RenderScript Runtime API Reference


RenderScript is a high-performance runtime that provides compute operations at the native level. RenderScript code is compiled on devices at runtime to allow platform-independence as well.

This reference documentation describes the RenderScript runtime APIs, which you can utilize to write RenderScript code in C99. The RenderScript compute header files are automatically included for you.

To use RenderScript, you need to utilize the RenderScript runtime APIs documented here as well as the Android framework APIs for RenderScript. For documentation on the Android framework APIs, see the android.renderscript package reference.

For more information on how to develop with RenderScript and how the runtime and Android framework APIs interact, see the RenderScript developer guide and the RenderScript samples.

Numerical Types


RenderScript supports the following scalar numerical types:

8 bits 16 bits 32 bits 64 bits
Integer: char, int8_t short, int16_t int32_t long, long long, int64_t
Unsigned integer: uchar, uint8_t ushort, uint16_t uint, uint32_t ulong, uint64_t
Floating point: float double


RenderScript supports fixed size vectors of length 2, 3, and 4. Vectors are declared using the common type name followed by a 2, 3, or 4. E.g. float4, int3, double2, ulong4.

To create vector literals, use the vector type followed by the values enclosed between parentheses, e.g. (float3)(1.0f, 2.0f, 3.0f).

Entries of a vector can be accessed using different naming styles.

Single entries can be accessed by following the variable name with a dot and:

  • The letters x, y, z, and w,
  • The letters r, g, b, and a,
  • The letter s or S, followed by a zero based index.

For example, with int4 myVar; the following are equivalent:
myVar.x == myVar.r == myVar.s0 == myVar.S0
myVar.y == myVar.g == myVar.s1 == myVar.S1
myVar.z == myVar.b == myVar.s2 == myVar.S2
myVar.w == myVar.a == myVar.s3 == myVar.S3

Multiple entries of a vector can be accessed at once by using an identifier that is the concatenation of multiple letters or indices. The resulting vector has a size equal to the number of entries named.

With the example above, the middle two entries can be accessed using myVar.yz,, myVar.s12, and myVar.S12.

The entries don't have to be contiguous or in increasing order. Entries can even be repeated, as long as we're not trying to assign to it. You also can't mix the naming styles.

Here are examples of what can or can't be done:
float4 v4;
float3 v3;
float2 v2;
v2 = v4.xx; // Valid
v3 = v4.zxw; // Valid
v3 = v4.bba; // Valid
v3 = v4.s032; // Valid
v3.s120 = v4.S233; // Valid
v4.yz = v3.rg; // Valid
v4.yzx = v3.rg; // Invalid: mismatched sizes
v4.yzz = v3; // Invalid: z appears twice in an assignment
v3 = v3.xas0; // Invalid: can't mix xyzw with rgba nor s0...
v3 = v4.s034; // Invalid: the digit can only be 0, 1, 2, or 3

Matrices and Quaternions:

RenderScript supports fixed size square matrices of floats of size 2x2, 3x3, and 4x4. The types are named rs_matrix2x2, rs_matrix3x3, and rs_matrix4x4. See Matrix Functions for the list of operations.

Quaternions are also supported via rs_quaternion. See Quaterion Functions for the list of operations.

char2 Two 8 bit signed integers
char3 Three 8 bit signed integers
char4 Four 8 bit signed integers
double2 Two 64 bit floats
double3 Three 64 bit floats
double4 Four 64 bit floats
float2 Two 32 bit floats
float3 Three 32 bit floats
float4 Four 32 bit floats
int16_t 16 bit signed integer
int2 Two 32 bit signed integers
int3 Three 32 bit signed integers
int32_t 32 bit signed integer
int4 Four 32 bit signed integers
int64_t 64 bit signed integer
int8_t 8 bit signed integer
long2 Two 64 bit signed integers
long3 Three 64 bit signed integers
long4 Four 64 bit signed integers
rs_matrix2x2 2x2 matrix of 32 bit floats
rs_matrix3x3 3x3 matrix of 32 bit floats
rs_matrix4x4 4x4 matrix of 32 bit floats
rs_quaternion Quaternion
short2 Two 16 bit signed integers
short3 Three 16 bit signed integers
short4 Four 16 bit signed integers
size_t Unsigned size type
ssize_t Signed size type
uchar 8 bit unsigned integer
uchar2 Two 8 bit unsigned integers
uchar3 Three 8 bit unsigned integers
uchar4 Four 8 bit unsigned integers
uint 32 bit unsigned integer
uint16_t 16 bit unsigned integer
uint2 Two 32 bit unsigned integers
uint3 Three 32 bit unsigned integers
uint32_t 32 bit unsigned integer
uint4 Four 32 bit unsigned integers
uint64_t 64 bit unsigned integer
uint8_t 8 bit unsigned integer
ulong 64 bit unsigned integer
ulong2 Two 64 bit unsigned integers
ulong3 Three 64 bit unsigned integers
ulong4 Four 64 bit unsigned integers
ushort 16 bit unsigned integer
ushort2 Two 16 bit unsigned integers
ushort3 Three 16 bit unsigned integers
ushort4 Four 16 bit unsigned integers

