Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml: add gfx90c HIP support (#26454)

Browse Source 
* ggml: add gfx90c HIP support

* ggml: make gfx90c HIP support compliant with specifications
This commit is contained in:
AuroraRAS
2026-09-07 15:21:42 +02:00
committed by GitHub
parent d23c47f2a9
commit 4735997382
4 changed files with 14 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/common.cuh
+9 -6
View File
@@ -69,6 +69,8 @@
#define GGML_CUDA_CC_GCN4 (GGML_CUDA_CC_OFFSET_AMD + 0x803) // Tonga, Fiji, Polaris, minimum for fast fp16 #define GGML_CUDA_CC_GCN4 (GGML_CUDA_CC_OFFSET_AMD + 0x803) // Tonga, Fiji, Polaris, minimum for fast fp16
#define GGML_CUDA_CC_VEGA (GGML_CUDA_CC_OFFSET_AMD + 0x900) // Vega56/64, minimum for fp16 dual issue #define GGML_CUDA_CC_VEGA (GGML_CUDA_CC_OFFSET_AMD + 0x900) // Vega56/64, minimum for fp16 dual issue
#define GGML_CUDA_CC_VEGA20 (GGML_CUDA_CC_OFFSET_AMD + 0x906) // MI50/Radeon VII, minimum for dp4a #define GGML_CUDA_CC_VEGA20 (GGML_CUDA_CC_OFFSET_AMD + 0x906) // MI50/Radeon VII, minimum for dp4a
#define GGML_CUDA_CC_GFX909 (GGML_CUDA_CC_OFFSET_AMD + 0x909) // GCN APU
#define GGML_CUDA_CC_GFX90C (GGML_CUDA_CC_OFFSET_AMD + 0x90c) // GCN APU
#define GGML_CUDA_CC_CDNA1 (GGML_CUDA_CC_OFFSET_AMD + 0x908) // MI100, minimum for MFMA, acc registers #define GGML_CUDA_CC_CDNA1 (GGML_CUDA_CC_OFFSET_AMD + 0x908) // MI100, minimum for MFMA, acc registers
#define GGML_CUDA_CC_CDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x90a) // MI210 (gfx90a), minimum acc register renaming #define GGML_CUDA_CC_CDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x90a) // MI210 (gfx90a), minimum acc register renaming
#define GGML_CUDA_CC_CDNA3 (GGML_CUDA_CC_OFFSET_AMD + 0x942) // MI300 #define GGML_CUDA_CC_CDNA3 (GGML_CUDA_CC_OFFSET_AMD + 0x942) // MI300
@@ -89,12 +91,13 @@
#define GGML_CUDA_CC_IS_RDNA3_5(cc) (cc >= GGML_CUDA_CC_RDNA3_5 && cc < GGML_CUDA_CC_RDNA4) #define GGML_CUDA_CC_IS_RDNA3_5(cc) (cc >= GGML_CUDA_CC_RDNA3_5 && cc < GGML_CUDA_CC_RDNA4)
#define GGML_CUDA_CC_IS_RDNA3(cc) (GGML_CUDA_CC_IS_RDNA3_0(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc)) #define GGML_CUDA_CC_IS_RDNA3(cc) (GGML_CUDA_CC_IS_RDNA3_0(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc))
#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4) #define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
#define GGML_CUDA_CC_IS_GCN(cc) (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1) #define GGML_CUDA_CC_IS_GCN_APU(cc) ((cc) == GGML_CUDA_CC_GFX909 || (cc) == GGML_CUDA_CC_GFX90C)
#define GGML_CUDA_CC_IS_CDNA(cc) (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1) #define GGML_CUDA_CC_IS_GCN(cc) ((cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1) || GGML_CUDA_CC_IS_GCN_APU(cc))
#define GGML_CUDA_CC_IS_CDNA1(cc) (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2) #define GGML_CUDA_CC_IS_CDNA(cc) (!GGML_CUDA_CC_IS_GCN_APU(cc) && cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)
#define GGML_CUDA_CC_IS_CDNA2(cc) (cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3) #define GGML_CUDA_CC_IS_CDNA1(cc) (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2)
#define GGML_CUDA_CC_IS_CDNA3(cc) (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_CDNA4) #define GGML_CUDA_CC_IS_CDNA2(cc) (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)
#define GGML_CUDA_CC_IS_CDNA4(cc) (cc >= GGML_CUDA_CC_CDNA4 && cc < GGML_CUDA_CC_RDNA1) #define GGML_CUDA_CC_IS_CDNA3(cc) (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_CDNA4)
#define GGML_CUDA_CC_IS_CDNA4(cc) (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA4 && cc < GGML_CUDA_CC_RDNA1)
// Moore Threads // Moore Threads
#define MUSART_HMASK 40300 // MUSA rc4.3, min. ver. for half2 -> uint mask comparisons #define MUSART_HMASK 40300 // MUSA rc4.3, min. ver. for half2 -> uint mask comparisons
ggml/src/ggml-cuda/ggml-cuda.cu
+1
View File
@@ -212,6 +212,7 @@ static int ggml_cuda_parse_id(char devName[]) {
} }
archNum += archMajor * 0x100; archNum += archMajor * 0x100;
archNum += archMinor; archNum += archMinor;
return archNum; return archNum;
} }
#endif // defined(GGML_USE_HIP) #endif // defined(GGML_USE_HIP)
ggml/src/ggml-cuda/mmq.cu
+2 -2
View File
@@ -375,10 +375,10 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
return true; return true;
} }
// gfx900 (Vega 10) lacks native dp4a, loses to dequant + hipBLAS // gfx900 (Vega 10), gfx909, and gfx90c lack native dp4a, losing to dequant + hipBLAS
// for dense matrices; keep MMQ only for MoE, where the // for dense matrices; keep MMQ only for MoE, where the
// hipBLAS path is much slower. // hipBLAS path is much slower.
if (cc == GGML_CUDA_CC_VEGA) { if (cc == GGML_CUDA_CC_VEGA || GGML_CUDA_CC_IS_GCN_APU(cc)) {
return n_experts > 0; return n_experts > 0;
} }
ggml/src/ggml-cuda/vendors/hip.h
Vendored
+2 -2
View File
@@ -176,9 +176,9 @@
#define __CUDA_ARCH__ 1300 #define __CUDA_ARCH__ 1300
#if defined(__gfx900__) || defined(__gfx906__) #if defined(__gfx900__) || defined(__gfx906__) || defined(__gfx909__) || defined(__gfx90c__)
#define GCN5 #define GCN5
#endif // defined(__gfx900__) || defined(__gfx906__) #endif // defined(__gfx900__) || defined(__gfx906__) || defined(__gfx909__) || defined(__gfx90c__)
#if defined(__gfx803__) #if defined(__gfx803__)
#define GCN4 #define GCN4