diff options
Diffstat (limited to 'src/cpu/kernels/activation/generic/sve2/qasymm8.cpp')
| -rw-r--r-- | src/cpu/kernels/activation/generic/sve2/qasymm8.cpp | 264 |
1 files changed, 144 insertions, 120 deletions
diff --git a/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp b/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp index bc9bc7aa3c..7efa9e4b72 100644 --- a/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp +++ b/src/cpu/kernels/activation/generic/sve2/qasymm8.cpp @@ -26,18 +26,21 @@ #include "arm_compute/core/Window.h" #include "arm_compute/function_info/ActivationLayerInfo.h" -#include <cmath> -#include <cstddef> - #include "src/core/NEON/SVEAsymm.h" #include "src/core/NEON/SVEMath.h" + #include <arm_sve.h> +#include <cmath> +#include <cstddef> namespace arm_compute { namespace cpu { -void sve2_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window) +void sve2_qasymm8_activation(const ITensor *src, + ITensor *dst, + const ActivationLayerInfo &act_info, + const Window &window) { const auto window_start_x = static_cast<int>(window.x().start()); const auto window_end_x = static_cast<int>(window.x().end()); @@ -61,7 +64,7 @@ void sve2_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationL // Initialise scale/offset for re-quantization bool requant = true; - if(qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset) + if (qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset) { requant = false; } @@ -78,139 +81,160 @@ void sve2_qasymm8_activation(const ITensor *src, ITensor *dst, const ActivationL const auto vo_s32 = svdup_n_s32(o_s32); // Initialise scale/offset for re-quantization for leaky relu - int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); - int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8), - arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + int32_t s_leaky_s32 = round(s * act_info.a() * (1 << 8), arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + int32_t o_leaky_s32 = round((-qi_in.offset * s * act_info.a() + qi_out.offset) * (1 << 8), + arm_compute::RoundingPolicy::TO_NEAREST_EVEN); const auto vs_leaky_s32 = svdup_n_s32(s_leaky_s32); const auto vo_leaky_s32 = svdup_n_s32(o_leaky_s32); - execute_window_loop(win_collapsed, [&](const Coordinates &) - { - const auto input_ptr = reinterpret_cast<const uint8_t *>(input.ptr()); - const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr()); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const uint8_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr()); - svuint8_t tmp; + svuint8_t tmp; - int x = window_start_x; - svbool_t pg = svwhilelt_b8(x, window_end_x); - do - { - const auto vin = svld1_u8(pg, input_ptr + x); - if(act == ActivationLayerInfo::ActivationFunction::RELU) - { - // Perform activation - tmp = svmax_u8_z(pg, vconst_0, vin); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) - { - // Perform activation - tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vconst_0, vin)); - // Re-quantize to new output space - tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; - } - else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) - { - // Perform activation - tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vb, vin)); - // Re-quantize to new output space - tmp = svmla_qasymm8_z(pg, tmp, vs, vo); - } - else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC) - { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = svcreate4_f32(svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))), - svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3)))))); - - // Re-quantize to new output space - tmp = svquantize_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::TANH) + int x = window_start_x; + svbool_t pg = svwhilelt_b8(x, window_end_x); + do { - // De-quantize - const auto vin_deq = svdequantize_z(pg, vin, qi_in); - // Perform activation - const svfloat32x4_t tmp_dep = svcreate4_f32(svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))), - svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32)))); - - // Re-quantize to new output space - tmp = svquantize_z(pg, tmp_dep, qi_out); - } - else if(act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) - { - svbool_t p0, p1, p2, p3; - svint32x4_t tmp_dep; - - // Expand to int32 - const svint32x4_t vin_s32 = svcreate4_s32( - svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(vin))), - svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(vin))), - svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(vin))), - svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(vin)))); - - // Compare elements to input offset - if(qi_in.scale >= 0) + const auto vin = svld1_u8(pg, input_ptr + x); + if (act == ActivationLayerInfo::ActivationFunction::RELU) { - p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); - p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); - p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); - p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + // Perform activation + tmp = svmax_u8_z(pg, vconst_0, vin); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; } - else + else if (act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) { - p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); - p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); - p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); - p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + // Perform activation + tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vconst_0, vin)); + // Re-quantize