Skip to content

Commit 3553833

Browse files
committed
Revert "CUDA: refactor and deduplicate vector FA kernels (ggml-org#16208)"
This reverts commit 75a3a6c.
1 parent 4dfca5a commit 3553833

File tree

129 files changed

+1902
-1307
lines changed

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

129 files changed

+1902
-1307
lines changed

ggml/src/ggml-cuda/common.cuh

Lines changed: 9 additions & 34 deletions
Original file line numberDiff line numberDiff line change
@@ -592,42 +592,17 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v,
592592
#endif // defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(GCN5) || defined(CDNA))
593593
}
594594

595-
static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v, const half2 u) {
596-
#ifdef FAST_FP16_AVAILABLE
597-
acc += v*u;
598-
#else
599-
const float2 tmpv = __half22float2(v);
600-
const float2 tmpu = __half22float2(u);
601-
float2 tmpacc = __half22float2(acc);
602-
tmpacc.x += tmpv.x * tmpu.x;
603-
tmpacc.y += tmpv.y * tmpu.y;
604-
acc = make_half2(tmpacc.x, tmpacc.y);
605-
#endif // FAST_FP16_AVAILABLE
606-
}
607-
608595
// Aligned memory transfers of 8/16 bytes can be faster than 2 transfers with 4 bytes, especially on AMD.
609-
template <int nbytes, int alignment = 0>
596+
template <int nbytes>
610597
static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) {
611-
if constexpr (alignment != 0) {
612-
static_assert(nbytes % alignment == 0, "bad alignment");
613-
}
614-
constexpr int nb_per_cpy = alignment == 0 ? nbytes : alignment;
615-
616-
#pragma unroll
617-
for (int i = 0; i < nbytes/nb_per_cpy; ++i) {
618-
if constexpr (nb_per_cpy == 1) {
619-
((char *) dst)[i] = ((const char *) src)[i];
620-
} else if constexpr (nb_per_cpy == 2) {
621-
((short *) dst)[i] = ((const short *) src)[i];
622-
} else if constexpr (nb_per_cpy == 4) {
623-
((int *) dst)[i] = ((const int *) src)[i];
624-
} else if constexpr (nb_per_cpy == 8) {
625-
((int2 *) dst)[i] = ((const int2 *) src)[i];
626-
} else if constexpr (nb_per_cpy == 16) {
627-
((int4 *) dst)[i] = ((const int4 *) src)[i];
628-
} else {
629-
static_assert(nbytes == 0 && nbytes == -1, "bad nbytes");
630-
}
598+
if constexpr (nbytes == 4) {
599+
*(int *) dst = *(const int *) src;
600+
} else if constexpr (nbytes == 8) {
601+
*(int2 *) dst = *(const int2 *) src;
602+
} else if constexpr (nbytes == 16) {
603+
*(int4 *) dst = *(const int4 *) src;
604+
} else {
605+
static_assert(nbytes == 0 && nbytes == -1, "bad nbytes");
631606
}
632607
}
633608

0 commit comments

Comments
 (0)