Object Types

The types below are used to manipulate RenderScript objects like allocations, samplers, elements, and scripts. Most of these object are created using the Java RenderScript APIs.

rs_allocation Handle to an allocation
rs_allocation_cubemap_face Enum for selecting cube map faces
rs_allocation_usage_type Bitfield to specify how an allocation is used
rs_data_kind Element data kind
rs_data_type Element basic data type
rs_element Handle to an element
rs_sampler Handle to a Sampler
rs_sampler_value Sampler wrap T value
rs_script Handle to a Script
rs_type Handle to a Type

Conversion Functions

The functions below convert from a numerical vector type to another, of from one color representation to another.

convert Convert numerical vectors
rsPackColorTo8888 Create a uchar4 RGBA from floats
rsUnpackColor8888 Create a float4 RGBA from uchar4
rsYuvToRGBA Convert a YUV value to RGBA

Mathematical Constants and Functions

The mathematical functions below can be applied to scalars and vectors. When applied to vectors, the returned value is a vector of the function applied to each entry of the input.

For example:
float3 a, b;
// The following call sets
// a.x to sin(b.x),
// a.y to sin(b.y), and
// a.z to sin(b.z).
a = sin(b);

See Vector Math Functions for functions like distance() and length() that interpret instead the input as a single vector in n-dimensional space.

The precision of the mathematical operations on 32 bit floats is affected by the pragmas rs_fp_relaxed and rs_fp_full. Under rs_fp_relaxed, subnormal values may be flushed to zero and rounding may be done towards zero. In comparison, rs_fp_full requires correct handling of subnormal values, i.e. smaller than 1.17549435e-38f. rs_fp_rull also requires round to nearest with ties to even.

Different precision/speed tradeoffs can be achieved by using variants of the common math functions. Functions with a name starting with

  • native_: May have custom hardware implementations with weaker precision. Additionally, subnormal values may be flushed to zero, rounding towards zero may be used, and NaN and infinity input may not be handled correctly.
  • half_: May perform internal computations using 16 bit floats. Additionally, subnormal values may be flushed to zero, and rounding towards zero may be used.

