From 605a928960c0c36384925a9064d12addc8f43a41 Mon Sep 17 00:00:00 2001 From: Omar Al Khatib Date: Tue, 1 Nov 2022 17:01:24 +0000 Subject: Optimize CPU mul layer on quantized data Resolves : [COMPMID-5461] Signed-off-by: Omar Al Khatib Change-Id: I89b99d267c32b00ef44f9bb6e7c714dfe4a0d29d Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8420 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gunes Bayir Benchmark: Arm Jenkins --- src/core/NEON/wrapper/intrinsics/shr.h | 28 ++- src/core/NEON/wrapper/intrinsics/store.h | 6 +- src/core/NEON/wrapper/intrinsics/sub.h | 17 +- src/cpu/kernels/CpuMulKernel.cpp | 328 ++++++++++++++++++++++++++++--- 4 files changed, 345 insertions(+), 34 deletions(-) diff --git a/src/core/NEON/wrapper/intrinsics/shr.h b/src/core/NEON/wrapper/intrinsics/shr.h index d740091464..e41e9b8b31 100644 --- a/src/core/NEON/wrapper/intrinsics/shr.h +++ b/src/core/NEON/wrapper/intrinsics/shr.h @@ -25,21 +25,19 @@ #ifndef ARM_COMPUTE_WRAPPER_SHR_H #define ARM_COMPUTE_WRAPPER_SHR_H -#include #include +#include namespace arm_compute { namespace wrapper { - #define VQRSHRN_IMPL(half_vtype, vtype, prefix, postfix) \ template \ inline half_vtype vqrshrn(const vtype &a) \ { \ return prefix##_##postfix(a, b); \ } - VQRSHRN_IMPL(int8x8_t, int16x8_t, vqrshrn_n, s16) VQRSHRN_IMPL(uint8x8_t, uint16x8_t, vqrshrn_n, u16) VQRSHRN_IMPL(int16x4_t, int32x4_t, vqrshrn_n, s32) @@ -77,20 +75,38 @@ VQRSHRN_SCALAR_IMPL(uint32_t, uint64_t, vqrshrnd_n, u64) { \ return prefix_signed##_##postfix(a, b); \ } \ - \ + \ template \ inline typename std::enable_if::value && !std::is_signed::value, u##half_vtype>::type \ vqrshrn_ex(const vtype &a) \ { \ return prefix_unsigned##_##postfix(a, b); \ } - VQRSHRN_EX_IMPL(int8x8_t, int16x8_t, vqrshrn_n, vqrshrun_n, s16) VQRSHRN_EX_IMPL(int16x4_t, int32x4_t, vqrshrn_n, vqrshrun_n, s32) VQRSHRN_EX_IMPL(int32x2_t, int64x2_t, vqrshrn_n, vqrshrun_n, s64) - #undef VQRSHRN_EX_IMPL +#define VSHR_IMPL(vtype, prefix, postfix) \ + template \ + inline vtype vshr_n(const vtype &a) \ + { \ + return prefix##_##postfix(a, b); \ + } +VSHR_IMPL(uint8x8_t, vshr_n, u8) +VSHR_IMPL(int8x8_t, vshr_n, s8) +#undef VSHR_IMPL + +#define VSHRQ_IMPL(vtype, prefix, postfix) \ + template \ + inline vtype vshrq_n(const vtype &a) \ + { \ + return prefix##_##postfix(a, b); \ + } +VSHRQ_IMPL(uint32x4_t, vshrq_n, u32) +VSHRQ_IMPL(int32x4_t, vshrq_n, s32) +#undef VSHRQ_IMPL + #ifdef __aarch64__ #define VQRSHRN_EX_SCALAR_IMPL(half_vtype, vtype, prefix_signed, prefix_unsigned, postfix) \ template \ diff --git a/src/core/NEON/wrapper/intrinsics/store.h b/src/core/NEON/wrapper/intrinsics/store.h index 6dda432ea9..ce1b9a554e 100644 --- a/src/core/NEON/wrapper/intrinsics/store.h +++ b/src/core/NEON/wrapper/intrinsics/store.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,8 +44,6 @@ VSTORE_IMPL(uint16_t, uint16x4_t, vst1, u16) VSTORE_IMPL(int16_t, int16x4_t, vst1, s16) VSTORE_IMPL(uint32_t, uint32x2_t, vst1, u32) VSTORE_IMPL(int32_t, int32x2_t, vst1, s32) -//VSTORE_IMPL(uint64_t, 1, vst1, u64) -//VSTORE_IMPL(int64_t, 1, vst1, s64) VSTORE_IMPL(float, float32x2_t, vst1, f32) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC VSTORE_IMPL(float16_t, float16x4_t, vst1, f16) @@ -57,8 +55,6 @@ VSTORE_IMPL(uint16_t, uint16x8_t, vst1q, u16) VSTORE_IMPL(int16_t, int16x8_t, vst1q, s16) VSTORE_IMPL(uint32_t, uint32x4_t, vst1q, u32) VSTORE_IMPL(int32_t, int32x4_t, vst1q, s32) -//VSTORE_IMPL(uint64_t, 2, vst1q, u64) -//VSTORE_IMPL(int64_t, 2, vst1q, s64) VSTORE_IMPL(float, float32x4_t, vst1q, f32) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC VSTORE_IMPL(float16_t, float16x8_t, vst1q, f16) diff --git a/src/core/NEON/wrapper/intrinsics/sub.h b/src/core/NEON/wrapper/intrinsics/sub.h index 475986d0f6..20436714ef 100644 --- a/src/core/NEON/wrapper/intrinsics/sub.h +++ b/src/core/NEON/wrapper/intrinsics/sub.