| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // | ||
| 2 | // SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 3 | // | ||
| 4 | // SPDX-License-Identifier: Apache-2.0 | ||
| 5 | // | ||
| 6 | #pragma once | ||
| 7 | |||
| 8 | #if defined(__ARM_NEON) | ||
| 9 | #include <arm_neon.h> | ||
| 10 | #endif // defined(__ARM_NEON) | ||
| 11 | |||
| 12 | #include <stddef.h> | ||
| 13 | #include <stdint.h> | ||
| 14 | #include <stdio.h> | ||
| 15 | #include <stdlib.h> | ||
| 16 | #include <string.h> | ||
| 17 | |||
| 18 | #ifdef __cplusplus | ||
| 19 | extern "C" { | ||
| 20 | #endif | ||
| 21 | |||
| 22 | // NOLINTBEGIN(cppcoreguidelines-avoid-do-while,cppcoreguidelines-pro-type-vararg,cert-err33-c) | ||
| 23 | // | ||
| 24 | // * cppcoreguidelines-avoid-do-while: do-while is necessary for macros. | ||
| 25 | // * cppcoreguidelines-pro-type-vararg: use of variadic arguments in fprintf is expected. | ||
| 26 | // * cert-err33-c: checking the output of fflush and fprintf is not necessary for error reporting. | ||
| 27 | #define KAI_ERROR(msg) \ | ||
| 28 | do { \ | ||
| 29 | fflush(stdout); \ | ||
| 30 | fprintf(stderr, "%s:%d %s", __FILE__, __LINE__, msg); \ | ||
| 31 | abort(); \ | ||
| 32 | } while (0) | ||
| 33 | |||
| 34 | #define KAI_ASSERT_MSG(cond, msg) \ | ||
| 35 | do { \ | ||
| 36 | if (!(cond)) { \ | ||
| 37 | KAI_ERROR(msg); \ | ||
| 38 | } \ | ||
| 39 | } while (0) | ||
| 40 | |||
| 41 | // NOLINTEND(cppcoreguidelines-avoid-do-while,cppcoreguidelines-pro-type-vararg,cert-err33-c) | ||
| 42 | |||
| 43 | #define KAI_ASSERT(cond) KAI_ASSERT_MSG(cond, #cond) | ||
| 44 | |||
| 45 | #define KAI_ASSERT_IF_MSG(precond, cond, msg) KAI_ASSERT_MSG(!(precond) || (cond), msg) | ||
| 46 | #define KAI_ASSERT_IF(precond, cond) KAI_ASSERT_IF_MSG(precond, cond, #precond " |-> " #cond) | ||
| 47 | |||
| 48 | #define KAI_ASSUME_MSG KAI_ASSERT_MSG | ||
| 49 | #define KAI_ASSUME KAI_ASSERT | ||
| 50 | #define KAI_ASSUME_IF_MSG KAI_ASSERT_IF_MSG | ||
| 51 | #define KAI_ASSUME_IF KAI_ASSERT_IF | ||
| 52 | |||
| 53 | #define KAI_UNUSED(x) (void)(x) | ||
| 54 | #define KAI_MIN(a, b) (((a) < (b)) ? (a) : (b)) | ||
| 55 | #define KAI_MAX(a, b) (((a) > (b)) ? (a) : (b)) | ||
| 56 | |||
| 57 | /// Largest supported SME vector length in bytes | ||
| 58 | #define KAI_SME_VEC_LENGTH_MAX_BYTES 256 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum) | ||
| 59 | |||
| 60 | /// Gets the version of the project in the Major.Minor.Patch semantic versioning format. | ||
| 61 | /// | ||
| 62 | /// @return Project version as a string literal. | ||
| 63 | 1 | inline const char* kai_get_version(void) { | |
| 64 | 1 | return "1.15.1"; | |
| 65 | } | ||
| 66 | |||
| 67 | /// KleidiAI data types | ||
| 68 | /// Format: <byte 3>(reserved)|<byte 2>(num-bytes)|<byte 1>(type)|<byte 0>(variant-type) | ||
| 69 | enum kai_datatype { | ||
