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

HIP: RDNA4 tensor core support for MMF (#17077)

Browse Source 
* mmf for rdna4

* align the padding for rdna4

* forbit mul_mat_f for rdna4

* fix as comment

* remove device kernels

* add constexpr for early return

* update based on review comment

* change based on the review comment

* pass compile error

* keep code consistency

---------

Co-authored-by: zhang hui <you@example.com>
This commit is contained in:
yulo
2025-11-22 00:03:24 +01:00
committed by GitHub
co-authored by zhang hui
parent 8e9ddba610
commit 028f93ef98
5 changed files with 180 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/mma.cuh
+107
View File
@@ -74,6 +74,33 @@ namespace ggml_cuda_mma {
static constexpr int J = J_;
#if defined(GGML_USE_HIP)
#if defined(RDNA4)
static constexpr int ne = I * J / 32;
T x[ne] = {0};
static constexpr __device__ bool supported() {
if (I == 16 && J == 16) return true;
return false;
}
static __device__ __forceinline__ int get_i(const int l) {
if constexpr (I == 16 && J == 16) {
return 8 * (threadIdx.x / 16) + l;
} else {
NO_DEVICE_CODE;
return -1;
}
}
static __device__ __forceinline__ int get_j(const int l) {
if constexpr (I == 16 && J == 16) {
return threadIdx.x % 16;
} else {
NO_DEVICE_CODE;
return -1;
}
}
#else
static constexpr int ne = I * J / 64;
T x[ne] = {0};
@@ -119,6 +146,7 @@ namespace ggml_cuda_mma {
return -1;
}
}
#endif // defined(RDNA4)
#elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
static constexpr int ne = I * J / 32;
T x[ne] = {0};
@@ -236,6 +264,32 @@ namespace ggml_cuda_mma {
return -1;
}
}
#elif defined(AMD_WMMA_AVAILABLE)
static constexpr int ne = I * J / 32;
half2 x[ne] = {{0.0f, 0.0f}};
static constexpr __device__ bool supported() {
if (I == 16 && J == 8) return true;
return false;
}
static __device__ __forceinline__ int get_i(const int l) {
if constexpr (I == 16 && J == 8) {
return threadIdx.x % 16;
} else {
NO_DEVICE_CODE;
return -1;
}
}
static __device__ __forceinline__ int get_j(const int l) {
if constexpr (I == 16 && J == 8) {
return 4 * (threadIdx.x / 16) + l;
} else {
NO_DEVICE_CODE;
return -1;
}
}
#else
static constexpr int ne = I * J / WARP_SIZE;
half2 x[ne] = {{0.0f, 0.0f}};
@@ -285,6 +339,34 @@ namespace ggml_cuda_mma {
struct tile<I_, J_, nv_bfloat162> {
static constexpr int I = I_;
static constexpr int J = J_;
#if defined(AMD_WMMA_AVAILABLE)
static constexpr int ne = I * J / 32;
nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
static constexpr __device__ bool supported() {
if (I == 16 && J == 8) return true;
return false;
}
static __device__ __forceinline__ int get_i(const int l) {
if constexpr (I == 16 && J == 8) {
return threadIdx.x % 16;
} else {
NO_DEVICE_CODE;
return -1;
}
}
static __device__ __forceinline__ int get_j(const int l) {
if constexpr (I == 16 && J == 8) {
return 4 * (threadIdx.x / 16) + l;
} else {
NO_DEVICE_CODE;
return -1;
}
}
#else
static constexpr int ne = I * J / WARP_SIZE;
nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
@@ -320,6 +402,7 @@ namespace ggml_cuda_mma {
return -1;
}
}
#endif // defined(AMD_WMMA_AVAILABLE)
};
template <int I, int J>
@@ -353,6 +436,8 @@ namespace ggml_cuda_mma {
const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 2 * (threadIdx.x / t.I));
xi[0] = xs[0];
}
#elif defined(AMD_WMMA_AVAILABLE)
ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
#else
#pragma unroll
for (int l = 0; l < t.ne; ++l) {
@@ -639,12 +724,34 @@ namespace ggml_cuda_mma {
: "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
: "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
#elif defined(AMD_WMMA_AVAILABLE)
using halfx8_t = __attribute__((ext_vector_type(8))) _Float16;
using floatx8_t = __attribute__((ext_vector_type(8))) float;
floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
const halfx8_t& a_frag = reinterpret_cast<const halfx8_t&>(A.x[0]);
const halfx8_t& b_frag = reinterpret_cast<const halfx8_t&>(B.x[0]);
acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32_gfx12(a_frag, b_frag, acc_frag);
#else
GGML_UNUSED_VARS(D, A, B);
NO_DEVICE_CODE;
#endif // TURING_MMA_AVAILABLE
}
static __device__ __forceinline__ void mma(
tile<16, 16, float> & D, const tile<16, 8, nv_bfloat162> & A, const tile<16, 8, nv_bfloat162> & B) {
#if defined(AMD_WMMA_AVAILABLE)
using bf16x8_t = __attribute__((ext_vector_type(8))) __bf16;
using floatx8_t = __attribute__((ext_vector_type(8))) float;
floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
const bf16x8_t& a_frag = reinterpret_cast<const bf16x8_t&>(A.x[0]);
const bf16x8_t& b_frag = reinterpret_cast<const bf16x8_t&>(B.x[0]);
acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12(a_frag, b_frag, acc_frag);
#else
GGML_UNUSED_VARS(D, A, B);
NO_DEVICE_CODE;
#endif // AMPERE_MMA_AVAILABLE
}
static __device__ __forceinline__ void mma(
tile<16, 16, int> & D, const tile<16, 8, int> & A, const tile<16, 8, int> & B) {
#if defined(AMD_MFMA_AVAILABLE)