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -98,6 +98,21 @@ VQSUB_IMPL(float16x8_t, float16x8_t, vsubq, f16) #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC #undef VQSUB_IMPL +#define VSUBL_IMPL(rtype, vtype, prefix, postfix) \ + inline rtype vsubl(const vtype &a, const vtype &b) \ + { \ + return prefix##_##postfix(a, b); \ + } + +VSUBL_IMPL(int16x8_t, int8x8_t, vsubl, s8) +VSUBL_IMPL(int32x4_t, int16x4_t, vsubl, s16) +VSUBL_IMPL(int64x2_t, int32x2_t, vsubl, s32) +VSUBL_IMPL(uint16x8_t, uint8x8_t, vsubl, u8) +VSUBL_IMPL(uint32x4_t, uint16x4_t, vsubl, u16) +VSUBL_IMPL(uint64x2_t, uint32x2_t, vsubl, u32) + +#undef VSUB_IMPL + } // namespace wrapper } // namespace arm_compute #endif /* ARM_COMPUTE_WRAPPER_SUB_H */ diff --git a/src/cpu/kernels/CpuMulKernel.cpp b/src/cpu/kernels/CpuMulKernel.cpp index 2f04bf9f26..487954b889 100644 --- a/src/cpu/kernels/CpuMulKernel.cpp +++ b/src/cpu/kernels/CpuMulKernel.cpp @@ -241,7 +241,8 @@ void mul_saturate_quantized_8(const ITensor *src1, const ITensor *src2, ITensor Iterator input2(src2, input2_win); Iterator dst(out, win); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto input1_ptr = reinterpret_cast(input1.ptr()); const auto input2_ptr = reinterpret_cast(input2.ptr()); @@ -290,6 +291,262 @@ void mul_saturate_quantized_8(const ITensor *src1, const ITensor *src2, ITensor } } +bool mul_q8_neon_fixedpoint_possible(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, float scale) +{ + const auto iq0 = src0->quantization_info().uniform(); + const auto iq1 = src1->quantization_info().uniform(); + const auto oq = dst->quantization_info().uniform(); + + const auto multiplier = ((iq0.scale * iq1.scale) / oq.scale) * scale; + + if(multiplier < -8191.f || multiplier > 8191.f) + { + //The multiplier cannot be stored as a 14.18 signed fixed-point number + return false; + } + + const auto offset_out = float(oq.offset); + + const auto max_result = multiplier * (256) * (256) + offset_out; + + if(max_result > 8191.f) + { + //It might not be possible to store the result as a 14.18 signed fixed-point number. + return false; + } + + return true; +} + +template +void mul_q8_neon_fixedpoint(const ITensor *src0, const ITensor *src1, ITensor *dst, const Window &window, float scale) +{ + const auto in0_info = src0->info(); + const auto in1_info = src1->info(); + + const auto &in0_shape = in0_info->tensor_shape(); + const auto &in1_shape = in1_info->tensor_shape(); + + // Create input windows. + Window in0_win = window.broadcast_if_dimension_le_one(in0_shape); + Window in1_win = window.broadcast_if_dimension_le_one(in1_shape); + + // Clear the x dimension on the execution window as we process the whole row each iteration. + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + constexpr int window_step_x = 16; + const auto window_start_x = window.x().start(); + const auto window_end_x = window.x().end(); + const auto is_broadcast_across_x = in0_shape.x() != in1_shape.x(); + + const auto iq0_info = in0_info->quantization_info().uniform(); + const auto iq1_info = in1_info->quantization_info().uniform(); + const auto oq_info = dst->info()->quantization_info().uniform(); + + const auto in0_offset = iq0_info.offset; + const auto in1_offset = iq1_info.offset; + const auto out_offset = oq_info.offset; + const auto multiplier = ((iq0_info.scale * iq1_info.scale) / oq_info.scale) * scale; + + constexpr int32_t two_pwr18i = 262144; + constexpr float two_pwr18f = 262144.f; + + const auto in0_offset_16p0 = static_cast(in0_offset); + const auto in1_offset_16p0 = static_cast(in1_offset); + const auto out_offset_14p18 = static_cast(out_offset * two_pwr18i); + const auto multiplier_14p18 = static_cast(multiplier * two_pwr18f); + + if(is_broadcast_across_x) + { + // Prefix: a = non-broadcast, b = broadcast. + + const auto is_broadcast_input_1 = in1_win.x().step() == 0; + auto a_win = is_broadcast_input_1 ? in0_win : in1_win; + auto b_win = is_broadcast_input_1 ? in1_win : in0_win; + const auto a_tensor = is_broadcast_input_1 ? src0 : src1; + const auto b_tensor = is_broadcast_input_1 ? src1 : src0; + + const auto a_offset_16p0 = is_broadcast_input_1 ? in0_offset_16p0 : in1_offset_16p0; + const auto b_offset_16p0 = is_broadcast_input_1 ? in1_offset : in0_offset; +#ifndef __aarch64__ + const auto a_offset = is_broadcast_input_1 ? in0_offset : in1_offset; + const auto b_offset = is_broadcast_input_1 ? in1_offset : in0_offset; +#endif //__aarch64__ + const auto a_voffset_16p0 = wrapper::vdup_n(a_offset_16p0, wrapper::traits::vector_64_tag()); + + // Clear the x dimension on the execution window as we process the whole row each iteration. + a_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator a_input_it(a_tensor, a_win); + Iterator b_input_it(b_tensor, b_win); + Iterator out_it(dst, win); + + execute_window_loop( + win, [&](const Coordinates &) + { + const auto a_ptr = reinterpret_cast(a_input_it.ptr()); + const auto b_ptr = reinterpret_cast(b_input_it.ptr()); + const auto out_ptr = reinterpret_cast(out_it.ptr()); + + const auto b_val = *b_ptr; + const auto b_offseted_32p0 = static_cast(b_val - b_offset_16p0); + const auto b_voffseted_32p0 = wrapper::vdup_n(b_offseted_32p0, wrapper::traits::vector_128_tag()); + + const auto vmultiplier_14p18 = wrapper::vdup_n(multiplier_14p18, wrapper::traits::vector_128_tag()); + const auto voffsetout_14p18 = wrapper::vdup_n(out_offset_14p18, wrapper::traits::vector_128_tag()); + + int x = window_start_x; + + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + // Load the inputs. + const auto a_vin_8p0 = wrapper::vloadq(a_ptr + x); + + // Widen the non-broadcast elements to signed 16-bit regardless of the input signedness. + const auto a_vin_16p0_0 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(a_vin_8p0))); + const auto a_vin_16p0_1 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(a_vin_8p0))); + + const auto voffseted_32p0_00 = wrapper::vsubl(wrapper::vgetlow(a_vin_16p0_0), a_voffset_16p0); + const auto voffseted_32p0_01 = wrapper::vsubl(wrapper::vgethigh(a_vin_16p0_0), a_voffset_16p0); + const auto voffseted_32p0_10 = wrapper::vsubl(wrapper::vgetlow(a_vin_16p0_1), a_voffset_16p0); + const auto voffseted_32p0_11 = wrapper::vsubl(wrapper::vgethigh(a_vin_16p0_1), a_voffset_16p0); + + const auto vinnermul_32p0_00 = wrapper::vmul(voffseted_32p0_00, b_voffseted_32p0); + const auto vinnermul_32p0_01 = wrapper::vmul(voffseted_32p0_01, b_voffseted_32p0); + const