| 70 | kai_dt_unknown = 0x0000, | ||
| 71 | kai_dt_f32 = 0x0411, | ||
| 72 | kai_dt_f16 = 0x0212, | ||
| 73 | kai_dt_bf16 = 0x0213, | ||
| 74 | kai_dt_int32 = 0x0421, | ||
| 75 | kai_dt_int16 = 0x0222, | ||
| 76 | kai_dt_int8 = 0x0124, | ||
| 77 | kai_dt_uint32 = 0x0431, | ||
| 78 | kai_dt_uint16 = 0x0232, | ||
| 79 | kai_dt_uint8 = 0x0134, | ||
| 80 | kai_dt_bool = 0x0441 | ||
| 81 | }; | ||
| 82 | |||
| 83 | /// Gets number of bytes for a given data type | ||
| 84 | /// @param[in] dt KleidiAI data type | ||
| 85 | /// | ||
| 86 | /// @return the numbers of bytes for the data type | ||
| 87 | 9126 | inline static size_t kai_get_datatype_size_in_bytes(enum kai_datatype dt) { | |
| 88 | 9126 | return (size_t)(dt >> 8); | |
| 89 | } | ||
| 90 | |||
| 91 | /// Converts a scalar f16 value to f32 | ||
| 92 | /// @param[in] f16 The f16 value | ||
| 93 | /// | ||
| 94 | /// @return the f32 value | ||
| 95 | #if defined(__ARM_NEON) | ||
| 96 | inline static float kai_cast_f32_f16(uint16_t f16) { | ||
| 97 | float16_t f32 = 0; | ||
| 98 | memcpy(&f32, &f16, sizeof(uint16_t)); | ||
| 99 | return (float)f32; | ||
| 100 | } | ||
| 101 | #endif | ||
| 102 | |||
| 103 | /// Converts a scalar bf16 value to f32 | ||
| 104 | /// @param[in] bf16 The f16 value | ||
| 105 | /// | ||
| 106 | /// @return the f32 value | ||
| 107 | 1685863882 | inline static float kai_cast_f32_bf16(uint16_t bf16) { | |
| 108 | 1685863882 | const uint32_t i32 = (bf16 << 16); | |
| 109 | 1685863882 | float f32 = 0; | |
| 110 | 1685863882 | memcpy(&f32, &i32, sizeof(i32)); | |
| 111 | 3371727764 | return f32; | |
| 112 | 1685863882 | } | |
| 113 | |||
| 114 | /// Converts a f32 value to bf16 | ||
| 115 | /// @param[in] f32 The f32 value | ||
| 116 | /// | ||
| 117 | /// @return the bf16 value | ||
| 118 | 128 | inline static uint16_t kai_cast_bf16_f32(float f32) { | |
| 119 | 128 | uint16_t bf16 = 0; | |
| 120 | #ifdef __ARM_FEATURE_BF16 | ||
| 121 | 128 | __asm__ __volatile__("bfcvt %h[output], %s[input]" : [output] "=w"(bf16) : [input] "w"(f32)); | |
| 122 | #else | ||
| 123 | const uint32_t* i32 = (uint32_t*)(&f32); | ||
| 124 | bf16 = (*i32 >> 16); | ||
| 125 | #endif | ||
| 126 | 256 | return bf16; | |
| 127 | 128 | } | |
| 128 | |||
| 129 | /// Converts a scalar f32 value to f16 | ||
| 130 | /// @param[in] f32 The f32 value | ||
| 131 | /// | ||
| 132 | /// @return the f16 value | ||
| 133 | #if defined(__ARM_NEON) | ||
| 134 | 4540 | inline static uint16_t kai_cast_f16_f32(float f32) { | |
| 135 | 4540 | uint16_t f16 = 0; | |
| 136 | 4540 | float16_t tmp = (float16_t)f32; | |
| 137 | 4540 | memcpy(&f16, &tmp, sizeof(uint16_t)); | |
| 138 | 9080 | return f16; | |
| 139 | 4540 | } | |
| 140 | #endif | ||
| 141 | |||
| 142 | 10736077 | inline static size_t kai_roundup(size_t a, size_t b) { | |
| 143 | 10736077 | return ((a + b - 1) / b) * b; | |
