|
| 1 | +// Copyright (C) 2024 Intel Corporation |
| 2 | +// SPDX-License-Identifier: Apache-2.0 |
| 3 | +// |
| 4 | + |
| 5 | +#include "jit_conversion_emitters.hpp" |
| 6 | +#include "emitters/utils.hpp" |
| 7 | + |
| 8 | +using namespace dnnl::impl::cpu::aarch64; |
| 9 | +using namespace Xbyak_aarch64; |
| 10 | + |
| 11 | +namespace ov { |
| 12 | +namespace intel_cpu { |
| 13 | +namespace aarch64 { |
| 14 | + |
| 15 | +// In aarch64, conversion between f16 and i16/u16 can be done with single instruction. The supported |
| 16 | +// conversion precicions are f32, i32, f16, i8 (byte), u8 (byte). If we introduce an intermediate |
| 17 | +// precision i16 in the following graph. Then the conversion between each pair of |
| 18 | +// neighbors in this graph will be done with single instruction. |
| 19 | +// f16 - f32 - i32 - i16 - byte |
| 20 | +// | | |
| 21 | +// - - - - - - - - - - |
| 22 | +// Note that using single instruction for conversion between f16 and i16 is only available for |
| 23 | +// architecture ARMv8.2-A or later versions. So ARM platforms like Raspberry (Model name Cortex-A72) |
| 24 | +// with architecture ARMv8 do not support such instructions. And as the isa asimd we supported |
| 25 | +// does not distinguish ARMv8.2 with ARMv8.2-A, conversion between f16 and i16 will still use three |
| 26 | +// instructions f16 -> f32 -> i32 -> i16 (f16 <- f32 <- i32 <- i16). |
| 27 | +template <typename TReg> |
| 28 | +inline void jit_convert_emitter::cvt_f16_to_f32(const TReg &src, const TReg &dst) const { |
| 29 | + h->fcvtl(dst.s4, src.h4); |
| 30 | +} |
| 31 | + |
| 32 | +template <typename TReg> |
| 33 | +inline void jit_convert_emitter::cvt_f32_to_f16(const TReg &src, const TReg &dst) const { |
| 34 | + h->fcvtn(dst.h4, src.s4); |
| 35 | +} |
| 36 | + |
| 37 | +template <typename TReg> |
| 38 | +inline void jit_convert_emitter::cvt_f32_to_i32(const TReg &src, const TReg &dst) const { |
| 39 | + h->fcvtzs(dst.s, src.s); |
| 40 | +} |
| 41 | + |
| 42 | +template <typename TReg> |
| 43 | +inline void jit_convert_emitter::cvt_i32_to_f32(const TReg &src, const TReg &dst) const { |
| 44 | + h->scvtf(dst.s, src.s); |
| 45 | +} |
| 46 | + |
| 47 | +template <typename TReg> |
| 48 | +inline void jit_convert_emitter::cvt_i32_to_i16(const TReg &src, const TReg &dst, bool is_saturated) const { |
| 49 | + if (is_saturated) { |
| 50 | + h->sqxtn(dst.h4, src.s4); |
| 51 | + } else { |
| 52 | + h->xtn(dst.h4, src.s4); |
| 53 | + } |
| 54 | +} |
| 55 | + |
| 56 | +template <typename TReg> |
| 57 | +inline void jit_convert_emitter::cvt_i16_to_i32(const TReg &src, const TReg &dst) const { |
| 58 | + h->sxtl(dst.s4, src.h4); |
| 59 | +} |
| 60 | + |
| 61 | +template <typename TReg> |
| 62 | +inline void jit_convert_emitter::cvt_f16_to_i16(const TReg &src, const TReg &dst) const { |
| 63 | + h->fcvtzs(dst.h4, src.h4); |
| 64 | +} |
| 65 | + |
| 66 | +template <typename TReg> |
| 67 | +inline void jit_convert_emitter::cvt_i16_to_f16(const TReg &src, const TReg &dst) const { |
| 68 | + h->scvtf(dst.h4, src.h4); |
| 69 | +} |
| 70 | + |
| 71 | +template <typename TReg> |
