psadbw

Compute Sum of Absolute Differences

PSADBW xmm1, xmm2/m128

Computes absolute differences of bytes and sums them to words.

Details

Computes the absolute value of the difference for each of 16 packed unsigned bytes in the destination and source, then sums those differences within 64-bit groups and stores the results as two 64-bit values. Useful for motion estimation and SAD (Sum of Absolute Differences) operations. No flags are affected.

Pseudocode Operation

sum_low ← 0
sum_high ← 0
for i in 0..7:
  sum_low ← sum_low + abs(dest[i*8 + 7:i*8] - src[i*8 + 7:i*8])
for i in 8..15:
  sum_high ← sum_high + abs(dest[i*8 + 7:i*8] - src[i*8 + 7:i*8])
dest[0:63] ← sum_low
dest[64:127] ← sum_high

Example

PSADBW xmm1, xmm2/m128

Encoding

Binary Layout
66
+0
0F
+1
F6
+2
 
Format SSE2
Opcode 66 0F F6
Extension SSE2

Operands

  • dest
    XMM
  • src
    XMM/Mem

Reference (Intel® SDM)

Instruction Forms

Opcode Instruction Op/En 64/32-bit Mode CPUID Description
NP 0F F6 /r1 PSADBW mm1, mm2/m64 A V/V SSE Computes the absolute differences of the packed unsigned byte integers from mm2 /m64 and mm1; differences are then summed to produce an unsigned word integer result.
66 0F F6 /r PSADBW xmm1, xmm2/m128 A V/V SSE2 Computes the absolute differences of the packed unsigned byte integers from xmm2 /m128 and xmm1; the 8 low differences and 8 high differences are then summed separately to produce two unsigned word integer results.
VEX.128.66.0F.WIG F6 /r VPSADBW xmm1, xmm2, xmm3/m128 B V/V AVX Computes the absolute differences of the packed unsigned byte integers from xmm3 /m128 and xmm2; the 8 low differences and 8 high differences are then summed separately to produce two unsigned word integer results.
VEX.256.66.0F.WIG F6 /r VPSADBW ymm1, ymm2, ymm3/m256 B V/V AVX2 Computes the absolute differences of the packed unsigned byte integers from ymm3 /m256 and ymm2; then each consecutive 8 differences are summed separately to produce four unsigned word integer results.
EVEX.128.66.0F.WIG F6 /r VPSADBW xmm1, xmm2, xmm3/m128 C V/V (AVX512VL AND AVX512BW) OR AVX10.1 Computes the absolute differences of the packed unsigned byte integers from xmm3 /m128 and xmm2; then each consecutive 8 differences are summed separately to produce two unsigned word integer results.
EVEX.256.66.0F.WIG F6 /r VPSADBW ymm1, ymm2, ymm3/m256 C V/V (AVX512VL AND AVX512BW) OR AVX10.1 Computes the absolute differences of the packed unsigned byte integers from ymm3 /m256 and ymm2; then each consecutive 8 differences are summed separately to produce four unsigned word integer results.
EVEX.512.66.0F.WIG F6 /r VPSADBW zmm1, zmm2, zmm3/m512 C V/V AVX512BW OR AVX10.1 Computes the absolute differences of the packed unsigned byte integers from zmm3 /m512 and zmm2; then each consecutive 8 differences are summed separately to produce eight unsigned word integer results.

Instruction Operand Encoding

Op/En Tuple Type Operand 1 Operand 2 Operand 3 Operand 4
A N/A ModRM:reg (r, w) ModRM:r/m (r) N/A N/A
B N/A ModRM:reg (w) VEX.vvvv (r) ModRM:r/m (r) N/A
C Full Mem ModRM:reg (w) EVEX.vvvv ModRM:r/m (r) N/A

Description

Computes the absolute value of the difference of 8 unsigned byte integers from the source operand (second operand) and from the destination operand (first operand). These 8 differences are then summed to produce an unsigned word integer result that is stored in the destination operand. Figure 4-14 shows the operation of the PSADBW instruction when using 64-bit operands. When operating on 64-bit operands, the word integer result is stored in the low word of the destination operand, and the remaining bytes in the destination operand are cleared to all 0s. When operating on 128-bit operands, two packed results are computed. Here, the 8 low-order bytes of the source and destination operands are operated on to produce a word result that is stored in the low word of the destination operand, and the 8 high-order bytes are operated on to produce a word result that is stored in bits 64 through 79 of the destination operand. The remaining bytes of the destination operand are cleared. For 256-bit version, the third group of 8 differences are summed to produce an unsigned word in bits[143:128] of the destination register and the fourth group of 8 differences are summed to produce an unsigned word in bits[207:192] of the destination register. The remaining words of the destination are set to 0. For 512-bit version, the fifth group result is stored in bits [271:256] of the destination. The result from the sixth group is stored in bits [335:320]. The results for the seventh and eighth group are stored respectively in bits [399:384] and bits [463:447], respectively. The remaining bits in the destination are set to 0. In 64-bit mode and not encoded by VEX/EVEX prefix, using a REX prefix in the form of REX.R permits this instruction to access additional registers (XMM8-XMM15). Legacy SSE version: The source operand can be an MMX technology register or a 64-bit memory location. The destination operand is an MMX technology register. 128-bit Legacy SSE version: The first source operand and destination register are XMM registers. The second source operand is an XMM register or a 128-bit memory location. Bits (MAXVL-1:128) of the corresponding ZMM destination register remain unchanged. VEX.128 and EVEX.128 encoded versions: The first source operand and destination register are XMM registers. The second source operand is an XMM register or a 128-bit memory location. Bits (MAXVL-1:128) of the corresponding ZMM register are zeroed. VEX.256 and EVEX.256 encoded versions: The first source operand and destination register are YMM registers. The second source operand is an YMM register or a 256-bit memory location. Bits (MAXVL-1:256) of the corresponding ZMM register are zeroed. EVEX.512 encoded version: The first source operand and destination register are ZMM registers. The second source operand is a ZMM register or a 512-bit memory location. SRC X7 X6 X5 X4 X3 X2 X1 X0 DEST Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 TEMP ABS(X7:Y7) ABS(X6:Y6) ABS(X5:Y5) ABS(X4:Y4) ABS(X3:Y3) ABS(X2:Y2) ABS(X1:Y1) ABS(X0:Y0) DEST 00H 00H 00H 00H 00H 00H SUM(TEMP7...TEMP0) Figure 4-14. PSADBW Instruction Operation Using 64-bit Operands PSADBW—Compute Sum of Absolute Differences Vol. 2B 4-419

