "op_S16_S16_S16", &elementwise_arithm_op }, + { "op_S32_S32_S32", &elementwise_arithm_op }, + { "op_QASYMM8_QASYMM8_QASYMM8", &elementwise_arithm_op_quantized } + }; +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + map_function["op_F16_F16_F16"] = &elementwise_arithm_op; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + + return configure_func(input1, input2, output, map_function); +} + +template +std::function +configure_comp_func(const ITensor *input1, const ITensor *input2, ITensor *output) +{ + static std::map map_function = + { + { "op_F32_F32_U8", &elementwise_comp_op_32 }, + { "op_S16_S16_U8", &elementwise_comp_op_16 }, + { "op_S32_S32_U8", &elementwise_comp_op_32 }, + { "op_QASYMM8_QASYMM8_U8", &elementwise_comp_op_quantized } + }; +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + map_function["op_F16_F16_U8"] = &elementwise_comp_op_16; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + + return configure_func(input1, input2, output, map_function); +} +} // namespace + +NEElementwiseOperationKernel::NEElementwiseOperationKernel() + : _function(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr) +{ +} + +Status NEElementwiseOperationKernel::validate_arguments_common(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) { - ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input1); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::S32, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::S32, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input1); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &input2); const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape()); @@ -382,24 +756,16 @@ Status validate_arguments_arithmetic(const ITensorInfo &input1, const ITensorInf // Validate in case of configured output if(output.total_size() > 0) { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &output); ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0), "Wrong shape for output"); } return Status{}; } -} // namespace -NEElementwiseOperationKernel::NEElementwiseOperationKernel() - : _op(), _func(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr) -{ -} -template void NEElementwiseOperationKernel::configure_common(const ITensor *input1, const ITensor *input2, ITensor *output) { ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1->info(), *input2->info(), *output->info())); // Configure kernel window const std::pair broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input1->info(), *input2->info()); @@ -411,76 +777,108 @@ void NEElementwiseOperationKernel::configure_common(const ITensor *input1, const Window win = calculate_max_window(valid_region); - static std::map map_function = - { - { "op_F32_F32_F32", &elementwise_op }, - { "op_S16_S16_S16", &elementwise_op }, - { "op_S32_S32_S32", &elementwise_op }, - { "op_QASYMM8_QASYMM8_QASYMM8", &elementwise_op_quantized } - }; -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - map_function["op_F16_F16_F16"] = &elementwise_op; -#endif /* ARM_COMPUTE_AARCH64_V8_2 */ _input1 = input1; _input2 = input2; _output = output; - std::string function_to_call("op_"); - function_to_call += string_from_data_type(input1->info()->data_type()) + "_"; - function_to_call += string_from_data_type(input2->info()->data_type()) + "_"; - function_to_call += string_from_data_type(output->info()->data_type()); - auto it = map_function.find(function_to_call); - - if(it != map_function.end()) - { - _func = it->second; - } - INEKernel::configure(win); } void NEElementwiseOperationKernel::run(const Window &window, const ThreadInfo &info) { - ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_UNUSED(info, window); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - ARM_COMPUTE_ERROR_ON(_func == nullptr); - - (*_func)(_input1, _input2, _output, window); + ARM_COMPUTE_ERROR_ON(_function == nullptr); + _function(_input1, _input2, _output, window); } /** Arithmetic operators (min, max, squared_diff) */ void NEArithmeticOperationKernel::configure(ArithmeticOperation op, const ITensor *input1, const ITensor *input2, ITensor *output) { - _op = op; + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1->info(), *input2->info(), *output->info())); + configure_common(input1, input2, output); switch(op) { case ArithmeticOperation::MAX: - configure_common(input1, input2, output); + _function = configure_arithm_func(input1, input2, output); break; case ArithmeticOperation::MIN: - configure_common(input1, input2, output); + _function = configure_arithm_func(input1, input2, output); break; case ArithmeticOperation::SQUARED_DIFF: - configure_common(input1, input2, output); + _function = configure_arithm_func(input1, input2, output); break; default: ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); } } +Status NEArithmeticOperationKernel::validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) +{ + // Validate in case of configured output + if(output.total_size() > 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &output); + } + return validate_arguments_common(input1, input2, output); +} + Status NEArithmeticOperationKernel::validate(ArithmeticOperation op, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) { ARM_COMPUTE_UNUSED(op); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output); - - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_arithmetic(*input1, *input2, *output)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output)); return Status{}; } -Status NEArithmeticOperationKernel::validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) +/** Comparison operators (equal, not equal, less than, greater than, less than or equal, greater than or equal) */ + +void NEComparisonOperationKernel::configure(ComparisonOperation op, const ITensor *input1, const ITensor *input2, ITensor *output) { - return validate_arguments_arithmetic(input1, input2, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1->info(), *input2->info(), *output->info())); + configure_common(input1, input2, output); + switch(op) + { + case ComparisonOperation::Equal: + _function = configure_comp_func(input1, input2, output); + break; + case ComparisonOperation::NotEqual: + _function = configure_comp_func(input1, input2, output); + break; + case ComparisonOperation::Greater: + _function = configure_comp_func(input1, input2, output); + break; + case ComparisonOperation::GreaterEqual: + _function = configure_comp_func(input1, input2, output); + break; + case ComparisonOperation::Less: + _function = configure_comp_func(input1, input2, output); + break; + case ComparisonOperation::LessEqual: + _function = configure_comp_func(input1, input2, output); + break; + default: + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); + } +} + +Status NEComparisonOperationKernel::validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) +{ + // Validate in case of configured output + if(output.total_size() > 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8); + } + return validate_arguments_common(input1, input2, output); +} + +Status NEComparisonOperationKernel::validate(ComparisonOperation op, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) +{ + ARM_COMPUTE_UNUSED(op); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output)); + return Status{}; } } // namespace arm_compute -- cgit v1.2.1