| 72 | +inline void jit_convert_emitter::cvt_i16_to_byte(const TReg &src, const TReg &dst, bool is_signed, bool is_saturated) const { |
| 73 | + if (is_saturated) { |
| 74 | + if (is_signed) { |
| 75 | + h->sqxtn(dst.b8, src.h8); |
| 76 | + } else { |
| 77 | + h->uqxtn(dst.b8, src.h8); |
| 78 | + } |
| 79 | + } else { |
| 80 | + h->xtn(dst.b8, src.h8); |
| 81 | + } |
| 82 | +} |
| 83 | + |
| 84 | +template <typename TReg> |
| 85 | +inline void jit_convert_emitter::cvt_byte_to_i16(const TReg &src, const TReg &dst, bool is_signed) const { |
| 86 | + if (is_signed) { |
| 87 | + h->sxtl(dst.h8, src.b8); |
| 88 | + } else { |
| 89 | + h->uxtl(dst.h8, src.b8); |
| 90 | + } |
| 91 | +} |
| 92 | + |
| 93 | +template <typename TReg> |
| 94 | +void jit_convert_emitter::jit_convert_process(const TReg &src, const TReg &dst, ov::element::Type input_type, ov::element::Type output_type, |
| 95 | + bool is_saturated) const { |
| 96 | + if (input_type == output_type || (!is_saturated && |
| 97 | + one_of(input_type, ov::element::i8, ov::element::u8) && one_of(output_type, ov::element::i8, ov::element::u8))) { |
| 98 | + if (src.getIdx() != dst.getIdx()) { |
| 99 | + h->mov(dst.b16, src.b16); |
| 100 | + } |
| 101 | + return; |
| 102 | + } |
| 103 | + |
| 104 | + switch (output_type) { |
| 105 | + case ov::element::f32: |
| 106 | + switch (input_type) { |
| 107 | + case ov::element::i32: |
| 108 | + cvt_i32_to_f32<TReg>(src, dst); |
| 109 | + break; |
| 110 | + case ov::element::f16: |
| 111 | + cvt_f16_to_f32<TReg>(src, dst); |
| 112 | + break; |
| 113 | + case ov::element::i8: |
| 114 | + case ov::element::u8: |
| 115 | + cvt_byte_to_i16<TReg>(src, dst, input_type.is_signed()); |
| 116 | + cvt_i16_to_i32<TReg>(dst, dst); |
| 117 | + cvt_i32_to_f32<TReg>(dst, dst); |
| 118 | + break; |
| 119 | + default: |
| 120 | + OV_CPU_JIT_EMITTER_THROW("Unsupported input type: ", input_type.get_type_name()); |
| 121 | + } |
| 122 | + break; |
| 123 | + case ov::element::i32: |
| 124 | + switch (input_type) { |
| 125 | + case ov::element::f32: |
| 126 | + cvt_f32_to_i32<TReg>(src, dst); |
| 127 | + break; |
| 128 | + case ov::element::f16: |
| 129 | + cvt_f16_to_f32<TReg>(src, dst); |
| 130 | + cvt_f32_to_i32<TReg>(dst, dst); |
| 131 | + break; |
| 132 | + case ov::element::i8: |
| 133 | + case ov::element::u8: |
| 134 | + cvt_byte_to_i16<TReg>(src, dst, input_type.is_signed()); |
| 135 | + cvt_i16_to_i32<TReg>(dst, dst); |
| 136 | + break; |
| 137 | + default: |
| 138 | + OV_CPU_JIT_EMITTER_THROW("Unsupported input type: ", input_type.get_type_name()); |
| 139 | + } |
| 140 | + break; |
| 141 | + case ov::element::f16: |
| 142 | + switch (input_type) { |
| 143 | + case ov::element::f32: |
| 144 | + cvt_f32_to_f16<TReg>(src, dst); |
| 145 | + break; |
| 146 | + case ov::element::i32: |
| 147 | + cvt_i32_to_f32<TReg>(src, dst); |
| 148 | + cvt_f32_to_f16<TReg>(dst, dst); |
| 149 | + break; |
| 150 | + case ov::element::i8: |
| 151 | + case ov::element::u8: |
| 152 | + cvt_byte_to_i16<TReg>(src, dst, input_type.is_signed()); |
| 153 | + cvt_i16_to_i32<TReg>(dst, dst); |