M_1_PI 1 / pi, as a 32 bit float
M_2_PI 2 / pi, as a 32 bit float
M_2_SQRTPI 2 / sqrt(pi), as a 32 bit float
M_E e, as a 32 bit float
M_LN10 log_e(10), as a 32 bit float
M_LN2 log_e(2), as a 32 bit float
M_LOG10E log_10(e), as a 32 bit float
M_LOG2E log_2(e), as a 32 bit float
M_PI pi, as a 32 bit float
M_PI_2 pi / 2, as a 32 bit float
M_PI_4 pi / 4, as a 32 bit float
M_SQRT1_2 1 / sqrt(2), as a 32 bit float
M_SQRT2 sqrt(2), as a 32 bit float
abs Absolute value of an integer
acos Inverse cosine
acosh Inverse hyperbolic cosine
acospi Inverse cosine divided by pi
asin Inverse sine
asinh Inverse hyperbolic sine
asinpi Inverse sine divided by pi
atan Inverse tangent
atan2 Inverse tangent of a ratio
atan2pi Inverse tangent of a ratio, divided by pi
atanh Inverse hyperbolic tangent
atanpi Inverse tangent divided by pi
cbrt Cube root
ceil Smallest integer not less than a value
clamp Restrain a value to a range
clz Number of leading 0 bits
copysign Copies the sign of a number to another
cos Cosine
cosh Hypebolic cosine
cospi Cosine of a number multiplied by pi
degrees Converts radians into degrees
erf Mathematical error function
erfc Mathematical complementary error function
exp e raised to a number
exp10 10 raised to a number
exp2 2 raised to a number
expm1 e raised to a number minus one
fabs Absolute value of a float
fdim Positive difference between two values
floor Smallest integer not greater than a value
fma Multiply and add
fmax Maximum of two floats
fmin Minimum of two floats
fmod Modulo
fract Positive fractional part
frexp Binary mantissa and exponent
half_recip Reciprocal computed to 16 bit precision
half_rsqrt Reciprocal of a square root computed to 16 bit precision
half_sqrt Square root computed to 16 bit precision
hypot Hypotenuse
ilogb Base two exponent
ldexp Creates a floating point from mantissa and exponent
lgamma Natural logarithm of the gamma function
log Natural logarithm
log10 Base 10 logarithm
log1p Natural logarithm of a value plus 1
log2 Base 2 logarithm
logb Base two exponent
mad Multiply and add
max Maximum
min Minimum
mix Mixes two values
modf Integral and fractional components
nan Not a Number
native_acos Approximate inverse cosine
native_acosh Approximate inverse hyperbolic cosine
native_acospi Approximate inverse cosine divided by pi
native_asin Approximate inverse sine
native_asinh Approximate inverse hyperbolic sine
native_asinpi Approximate inverse sine divided by pi
native_atan Approximate inverse tangent
native_atan2 Approximate inverse tangent of a ratio
native_atan2pi Approximate inverse tangent of a ratio, divided by pi
native_atanh Approximate inverse hyperbolic tangent
native_atanpi Approximate inverse tangent divided by pi
native_cbrt Approximate cube root
native_cos Approximate cosine
native_cosh Approximate hypebolic cosine
native_cospi Approximate cosine of a number multiplied by pi
native_divide Approximate division
native_exp Approximate e raised to a number
native_exp10 Approximate 10 raised to a number
native_exp2 Approximate 2 raised to a number
native_expm1 Approximate e raised to a number minus one
native_hypot Approximate hypotenuse
native_log Approximate natural logarithm
native_log10 Approximate base 10 logarithm
native_log1p Approximate natural logarithm of a value plus 1
native_log2 Approximate base 2 logarithm
native_powr Approximate positive base raised to an exponent
native_recip Approximate reciprocal
native_rootn Approximate nth root
native_rsqrt Approximate reciprocal of a square root
native_sin Approximate sine
native_sincos Approximate sine and cosine
native_sinh Approximate hyperbolic sine
native_sinpi Approximate sine of a number multiplied by pi
native_sqrt Approximate square root
native_tan Approximate tangent
native_tanh Approximate hyperbolic tangent
native_tanpi Approximate tangent of a number multiplied by pi
nextafter Next floating point number
pow Base raised to an exponent
pown Base raised to an integer exponent
powr Positive base raised to an exponent
radians Converts degrees into radians
remainder Remainder of a division
remquo Remainder and quotient of a division
rint Round to even
rootn Nth root
round Round away from zero
rsRand Pseudo-random number
rsqrt Reciprocal of a square root
sign Sign of a value
sin Sine
sincos Sine and cosine
sinh Hyperbolic sine
sinpi Sine of a number multiplied by pi
sqrt Square root
step 0 if less than a value, 0 otherwise
tan Tangent
tanh Hyperbolic tangent
tanpi Tangent of a number multiplied by pi
tgamma Gamma function
trunc Truncates a floating point

Vector Math Functions

These functions interpret the input arguments as representation of vectors in n-dimensional space.

The precision of the mathematical operations on 32 bit floats is affected by the pragmas rs_fp_relaxed and rs_fp_full. See Mathematical Constants and Functions for details.

Different precision/speed tradeoffs can be achieved by using variants of the common math functions. Functions with a name starting with

  • native_: May have custom hardware implementations with weaker precision. Additionally, subnormal values may be flushed to zero, rounding towards zero may be used, and NaN and infinity input may not be handled correctly.
  • fast_: May perform internal computations using 16 bit floats. Additionally, subnormal values may be flushed to zero, and rounding towards zero may be used.

cross Cross product of two vectors
distance Distance between two points
dot Dot product of two vectors
fast_distance Approximate distance between two points
fast_length Approximate length of a vector
fast_normalize Approximate normalized vector
length Length of a vector
native_distance Approximate distance between two points
native_length Approximate length of a vector
native_normalize Approximately normalize a vector
normalize Normalize a vector

Matrix Functions

These functions let you manipulate square matrices of rank 2x2, 3x3, and 4x4. They are particularly useful for graphical transformations and are compatible with OpenGL.

