Simple macro translation

For many programs, it is possible to use a simple macro translation strategy for developing code that is portable between the Vector/SIMD Multimedia Extension and SPU instruction sets.

The keys to simple macro translation are:

Use a Compatible Vector-Literal Construction Format — The PPE Vector/SIMD Multimedia Extension and the SPE's SPU instruction set specifies two styles of constructing literal vectors: curly brace and parenthesis. Some compilers support both styles. A set of construction macros can be used to insulate programs from any differences in the tools.

Use Single-Token Vector Data Types — The C/C++ Language Extensions for Cell Broadband Engine Architecture document specifies a set of single-token vector data types. Because these are single-token, the data types can be easily redefined by a preprocessor to the desired target processor. Additional single-token data types must be standardized for the unique Vector/SIMD Multimedia Extension data types. Table 1 lists the data types. See also:

Table 1.
Table 2.

Table 1. Vector/SIMD Multimedia Extension Single-Token Data Types
Vector Data Type	Single-Token Data Type
vector bool char	vec_bchar16
vector bool short	vec_bshort8
vector bool int	vec_bint4
vector pixel	vec_pixel8

Use Intrinsics that Map One-to-One — Regardless of the technique used to provide portability, performance will be optimized if the operations map one-to-one between Vector/SIMD Multimedia Extension intrinsics and SPU intrinsics.

The SPU intrinsics that map one-to-one with Vector/SIMD Multimedia Extension (except for the specific intrinsics described in Specific intrinsics) are shown in Table 2.

The Vector/SIMD Multimedia Extension intrinsics that map one-to-one with SPU are shown in Table 3.

Table 2. SPU Intrinsics with One-to-One Vector/SIMD Multimedia Extension Mapping
SPU Intrinsic	Vector/SIMD Multimedia Extension Intrinsic	For Data Types
spu_add	vec_add	vector operands only, no scalar operands
spu_and	vec_and	vector operands only, no scalar operands
spu_andc	vec_andc	all
spu_avg	vec_avg	all
spu_cmpeq	vec_cmpeq	vector operands only, no scalar operands
spu_cmpgt	vec_cmpgt	vector operands only, no scalar operands
spu_convtf	vec_ctf	limited scale range (5 bits)
spu_convts	vec_cts	limited scale range (5 bits)
spu_convtu	vec_ctu	limited scale range (5 bits)
spu_extract	vec_extract	all
spu_genc	vec_addc	all
spu_insert	vec_insert	all
spu_madd	vec_madd	float only
spu_mulhh	vec_mule	all
spu_mulo	vec_mulo	halfword vector operands only, no scalar operands
spu_nmsub	vec_nmsub	float only
spu_nor	vec_nor	all
spu_or	vec_or	vector operands only, no scalar operands
spu_promote	vec_promote	all
spu_re	vec_re	all
spu_rl	vec_rl	vector operands only, no scalar operands
spu_rsqrte	vec_rsqrte	all
spu_sel	vec_sel	all
spu_splats	vec_splats	all
spu_sub	vec_sub	vector operands only, no scalar operands
spu_genb	vec_subc	vector operands only, no scalar operands
spu_xor	vec_xor	vector operands only, no scalar operands

Table 3. Vector/SIMD Multimedia Extension Intrinsics with One-to-One SPU Mapping
Vector/SIMD Multimedia Extension Intrinsic	SPU Intrinsic	For Data Types
vec_add	spu_add	halfwords, words, and floats only (not bytes)
vec_addc	spu_genc	all
vec_and	spu_and	all
vec_andc	spu_andc	all
vec_avg	spu_avg	unsigned chars only
vec_cmpeq	spu_cmpeq	all
vec_cmpgt	spu_cmpgt	all
vec_ctf	spu_convtf	all
vec_cts	spu_convts	all
vec_ctu	spu_convtu	all
vec_extract	spu_extract	all
vec_insert	spu_insert	all
vec_madd	spu_madd	all
vec_mulo	spu_mulo	halfwords only (not bytes)
vec_nmsub	spu_nmsub	all
vec_nor	spu_nor	all
vec_or	spu_or	all
vec_promote	spu_promote	all
vec_re	spu_re	all
vec_rl	spu_rl	halfwords and words only (not bytes)
vec_rsqrte	spu_rsqrte	all
vec_sel	spu_sel	all
vec_splats	spu_splats	all
vec_sub	spu_sub	halfwords, words, and floats only
vec_subc	spu_genb	all
vec_xor	spu_xor	all

Note: The toolchain contains headers files of overloaded C++ functions that can used to assist in mapping or porting of Vector/SIMD Multimedia Extension intrinsics to SPU intrinsics, and vice-versa.