Michalis Spyrou | aa51a5b | 2020-11-22 00:49:42 +0000 | [diff] [blame] | 1 | /* |
Michalis Spyrou | 168d6a8 | 2022-05-03 17:15:42 +0100 | [diff] [blame] | 2 | * Copyright (c) 2020-2022 Arm Limited. |
Michalis Spyrou | aa51a5b | 2020-11-22 00:49:42 +0000 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #ifndef ARM_COMPUTE_SVEASYMM_H |
| 25 | #define ARM_COMPUTE_SVEASYMM_H |
| 26 | |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 27 | #if defined(ARM_COMPUTE_ENABLE_SVE2) |
Michalis Spyrou | aa51a5b | 2020-11-22 00:49:42 +0000 | [diff] [blame] | 28 | #include "src/core/NEON/SVEMath.h" |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 29 | |
Michalis Spyrou | aa51a5b | 2020-11-22 00:49:42 +0000 | [diff] [blame] | 30 | #include <arm_sve.h> |
| 31 | |
| 32 | namespace arm_compute |
| 33 | { |
| 34 | /** Perform a multiply-accumulate on all components of a QASYMM8 vector |
| 35 | * |
| 36 | * vd*vs + vo |
| 37 | * |
| 38 | * @param[in] pg Predicate value. |
| 39 | * @param[in] vd Input vector value in QASYMM8 format |
| 40 | * @param[in] vs Vector multiplier in F32 format. The multiplier value must be duplicated across all four lanes. |
| 41 | * @param[in] vo Vector addend in F32 format. The addend value must be duplicated across all four lanes. |
| 42 | * |
| 43 | * @return A vector in QASYMM8 format, saturated to fit |
| 44 | */ |
| 45 | svuint8_t svmla_qasymm8_z(svbool_t pg, svuint8_t vd, svfloat32_t vs, svfloat32_t vo); |
| 46 | |
| 47 | /** Perform a multiply-accumulate on all components of a QASYMM8_SIGNED vector |
| 48 | * |
| 49 | * vd*vs + vo |
| 50 | * |
| 51 | * @param[in] pg Predicate value. |
| 52 | * @param[in] vd Input vector value in QASYMM8_SIGNED format |
| 53 | * @param[in] vs Vector multiplier in F32 format. The multiplier value must be duplicated across all four lanes. |
| 54 | * @param[in] vo Vector addend in F32 format. The addend value must be duplicated across all four lanes. |
| 55 | * |
| 56 | * @return A vector in QASYMM8_SIGNED format, saturated to fit |
| 57 | */ |
| 58 | svint8_t svmla_qasymm8_signed_z(svbool_t pg, svint8_t vd, svfloat32_t vs, svfloat32_t vo); |
| 59 | |
| 60 | /** Dequantize following an asymmetric quantization scheme a sve vector. |
| 61 | * |
| 62 | * @param[in] pg Predicate value. |
| 63 | * @param[in] qv Input values to be dequantized. |
| 64 | * @param[in] scale Quantization scaling factor. |
| 65 | * @param[in] offset Zero quantization offset. |
| 66 | * |
| 67 | * @return Dequantized values in an sve vector |
| 68 | */ |
| 69 | inline svfloat32x4_t svdequantize_z(svbool_t pg, const svuint8_t &qv, float scale, int32_t offset) |
| 70 | { |
Michalis Spyrou | 168d6a8 | 2022-05-03 17:15:42 +0100 | [diff] [blame] | 71 | const auto voffset = svdup_n_s32(offset); |
| 72 | const auto vscale = svdup_n_f32(scale); |
| 73 | const svfloat32x4_t vdequantized_input = svcreate4_f32( |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 74 | svmul_f32_z(pg, |
| 75 | svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(qv))), voffset)), |
| 76 | vscale), |
| 77 | svmul_f32_z(pg, |
| 78 | svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(qv))), voffset)), |
| 79 | vscale), |
| 80 | svmul_f32_z(pg, |
| 81 | svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(qv))), voffset)), |
| 82 | vscale), |
| 83 | svmul_f32_z(pg, |
| 84 | svcvt_f32_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(qv))), voffset)), |
| 85 | vscale)); |
Michalis Spyrou | aa51a5b | 2020-11-22 00:49:42 +0000 | [diff] [blame] | 86 | return vdequantized_input; |
| 87 | } |
| 88 | |
| 89 | /** Dequantize an sve vector |
| 90 | * |
| 91 | * @param[in] pg Predicate value. |
| 92 | * @param[in] qv Input values to be dequantized. |
| 93 | * @param[in] qi Quantization information to be used in the computation. |
| 94 | * |
| 95 | * @return Dequantized values in an sve vector |
| 96 | */ |
| 97 | inline svfloat32x4_t svdequantize_z(svbool_t pg, const svuint8_t &qv, const UniformQuantizationInfo &qi) |
| 98 | { |
| 99 | return svdequantize_z(pg, qv, qi.scale, qi.offset); |
| 100 | } |
| 101 | |
| 102 | /** Dequantize an sve vector stored as signed asymmetric. |
| 103 | * |
