Skip to content

Commit 0e6ff00

Browse files
JohannesGaesslerJohannesGaessler
authored
CUDA: larger SRAM reads for tile FA, AMD FP16 dot (ggml-org#15927)
* CUDA: larger SRAM reads for tile FA, AMD FP16 dot * fix logic for availability of v_dot2_f32_f16
1 parent df082f5 commit 0e6ff00

File tree

3 files changed

+124
-33
lines changed

3 files changed

+124
-33
lines changed

‎ggml/src/ggml-cuda/common.cuh‎

Lines changed: 16 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -555,7 +555,7 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float2 v
555555
}
556556

557557
static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
558-
#if defined(GGML_USE_HIP) && defined(GCN)
558+
#if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
559559
asm volatile("v_dot2_f32_f16 %0, %1, %2, %0" : "+v"(acc) : "v"(v), "v"(u));
560560
#else
561561
#ifdef FAST_FP16_AVAILABLE
@@ -567,7 +567,21 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v,
567567
acc += tmpv.x * tmpu.x;
568568
acc += tmpv.y * tmpu.y;
569569
#endif // FAST_FP16_AVAILABLE
570-
#endif // defined(GGML_USE_HIP) && defined(GCN)
570+
#endif // defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(GCN5) || defined(CDNA))
571+
}
572+
573+
// Aligned memory transfers of 8/16 bytes can be faster than 2 transfers with 4 bytes, especially on AMD.
574+
template <int nbytes>
575+
static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) {
576+
if constexpr (nbytes == 4) {
577+
*(int *) dst = *(const int *) src;
578+
} else if constexpr (nbytes == 8) {
579+
*(int2 *) dst = *(const int2 *) src;
580+
} else if constexpr (nbytes == 16) {
581+
*(int4 *) dst = *(const int4 *) src;
582+
} else {
583+
static_assert(nbytes == 0 && nbytes == -1, "bad nbytes");
584+
}
571585
}
572586

573587
static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {

‎ggml/src/ggml-cuda/fattn-tile.cu‎

Lines changed: 100 additions & 31 deletions
Original file line numberDiff line numberDiff line change
@@ -8,11 +8,14 @@ static int fattn_tile_get_kq_stride_host(const int D, const int ncols, const int
88
if (GGML_CUDA_CC_IS_AMD(cc)) {
99
switch (D) {
1010
case 64:
11-
return ncols <= 16 ? 32 : 64;
11+
return 64;
1212
case 128:
13-
return ncols <= 16 ? 64 : warp_size;
1413
case 256:
15-
return 64;
14+
if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) {
15+
return ncols <= 16 ? 64 : 32;
16+
} else {
17+
return 64;
18+
}
1619
default:
1720
GGML_ABORT("fatal error");
1821
return -1;
@@ -41,17 +44,26 @@ static int fattn_tile_get_kq_stride_host(const int D, const int ncols, const int
4144
GGML_ABORT("fatal error");
4245
return -1;
4346
}
47+
GGML_UNUSED(warp_size);
4448
}
4549

4650
static constexpr __device__ int fattn_tile_get_kq_stride_device(int D, int ncols, int warp_size) {
4751
#ifdef GGML_USE_HIP
4852
switch (D) {
4953
case 64:
50-
return ncols <= 16 ? 32 : 64;
54+
return 64;
5155
case 128:
52-
return ncols <= 16 ? 64 : warp_size;
56+
#if defined(GCN) || defined(CDNA)
57+
return ncols <= 16 ? 64 : 32;
58+
#else
59+
return 64;
60+
#endif // defined(GCN) || defined(CDNA)
5361
case 256:
62+
#if defined(GCN) || defined(CDNA)
63+
return ncols <= 16 ? 64 : 32;
64+
#else
5465
return 64;
66+
#endif // defined(GCN) || defined(CDNA)
5567
default:
5668
return -1;
5769
}
@@ -88,9 +100,17 @@ static constexpr __device__ int fattn_tile_get_kq_nbatch_device(int D, int ncols
88100
case 64:
89101
return 64;
90102
case 128:
91-
return ncols <= 16 ? 2*warp_size : 128;
103+
#if defined(GCN) || defined(CDNA)
104+
return ncols <= 16 ? 64 : 128;
105+
#else
106+
return 64;
107+
#endif // defined(GCN) || defined(CDNA)
92108
case 256:
93-
return ncols <= 16 ? 128 : 2*warp_size;
109+
#if defined(GCN) || defined(CDNA)
110+
return ncols <= 16 ? 64 : 128;
111+
#else
112+
return ncols <= 16 ? 64 : 256;
113+
#endif // defined(GCN) || defined(CDNA)
94114
default:
95115
return -1;
96116
}
@@ -196,14 +216,21 @@ static __global__ void flash_attn_tile(
196216

