{"id":701,"date":"2010-12-01T15:26:32","date_gmt":"2010-12-01T05:26:32","guid":{"rendered":"http:\/\/brnz.org\/hbr\/?p=701"},"modified":"2015-01-25T09:39:42","modified_gmt":"2015-01-24T23:39:42","slug":"assembly-primer-part-6-moving-data-spu","status":"publish","type":"post","link":"https:\/\/brnz.org\/hbr\/?p=701","title":{"rendered":"Assembly Primer Part 6 — Moving Data — SPU"},"content":{"rendered":"

These are my notes for where I can see SPU varying from ia32, as presented in the video Part 6 \u2014 Moving Data<\/a>.<\/p>\n

SPU and ia32 differ significantly when it comes to moving\/copying data around, in terms of the ways things can be copied, the alignment of data in memory and the vector nature of SPU registers.<\/p>\n

This is the quick SPU instruction reference I use<\/a>.\u00a0 The SPU ISA doc<\/a> is worth having nearby if trying to do silly tricks with SPU instructions.<\/p>\n

Moving Data<\/h2>\n

Lets consider what MovDemo.s<\/a> might look like for SPU, piece by piece.<\/p>\n

First, the storage:<\/p>\n

# Demo program to show how to use Data types and MOVx instructions \r\n\r\n.data\r\n    HelloWorld:\r\n        .ascii \"Hello World!\"\r\n\r\n    ByteLocation:\r\n        .byte 10\r\n\r\n    Int32:\r\n        .int 2\r\n    Int16:\r\n        .short 3\r\n    Float:\r\n        .float 10.23\r\n\r\n    IntegerArray:\r\n        .int 10,20,30,40,50\r\n<\/pre>\n
Icky.\u00a0 Not naturally aligned for size, and crossing qword boundaries.\u00a0 This makes the following code particularly messy, because the SPU really<\/em> doesn’t like unaligned data.\u00a0 Oh well, let’s begin.<\/p>\n
1. Immediate value to register<\/h3>\n
    .align 3  # ensure code is aligned after awkwardly arranged data\r\n.text\r\n    .globl _start\r\n    _start:\r\n        #movl $10, %eax\r\n        il $5, 10\r\n<\/pre>\n
Well, that was easy enough.<\/p>\n
It does get a little trickier if trying to load more complex values.\u00a0 To load a full 32-bits of immediate data into a register requires two half-word load\u00a0 instructions for the upper and lower parts.<\/p>\n
Loading arbitrary values spanning multiple words is more complex, and is often able to be done simply by storing the constant in .rodata and loading it when needed, although fsmbi<\/em> can be useful on occasion.<\/p>\n
2. Immediate value to local store<\/h3>\n
#movw $50, Int16\r\n\r\n# Write first byte\r\nila $6,Int16        # load the address of the target\r\nlqa $7,Int16        # load the qword\r\nil $8,0             # load the upper byte of the constant into preferred slot\r\ncbd $9,0($6)        # insertion mask for higher byte\r\nshufb $10,$8,$7,$9  # shuffle the upper byte into the qword\r\nstqa $10,Int16      # write the word back\r\n# Write second byte\r\nil $11,1            # load 1 for address offsetting\r\ncbd $12,1($6)       # insertion mask for lower byte\r\nil $13,50           # load the lower byte of the constant into preferred slot\r\nlqx $14,$6,$11      # load Int16+1\r\nshufb $15,$13,$14,$12 # shuffle lower byte into qword\r\nstqx $15,$6,$11     # write back qword\r\n<\/pre>\n
Writing two bytes to a non-aligned memory location is a messy business.<\/p>\n