Vesta
English
Source code

VFCMULCSH, VFMULCSH

Complex Multiply Scalar FP16 Values

stableVMJITAOTinstruction

Encodings

OpcodeInstructionOp/En64-bitCompat/LegacyDescription
EVEX.LLIG.F2.MAP6.W0 D7 /rVFCMULCSH xmm1{k1}{z}, xmm2, xmm3/m32 {er}AValidValidComplex multiply a pair of FP16 values from OR AVX10.1 xmm2 and xmm3/m32, and store the result in xmm1 subject to writemask k1. Bits 127:32 of xmm2 are copied to xmm1[127:32].
EVEX.LLIG.F3.MAP6.W0 D7 /rVFMULCSH xmm1{k1}{z}, xmm2, xmm3/m32 {er}AValidValidComplex multiply a pair of FP16 values from OR AVX10.1 xmm2 and the complex conjugate of xmm3/m32, and store the result in xmm1 subject to writemask k1. Bits 127:32 of xmm2 are copied to xmm1[127:32].

Operand encoding

Each mode is a value of the Op/En column above. It says which field of the encoded instruction carries each operand, in the order they are written, and whether the instruction reads it, writes it or both.

A

  1. modrm.reg escrituraModRM byte, reg field (bits 5-3)
  2. vex.vvvv lecturaVEX prefix, vvvv field (inverted)
  3. modrm.rm lecturaModRM byte, r/m field (bits 2-0); with the SIB byte and the displacement when the mod field asks for them

Tupla: Scalar

Measured cost

Loading measurements from arch-data...

Description

This instruction performs a complex multiply operation. There are normal and complex conjugate forms of the operation. The masking for this operation is done on 32-bit quantities representing a pair of FP16 values.

Bits 127:32 of the destination operand are copied from the corresponding bits of the first source operand. Bits MAXVL-1:128 of the destination operand are zeroed. The low FP16 element of the destination is updated according to the writemask.

Rounding is performed at every FMA (fused multiply and add) boundary. Execution occurs as if all MXCSR exceptions are masked. MXCSR status bits are updated to reflect exceptional conditions.

Operation

VFCMULCSH dest{k1}, src1, src2 (AVX512)
KL := VL / 32

IF k1[0] or *no writemask*:
    tmp.fp16[0] := src1.fp16[0] * src2.fp16[0]
    tmp.fp16[1] := src1.fp16[1] * src2.fp16[0]

    // conjugate version subtracts odd final term
    dest.fp16[0] := tmp.fp16[0] + src1.fp16[1] * src2.fp16[1]
    dest.fp16[1] := tmp.fp16[1] - src1.fp16[0] * src2.fp16[1]
ELSE IF *zeroing*:
    dest.fp16[0] := 0
    dest.fp16[1] := 0

DEST[127:32] := src1[127:32] // copy upper part of src1
DEST[MAXVL-1:128] := 0


VFMULCSH dest{k1}, src1, src2 (AVX512)
KL := VL / 32

IF k1[0] or *no writemask*:
    // non-conjugate version subtracts last even term
    tmp.fp16[0] := src1.fp16[0] * src2.fp16[0]
    tmp.fp16[1] := src1.fp16[1] * src2.fp16[0]
    dest.fp16[0] := tmp.fp16[0] - src1.fp16[1] * src2.fp16[1]
    dest.fp16[1] := tmp.fp16[1] + src1.fp16[0] * src2.fp16[1]

ELSE IF *zeroing*:
    dest.fp16[0] := 0
    dest.fp16[1] := 0

DEST[127:32] := src1[127:32] // copy upper part of src1
DEST[MAXVL-1:128] := 0

Intel C/C++ compiler intrinsics

VFCMULCSH __m128h _mm_cmul_round_sch (__m128h a, __m128h b, const int rounding);
VFCMULCSH __m128h _mm_mask_cmul_round_sch (__m128h src, __mmask8 k, __m128h a, __m128h b, const int rounding);
VFCMULCSH __m128h _mm_maskz_cmul_round_sch (__mmask8 k, __m128h a, __m128h b, const int rounding);
VFCMULCSH __m128h _mm_cmul_sch (__m128h a, __m128h b);
VFCMULCSH __m128h _mm_mask_cmul_sch (__m128h src, __mmask8 k, __m128h a, __m128h b);
VFCMULCSH __m128h _mm_maskz_cmul_sch (__mmask8 k, __m128h a, __m128h b);
VFCMULCSH __m128h _mm_fcmul_round_sch (__m128h a, __m128h b, const int rounding);
VFCMULCSH __m128h _mm_mask_fcmul_round_sch (__m128h src, __mmask8 k, __m128h a, __m128h b, const int rounding);
VFCMULCSH __m128h _mm_maskz_fcmul_round_sch (__mmask8 k, __m128h a, __m128h b, const int rounding);
VFCMULCSH __m128h _mm_fcmul_sch (__m128h a, __m128h b);
VFCMULCSH __m128h _mm_mask_fcmul_sch (__m128h src, __mmask8 k, __m128h a, __m128h b);
VFCMULCSH __m128h _mm_maskz_fcmul_sch (__mmask8 k, __m128h a, __m128h b);
VFMULCSH __m128h _mm_fmul_round_sch (__m128h a, __m128h b, const int rounding);
VFMULCSH __m128h _mm_mask_fmul_round_sch (__m128h src, __mmask8 k, __m128h a, __m128h b, const int rounding);
VFMULCSH __m128h _mm_maskz_fmul_round_sch (__mmask8 k, __m128h a, __m128h b, const int rounding);
VFMULCSH __m128h _mm_fmul_sch (__m128h a, __m128h b);
VFMULCSH __m128h _mm_mask_fmul_sch (__m128h src, __mmask8 k, __m128h a, __m128h b);
VFMULCSH __m128h _mm_maskz_fmul_sch (__mmask8 k, __m128h a, __m128h b);
VFMULCSH __m128h _mm_mask_mul_round_sch (__m128h src, __mmask8 k, __m128h a, __m128h b, const int rounding);
VFMULCSH __m128h _mm_maskz_mul_round_sch (__mmask8 k, __m128h a, __m128h b, const int rounding);
VFMULCSH __m128h _mm_mul_round_sch (__m128h a, __m128h b, const int rounding);
VFMULCSH __m128h _mm_mask_mul_sch (__m128h src, __mmask8 k, __m128h a, __m128h b);
VFMULCSH __m128h _mm_maskz_mul_sch (__mmask8 k, __m128h a, __m128h b);
VFMULCSH __m128h _mm_mul_sch (__m128h a, __m128h b);

SIMD Floating-Point Exceptions

Invalid, Underflow, Overflow, Precision, Denormal.

Other Exceptions

EVEX-encoded instructions, see Table 2-60, "Type E10 Class Exception Conditions."

Additionally:

#UD               If (dest_reg == src1_reg) or (dest_reg == src2_reg).

Sources