auto vinnermul_32p0_10 = wrapper::vmul(voffseted_32p0_10, b_voffseted_32p0); + const auto vinnermul_32p0_11 = wrapper::vmul(voffseted_32p0_11, b_voffseted_32p0); + + const auto vout_14p18_00 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_00, vmultiplier_14p18); + const auto vout_14p18_01 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_01, vmultiplier_14p18); + const auto vout_14p18_10 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_10, vmultiplier_14p18); + const auto vout_14p18_11 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_11, vmultiplier_14p18); + + // These shift rights are to revert the multiplication by twopwr18. Hard limit of a maximum shift by 8 requires multiple shift instructions to achieve this. + const auto vout_15p1_00 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_00)); + const auto vout_15p1_01 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_01)); + const auto vout_15p1_10 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_10)); + const auto vout_15p1_11 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_11)); + + const auto vout_15p1_0 = wrapper::vcombine( + vout_15p1_00, + vout_15p1_01); + + const auto vout_15p1_1 = wrapper::vcombine( + vout_15p1_10, + vout_15p1_11); + const auto out_ptr = reinterpret_cast(out_it.ptr()); + + const auto vout_8p0 = wrapper::vcombine( + wrapper::vqrshrn<2>(vout_15p1_0), + wrapper::vqrshrn<2>(vout_15p1_1)); + wrapper::vstore(out_ptr + x, vout_8p0); + } + + //Process the left-over elements. + for(; x < window_end_x; ++x) + { +#ifdef __aarch64__ + out_ptr[x] = wrapper::vqrshrn<2>(wrapper::vqrshrn<8>(wrapper::vqrshrn_ex<8, ScalarType>((multiplier_14p18 * (int32_t(a_ptr[x]) - a_offset_16p0) * (int32_t( + b_val) - b_offset_16p0)) + out_offset_14p18))); +#else //__aarch64__ + out_ptr[x] = utility::clamp(support::cpp11::lround(multiplier * ((float(a_ptr[x]) - a_offset) * (float(b_val) - b_offset)) + float(out_offset))); +#endif //__aarch64__ + } + }, + a_input_it, b_input_it, out_it); + } + else + { + const auto voffset0_16p0 = wrapper::vdup_n(in0_offset_16p0, wrapper::traits::vector_64_tag()); + const auto voffset1_16p0 = wrapper::vdup_n(in1_offset_16p0, wrapper::traits::vector_64_tag()); + const auto voffsetout_14p18 = wrapper::vdup_n(out_offset_14p18, wrapper::traits::vector_128_tag()); + const auto vmultiplier_14p18 = wrapper::vdup_n(multiplier_14p18, wrapper::traits::vector_128_tag()); + + // Clear the x dimension on the execution window as we process the whole row each iteration. + in0_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + in1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator in0_it(src0, in0_win); + Iterator in1_it(src1, in1_win); + Iterator out_it(dst, win); + + execute_window_loop( + win, [&](const Coordinates &) + { + const auto in0_ptr = reinterpret_cast(in0_it.ptr()); + const auto in1_ptr = reinterpret_cast(in1_it.ptr()); + const auto out_ptr = reinterpret_cast(out_it.ptr()); + + int x = window_start_x; + + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + // Load the inputs. + const auto vin0_8p0 = wrapper::vloadq(in0_ptr + x); + const auto vin1_8p0 = wrapper::vloadq(in1_ptr + x); + + // Widen the input elements to signed 16-bit regardless of the input signedness. + const auto vin0_16p0_0 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(vin0_8p0))); + const auto vin0_16p0_1 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(vin0_8p0))); + const