| 144 | } | ||
| 145 | |||
| 146 | #if defined(__ARM_FEATURE_SVE2) || defined(_M_ARM64) | ||
| 147 | /// Gets the SME vector length for 8-bit elements. | ||
| 148 | uint64_t kai_get_sme_vector_length_u8(void); | ||
| 149 | |||
| 150 | /// Gets the SME vector length for 16-bit elements. | ||
| 151 | 230550 | inline static uint64_t kai_get_sme_vector_length_u16(void) { | |
| 152 | 230550 | return kai_get_sme_vector_length_u8() / 2; | |
| 153 | } | ||
| 154 | |||
| 155 | /// Gets the SME vector length for 32-bit elements. | ||
| 156 | 273336 | inline static uint64_t kai_get_sme_vector_length_u32(void) { | |
| 157 | 273336 | return kai_get_sme_vector_length_u8() / 4; | |
| 158 | } | ||
| 159 | #endif // defined(__ARM_FEATURE_SVE2) || defined(_M_ARM64) | ||
| 160 | |||
| 161 | /// Extends the sign bit of int 4-bit value (stored in int8_t variable) | ||
| 162 | /// @param[in] value The 4-bit int value | ||
| 163 | /// | ||
| 164 | /// @return the int8_t value with sign extended | ||
| 165 | 5788242 | inline static int8_t kai_ext_sign_i8_i4(int8_t value) { | |
| 166 | // Make sure value holds correct int4 value | ||
| 167 | − | KAI_ASSERT(value <= 0xF); | |
| 168 | |||
| 169 | 5788242 | return (value ^ 0x8) - 8; // NOLINT(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) | |
| 170 | } | ||
| 171 | |||
| 172 | /// Parameter struct for RHS matrix packing (Quantized Symmetric Integer 8-bit with per-channel quantization) | ||
| 173 | struct kai_rhs_pack_qsi8cx_params { | ||
| 174 | int32_t lhs_zero_point; ///< LHS Matrix quantization zero-point | ||
| 175 | float scale_multiplier; ///< Product of input (refers to lhs and rhs) and output quantization scales. | ||
| 176 | }; | ||
| 177 | |||
| 178 | /// Parameter struct for RHS matrix packing (Quantized Symmetric Integer 4-bit with per-block quantizatio and s1s0 | ||
| 179 | /// nibble ordering) | ||
| 180 | struct kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0_params { | ||
| 181 | int8_t lhs_zero_point; | ||
| 182 | uint8_t rhs_zero_point; | ||
| 183 | enum kai_datatype scale_dt; | ||
| 184 | }; | ||
| 185 | |||
| 186 | /// Parameter struct for RHS matrix packing | ||
| 187 | struct kai_rhs_pack_qs4cxs1s0_param { | ||
| 188 | int8_t lhs_zero_point; ///< LHS Matrix quantization zero-point | ||
| 189 | uint8_t rhs_zero_point; ///< RHS Matrix quantization zero-point | ||
| 190 | }; | ||
| 191 | |||
| 192 | /// Requantization and clamp parameters for GEMM/GEMV output stage. | ||
| 193 | struct kai_matmul_requantize32_params { | ||
| 194 | int32_t min_value; ///< Minimum output value. | ||
| 195 | int32_t max_value; ///< Maximum output value. | ||
| 196 | int32_t output_zero_point; ///< Output quantization zero point. | ||
| 197 | }; | ||
| 198 | |||
| 199 | #ifdef __cplusplus | ||
| 200 | } | ||
| 201 | #endif | ||
| 202 |