Intel® IXP42X Product Line of Network Processors and IXC1100 Control Plane Processor
2.2.8
Fill Buffer (FB) and Pend Buffer (PB)
The four-entry fill buffer (FB) works with the core to hold non-cacheable loads until the bus
controller can act on them. The FB and the four-entry pend buffer (PB) work with the D-cache and
mini-data cache to provide “hit-under-miss” capability, allowing the core to seek other data in the
caches while “miss” data is being fetched from memory.
The FB can contain up to four unique “miss” addresses (logical), allowing four “misses” before the
core is stalled. The PB holds up to four addresses (logical) for additional “misses” to those
addresses that are already in the FB. A coprocessor register can specify draining of the fill and pend
(write) buffers.
2.2.9
Write Buffer (WB)
The write buffer (WB) holds data for storage to memory until the bus controller can act on it. The
WB is eight entries deep, where each entry holds 16 bytes. The WB is constantly enabled and
accepts data from the core, D-cache, or mini-data cache.
Coprocessor 15, Register 1 specifies whether WB coalescing is enabled or disabled. When
coalescing is disabled, stores to memory occur in program order regardless of the attribute bits
within the descriptors located in the DTLB. When coalescing is enabled, the attribute bits within
the descriptors located in the DTLB are examined to determine when coalescing is enabled for the
destination region of memory. When coalescing is enabled in both CP15, R1 and the DTLB, data
entering the WB can coalesce with any of the eight entries (16 bytes) and be stored to the
destination memory region, but possibly out of program order.
Stores to a memory region specified to be non-cacheable and non-bufferable by the attribute bits
within the descriptors located in the DTLB causes the core to stall until the store completes. A
coprocessor register can specify draining of the write buffer.
2.2.10
Multiply-Accumulate Coprocessor (CP0)
For efficient processing of high-quality, media-and-signal-processing algorithms, CP0 provides
40-bit accumulation of 16 x 16, dual-16 x 16 (SIMD), and 32 x 32 signed multiplies. Special MAR
and MRA instructions are implemented to move the 40-bit accumulator to two core-general
registers (MAR) and move two core-general registers to the 40-bit accumulator (MRA). The 40-bit
accumulator can be stored or loaded to or from D-cache, mini-data cache, or memory using two
STC or LDC instructions.
The 16 x 16 signed multiply-accumulates (MIAxy) multiply either the high/high, low/low, high/
low, or low/high 16 bits of a 32-bit core general register (multiplier) and another 32-bit core
general register (multiplicand) to produce a full, 32-bit product that is sign-extended to 40 bits and
added to the 40-bit accumulator.
Dual-signed, 16 x 16 (SIMD) multiply-accumulates (MIAPH) multiply the high/high and low/low
16-bits of a packed 32-bit, core-general register (multiplier) and another packed 32-bit, core-
general register (multiplicand) to produce two 16-bits products that are both sign-extended to
40 bits and added to the 40-bit accumulator.
The 32 x 32 signed multiply-accumulates (MIA) multiply a 32-bit, core-general register
(multiplier) and another 32-bit, core-general register (multiplicand) to produce a 64-bit product
where the 40 LSBs are added to the 40-bit accumulator. The 16 x 32 versions of the 32 x 32
multiply-accumulate instructions complete in a single cycle.
Datasheet
March 2005
Document Number: 252479, Revision: 005
31
INTEL [ INTEL ]
