| 1 | #include "unary-ops.h" |
| 2 | |
| 3 | static inline float op_abs(float x) { |
| 4 | return fabsf(x: x); |
| 5 | } |
| 6 | |
| 7 | static inline float op_sgn(float x) { |
| 8 | return (x > 0.f) ? 1.f : ((x < 0.f) ? -1.f : 0.f); |
| 9 | } |
| 10 | |
| 11 | static inline float op_neg(float x) { |
| 12 | return -x; |
| 13 | } |
| 14 | |
| 15 | static inline float op_step(float x) { |
| 16 | return (x > 0.f) ? 1.f : 0.f; |
| 17 | } |
| 18 | |
| 19 | static inline float op_tanh(float x) { |
| 20 | return tanhf(x: x); |
| 21 | } |
| 22 | |
| 23 | static inline float op_elu(float x) { |
| 24 | return (x > 0.f) ? x : expm1f(x: x); |
| 25 | } |
| 26 | |
| 27 | static inline float op_relu(float x) { |
| 28 | return (x > 0.f) ? x : 0.f; |
| 29 | } |
| 30 | |
| 31 | static inline float op_sigmoid(float x) { |
| 32 | return 1.f / (1.f + expf(x: -x)); |
| 33 | } |
| 34 | |
| 35 | static inline float op_hardsigmoid(float x) { |
| 36 | return fminf(x: 1.0f, y: fmaxf(x: 0.0f, y: (x + 3.0f) / 6.0f)); |
| 37 | } |
| 38 | |
| 39 | static inline float op_exp(float x) { |
| 40 | return expf(x: x); |
| 41 | } |
| 42 | |
| 43 | static inline float op_hardswish(float x) { |
| 44 | return x * fminf(x: 1.0f, y: fmaxf(x: 0.0f, y: (x + 3.0f) / 6.0f)); |
| 45 | } |
| 46 | |
| 47 | static inline float op_sqr(float x) { |
| 48 | return x * x; |
| 49 | } |
| 50 | |
| 51 | static inline float op_sqrt(float x) { |
| 52 | return sqrtf(x: x); |
| 53 | } |
| 54 | |
| 55 | static inline float op_xielu(float x, float alpha_n, float alpha_p, float beta, float eps) { |
| 56 | if (x > 0.0f) { |
| 57 | return alpha_p * x * x + beta * x; |
| 58 | } else { |
| 59 | const float min_x_eps = fminf(x: x, y: eps); |
| 60 | return (expm1f(x: min_x_eps) - x) * alpha_n + beta * x; |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | static inline float op_sin(float x) { |
| 65 | return sinf(x: x); |
| 66 | } |
| 67 | |
| 68 | static inline float op_cos(float x) { |
| 69 | return cosf(x: x); |
| 70 | } |
| 71 | |
| 72 | static inline float op_log(float x) { |
| 73 | return logf(x: x); |
| 74 | } |
| 75 | |
| 76 | static inline float op_floor(float x) { |
| 77 | return floorf(x: x); |
| 78 | } |
| 79 | |
| 80 | static inline float op_ceil(float x) { |
| 81 | return ceilf(x: x); |
| 82 | } |
| 83 | |
| 84 | static inline float op_round(float x) { |
| 85 | return roundf(x: x); |
| 86 | } |
| 87 | |
| 88 | static inline float op_trunc(float x) { |
| 89 | return truncf(x: x); |
| 90 | } |
| 91 | |
| 92 | template <float (*op)(float), typename src0_t, typename dst_t> |
| 93 | static inline void vec_unary_op(int64_t n, dst_t * y, const src0_t * x) { |
| 94 | constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32; |
| 95 | constexpr auto f32_to_dst = type_conversion_table<dst_t >::from_f32; |
| 96 | |
| 97 | for (int i = 0; i < n; i++) { |
| 98 | y[i] = f32_to_dst(op(src0_to_f32(x[i]))); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | template <float (*op)(float), typename src0_t, typename dst_t> |
| 103 | static void apply_unary_op(const ggml_compute_params * params, ggml_tensor * dst) { |
| 104 | const ggml_tensor * src0 = dst->src[0]; |
| 105 | |
| 106 | GGML_ASSERT(ggml_is_contiguous_1(src0) && ggml_is_contiguous_1(dst) && ggml_are_same_shape(src0, dst)); |
| 107 | |