Operation

VPSADBW (EVEX Encoded Versions)
VL = 128, 256, 512
TEMP0 := ABS(SRC1[7:0] - SRC2[7:0])
(* Repeat operation for bytes 1 through 15 *)
TEMP15 := ABS(SRC1[127:120] - SRC2[127:120])
DEST[15:0] := SUM(TEMP0:TEMP7)
DEST[63:16] := 000000000000H
DEST[79:64] := SUM(TEMP8:TEMP15)
DEST[127:80] := 00000000000H

IF VL >= 256
(* Repeat operation for bytes 16 through 31*)
TEMP31 := ABS(SRC1[255:248] - SRC2[255:248])
DEST[143:128] := SUM(TEMP16:TEMP23)
DEST[191:144] := 000000000000H
DEST[207:192] := SUM(TEMP24:TEMP31)
DEST[223:208] := 00000000000H
FI;
IF VL >= 512
(* Repeat operation for bytes 32 through 63*)
TEMP63 := ABS(SRC1[511:504] - SRC2[511:504])
DEST[271:256] := SUM(TEMP0:TEMP7)
DEST[319:272] := 000000000000H
DEST[335:320] := SUM(TEMP8:TEMP15)
DEST[383:336] := 00000000000H
DEST[399:384] := SUM(TEMP16:TEMP23)
DEST[447:400] := 000000000000H
DEST[463:448] := SUM(TEMP24:TEMP31)
DEST[511:464] := 00000000000H
FI;
DEST[MAXVL-1:VL] := 0

VPSADBW (VEX.256 Encoded Version)
TEMP0 := ABS(SRC1[7:0] - SRC2[7:0])
(* Repeat operation for bytes 2 through 30*)
TEMP31 := ABS(SRC1[255:248] - SRC2[255:248])
DEST[15:0] := SUM(TEMP0:TEMP7)
DEST[63:16] := 000000000000H
DEST[79:64] := SUM(TEMP8:TEMP15)
DEST[127:80] := 00000000000H
DEST[143:128] := SUM(TEMP16:TEMP23)
DEST[191:144] := 000000000000H
DEST[207:192] := SUM(TEMP24:TEMP31)
DEST[223:208] := 00000000000H
DEST[MAXVL-1:256] := 0





PSADBW—Compute Sum of Absolute Differences                                                                                               Vol. 2B 4-420
VPSADBW (VEX.128 Encoded Version)
TEMP0 := ABS(SRC1[7:0] - SRC2[7:0])
(* Repeat operation for bytes 2 through 14 *)
TEMP15 := ABS(SRC1[127:120] - SRC2[127:120])
DEST[15:0] := SUM(TEMP0:TEMP7)
DEST[63:16] := 000000000000H
DEST[79:64] := SUM(TEMP8:TEMP15)
DEST[127:80] := 00000000000H
DEST[MAXVL-1:128] := 0

PSADBW (128-bit Legacy SSE Version)
TEMP0 := ABS(DEST[7:0] - SRC[7:0])
(* Repeat operation for bytes 2 through 14 *)
TEMP15 := ABS(DEST[127:120] - SRC[127:120])
DEST[15:0] := SUM(TEMP0:TEMP7)
DEST[63:16] := 000000000000H
DEST[79:64] := SUM(TEMP8:TEMP15)
DEST[127:80] := 00000000000
DEST[MAXVL-1:128] (Unmodified)

PSADBW (64-bit Operand)
TEMP0 := ABS(DEST[7:0] - SRC[7:0])
(* Repeat operation for bytes 2 through 6 *)
TEMP7 := ABS(DEST[63:56] - SRC[63:56])
DEST[15:0] := SUM(TEMP0:TEMP7)
DEST[63:16] := 000000000000H

Intel C/C++ Compiler Intrinsic Equivalent

VPSADBW __m512i _mm512_sad_epu8( __m512i a, __m512i b)
PSADBW __m64 _mm_sad_pu8(__m64 a,__m64 b)
(V)PSADBW __m128i _mm_sad_epu8(__m128i a, __m128i b)
VPSADBW __m256i _mm256_sad_epu8( __m256i a, __m256i b)

Flags Affected

None.

Exceptions

SIMD Floating-Point Exceptions

None.

Other Exceptions

Non-EVEX-encoded instruction, see Table 2-21, “Type 4 Class Exception Conditions.” EVEX-encoded instruction, see Exceptions Type E4NF.nb in Table 2-52, “Type E4NF Class Exception Conditions.” PSADBW—Compute Sum of Absolute Differences Vol. 2B 4-421