LLM inference in C/C++
C++ ggml
ggml: add gfx90c HIP support (#26454)
AuroraRAS4735997382b5
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuhindex 7fb04c95f9..7d14ce9067 100644--- a/ggml/src/ggml-cuda/common.cuh+++ b/ggml/src/ggml-cuda/common.cuh@@ -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_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_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_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@@ -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(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_GCN(cc) (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1)-#define GGML_CUDA_CC_IS_CDNA(cc) (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)-#define GGML_CUDA_CC_IS_CDNA1(cc) (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2)-#define GGML_CUDA_CC_IS_CDNA2(cc) (cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)-#define GGML_CUDA_CC_IS_CDNA3(cc) (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_CDNA4)-#define GGML_CUDA_CC_IS_CDNA4(cc) (cc >= GGML_CUDA_CC_CDNA4 && cc < GGML_CUDA_CC_RDNA1)+#define GGML_CUDA_CC_IS_GCN_APU(cc) ((cc) == GGML_CUDA_CC_GFX909 || (cc) == GGML_CUDA_CC_GFX90C)+#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_CDNA(cc) (!GGML_CUDA_CC_IS_GCN_APU(cc) && cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)+#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_CDNA2(cc) (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)+#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#define MUSART_HMASK 40300 // MUSA rc4.3, min. ver. for half2 -> uint mask comparisonsdiff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cuindex 45e9537f0e..0e6601f034 100644--- a/ggml/src/ggml-cuda/ggml-cuda.cu+++ b/ggml/src/ggml-cuda/ggml-cuda.cu@@ -212,6 +212,7 @@ static int ggml_cuda_parse_id(char devName[]) {}archNum += archMajor * 0x100;archNum += archMinor;+return archNum;}#endif // defined(GGML_USE_HIP)diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cuindex 7fb4401489..9beff0d9b7 100644--- a/ggml/src/ggml-cuda/mmq.cu+++ b/ggml/src/ggml-cuda/mmq.cu@@ -375,10 +375,10 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_treturn 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// 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;}diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.hindex 9aa558f3f4..2fc0fe9fdb 100644--- a/ggml/src/ggml-cuda/vendors/hip.h+++ b/ggml/src/ggml-cuda/vendors/hip.h@@ -176,9 +176,9 @@#define __CUDA_ARCH__ 1300-#if defined(__gfx900__) || defined(__gfx906__)+#if defined(__gfx900__) || defined(__gfx906__) || defined(__gfx909__) || defined(__gfx90c__)#define GCN5-#endif // defined(__gfx900__) || defined(__gfx906__)+#endif // defined(__gfx900__) || defined(__gfx906__) || defined(__gfx909__) || defined(__gfx90c__)#if defined(__gfx803__)#define GCN4