| 108 | GGML_TENSOR_UNARY_OP_LOCALS |
| 109 | |
| 110 | GGML_ASSERT( nb0 == sizeof(dst_t)); |
| 111 | GGML_ASSERT(nb00 == sizeof(src0_t)); |
| 112 | |
| 113 | const auto [ir0, ir1] = get_thread_range(params, src0); |
| 114 | |
| 115 | for (int64_t ir = ir0; ir < ir1; ++ir) { |
| 116 | const int64_t i03 = ir/(ne02*ne01); |
| 117 | const int64_t i02 = (ir - i03*ne02*ne01)/ne01; |
| 118 | const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); |
| 119 | |
| 120 | dst_t * dst_ptr = (dst_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); |
| 121 | const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); |
| 122 | |
| 123 | vec_unary_op<op>(ne0, dst_ptr, src0_ptr); |
| 124 | } |
| 125 | } |
| 126 | |
| 127 | // TODO: Use the 'traits' lookup table (for type conversion fns), instead of a mass of 'if' conditions with long templates |
| 128 | template <float (*op)(float)> |
| 129 | static void unary_op(const ggml_compute_params * params, ggml_tensor * dst) { |
| 130 | const ggml_tensor * src0 = dst->src[0]; |
| 131 | |
| 132 | /* */ if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32 |
| 133 | apply_unary_op<op, float, float>(params, dst); |
| 134 | } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16 |
| 135 | apply_unary_op<op, ggml_fp16_t, ggml_fp16_t>(params, dst); |
| 136 | } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16 |
| 137 | apply_unary_op<op, ggml_bf16_t, ggml_bf16_t>(params, dst); |
| 138 | } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) { |
| 139 | apply_unary_op<op, ggml_bf16_t, float>(params, dst); |
| 140 | } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { |
| 141 | apply_unary_op<op, ggml_fp16_t, float>(params, dst); |
| 142 | } else { |
| 143 | fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n" , __func__, |
| 144 | ggml_type_name(type: dst->type), ggml_type_name(type: src0->type)); |
| 145 | GGML_ABORT("fatal error" ); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | template <float (*op)(float, ggml_tensor *)> |
| 150 | static void unary_op_params(const ggml_compute_params * params, ggml_tensor * dst) { |
| 151 | const ggml_tensor * src0 = dst->src[0]; |
| 152 | |
| 153 | /* */ if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32 |
| 154 | apply_unary_op<op, float, float>(params, dst); |
| 155 | } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16 |
| 156 | apply_unary_op<op, ggml_fp16_t, ggml_fp16_t>(params, dst); |
| 157 | } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16 |
| 158 | apply_unary_op<op, ggml_bf16_t, ggml_bf16_t>(params, dst); |
| 159 | } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) { |
| 160 | apply_unary_op<op, ggml_bf16_t, float>(params, dst); |
| 161 | } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { |
| 162 | apply_unary_op<op, ggml_fp16_t, float>(params, dst); |
| 163 | } else { |
| 164 | fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n" , __func__, |
| 165 | ggml_type_name(type: dst->type), ggml_type_name(type: src0->type)); |
| 166 | GGML_ABORT("fatal error" ); |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | // Extend vec_unary_op to support functors |
| 171 | template <typename Op, typename src0_t, typename dst_t> |
| 172 | static inline void vec_unary_op_functor(int64_t n, dst_t * y, const src0_t * x, Op op) { |