We use a zero-based index for rows and columns. E.g. the last element of a rs_matrix4x4 is found at (3, 3).

RenderScript uses column-major matrices and column-based vectors. Transforming a vector is done by postmultiplying the vector, e.g. (matrix * vector), as provided by rsMatrixMultiply().

To create a transformation matrix that performs two transformations at once, multiply the two source matrices, with the first transformation as the right argument. E.g. to create a transformation matrix that applies the transformation s1 followed by s2, call rsMatrixLoadMultiply(&combined, &s2, &s1). This derives from s2 * (s1 * v), which is (s2 * s1) * v.

We have two style of functions to create transformation matrices: rsMatrixLoadTransformation and rsMatrixTransformation. The former style simply stores the transformation matrix in the first argument. The latter modifies a pre-existing transformation matrix so that the new transformation happens first. E.g. if you call rsMatrixTranslate() on a matrix that already does a scaling, the resulting matrix when applied to a vector will first do the translation then the scaling.

rsExtractFrustumPlanes Compute frustum planes
rsIsSphereInFrustum Checks if a sphere is within the frustum planes
rsMatrixGet Get one element
rsMatrixInverse Inverts a matrix in place
rsMatrixInverseTranspose Inverts and transpose a matrix in place
rsMatrixLoad Load or copy a matrix
rsMatrixLoadFrustum Load a frustum projection matrix
rsMatrixLoadIdentity Load identity matrix
rsMatrixLoadMultiply Multiply two matrices
rsMatrixLoadOrtho Load an orthographic projection matrix
rsMatrixLoadPerspective Load a perspective projection matrix
rsMatrixLoadRotate Load a rotation matrix
rsMatrixLoadScale Load a scaling matrix
rsMatrixLoadTranslate Load a translation matrix
rsMatrixMultiply Multiply a matrix by a vector or another matrix
rsMatrixRotate Apply a rotation to a transformation matrix
rsMatrixScale Apply a scaling to a transformation matrix
rsMatrixSet Set one element
rsMatrixTranslate Apply a translation to a transformation matrix
rsMatrixTranspose Transpose a matrix place

Quaternion Functions

The following functions manipulate quaternions.

rsQuaternionAdd Add two quaternions
rsQuaternionConjugate Conjugate a quaternion
rsQuaternionDot Dot product of two quaternions
rsQuaternionGetMatrixUnit Get a rotation matrix from a quaternion
rsQuaternionLoadRotate Create a rotation quaternion
rsQuaternionLoadRotateUnit Quaternion that represents a rotation about an arbitrary unit vector
rsQuaternionMultiply Multiply a quaternion by a scalar or another quaternion
rsQuaternionNormalize Normalize a quaternion
rsQuaternionSet Create a quaternion
rsQuaternionSlerp Spherical linear interpolation between two quaternions

Atomic Update Functions

To update values shared between multiple threads, use the functions below. They ensure that the values are atomically updated, i.e. that the memory reads, the updates, and the memory writes are done in the right order.

These functions are slower than their non-atomic equivalents, so use them only when synchronization is needed.

Note that in RenderScript, your code is likely to be running in separate threads even though you did not explicitely create them. The RenderScript runtime will very often split the execution of one kernel across multiple threads. Updating globals should be done with atomic functions. If possible, modify your algorithm to avoid them altogether.

rsAtomicAdd Thread-safe addition
rsAtomicAnd Thread-safe bitwise and
rsAtomicCas Thread-safe compare and set
rsAtomicDec Thread-safe decrement
rsAtomicInc Thread-safe increment
rsAtomicMax Thread-safe maximum
rsAtomicMin Thread-safe minimum
rsAtomicOr Thread-safe bitwise or
rsAtomicSub Thread-safe subtraction
rsAtomicXor Thread-safe bitwise exclusive or

Time Functions and Types

The functions below can be used to tell the current clock time and the current system up time. It is not recommended to call these functions inside of a kernel.

rs_time_t Seconds since January 1, 1970
rs_tm Date and time structure
rsGetDt Elapsed time since last call
rsLocaltime Convert to local time
rsTime Seconds since January 1, 1970
rsUptimeMillis System uptime in milliseconds
rsUptimeNanos System uptime in nanoseconds

Allocation Data Access Functions

The functions below can be used to get and set the cells that comprise an allocation.

