_mm_maskz_reduce_sh
Classification
AVX-512, Special Math Functions, CPUID Test: AVX512_FP16
Header File
Instruction
VREDUCESH xmm {z}, xmm, xmm, imm8
Synopsis
_mm_maskz_reduce_sh(__mmask8 k, __m128h a, __m128h b, int imm8);
Description
Extract the reduced argument of the lower half-precision (16-bit) floating-point element in "b" by the number of bits specified by "imm8", store the result in the lower element of "dst" using zeromask "k" (the element is zeroed out when mask bit 0 is not set), and copy the upper 7 packed elements from "a" to the upper elements of "dst". [round_imm_note]
Operation
DEFINE ReduceArgumentFP16(src[15:0], imm8[7:0]) {
m[15:0] := FP16(imm8[7:4]) // number of fraction bits after the binary point to be preserved
tmp[15:0] := POW(2.0, FP16(-m)) * ROUND(POW(2.0, FP16(m)) * src[15:0], imm8[3:0])
tmp[15:0] := src[15:0] - tmp[15:0]
IF IsInf(tmp[15:0])
tmp[15:0] := FP16(0.0)
FI
RETURN tmp[15:0]
}
IF k[0]
dst.fp16[0] := ReduceArgumentFP16(b.fp16[0], imm8)
ELSE
dst.fp16[0] := 0
FI
dst[127:16] := a[127:16]
dst[MAX:128] := 0