Skip to content

CUDA: Volta tensor core support for MMF #16843

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Oct 31, 2025
Merged

CUDA: Volta tensor core support for MMF #16843

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
5d14386
CUDA: Volta tensor core support for MMF
JohannesGaessler Oct 28, 2025
7efb6ac
more generic checks for hardware support
JohannesGaessler Oct 30, 2025
53badac
Update ggml/src/ggml-cuda/mmf.cuh
JohannesGaessler Oct 31, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 9 additions & 1 deletion ggml/src/ggml-cuda/common.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -224,6 +224,11 @@ static const char * cu_get_error_str(CUresult err) {
#define AMD_MFMA_AVAILABLE
#endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)

// The Volta instructions are in principle available on Turing or newer but they are effectively unusable:
#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
#define VOLTA_MMA_AVAILABLE
#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA

#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
#define TURING_MMA_AVAILABLE
#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
Expand Down Expand Up @@ -278,7 +283,10 @@ static bool amd_mfma_available(const int cc) {
#endif //!defined(GGML_HIP_NO_MMQ_MFMA)
}

// Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
static bool volta_mma_available(const int cc) {
return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_VOLTA;
}

static bool turing_mma_available(const int cc) {
return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
}
Expand Down
237 changes: 213 additions & 24 deletions ggml/src/ggml-cuda/mma.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,10 @@

#include "common.cuh"

// On Volta each warp is doing 4 8x8 mma operations in parallel.
// The basic memory layout for a 32x8 output tile is to stack 4 input tiles in I direction and to mirror the B tile.
// However, the i indices in this file are by default permuted to simplify the index calculations.
// #define GGML_CUDA_MMA_NO_VOLTA_PERM

#if CUDART_VERSION >= 11080

Expand Down Expand Up @@ -73,6 +77,15 @@ namespace ggml_cuda_mma {
static constexpr int ne = I * J / 64;
T x[ne] = {0};

static constexpr __device__ bool supported() {
if (I == 64 && J == 2) return true;
if (I == 16 && J == 8) return true;
if (I == 32 && J == 4) return true;
if (I == 16 && J == 16) return true;
if (I == 32 && J == 32) return true;
return false;
}

static __device__ __forceinline__ int get_i(const int l) {
if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
return threadIdx.x % 16;
Expand All @@ -85,7 +98,8 @@ namespace ggml_cuda_mma {
} else if constexpr (I == 32 && J == 32) {
return 4 * (threadIdx.x / 32) + 8 * (l / 4) + (l % 4);
} else {
static_assert(I == -1 && J == -1, "template specialization not implemented");
NO_DEVICE_CODE;
return -1;
}
}

Expand All @@ -101,36 +115,84 @@ namespace ggml_cuda_mma {
} else if constexpr (I == 32 && J == 32) {
return threadIdx.x % 32;
} else {
static_assert(I == -1 && J == -1, "template specialization not implemented");
NO_DEVICE_CODE;
return -1;
}
}
#elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
static constexpr int ne = I * J / 32;
T x[ne] = {0};

static constexpr __device__ bool supported() {
if (I == 32 && J == 8) return true;
return false;
}

static __device__ __forceinline__ int get_i(const int l) {
if constexpr (I == 32 && J == 8) {
#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (l & 2) | (threadIdx.x % 2);
#else
return (l & 2) | (threadIdx.x & ~2);
#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
} else {
NO_DEVICE_CODE;
return -1;
}
}

static __device__ __forceinline__ int get_j(const int l) {
if constexpr (I == 32 && J == 8) {
return (threadIdx.x & 2) | (l & (4 + 1));
} else {
NO_DEVICE_CODE;
return -1;
}
}
#else
static constexpr int ne = I * J / 32;
T x[ne] = {0};

static constexpr __device__ bool supported() {
if (I == 8 && J == 4) return true;
if (I == 8 && J == 8) return true;
if (I == 16 && J == 8) return true;
if (I == 16 && J == 16) return true;
if (I == 32 && J == 8) return true;
return false;
}

static __device__ __forceinline__ int get_i(const int l) {
if constexpr (I == 8 && (J == 4 || J == 8)) {
if constexpr (I == 8 && J == 4) {
return threadIdx.x / 4;
} else if constexpr (I == 8 && J == 8) {
return threadIdx.x / 4;
} else if constexpr (I == 16 && J == 8) {
return (l / 2) * 8 + threadIdx.x / 4;
return ((l / 2) * 8) | (threadIdx.x / 4);
} else if constexpr (I == 16 && J == 16) {
return ((l / 2) % 2) * 8 + threadIdx.x / 4;
return (((l / 2) % 2) * 8) | (threadIdx.x / 4);
} else if constexpr (I == 32 && J == 8) {
return tile<16, 8, T>::get_i(l); // Memory layout simply repeated with same pattern in i direction.
} else {
static_assert(I == -1 && J == -1, "template specialization not implemented");
NO_DEVICE_CODE;
return -1;
}
}

