From 5adccee9658f2e8f30e69f5545162b1b271405a1 Mon Sep 17 00:00:00 2001 From: Wilson Fung Date: Thu, 17 May 2012 00:54:12 -0800 Subject: Updating the CHANGES file to describe various changes I made since last release. [git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 12621] --- CHANGES | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/CHANGES b/CHANGES index b677137..acbc2de 100644 --- a/CHANGES +++ b/CHANGES @@ -13,6 +13,8 @@ Version 3.1.0 versus 3.0.2 - Support for configurable instruction latency and initiation interval - Added corresponding configuration to Fermi and Quadro config files according to benchmark results +- Support for 64-bit atomicAdd, atomicExch, and atomicCAS. +- Support for 32-bit floating point atomicAdd. - New installation instructions for AerialVision, removed deprecated install script. - Rework for some makefiles @@ -34,6 +36,20 @@ Version 3.1.0 versus 3.0.2 - Fixed a bug that caused deadlock check to be omitted - Added a missed support for the perfect memory configuration option which models a memory system with 1 cycle latency per request. + - Updated the Fermi config files such that when an access misses the L1 data + cache, it allocates a line immediately before sending a data fetch request + out to the memory partition. + - Changed the writeback arbitration among multiple clients in the LDST unit + to round-robin. + - Fixed a rare pathological case that caused LD instructions with divergent + memory accesses to be overcounted. + - Fixed linear_to_raw_address_translation::partition_address() so that it + works for non-power-of-two number of memory partitions. Before, it just + returns the input memory address without any change. + - Fixed warp_inst_t::memory_coalescing_arch_13() to use the + gpgpu_shmem_warp_parts option to divide up the warps in the coalescing + logic. It was hardcoded to 2, which causes coalescing to always operate + on half-warps. Version 3.0.2 versus 3.0.1 - Added Fermi configuration -- cgit v1.3