| 173 | constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32; |
| 174 | constexpr auto f32_to_dst = type_conversion_table<dst_t >::from_f32; |
| 175 | |
| 176 | for (int i = 0; i < n; i++) { |
| 177 | y[i] = f32_to_dst(op(src0_to_f32(x[i]))); |
| 178 | } |
| 179 | } |
| 180 | |
| 181 | // Extend apply_unary_op to support functors |
| 182 | template <typename Op, typename src0_t, typename dst_t> |
| 183 | static void apply_unary_op_functor(const ggml_compute_params * params, ggml_tensor * dst, Op op) { |
| 184 | const ggml_tensor * src0 = dst->src[0]; |
| 185 | |
| 186 | GGML_ASSERT(ggml_is_contiguous_1(src0) && ggml_is_contiguous_1(dst) && ggml_are_same_shape(src0, dst)); |
| 187 | |
| 188 | GGML_TENSOR_UNARY_OP_LOCALS |
| 189 | |
| 190 | GGML_ASSERT( nb0 == sizeof(dst_t)); |
| 191 | GGML_ASSERT(nb00 == sizeof(src0_t)); |
| 192 | |
| 193 | const auto [ir0, ir1] = get_thread_range(params, src0); |
| 194 | |
| 195 | for (int64_t ir = ir0; ir < ir1; ++ir) { |
| 196 | const int64_t i03 = ir/(ne02*ne01); |
| 197 | const int64_t i02 = (ir - i03*ne02*ne01)/ne01; |
| 198 | const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); |
| 199 | |
| 200 | dst_t * dst_ptr = (dst_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); |
| 201 | const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); |
| 202 | |
| 203 | vec_unary_op_functor(ne0, dst_ptr, src0_ptr, op); |
| 204 | } |
| 205 | } |
| 206 | |
| 207 | // Generic dispatcher for functors |
| 208 | template <typename Op> |
| 209 | static void unary_op_functor(const ggml_compute_params * params, ggml_tensor * dst, Op op) { |
| 210 | const ggml_tensor * src0 = dst->src[0]; |
| 211 | |
| 212 | /* */ if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32 |
| 213 | apply_unary_op_functor<Op, float, float>(params, dst, op); |
| 214 | } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16 |
| 215 | apply_unary_op_functor<Op, ggml_fp16_t, ggml_fp16_t>(params, dst, op); |
| 216 | } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16 |
| 217 | apply_unary_op_functor<Op, ggml_bf16_t, ggml_bf16_t>(params, dst, op); |
| 218 | } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) { |
| 219 | apply_unary_op_functor<Op, ggml_bf16_t, float>(params, dst, op); |
| 220 | } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { |
| 221 | apply_unary_op_functor<Op, ggml_fp16_t, float>(params, dst, op); |
| 222 | } else { |
| 223 | fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n" , __func__, |
| 224 | ggml_type_name(type: dst->type), ggml_type_name(type: src0->type)); |
| 225 | GGML_ABORT("fatal error" ); |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | void ggml_compute_forward_abs(const ggml_compute_params * params, ggml_tensor * dst) { |
| 230 | unary_op<op_abs>(params, dst); |
| 231 | } |
| 232 | |
| 233 | void ggml_compute_forward_sgn(const ggml_compute_params * params, ggml_tensor * dst) { |
| 234 | unary_op<op_sgn>(params, dst); |
| 235 | } |
| 236 | |
| 237 | void ggml_compute_forward_neg(const ggml_compute_params * params, ggml_tensor * dst) { |
| 238 | unary_op<op_neg>(params, dst); |
| 239 | } |
| 240 | |
| 241 | void ggml_compute_forward_step(const ggml_compute_params * params, ggml_tensor * dst) { |
| 242 | unary_op<op_step>(params, dst); |