| 104 | * @param[in] pg Predicate value. |
| 105 | * @param[in] qv Input values to be dequantized. |
| 106 | * @param[in] scale Quantization scaling factor. |
| 107 | * @param[in] offset Zero quantization offset. |
| 108 | * |
| 109 | * @return Dequantized values in a sve vector |
| 110 | */ |
| 111 | inline svfloat32x4_t svdequantize_z(svbool_t pg, const svint8_t &qv, float scale, int32_t offset) |
| 112 | { |
Michalis Spyrou | 168d6a8 | 2022-05-03 17:15:42 +0100 | [diff] [blame] | 113 | const auto voffset = svdup_n_s32(offset); |
| 114 | const auto vscale = svdup_n_f32(scale); |
| 115 | const svfloat32x4_t vdequantized_input = svcreate4_f32( |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 116 | svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlb_s16(qv)), voffset)), vscale), |
| 117 | svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlb_s16(qv)), voffset)), vscale), |
| 118 | svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlt_s16(qv)), voffset)), vscale), |
| 119 | svmul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlt_s16(qv)), voffset)), vscale)); |
Michalis Spyrou | 168d6a8 | 2022-05-03 17:15:42 +0100 | [diff] [blame] | 120 | |
Michalis Spyrou | aa51a5b | 2020-11-22 00:49:42 +0000 | [diff] [blame] | 121 | return vdequantized_input; |
| 122 | } |
| 123 | |
| 124 | /** Dequantize an sve vector. |
| 125 | * |
| 126 | * @param[in] pg Predicate value. |
| 127 | * @param[in] qv Input values to be dequantized. |
| 128 | * @param[in] qi Quantization information to be used in the computation. |
| 129 | * |
| 130 | * @return Dequantized values in an sve vector |
| 131 | */ |
| 132 | inline svfloat32x4_t svdequantize_z(svbool_t pg, const svint8_t &qv, const UniformQuantizationInfo &qi) |
| 133 | { |
| 134 | return svdequantize_z(pg, qv, qi.scale, qi.offset); |
| 135 | } |
| 136 | |
| 137 | /** Dequantize following symmetric quantization scheme on an sve vector. |
| 138 | * |
| 139 | * @param[in] pg Predicate value. |
| 140 | * @param[in] qv Input values to be dequantized. |
| 141 | * @param[in] vscale Vector containing quantization scaling factors. |
| 142 | * |
| 143 | * @return Dequantized values in a sve vector |
| 144 | */ |
| 145 | inline svfloat32x4_t svdequantize_z(svbool_t pg, const svint8_t &qv, const svfloat32x4_t vscale) |
| 146 | { |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 147 | const svfloat32x4_t vdequantized_input = |
| 148 | svcreate4_f32(svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(svmovlb_s16(qv))), svget4_f32(vscale, 0)), |
| 149 | svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(svmovlb_s16(qv))), svget4_f32(vscale, 1)), |
| 150 | svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(svmovlt_s16(qv))), svget4_f32(vscale, 2)), |
| 151 | svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(svmovlt_s16(qv))), svget4_f32(vscale, 3))); |
Michalis Spyrou | 168d6a8 | 2022-05-03 17:15:42 +0100 | [diff] [blame] | 152 | |
Michalis Spyrou | aa51a5b | 2020-11-22 00:49:42 +0000 | [diff] [blame] | 153 | return vdequantized_input; |
| 154 | } |
| 155 | |
| 156 | /** Dequantize following a symmetric quantization scheme an sve vector. |
| 157 | * |
| 158 | * @param[in] qv Input values to be dequantized. |
| 159 | * @param[in] scale Quantization scaling factor. |
| 160 | * |
| 161 | * @return Dequantized values in a sve vector |
| 162 | */ |
| 163 | inline svfloat32x4_t svdequantize_z(svbool_t pg, const svint8_t &qv, float scale) |
| 164 | { |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 165 | const auto vscale = svdup_n_f32(scale); |
| 166 | const svfloat32x4_t vdequantized_input = |
| 167 | svcreate4_f32(svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(svmovlb_s16(qv))), vscale), |
| 168 | svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(svmovlb_s16(qv))), vscale), |
| 169 | svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlb_s32(svmovlt_s16(qv))), vscale), |
| 170 | svmul_f32_z(pg, svcvt_f32_s32_z(pg, svmovlt_s32(svmovlt_s16(qv))), vscale)); |
Michalis Spyrou | aa51a5b | 2020-11-22 00:49:42 +0000 | [diff] [blame] | 171 | return vdequantized_input; |
| 172 | } |
| 173 | |