to new output space + tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp; } - - // Multiply negative elements and requantize if necessary - if(requant) + else if (act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + // Perform activation + tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vb, vin)); + // Re-quantize to new output space + tmp = svmla_qasymm8_z(pg, tmp, vs, vo); + } + else if (act == ActivationLayerInfo::ActivationFunction::LOGISTIC) + { + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = svcreate4_f32( + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))), + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))), + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))), + svdiv_f32_z( + pg, vconst_1, + svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3)))))); + + // Re-quantize to new output space + tmp = svquantize_z(pg, tmp_dep, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::TANH) { - tmp_dep = svcreate4_s32( - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p0, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 0), svsel(p0, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p1, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 1), svsel(p1, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p2, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 2), svsel(p2, vs_leaky_s32, vs_s32)), 8), - svasr_n_s32_m(pg, svmla_s32_m(pg, svsel(p3, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 3), svsel(p3, vs_leaky_s32, vs_s32)), 8)); + // De-quantize + const auto vin_deq = svdequantize_z(pg, vin, qi_in); + // Perform activation + const svfloat32x4_t tmp_dep = svcreate4_f32( + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))), + svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32)))); + + // Re-quantize to new output space + tmp = svquantize_z(pg, tmp_dep, qi_out); + } + else if (act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + { + svbool_t p0, p1, p2, p3; + svint32x4_t tmp_dep; + + // Expand to int32 + const svint32x4_t vin_s32 = svcreate4_s32(svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(vin))), + svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(vin))), + svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(vin))), + svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(vin)))); + + // Compare elements to input offset + if (qi_in.scale >= 0) + { + p0 = svcmplt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); + p1 = svcmplt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); + p2 = svcmplt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); + p3 = svcmplt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + } + else + { + p0 = svcmpgt_s32(pg, svget4_s32(vin_s32, 0), voffset_in); + p1 = svcmpgt_s32(pg, svget4_s32(vin_s32, 1), voffset_in); + p2 = svcmpgt_s32(pg, svget4_s32(vin_s32, 2), voffset_in); + p3 = svcmpgt_s32(pg, svget4_s32(vin_s32, 3), voffset_in); + } + + // Multiply negative elements and requantize if necessary + if (requant) + { + tmp_dep = svcreate4_s32( + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p0, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 0), + svsel(p0, vs_leaky_s32, vs_s32)), + 8), + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p1, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 1), + svsel(p1, vs_leaky_s32, vs_s32)), + 8), + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p2, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 2), + svsel(p2, vs_leaky_s32, vs_s32)), + 8), + svasr_n_s32_m(pg, + svmla_s32_m(pg, svsel(p3, vo_leaky_s32, vo_s32), svget4_s32(vin_s32, 3), + svsel(p3, vs_leaky_s32, vs_s32)), + 8)); + } + else + { + tmp_dep = svcreate4_s32( + svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8), + svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8)); + } + + // Convert uint32 vectors to uint16 vectors (with saturation) + const auto v_low_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1)); + const auto v_high_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3)); + + // convert uint16 vectors to uint8 vectors (with saturation) + tmp = svqxtnt_u16(svqxtnb_u16(v_low_u16), v_high_u16); } else { - tmp_dep = svcreate4_s32( - svasr_n_s32_m(p0, svmad_s32_m(p0, svget4_s32(vin_s32, 0), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p1, svmad_s32_m(p1, svget4_s32(vin_s32, 1), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p2, svmad_s32_m(p2, svget4_s32(vin_s32, 2), vs_leaky_s32, vo_leaky_s32), 8), - svasr_n_s32_m(p3, svmad_s32_m(p3, svget4_s32(vin_s32, 3), vs_leaky_s32, vo_leaky_s32), 8)); + ARM_COMPUTE_ERROR("Unsupported activation function"); } - // Convert uint32 vectors to uint16 vectors (with saturation) - const auto v_low_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 0)), svget4_s32(tmp_dep, 1)); - const auto v_high_u16 = svqxtunt_s32(svqxtunb_s32(svget4_s32(tmp_dep, 2)), svget4_s32(tmp_dep, 3)); - - // convert uint16 vectors to uint8 vectors (with saturation) - tmp = svqxtnt_u16(svqxtnb_u16(v_low_u16), v_high_u16); - } - else - { - ARM_COMPUTE_ERROR("Unsupported activation function"); - } - - svst1_u8(pg, output_ptr + x, tmp); - - x += svcntb(); - pg = svwhilelt_b8(x, window_end_x); + svst1_u8(pg, output_ptr + x, tmp); - } - while(svptest_any(svptrue_b8(), pg)); + x += svcntb(); + pg = svwhilelt_b8(x, window_end_x); - }, - input, output); + } while (svptest_any(svptrue_b8(), pg)); + }, + input, output); } } // namespace cpu } // namespace arm_compute |