RSQRTSS—Compute Reciprocal of Square Root of Scalar Single-Precision Floating-Point Value

Opcode*/Instruction Op/En 64/32 bit Mode Support CPUID Feature Flag Description

F3 0F 52 /r

RSQRTSS xmm1, xmm2/m32

RM V/V SSE Computes the approximate reciprocal of the square root of the low single-precision floating-point value in xmm2/m32 and stores the results in xmm1.
VEX.NDS.LIG.F3.0F.WIG 52 /r VRSQRTSS xmm1, xmm2, xmm3/m32 RVM V/V AVX Computes the approximate reciprocal of the square root of the low single precision floating-point value in xmm3/m32 and stores the results in xmm1. Also, upper single precision floating-point values (bits[127:32]) from xmm2 are copied to xmm1[127:32].

Instruction Operand Encoding

Op/En Operand 1 Operand 2 Operand 3 Operand 4
RM ModRM:reg (w) ModRM:r/m (r) NA NA
RVM ModRM:reg (w) VEX.vvvv (r) ModRM:r/m (r) NA

Description

Computes an approximate reciprocal of the square root of the low single-precision floating-point value in the source operand (second operand) stores the single-precision floating-point result in the destination operand. The source operand can be an XMM register or a 32-bit memory location. The destination operand is an XMM register. The three high-order doublewords of the destination operand remain unchanged. See Figure 10-6 in the Intel® 64 and IA-32 Architectures Software Developer’s Manual, Volume 1, for an illustration of a scalar single-precision floating-point operation.

The relative error for this approximation is:

|Relative Error| ≤ 1.5 ∗ 2−12

The RSQRTSS instruction is not affected by the rounding control bits in the MXCSR register. When a source value is a 0.0, an ∞ of the sign of the source value is returned. A denormal source value is treated as a 0.0 (of the same sign). When a source value is a negative value (other than −0.0), a floating-point indefinite is returned. When a source value is an SNaN or QNaN, the SNaN is converted to a QNaN or the source QNaN is returned.

In 64-bit mode, using a REX prefix in the form of REX.R permits this instruction to access additional registers (XMM8-XMM15).

128-bit Legacy SSE version: The first source operand and the destination operand are the same. Bits (VLMAX-1:32) of the corresponding YMM destination register remain unchanged.

VEX.128 encoded version: Bits (VLMAX-1:128) of the destination YMM register are zeroed.

Operation

RSQRTSS (128-bit Legacy SSE version)

DEST[31:0] ← APPROXIMATE(1/SQRT(SRC2[31:0]))
DEST[VLMAX-1:32] (Unmodified)

VRSQRTSS (VEX.128 encoded version)

DEST[31:0] ← APPROXIMATE(1/SQRT(SRC2[31:0]))
DEST[127:32] ← SRC1[127:32]
DEST[VLMAX-1:128] ← 0

Intel C/C++ Compiler Intrinsic Equivalent

RSQRTSS:

__m128 _mm_rsqrt_ss(__m128 a)

SIMD Floating-Point Exceptions

None.

Other Exceptions

See Exceptions Type 5.