auto vin1_16p0_0 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(vin1_8p0))); + const auto vin1_16p0_1 = wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(vin1_8p0))); + + const auto voffseted0_32p0_00 = wrapper::vsubl(wrapper::vgetlow(vin0_16p0_0), voffset0_16p0); + const auto voffseted0_32p0_01 = wrapper::vsubl(wrapper::vgethigh(vin0_16p0_0), voffset0_16p0); + const auto voffseted0_32p0_10 = wrapper::vsubl(wrapper::vgetlow(vin0_16p0_1), voffset0_16p0); + const auto voffseted0_32p0_11 = wrapper::vsubl(wrapper::vgethigh(vin0_16p0_1), voffset0_16p0); + + const auto voffseted1_32p0_00 = wrapper::vsubl(wrapper::vgetlow(vin1_16p0_0), voffset1_16p0); + const auto voffseted1_32p0_01 = wrapper::vsubl(wrapper::vgethigh(vin1_16p0_0), voffset1_16p0); + const auto voffseted1_32p0_10 = wrapper::vsubl(wrapper::vgetlow(vin1_16p0_1), voffset1_16p0); + const auto voffseted1_32p0_11 = wrapper::vsubl(wrapper::vgethigh(vin1_16p0_1), voffset1_16p0); + + const auto vinnermul_32p0_00 = wrapper::vmul(voffseted0_32p0_00, voffseted1_32p0_00); + const auto vinnermul_32p0_01 = wrapper::vmul(voffseted0_32p0_01, voffseted1_32p0_01); + const auto vinnermul_32p0_10 = wrapper::vmul(voffseted0_32p0_10, voffseted1_32p0_10); + const auto vinnermul_32p0_11 = wrapper::vmul(voffseted0_32p0_11, voffseted1_32p0_11); + + const auto vout_14p18_00 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_00, vmultiplier_14p18); + const auto vout_14p18_01 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_01, vmultiplier_14p18); + const auto vout_14p18_10 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_10, vmultiplier_14p18); + const auto vout_14p18_11 = wrapper::vmla(voffsetout_14p18, vinnermul_32p0_11, vmultiplier_14p18); + + // These shift rights are to revert the multiplication by twopwr18. Hard limit of a maximum shift by 8 requires multiple shift instructions to achieve this. + const auto vout_14p2_00 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_00)); + const auto vout_14p2_01 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_01)); + const auto vout_14p2_10 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_10)); + const auto vout_14p2_11 = wrapper::vqrshrn_ex<8, ScalarType>(wrapper::vshrq_n<8>(vout_14p18_11)); + + const auto vout_14p2_0 = wrapper::vcombine( + vout_14p2_00, + vout_14p2_01); + + const auto vout_14p2_1 = wrapper::vcombine( + vout_14p2_10, + vout_14p2_11); + + const auto vout_8p0 = wrapper::vcombine( + wrapper::vqrshrn<2>(vout_14p2_0), + wrapper::vqrshrn<2>(vout_14p2_1)); + wrapper::vstore(out_ptr + x, vout_8p0); + } + + //Process the left-over elements. + for(; x < window_end_x; ++x) + { +#ifdef __aarch64__ + out_ptr[x] = wrapper::vqrshrn<2>(wrapper::vqrshrn<8>(wrapper::vqrshrn_ex<8, ScalarType>((multiplier_14p18 * (int32_t(in0_ptr[x]) - in0_offset_16p0) * (int32_t( + in1_ptr[x]) - in1_offset_16p0)) + out_offset_14p18))); +#else //__aarch64__ + out_ptr[x] = utility::clamp(support::cpp11::lround(multiplier * ((float(in0_ptr[x]) - in0_offset) * (float(in1_ptr[x]) - in1_offset)) + float(out_offset))); +#endif //__aarch64__ + } + }, + in0_it, in1_it, out_it); + } +} + void mul_saturate_QSYMM16_QSYMM16_QSYMM16(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, float scale) { const UniformQuantizationInfo input1_qua_info = src1->info()->quantization_info().uniform(); @@ -316,7 +573,8 @@ void mul_saturate_QSYMM16_QSYMM16_QSYMM16(const