static __device__ __forceinline__ int get_j(const int l) {
if constexpr (I == 8 && J == 4) {
return threadIdx.x % 4;
} else if constexpr (I == 8 && J == 8) {
return 4 * l + threadIdx.x % 4;
return (l * 4) | (threadIdx.x % 4);
} else if constexpr (I == 16 && J == 8) {
return 2 * (threadIdx.x % 4) + l % 2;
return ((threadIdx.x % 4) * 2) | (l % 2);
} else if constexpr (I == 16 && J == 16) {
return 8 * (l / 4) + 2 * (threadIdx.x % 4) + l % 2;
return ((l / 4) * 8) | ((threadIdx.x % 4) * 2) | (l % 2);
} else if constexpr (I == 32 && J == 8) {
return tile<16, 8, T>::get_j(l); // Memory layout simply repeated with same pattern in i direction.
} else {
static_assert(I == -1 && J == -1, "template specialization not implemented");
NO_DEVICE_CODE;
return -1;
}
}
#endif // defined(GGML_USE_HIP)
Expand All @@ -140,32 +202,83 @@ namespace ggml_cuda_mma {
struct tile<I_, J_, half2> {
static constexpr int I = I_;
static constexpr int J = J_;

#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
static constexpr int ne = I == 8 && J == 8 ? I * J / (WARP_SIZE/4) : I * J / WARP_SIZE;
half2 x[ne] = {{0.0f, 0.0f}};

static constexpr __device__ bool supported() {
if (I == 8 && J == 8) return true;
if (I == 32 && J == 8) return true;
return false;
}

static __device__ __forceinline__ int get_i(const int l) {
if constexpr (I == 8 && J == 8) {
return ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
} else if constexpr (I == 32 && J == 8) {
#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
#else
return threadIdx.x;
#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
} else {
NO_DEVICE_CODE;
return -1;
}
}

static __device__ __forceinline__ int get_j(const int l) {
if constexpr ((I == 8 || I == 32) && J == 8) {
return l;
} else {
NO_DEVICE_CODE;
return -1;
}
}
#else
static constexpr int ne = I * J / WARP_SIZE;
half2 x[ne] = {{0.0f, 0.0f}};

static constexpr __device__ bool supported() {
if (I == 8 && J == 4) return true;
if (I == 8 && J == 8) return true;
if (I == 16 && J == 8) return true;
if (I == 16 && J == 16) return true;
if (I == 32 && J == 8) return true;
return false;
}

static __device__ __forceinline__ int get_i(const int l) {
if constexpr (I == 8 && J == 8) {
return threadIdx.x / 4;
} else if constexpr (I == 16 && J == 4) {
return l * 8 + threadIdx.x / 4;
return (l * 8) | (threadIdx.x / 4);
} else if constexpr (I == 16 && J == 8) {
return (l % 2) * 8 + threadIdx.x / 4;
return ((l % 2) * 8) | (threadIdx.x / 4);
} else if constexpr (I == 32 && J == 8) {
return ((l / 4) * 16) | ((l % 2) * 8) | (threadIdx.x / 4);
} else {
static_assert(I == -1 && J == -1, "template specialization not implemented");
NO_DEVICE_CODE;
return -1;
}
}

static __device__ __forceinline__ int get_j(const int l) {
if constexpr (I == 8 && J == 8) {
return l * 4 + threadIdx.x % 4;
return (l * 4) | (threadIdx.x % 4);
} else if constexpr (I == 16 && J == 4) {
return threadIdx.x % 4;
} else if constexpr (I == 16 && J == 8) {
return (l / 2) * 4 + threadIdx.x % 4;
return ((l / 2) * 4) | (threadIdx.x % 4);
} else if constexpr (I == 32 && J == 8) {
return ((l & 2) * 2) | (threadIdx.x % 4);
} else {
static_assert(I == -1 && J == -1, "template specialization not implemented");
NO_DEVICE_CODE;
return -1;
}
}
#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
};

template <int I_, int J_>
Expand All @@ -175,27 +288,36 @@ namespace ggml_cuda_mma {
static constexpr int ne = I * J / WARP_SIZE;
nv_bfloat162 x[ne] = {{0.0f, 0.0f}};

static constexpr __device__ bool supported() {
if (I == 8 && J == 8) return true;
if (I == 16 && J == 4) return true;
if (I == 16 && J == 8) return true;
return false;
}

static __device__ __forceinline__ int get_i(const int l) {
if constexpr (I == 8 && J == 8) {
return threadIdx.x / 4;
} else if constexpr (I == 16 && J == 4) {
return l * 8 + threadIdx.x / 4;
return (l * 8) | (threadIdx.x / 4);
Copy link
Collaborator

@am17an am17an Oct 31, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is this more performant? I find the earlier version easier to read

Copy link
Collaborator Author

@JohannesGaessler JohannesGaessler Oct 31, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's a good thing you're asking that because I misremembered the table from the CUDA documentation showing instruction throughput. The way I remembered it integer additions and binary operations had the same throughput but on a silicon level you would have lower power draw. In actually though the throughput of additions is twice that of binary operations.