| 154 | + cvt_i32_to_f32<TReg>(dst, dst); |
| 155 | + cvt_f32_to_f16<TReg>(dst, dst); |
| 156 | + break; |
| 157 | + default: |
| 158 | + OV_CPU_JIT_EMITTER_THROW("Unsupported input type: ", input_type.get_type_name()); |
| 159 | + } |
| 160 | + break; |
| 161 | + case ov::element::i8: |
| 162 | + case ov::element::u8: |
| 163 | + switch (input_type) { |
| 164 | + case ov::element::f32: |
| 165 | + cvt_f32_to_i32<TReg>(src, dst); |
| 166 | + cvt_i32_to_i16<TReg>(dst, dst, is_saturated); |
| 167 | + cvt_i16_to_byte<TReg>(dst, dst, output_type.is_signed(), is_saturated); |
| 168 | + break; |
| 169 | + case ov::element::i32: |
| 170 | + cvt_i32_to_i16<TReg>(src, dst, is_saturated); |
| 171 | + cvt_i16_to_byte<TReg>(dst, dst, output_type.is_signed(), is_saturated); |
| 172 | + break; |
| 173 | + case ov::element::f16: |
| 174 | + cvt_f16_to_f32<TReg>(src, dst); |
| 175 | + cvt_f32_to_i32<TReg>(dst, dst); |
| 176 | + cvt_i32_to_i16<TReg>(dst, dst, is_saturated); |
| 177 | + cvt_i16_to_byte<TReg>(dst, dst, output_type.is_signed(), is_saturated); |
| 178 | + break; |
| 179 | + case ov::element::i8: |
| 180 | + case ov::element::u8: |
| 181 | + cvt_byte_to_i16<TReg>(src, dst, input_type.is_signed()); |
| 182 | + cvt_i16_to_byte<TReg>(dst, dst, output_type.is_signed(), is_saturated); |
| 183 | + break; |
| 184 | + default: |
| 185 | + OV_CPU_JIT_EMITTER_THROW("Unsupported input type: ", input_type.get_type_name()); |
| 186 | + } |
| 187 | + break; |
| 188 | + default: |
| 189 | + OV_CPU_JIT_EMITTER_THROW("Unsupported output type: ", output_type.get_type_name()); |
| 190 | + } |
| 191 | +} |
| 192 | + |
| 193 | +jit_convert_emitter::jit_convert_emitter(jit_generator *host, cpu_isa_t host_isa, const std::shared_ptr<ov::Node>& node, ov::element::Type exec_prc) |
| 194 | +: jit_convert_emitter(host, host_isa, node->get_input_element_type(0), node->get_output_element_type(0), exec_prc) { |
| 195 | +} |
| 196 | + |
| 197 | +jit_convert_emitter::jit_convert_emitter(jit_generator *host, cpu_isa_t host_isa, |
| 198 | + ov::element::Type input_prc, |
| 199 | + ov::element::Type output_prc, |
| 200 | + ov::element::Type exec_prc) |
| 201 | +: jit_emitter(host, host_isa, exec_prc) { |
| 202 | + input_type = input_prc; |
| 203 | + output_type = output_prc; |
| 204 | +} |
| 205 | + |
| 206 | +void jit_convert_emitter::validate_types() const { |
| 207 | + OV_CPU_JIT_EMITTER_ASSERT(one_of(input_type, ov::element::f32, ov::element::i32, ov::element::f16, ov::element::i8, ov::element::u8), |
| 208 | + "Unsupported input type: ", input_type.get_type_name()); |
| 209 | + OV_CPU_JIT_EMITTER_ASSERT(one_of(output_type, ov::element::f32, ov::element::i32, ov::element::f16, ov::element::i8, ov::element::u8), |
| 210 | + "Unsupported output type: ", output_type.get_type_name()); |
| 211 | +} |
| 212 | + |
| 213 | +size_t jit_convert_emitter::get_inputs_count() const { return 1; } |
| 214 | + |
| 215 | +void jit_convert_emitter::emit_data() const { |
| 216 | + jit_emitter::emit_data(); |
| 217 | +} |
| 218 | + |
