vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Add some minimal optimizations for CDNA (#10498)

Browse source 
* Add some minimal optimizations for CDNA

* ggml_cuda: set launch bounds also for GCN as it helps there too
This commit is contained in:
uvos 2024-11-27 17:10:08 +01:00 • committed by GitHub
parent 46c69e0e75
commit 3ad5451f3b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 36 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

17
ggml/src/ggml-cuda/common.cuh
View file

@ -47,9 +47,20 @@
#define CC_TURING 750
#define CC_AMPERE 800
#define CC_OFFSET_AMD 1000000
#define CC_RDNA1 (CC_OFFSET_AMD + 1010)
#define CC_RDNA2 (CC_OFFSET_AMD + 1030)
#define CC_RDNA3 (CC_OFFSET_AMD + 1100)
// GCN/CNDA, wave size is 64
#define CC_GCN4 (CC_OFFSET_AMD + 803) // Tonga, Fiji, Polaris, minimum for fast fp16
#define CC_VEGA (CC_OFFSET_AMD + 900) // Vega56/64, minimum for fp16 dual issue
#define CC_VEGA20 (CC_OFFSET_AMD + 906) // MI50/Radeon VII, minimum for dp4a
#define CC_CDNA (CC_OFFSET_AMD + 908) // MI100, minimum for MFMA, acc registers
#define CC_CDNA2 (CC_OFFSET_AMD + 910) // MI210, minimum acc register renameing
#define CC_CDNA3 (CC_OFFSET_AMD + 942) // MI300
// RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32
#define CC_RDNA1 (CC_OFFSET_AMD + 1010) // RX 5000
#define CC_RDNA2 (CC_OFFSET_AMD + 1030) // RX 6000, minimum for dp4a
#define CC_RDNA3 (CC_OFFSET_AMD + 1100) // RX 7000, minimum for WMMA
#define CC_QY1 210
#define CC_QY2 220