} else if constexpr (I == 16 && J == 8) {
return (l % 2) * 8 + threadIdx.x / 4;
return ((l % 2) * 8) | (threadIdx.x / 4);
} else {
static_assert(I == -1 && J == -1, "template specialization not implemented");
NO_DEVICE_CODE;
return -1;
}
}

static __device__ __forceinline__ int get_j(const int l) {
if constexpr (I == 8 && J == 8) {
return l * 4 + threadIdx.x % 4;
return (l * 4) | (threadIdx.x % 4);
} else if constexpr (I == 16 && J == 4) {
return threadIdx.x % 4;
} else if constexpr (I == 16 && J == 8) {
return (l / 2) * 4 + threadIdx.x % 4;
return ((l / 2) * 4) | (threadIdx.x % 4);
} else {
static_assert(I == -1 && J == -1, "template specialization not implemented");
NO_DEVICE_CODE;
return -1;
}
}
};
Expand Down Expand Up @@ -263,8 +385,12 @@ namespace ggml_cuda_mma {
: "=r"(xi[0]), "=r"(xi[1])
: "l"(xs));
#else
load_generic(xs0, stride);
GGML_UNUSED(t);
#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
GGML_UNUSED_VARS(t, xs0, stride);
NO_DEVICE_CODE;
#else
load_generic(t, xs0, stride);
#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
#endif // TURING_MMA_AVAILABLE
}

Expand All @@ -277,11 +403,35 @@ namespace ggml_cuda_mma {
asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
: "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
: "l"(xs));
#else
#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
GGML_UNUSED_VARS(t, xs0, stride);
NO_DEVICE_CODE;
#else
load_generic(t, xs0, stride);
#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
#endif // TURING_MMA_AVAILABLE
}

template <typename T>
static __device__ __forceinline__ void load_ldmatrix(
tile<32, 8, T> & t, const T * __restrict__ xs0, const int stride) {
#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
#if 1
// TODO: more generic handling
static_assert(sizeof(T) == 4, "bad type size");
ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 4, xs0 + t.get_i(4)*stride + 4);
#else
load_generic(t, xs0, stride);
#endif // 1
#else
tile<16, 8, T> * t16 = (tile<16, 8, T> *) &t;
load_ldmatrix(t16[0], xs0 + 0*stride, stride);
load_ldmatrix(t16[1], xs0 + 16*stride, stride);
#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
}

template <typename T>
static __device__ __forceinline__ void load_ldmatrix_trans(
tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
Expand Down Expand Up @@ -546,4 +696,43 @@ namespace ggml_cuda_mma {
NO_DEVICE_CODE;
#endif // AMD_MFMA_AVAILABLE
}

template <typename T1, typename T2, int J, int K>
static __device__ __forceinline__ void mma(
tile<32, J, T1> & D, const tile<32, K, T2> & A, const tile<J, K, T2> & B) {
tile<16, J, T1> * D16 = (tile<16, J, T1> *) &D;
tile<16, K, T2> * A16 = (tile<16, K, T2> *) &A;
mma(D16[0], A16[0], B);
mma(D16[1], A16[1], B);
}

static __device__ __forceinline__ void mma(
tile<32, 8, float> & D, const tile<32, 8, half2> & A, const tile<8, 8, half2> & B) {
#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
const int * Axi = (const int *) A.x;
const int * Bxi = (const int *) B.x;
int * Dxi = (int *) D.x;
asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
"{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
: "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
: "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0]), "r"(Bxi[1]));
asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
"{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
: "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
: "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]), "r"(Bxi[3]));
asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
"{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
: "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
: "r"(Axi[4]), "r"(Axi[5]), "r"(Bxi[4]), "r"(Bxi[5]));
asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
"{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
: "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
: "r"(Axi[6]), "r"(Axi[7]), "r"(Bxi[6]), "r"(Bxi[7]));
#else
tile<16, 8, float> * D16 = (tile<16, 8, float> *) &D;
tile<16, 8, half2> * A16 = (tile<16, 8, half2> *) &A;
mma(D16[0], A16[0], B);
mma(D16[1], A16[1], B);
#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
}
}
2 changes: 1 addition & 1 deletion ggml/src/ggml-cuda/mmf.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -148,7 +148,7 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
case GGML_TYPE_F32:
return ampere_mma_available(cc);
case GGML_TYPE_F16:
return turing_mma_available(cc);
return volta_mma_available(cc) || turing_mma_available(cc);
case GGML_TYPE_BF16:
return ampere_mma_available(cc);
default:
Expand Down
Loading
Loading