| 243 | } |
| 244 | |
| 245 | void ggml_compute_forward_tanh(const ggml_compute_params * params, ggml_tensor * dst) { |
| 246 | unary_op<op_tanh>(params, dst); |
| 247 | } |
| 248 | |
| 249 | void ggml_compute_forward_elu(const ggml_compute_params * params, ggml_tensor * dst) { |
| 250 | unary_op<op_elu>(params, dst); |
| 251 | } |
| 252 | |
| 253 | void ggml_compute_forward_relu(const ggml_compute_params * params, ggml_tensor * dst) { |
| 254 | unary_op<op_relu>(params, dst); |
| 255 | } |
| 256 | |
| 257 | void ggml_compute_forward_sigmoid(const ggml_compute_params * params, ggml_tensor * dst) { |
| 258 | unary_op<op_sigmoid>(params, dst); |
| 259 | } |
| 260 | |
| 261 | void ggml_compute_forward_hardsigmoid(const ggml_compute_params * params, ggml_tensor * dst) { |
| 262 | unary_op<op_hardsigmoid>(params, dst); |
| 263 | } |
| 264 | |
| 265 | void ggml_compute_forward_exp(const ggml_compute_params * params, ggml_tensor * dst) { |
| 266 | unary_op<op_exp>(params, dst); |
| 267 | } |
| 268 | |
| 269 | void ggml_compute_forward_hardswish(const ggml_compute_params * params, ggml_tensor * dst) { |
| 270 | unary_op<op_hardswish>(params, dst); |
| 271 | } |
| 272 | |
| 273 | void ggml_compute_forward_sqr(const ggml_compute_params * params, ggml_tensor * dst) { |
| 274 | unary_op<op_sqr>(params, dst); |
| 275 | } |
| 276 | |
| 277 | void ggml_compute_forward_sqrt(const ggml_compute_params * params, ggml_tensor * dst) { |
| 278 | unary_op<op_sqrt>(params, dst); |
| 279 | } |
| 280 | |
| 281 | void ggml_compute_forward_sin(const ggml_compute_params * params, ggml_tensor * dst) { |
| 282 | unary_op<op_sin>(params, dst); |
| 283 | } |
| 284 | |
| 285 | void ggml_compute_forward_cos(const ggml_compute_params * params, ggml_tensor * dst) { |
| 286 | unary_op<op_cos>(params, dst); |
| 287 | } |
| 288 | |
| 289 | void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor * dst) { |
| 290 | unary_op<op_log>(params, dst); |
| 291 | } |
| 292 | |
| 293 | void ggml_compute_forward_floor(const ggml_compute_params * params, ggml_tensor * dst) { |
| 294 | unary_op<op_floor>(params, dst); |
| 295 | } |
| 296 | |
| 297 | void ggml_compute_forward_ceil(const ggml_compute_params * params, ggml_tensor * dst) { |
| 298 | unary_op<op_ceil>(params, dst); |
| 299 | } |
| 300 | |
| 301 | void ggml_compute_forward_round(const ggml_compute_params * params, ggml_tensor * dst) { |
| 302 | unary_op<op_round>(params, dst); |
| 303 | } |
| 304 | |
| 305 | void ggml_compute_forward_trunc(const ggml_compute_params * params, ggml_tensor * dst) { |
| 306 | unary_op<op_trunc>(params, dst); |
| 307 | } |
| 308 | |
| 309 | void ggml_compute_forward_xielu(const ggml_compute_params * params, ggml_tensor * dst) { |
| 310 | const float alpha_n = ggml_get_op_params_f32(tensor: dst, i: 1); |
| 311 | const float alpha_p = ggml_get_op_params_f32(tensor: dst, i: 2); |
| 312 | const float beta = ggml_get_op_params_f32(tensor: dst, i: 3); |
| 313 | const float eps = ggml_get_op_params_f32(tensor: dst, i: 4); |
| 314 | |
| 315 | const auto xielu_op_params = [alpha_n, alpha_p, beta, eps](float f) { |
| 316 | return op_xielu(x: f, alpha_n, alpha_p, beta, eps); |
| 317 | }; |
| 318 | |
| 319 | unary_op_functor(params, dst, op: xielu_op_params); |
| 320 | } |
| 321 | |
| 322 | |