| 174 | /** Quantize an sve vector holding floating point values. |
| 175 | * |
| 176 | * @param[in] pg Predicate value. |
| 177 | * @param[in] qv Input values to be quantized. |
| 178 | * @param[in] qi Quantization information to be used in the computation. |
| 179 | * |
| 180 | * @return An sve vector holding the quantized values |
| 181 | */ |
| 182 | inline svuint8_t svquantize_z(svbool_t pg, const svfloat32x4_t qv, const UniformQuantizationInfo &qi) |
| 183 | { |
| 184 | const float scale = qi.scale; |
| 185 | const int offset = qi.offset; |
| 186 | const auto voffset = svdup_n_f32(offset); |
| 187 | const auto vinvscale = svdup_n_f32(1.f / scale); |
| 188 | |
| 189 | const auto rf_0 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 0), vinvscale)); |
| 190 | const auto rf_1 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 1), vinvscale)); |
| 191 | const auto rf_2 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 2), vinvscale)); |
| 192 | const auto rf_3 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 3), vinvscale)); |
| 193 | |
| 194 | const auto pa = svqxtnt_u32(svqxtnb_u32(rf_0), rf_1); |
| 195 | const auto pb = svqxtnt_u32(svqxtnb_u32(rf_2), rf_3); |
| 196 | |
| 197 | return svqxtnt_u16(svqxtnb_u16(pa), pb); |
| 198 | } |
| 199 | |
| 200 | /** Signed quantize an sve vector holding floating point values. |
| 201 | * |
| 202 | * @param[in] pg Predicate value. |
| 203 | * @param[in] qv Input values to be quantized. |
| 204 | * @param[in] qi Quantization information to be used in the computation. |
| 205 | * |
| 206 | * @return An sve vector holding the quantized values |
| 207 | */ |
| 208 | inline svint8_t svquantize_signed_z(svbool_t pg, const svfloat32x4_t qv, const UniformQuantizationInfo &qi) |
| 209 | { |
| 210 | const float scale = qi.scale; |
| 211 | const int offset = qi.offset; |
| 212 | const auto voffset = svdup_n_f32(offset); |
| 213 | const auto vinvscale = svdup_n_f32(1.f / scale); |
| 214 | const auto rf_0 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 0), vinvscale)); |
| 215 | const auto rf_1 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 1), vinvscale)); |
| 216 | const auto rf_2 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 2), vinvscale)); |
| 217 | const auto rf_3 = svcvt_s32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 3), vinvscale)); |
| 218 | |
| 219 | const auto pa = svqxtnt_s32(svqxtnb_s32(rf_0), rf_1); |
| 220 | const auto pb = svqxtnt_s32(svqxtnb_s32(rf_2), rf_3); |
| 221 | |
| 222 | return svqxtnt_s16(svqxtnb_s16(pa), pb); |
| 223 | } |
| 224 | |
| 225 | /** Quantize to QASYMM16 an sve vector holding 16 floating point values. |
| 226 | * |
| 227 | * @param[in] pg Predicate value. |
| 228 | * @param[in] qv Input values to be quantized. |
| 229 | * @param[in] qi Quantization information to be used in the computation. |
| 230 | * |
| 231 | * @return An sve vector holding the quantized values |
| 232 | */ |
| 233 | inline svuint16x2_t svquantize_qasymm16_z(svbool_t pg, const svfloat32x4_t qv, const UniformQuantizationInfo &qi) |
| 234 | { |
| 235 | const float scale = qi.scale; |
| 236 | const int offset = qi.offset; |
| 237 | const auto voffset = svdup_n_f32(offset); |
| 238 | const auto vinvscale = svdup_n_f32(1.f / scale); |
| 239 | |
| 240 | const auto rf_0 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 0), vinvscale)); |
| 241 | const auto rf_1 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 1), vinvscale)); |
| 242 | const auto rf_2 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 2), vinvscale)); |
| 243 | const auto rf_3 = svcvt_u32_f32_z(pg, svmla_f32_z(pg, voffset, svget4_f32(qv, 3), vinvscale)); |
| 244 | |
| 245 | const auto pa = svqxtnt_u32(svqxtnb_u32(rf_0), rf_1); |
| 246 | const auto pb = svqxtnt_u32(svqxtnb_u32(rf_2), rf_3); |
| 247 | |
| 248 | return svcreate2_u16(pa, pb); |
| 249 | } |
| 250 | } // namespace arm_compute |
| 251 | #include "src/core/NEON/SVEAsymm.inl" |
Michalis Spyrou | 20fca52 | 2021-06-07 14:23:57 +0100 | [diff] [blame] | 252 | #endif /* defined(ARM_COMPUTE_ENABLE_SVE2) */ |
Michalis Spyrou | aa51a5b | 2020-11-22 00:49:42 +0000 | [diff] [blame] | 253 | #endif // ARM_COMPUTE_NEASYMM_H |