ITensor *src1, const ITensor *sr const UniformQuantizationInfo tmp_qua_info = { output_qua_info.scale / scale, output_qua_info.offset }; - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto input1_ptr = reinterpret_cast(input1.ptr()); const auto input2_ptr = reinterpret_cast(input2.ptr()); @@ -397,7 +655,8 @@ void mul_QSYMM16_QSYMM16_S32(const ITensor *src1, const ITensor *src2, ITensor * const auto window_start_x = static_cast(window.x().start()); const auto window_end_x = static_cast(window.x().end()); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto input1_ptr = reinterpret_cast(input1.ptr()); const auto input2_ptr = reinterpret_cast(input2.ptr()); @@ -488,7 +747,8 @@ void mul_U8_U8_U8(const ITensor *src1, const ITensor *src2, ITensor *out, const const auto window_start_x = static_cast(window.x().start()); const auto window_end_x = static_cast(window.x().end()); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto input1_ptr = reinterpret_cast(input1.ptr()); const auto input2_ptr = reinterpret_cast(input2.ptr()); @@ -653,7 +913,8 @@ void mul_S16_S16_S16(const ITensor *src1, const ITensor *src2, ITensor *out, con const auto window_start_x = static_cast(window.x().start()); const auto window_end_x = static_cast(window.x().end()); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto input1_ptr = reinterpret_cast(input1.ptr()); const auto input2_ptr = reinterpret_cast(input2.ptr()); @@ -716,7 +977,7 @@ void mul_S16_S16_S16(const ITensor *src1, const ITensor *src2, ITensor *out, con input1, input2, dst); } -template +template inline int32x4_t mul_S32_S32_S32_n_loop(const int32x4_t &src1, const int32x4_t &src2, int n) { const int32x2_t input1_1 = vget_low_s32(src1); @@ -756,7 +1017,7 @@ inline int32x4_t mul_S32_S32_S32_n_loop(const int32x4_t &src1, const int32x4_t & } } -template +template inline int32x4x2_t mul_S32_S32_S32_n_k(const int32x4x2_t &src1, const int32x4x2_t &src2, int n) { const int32x4x2_t result = @@ -803,7 +1064,8 @@ void mul_S32_S32_S32(const ITensor *src1, const ITensor *src2, ITensor *out, con Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); Iterator dst(out, win); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto non_broadcast_input_ptr = reinterpret_cast(non_broadcast_input.ptr()); const auto output_ptr = reinterpret_cast(dst.ptr()); @@ -868,7 +1130,8 @@ void mul_S32_S32_S32(const ITensor *src1, const ITensor *src2, ITensor *out, con Iterator input2(src2, input2_win); Iterator dst(out, win); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto input1_ptr = reinterpret_cast(input1.ptr()); const auto input2_ptr = reinterpret_cast(input2.ptr()); @@ -955,7 +1218,8 @@ void mul_F32_F32_F32(const ITensor *src1, const ITensor *src2, ITensor *out, con Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); Iterator dst(out, win); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto non_broadcast_input_ptr = reinterpret_cast(non_broadcast_input.ptr()); const auto output_ptr = reinterpret_cast(dst.ptr()); @@ -992,7 +1256,8 @@ void mul_F32_F32_F32(const ITensor *src1, const ITensor *src2, ITensor *out, con Iterator input2(src2, input2_win); Iterator dst(out, win); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto input1_ptr = reinterpret_cast(input1.ptr()); const auto input2_ptr = reinterpret_cast(input2.ptr()); @@ -1053,7 +1318,8 @@ void c_mul_F32_F32_F32_n(const ITensor *src1, const ITensor *src2, ITensor *out, Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); Iterator dst(out, win); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto non_broadcast_input_ptr = reinterpret_cast(non_broadcast_input.ptr()); const auto output_ptr = reinterpret_cast(dst.ptr()); @@ -1106,7 +1372,8 @@ void c_mul_F32_F32_F32_n(const ITensor *src1, const ITensor *src2, ITensor *out, Iterator input2(src2, input2_win); Iterator dst(out, win); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto input1_ptr = reinterpret_cast(input1.ptr()); const auto input2_ptr = reinterpret_cast(input2.ptr()); @@ -1180,12 +1447,13 @@ void mul_F16_F16_F16(const ITensor *src1, const ITensor *src2, ITensor *out, con Iterator broadcast_input(broadcast_tensor, broadcast_win); Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win); Iterator dst(out, win); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto non_broadcast_input_ptr = reinterpret_cast(non_broadcast_input.ptr()); const auto output_ptr = reinterpret_cast(dst.ptr()); const auto broadcast_value = *reinterpret_cast(broadcast_input.ptr()); - const float16x8x2_t broadcast_value_vec = + const float16x8x2_t broadcast_value_vec = { { vdupq_n_f16(broadcast_value), @@ -1230,7 +1498,8 @@ void mul_F16_F16_F16(const ITensor *src1, const ITensor *src2, ITensor *out, con Iterator input1(src1, input1_win); Iterator input2(src2, input2_win); Iterator dst(out, win); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto input1_ptr = reinterpret_cast(input1.ptr()); const auto input2_ptr = reinterpret_cast(input2.ptr()); @@ -1254,7 +1523,7 @@ void mul_F16_F16_F16(const ITensor *src1, const ITensor *src2, ITensor *out, con } }; const float16x8_t scale_vec = vdupq_n_f16(scale); - const float16x8x2_t result = + const float16x8x2_t result = { { vmulq_f16(vmulq_f16(ta1.val[0], ta2.val[0]), scale_vec), @@ -1298,7 +1567,8 @@ void mul_U8_U8_S16(const ITensor *src1, const ITensor *src2, ITensor *out, const const auto window_start_x = static_cast(window.x().start()); const auto window_end_x = static_cast(window.x().end()); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto input1_ptr = reinterpret_cast(input1.ptr()); const auto input2_ptr = reinterpret_cast(input2.ptr()); @@ -1396,7 +1666,8 @@ void mul_S16_U8_S16(const ITensor *src1, const ITensor *src2, ITensor *out, cons const auto window_start_x = static_cast(window.x().start()); const auto window_end_x = static_cast(window.x().end()); - execute_window_loop(win, [&](const Coordinates &) + execute_window_loop( + win, [&](const Coordinates &) { const auto input1_ptr = reinterpret_cast(input1.ptr()); const auto input2_ptr = reinterpret_cast(input2.ptr()); @@ -1520,14 +1791,27 @@ void CpuMulKernel::configure(ITensorInfo *src1, ITensorInfo *src2, ITensorInfo * case DataType::QASYMM8: if(dt_input2 == DataType::QASYMM8 && dt_output == DataType::QASYMM8) { - _func_quantized = &mul_saturate_quantized_8; + if(mul_q8_neon_fixedpoint_possible(src1, src2, dst, scale)) + { + _func_quantized = &mul_q8_neon_fixedpoint; + } + else + { + _func_quantized = &mul_saturate_quantized_8; + } } break; case DataType::QASYMM8_SIGNED: if(dt_input2 == DataType::QASYMM8_SIGNED) { - _func_quantized = &mul_saturate_quantized_8; - ; + if(mul_q8_neon_fixedpoint_possible(src1, src2, dst, scale)) + { + _func_quantized = &mul_q8_neon_fixedpoint; + } + else + { + _func_quantized = &mul_saturate_quantized_8; + } } break; case DataType::QSYMM16: -- cgit v1.2.1