CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-03 09:22:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

HIP: Add support for RDNA4 targets (#12372)

Browse Source
This commit is contained in:
Slobodan Josic
2025-03-26 23:46:30 +01:00
committed by GitHub
parent b3298fa47a
commit bd40678df7
7 changed files with 25 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-cuda/common.cuh
View File

@@ -52,7 +52,7 @@
#define GGML_CUDA_CC_IS_NVIDIA(cc) (cc < GGML_CUDA_CC_OFFSET_MTHREADS)
// AMD
// GCN/CNDA, wave size is 64
// GCN/CDNA, wave size is 64
#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_VEGA20 (GGML_CUDA_CC_OFFSET_AMD + 0x906) // MI50/Radeon VII, minimum for dp4a
@@ -60,16 +60,18 @@
#define GGML_CUDA_CC_CDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x910) // MI210, minimum acc register renameing
#define GGML_CUDA_CC_CDNA3 (GGML_CUDA_CC_OFFSET_AMD + 0x942) // MI300
// RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32
// RDNA removes MFMA, dp4a, xnack, acc registers, wave size is 32
#define GGML_CUDA_CC_RDNA1 (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
#define GGML_CUDA_CC_RDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
#define GGML_CUDA_CC_RDNA3 (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
#define GGML_CUDA_CC_RDNA4 (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000
#define GGML_CUDA_CC_IS_AMD(cc) (cc >= GGML_CUDA_CC_OFFSET_AMD)
#define GGML_CUDA_CC_IS_RDNA(cc) (cc >= GGML_CUDA_CC_RDNA1)
#define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
#define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3)
#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && 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_CDNA)
#define GGML_CUDA_CC_IS_CDNA(cc) (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
@@ -209,9 +211,9 @@ typedef float2 dfloat2;
#define FP16_MMA_AVAILABLE
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3))
#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
#define FP16_MMA_AVAILABLE
#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3))
#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
#define NEW_MMA_AVAILABLE
@@ -244,14 +246,14 @@ static bool fp16_mma_available(const int cc) {
return false;
#else
return (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc);
GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
}
// To be used for feature selection of external libraries, e.g. cuBLAS.
static bool fp16_mma_hardware_available(const int cc) {
return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) ||
GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc);
GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
}
// Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
@@ -409,7 +411,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#if defined(CDNA) || defined(RDNA2) || defined(__gfx906__)
c = __builtin_amdgcn_sdot4(a, b, c, false);
#elif defined(RDNA3)
#elif defined(RDNA3) || defined(RDNA4)
c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
#elif defined(RDNA1) || defined(__gfx900__)
int tmp1;