  • Individual cells are accessed using the rsGetElementAt* and rsSetElementAt functions.
  • Multiple cells can be copied using the rsAllocationCopy* and rsAllocationV* functions.
  • For getting values through a sampler, use rsSample.
The rsGetElementAt and rsSetElement* functions are somewhat misnamed. They don't get or set elements, which are akin to data types; they get or set cells. Think of them as rsGetCellAt and and rsSetCellAt.

rsAllocationCopy1DRange Copy consecutive cells between allocations
rsAllocationCopy2DRange Copy a rectangular region of cells between allocations
rsAllocationVLoadX Get a vector from an allocation of scalars
rsAllocationVStoreX Store a vector into an allocation of scalars
rsGetElementAt Return a cell from an allocation
rsGetElementAtYuv_uchar_U Get the U component of an allocation of YUVs
rsGetElementAtYuv_uchar_V Get the V component of an allocation of YUVs
rsGetElementAtYuv_uchar_Y Get the Y component of an allocation of YUVs
rsSample Sample a value from a texture allocation
rsSetElementAt Set a cell of an allocation

Object Characteristics Functions

The functions below can be used to query the characteristics of an Allocation, Element, or Sampler object. These objects are created from Java. You can't create them from a script.


Allocations are the primary method used to pass data to and from RenderScript kernels.

They are a structured collection of cells that can be used to store bitmaps, textures, arbitrary data points, etc.

This collection of cells may have many dimensions (X, Y, Z, Array0, Array1, Array2, Array3), faces (for cubemaps), and level of details (for mipmapping).

See the android.renderscript.Allocation for details on to create Allocations.


The term "element" is used a bit ambiguously in RenderScript, as both type information for the cells of an Allocation and the instantiation of that type. For example:

  • rs_element is a handle to a type specification, and
  • In functions like rsGetElementAt(), "element" means the instantiation of the type, i.e. a cell of an Allocation.

The functions below let you query the characteristics of the type specificiation.

An Element can specify a simple data types as found in C, e.g. an integer, float, or boolean. It can also specify a handle to a RenderScript object. See rs_data_type for a list of basic types.

Elements can specify fixed size vector (of size 2, 3, or 4) versions of the basic types. Elements can be grouped together into complex Elements, creating the equivalent of C structure definitions.

Elements can also have a kind, which is semantic information used to interpret pixel data. See rs_data_kind.

When creating Allocations of common elements, you can simply use one of the many predefined Elements like F32_2.

To create complex Elements, use the Element.Builder Java class.


Samplers objects define how Allocations can be read as structure within a kernel. See android.renderscript.S.

rsAllocationGetDimFaces Presence of more than one face
rsAllocationGetDimLOD Presence of levels of detail
rsAllocationGetDimX Size of the X dimension
rsAllocationGetDimY Size of the Y dimension
rsAllocationGetDimZ Size of the Z dimension
rsAllocationGetElement Get the object that describes the cell of an Allocation
rsClearObject Release an object
rsElementGetBytesSize Size of an Element
rsElementGetDataKind Kind of an Element
rsElementGetDataType Data type of an Element
rsElementGetSubElement Sub-element of a complex Element
rsElementGetSubElementArraySize Array size of a sub-element of a complex Element
rsElementGetSubElementCount Number of sub-elements
rsElementGetSubElementName Name of a sub-element
rsElementGetSubElementNameLength Length of the name of a sub-element
rsElementGetSubElementOffsetBytes Offset of the instantiated sub-element
rsElementGetVectorSize Vector size of the Element
rsIsObject Check for an empty handle
rsSamplerGetAnisotropy Anisotropy of the Sampler
rsSamplerGetMagnification Sampler magnification value
rsSamplerGetMinification Sampler minification value
rsSamplerGetWrapS Sampler wrap S value
rsSamplerGetWrapT Sampler wrap T value

Kernel Invocation Functions and Types

The rsForEach() function can be used to invoke the root kernel of a script.

rs_for_each_strategy_t Suggested cell processing order
rs_script_call_t Cell iteration information
rsForEach Invoke the root kernel of a script

Input/Output Functions

These functions are used to:

  • Send information to the Java client, and
  • Send the processed allocation or receive the next allocation to process.

rsAllocationIoReceive Receive new content from the queue
rsAllocationIoSend Send new content to the queue
rsSendToClient Send a message to the client, non-blocking
rsSendToClientBlocking Send a message to the client, blocking

Debugging Functions

The functions below are intended to be used during application developement. They should not be used in shipping applications.

rsDebug Log a message and values

Graphics Functions and Types

The graphics subsystem of RenderScript has been deprecated.