| 219 | +jit_convert_truncation_emitter::jit_convert_truncation_emitter(jit_generator *host, cpu_isa_t host_isa, |
| 220 | + const std::shared_ptr<ov::Node>& node, ov::element::Type exec_prc) |
| 221 | + : jit_convert_emitter(host, host_isa, node, exec_prc) { |
| 222 | +} |
| 223 | + |
| 224 | +jit_convert_truncation_emitter::jit_convert_truncation_emitter(jit_generator *host, cpu_isa_t host_isa, |
| 225 | + ov::element::Type input_prc, |
| 226 | + ov::element::Type output_prc, |
| 227 | + ov::element::Type exec_prc) |
| 228 | + : jit_convert_emitter(host, host_isa, input_prc, output_prc, exec_prc) { |
| 229 | +} |
| 230 | + |
| 231 | +void jit_convert_truncation_emitter::emit_impl(const std::vector<size_t> &in_idxs, const std::vector<size_t> &out_idxs) const { |
| 232 | + validate_types(); |
| 233 | + if (host_isa_ == dnnl::impl::cpu::aarch64::asimd) { |
| 234 | + emit_isa<dnnl::impl::cpu::aarch64::asimd>(in_idxs, out_idxs); |
| 235 | + } else { |
| 236 | + OV_CPU_JIT_EMITTER_THROW("Unsupported ISA ", host_isa_); |
| 237 | + } |
| 238 | +} |
| 239 | + |
| 240 | +template <cpu_isa_t isa> |
| 241 | +void jit_convert_truncation_emitter::emit_isa(const std::vector<size_t> &in_idxs, const std::vector<size_t> &out_idxs) const { |
| 242 | + using TReg = typename dnnl::impl::cpu::aarch64::cpu_isa_traits<isa>::TReg; |
| 243 | + TReg src = TReg(in_idxs[0]); |
| 244 | + TReg dst = TReg(out_idxs[0]); |
| 245 | + jit_convert_process<TReg>(src, dst, input_type, output_type, false); |
| 246 | +} |
| 247 | + |
| 248 | +jit_convert_saturation_emitter::jit_convert_saturation_emitter(jit_generator *host, cpu_isa_t host_isa, |
| 249 | + const std::shared_ptr<ov::Node>& node, ov::element::Type exec_prc) |
| 250 | + : jit_convert_emitter(host, host_isa, node, exec_prc) { |
| 251 | +} |
| 252 | + |
| 253 | +jit_convert_saturation_emitter::jit_convert_saturation_emitter(jit_generator *host, cpu_isa_t host_isa, |
| 254 | + ov::element::Type input_prc, |
| 255 | + ov::element::Type output_prc, |
| 256 | + ov::element::Type exec_prc) |
| 257 | + : jit_convert_emitter(host, host_isa, input_prc, output_prc, exec_prc) { |
| 258 | +} |
| 259 | + |
| 260 | +void jit_convert_saturation_emitter::emit_impl(const std::vector<size_t> &in_idxs, const std::vector<size_t> &out_idxs) const { |
| 261 | + validate_types(); |
| 262 | + if (host_isa_ == dnnl::impl::cpu::aarch64::asimd) { |
| 263 | + emit_isa<dnnl::impl::cpu::aarch64::asimd>(in_idxs, out_idxs); |
| 264 | + } else { |
| 265 | + OV_CPU_JIT_EMITTER_THROW("Unsupported ISA ", host_isa_); |
| 266 | + } |
| 267 | +} |
| 268 | + |
| 269 | +template <cpu_isa_t isa> |
| 270 | +void jit_convert_saturation_emitter::emit_isa(const std::vector<size_t> &in_idxs, const std::vector<size_t> &out_idxs) const { |
| 271 | + using TReg = typename dnnl::impl::cpu::aarch64::cpu_isa_traits<isa>::TReg; |
| 272 | + TReg src = TReg(in_idxs[0]); |
| 273 | + TReg dst = TReg(out_idxs[0]); |
| 274 | + jit_convert_process<TReg>(src, dst, input_type, output_type, true); |
| 275 | +} |
| 276 | + |
| 277 | +} // namespace aarch64 |
| 278 | +} // namespace intel_cpu |
| 279 | +} // namespace ov |
0 commit comments