PMADDUBSW — Multiply and Add Packed Signed and Unsigned Bytes

NOTES:

1. See note in Section 2.4, “AVX and SSE Instruction Exception Specification” in the Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 2A and Section 22.25.3, “Exception Conditions of Legacy SIMD Instructions Operating on MMX Registers” in the Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 3A.

Instruction Operand Encoding

Description

(V)PMADDUBSW multiplies vertically each unsigned byte of the destination operand (first operand) with the corre-sponding signed byte of the source operand (second operand), producing intermediate signed 16-bit integers. Each adjacent pair of signed words is added and the saturated result is packed to the destination operand. For example, the lowest-order bytes (bits 7-0) in the source and destination operands are multiplied and the intermediate signed word result is added with the corresponding intermediate result from the 2nd lowest-order bytes (bits 15-8) of the operands; the sign-saturated result is stored in the lowest word of the destination register (15-0). The same oper-ation is performed on the other pairs of adjacent bytes. Both operands can be MMX register or XMM registers. When the source operand is a 128-bit memory operand, the operand must be aligned on a 16-byte boundary or a general-protection exception (#GP) will be generated.

In 64-bit mode and not encoded with VEX/EVEX, use the REX prefix to access XMM8-XMM15.

128-bit Legacy SSE version: The first source and destination operands are XMM registers. The second source operand is an XMM register or a 128-bit memory location. Bits (MAX_VL-1:128) of the corresponding destination register remain unchanged.

VEX.128 and EVEX.128 encoded versions: The first source and destination operands are XMM registers. The second source operand is an XMM register or a 128-bit memory location. Bits (MAX_VL-1:128) of the corre-sponding destination register are zeroed.

VEX.256 and EVEX.256 encoded versions: The second source operand can be an YMM register or a 256-bit memory location. The first source and destination operands are YMM registers. Bits (MAX_VL-1:256) of the corresponding ZMM register are zeroed.

EVEX.512 encoded version: The second source operand can be an ZMM register or a 512-bit memory location. The first source and destination operands are ZMM registers.

Operation

DEST[15-0] = SaturateToSignedWord(SRC[15-8]*DEST[15-8]+SRC[7-0]*DEST[7-0]);
DEST[31-16] = SaturateToSignedWord(SRC[31-24]*DEST[31-24]+SRC[23-16]*DEST[23-16]);
DEST[47-32] = SaturateToSignedWord(SRC[47-40]*DEST[47-40]+SRC[39-32]*DEST[39-32]);
DEST[63-48] = SaturateToSignedWord(SRC[63-56]*DEST[63-56]+SRC[55-48]*DEST[55-48]);

DEST[15-0] = SaturateToSignedWord(SRC[15-8]* DEST[15-8]+SRC[7-0]*DEST[7-0]);
// Repeat operation for 2nd through 7th word
SRC1/DEST[127-112] = SaturateToSignedWord(SRC[127-120]*DEST[127-120]+ SRC[119-112]* DEST[119-112]);

DEST[15:0] (cid:197) SaturateToSignedWord(SRC2[15:8]* SRC1[15:8]+SRC2[7:0]*SRC1[7:0])
// Repeat operation for 2nd through 7th word
DEST[127:112] (cid:197) SaturateToSignedWord(SRC2[127:120]*SRC1[127:120]+ SRC2[119:112]* SRC1[119:112])
DEST[VLMAX-1:128] (cid:197) 0

DEST[15:0] (cid:197) SaturateToSignedWord(SRC2[15:8]* SRC1[15:8]+SRC2[7:0]*SRC1[7:0])
// Repeat operation for 2nd through 15th word
DEST[255:240] (cid:197) SaturateToSignedWord(SRC2[255:248]*SRC1[255:248]+ SRC2[247:240]* SRC1[247:240])
DEST[VLMAX-1:256] (cid:197) 0

(KL, VL) = (8, 128), (16, 256), (32, 512)
FOR j (cid:197) 0 TO KL-1
    i (cid:197) j * 16
    IF k1[j] OR *no writemask*
        THEN DEST[i+15:i] (cid:197) SaturateToSignedWord(SRC2[i+15:i+8]* SRC1[i+15:i+8] + SRC2[i+7:i]*SRC1[i+7:i])
        ELSE
        IF *merging-masking*
            ; merging-masking
            THEN *DEST[i+15:i] remains unchanged*
            ELSE *zeroing-masking*
            ; zeroing-masking
            DEST[i+15:i] = 0
        FI
    FI;
ENDFOR;
DEST[MAX_VL-1:VL] (cid:197) 0

Intel C/C++ Compiler Intrinsic Equivalent

VPMADDUBSW __m512i _mm512_mddubs_epi16( __m512i a, __m512i b);
VPMADDUBSW __m512i _mm512_mask_mddubs_epi16(__m512i s, __mmask32 k, __m512i a, __m512i b);
VPMADDUBSW __m512i _mm512_maskz_mddubs_epi16( __mmask32 k, __m512i a, __m512i b);
VPMADDUBSW __m256i _mm256_mask_mddubs_epi16(__m256i s, __mmask16 k, __m256i a, __m256i b);
VPMADDUBSW __m256i _mm256_maskz_mddubs_epi16( __mmask16 k, __m256i a, __m256i b);
VPMADDUBSW __m128i _mm_mask_mddubs_epi16(__m128i s, __mmask8 k, __m128i a, __m128i b);
VPMADDUBSW __m128i _mm_maskz_maddubs_epi16( __mmask8 k, __m128i a, __m128i b);
PMADDUBSW: __m64 _mm_maddubs_pi16 (__m64 a, __m64 b)
(V)PMADDUBSW: __m128i _mm_maddubs_epi16 (__m128i a, __m128i b)

VPMADDUBSW:
__m256i _mm256_maddubs_epi16 (__m256i a, __m256i b)

SIMD Floating-Point Exceptions

None.

Other Exceptions

Non-EVEX-encoded instruction, see Exceptions Type 4.

EVEX-encoded instruction, see Exceptions Type E4NF.nb.