 |
AI Engine-ML v2 Intrinsics User Guide
v2025.1
|
|
AI Engine-ML v2 Intrinsics User Guide
v2025.1
|
Functions | |
| v16bfloat16 | exp2 (v16accfloat a) |
| Base 2 exponential (fp32) | |
| v16bfloat16 | exp2 (v16float a) |
| v32bfloat16 | exp2 (v32accfloat a) |
| Base 2 exponential (fp32) | |
| v32bfloat16 | exp2 (v32float a) |
| v64bfloat16 | exp2 (v64accfloat a) |
| Base 2 exponential (fp32) | |
| v64bfloat16 | exp2 (v64float a) |
| v16accfloat | exp2_bf20 (v16accfloat a) |
| Base 2 exponential (bf16) | |
| v16accfloat | exp2_bf20 (v16float a) |
| v32accfloat | exp2_bf20 (v32accfloat a) |
| Base 2 exponential (bf16) | |
| v32accfloat | exp2_bf20 (v32float a) |
| v64accfloat | exp2_bf20 (v64accfloat a) |
| Base 2 exponential (bf16) | |
| v64accfloat | exp2_bf20 (v64float a) |
| v32bfloat16 | tanh (v32accfloat a) |
| Hyperbolic tangent. | |
| v32bfloat16 | tanh (v32float a) |
| v16bfloat16 exp2 | ( | v16accfloat | a | ) |
Base 2 exponential (fp32)
The exp2 function biased by 895 at the input side is approximated by this instruction. Logically the functionality can be devided in 2 steps. First, 895 is subtracted from the input. During this step, 11 fractional bits in respect to 895 are considered. The remaining bits in the 32-bit floating-point input are first truncated. Second, an approximation of the base 2 exponential function is applied.
| a | Single precision vector floating-point input value |
| v16bfloat16 exp2 | ( | v16float | a | ) |
| v32bfloat16 exp2 | ( | v32accfloat | a | ) |
Base 2 exponential (fp32)
The exp2 function biased by 895 at the input side is approximated by this instruction. Logically the functionality can be devided in 2 steps. First, 895 is subtracted from the input. During this step, 11 fractional bits in respect to 895 are considered. The remaining bits in the 32-bit floating-point input are first truncated. Second, an approximation of the base 2 exponential function is applied.v
| a | Single precision vector floating-point input value |
| v32bfloat16 exp2 | ( | v32float | a | ) |
| v64bfloat16 exp2 | ( | v64accfloat | a | ) |
Base 2 exponential (fp32)
The exp2 function biased by 895 at the input side is approximated by this instruction. Logically the functionality can be devided in 2 steps. First, 895 is subtracted from the input. During this step, 11 fractional bits in respect to 895 are considered. The remaining bits in the 32-bit floating-point input are first truncated. Second, an approximation of the base 2 exponential function is applied.
| a | Single precision vector floating-point input value |
| v64bfloat16 exp2 | ( | v64float | a | ) |
| v16accfloat exp2_bf20 | ( | v16accfloat | a | ) |
Base 2 exponential (bf16)
The exp2 function biased by 895 at the input side is approximated by this instruction. Logically the functionality can be devided in 2 steps. First, 895 is subtracted from the input. During this step, 11 fractional bits in respect to 895 are considered. The remaining bits in the 32-bit floating-point input are first truncated. Second, an approximation of the base 2 exponential function is applied.
| a | Single precision vector floating-point input value |
| v16accfloat exp2_bf20 | ( | v16float | a | ) |
| v32accfloat exp2_bf20 | ( | v32accfloat | a | ) |
Base 2 exponential (bf16)
The exp2 function biased by 895 at the input side is approximated by this instruction. Logically the functionality can be devided in 2 steps. First, 895 is subtracted from the input. During this step, 11 fractional bits in respect to 895 are considered. The remaining bits in the 32-bit floating-point input are first truncated. Second, an approximation of the base 2 exponential function is applied.
| a | Single precision vector floating-point input value |
| v32accfloat exp2_bf20 | ( | v32float | a | ) |
| v64accfloat exp2_bf20 | ( | v64accfloat | a | ) |
Base 2 exponential (bf16)
The exp2 function biased by 895 at the input side is approximated by this instruction. Logically the functionality can be devided in 2 steps. First, 895 is subtracted from the input. During this step, 11 fractional bits in respect to 895 are considered. The remaining bits in the 32-bit floating-point input are first truncated. Second, an approximation of the base 2 exponential function is applied.
| a | Single precision vector floating-point input value |
| v64accfloat exp2_bf20 | ( | v64float | a | ) |
| v32bfloat16 tanh | ( | v32accfloat | a | ) |
Hyperbolic tangent.
For 16 slices, the 32-bit floating-point input is first truncated to bfloat16 and second an approximation of the hyperbolic tangent function is applied.
| a | Single precision vector floating-point input value |
| v32bfloat16 tanh | ( | v32float | a | ) |