197217
const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
198218

219+
#if defined(GGML_USE_HIP)
220+
constexpr int cpy_nb = 16;
221+
#else
222+
constexpr int cpy_nb = 8;
223+
#endif // defined(GGML_USE_HIP) && defined(GCN)
224+
constexpr int cpy_ne = cpy_nb / 4;
225+
199226
__shared__ float KQ[ncols][kq_stride];
200227
#ifdef FAST_FP16_AVAILABLE
201228
__shared__ half2 Q_tmp[ncols][D/2];
202-
__shared__ half2 KV_tmp_h2[kq_stride * (kq_nbatch/2 + 1)]; // Padded to avoid memory bank conflicts.
229+
__shared__ half2 KV_tmp_h2[kq_stride * (kq_nbatch/2 + cpy_ne)]; // Padded to avoid memory bank conflicts.
203230
half2 VKQ[ncols/nwarps][D/(2*warp_size)] = {{{0.0f, 0.0f}}};
204231
#else
205232
__shared__ float Q_tmp[ncols][D];
206-
__shared__ float KV_tmp_f[kq_stride * (kq_nbatch + 1)]; // Padded to avoid memory bank conflicts.
233+
__shared__ float KV_tmp_f[kq_stride * (kq_nbatch + cpy_ne)]; // Padded to avoid memory bank conflicts.
207234
float2 * KV_tmp_f2 = (float2 *) KV_tmp_f;
208235
float2 VKQ[ncols/nwarps][D/(2*warp_size)] = {{{0.0f, 0.0f}}};
209236
#endif // FAST_FP16_AVAILABLE
@@ -256,11 +283,11 @@ static __global__ void flash_attn_tile(
256283
for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch/2; k_KQ_1 += warp_size) {
257284
const half2 tmp_h2 = K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + k_KQ_1 + threadIdx.x];
258285
#ifdef FAST_FP16_AVAILABLE
259-
KV_tmp_h2[i_KQ*(kq_nbatch/2 + 1) + k_KQ_1 + threadIdx.x] = tmp_h2;
286+
KV_tmp_h2[i_KQ*(kq_nbatch/2 + cpy_ne) + k_KQ_1 + threadIdx.x] = tmp_h2;
260287
#else
261288
const float2 tmp_f2 = __half22float2(tmp_h2);
262-
KV_tmp_f[i_KQ*(kq_nbatch + 1) + 2*k_KQ_1 + threadIdx.x] = tmp_f2.x;
263-
KV_tmp_f[i_KQ*(kq_nbatch + 1) + 2*k_KQ_1 + warp_size + threadIdx.x] = tmp_f2.y;
289+
KV_tmp_f[i_KQ*(kq_nbatch + cpy_ne) + 2*k_KQ_1 + threadIdx.x] = tmp_f2.x;
290+
KV_tmp_f[i_KQ*(kq_nbatch + cpy_ne) + 2*k_KQ_1 + warp_size + threadIdx.x] = tmp_f2.y;
264291
#endif // FAST_FP16_AVAILABLE
265292
}
266293
}
@@ -269,42 +296,45 @@ static __global__ void flash_attn_tile(
269296

270297
#ifdef FAST_FP16_AVAILABLE
271298
#pragma unroll
272-
for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch/2; ++k_KQ_1) {
273-
half2 K_k[kq_stride/warp_size];
274-
half2 Q_k[ncols/nwarps];
299+
for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch/2; k_KQ_1 += cpy_ne) {
300+
half2 K_k[kq_stride/warp_size][cpy_ne];
301+
half2 Q_k[ncols/nwarps][cpy_ne];
275302
#else
276303
#pragma unroll
277-
for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch; ++k_KQ_1) {
278-
float K_k[kq_stride/warp_size];
279-
float Q_k[ncols/nwarps];
304+
for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch; k_KQ_1 += cpy_ne) {
305+
float K_k[kq_stride/warp_size][cpy_ne];
306+
float Q_k[ncols/nwarps][cpy_ne];
280307
#endif // FAST_FP16_AVAILABLE
281308

282309
#pragma unroll
283310
for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
284311
const int i_KQ = i_KQ_0 + threadIdx.x;
285312

286313
#ifdef FAST_FP16_AVAILABLE
287-
K_k[i_KQ_0/warp_size] = KV_tmp_h2[i_KQ*(kq_nbatch/2 + 1) + k_KQ_1];
314+
ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/warp_size], &KV_tmp_h2[i_KQ*(kq_nbatch/2 + cpy_ne) + k_KQ_1]);
288315
#else
289-
K_k[i_KQ_0/warp_size] = KV_tmp_f [i_KQ*(kq_nbatch + 1) + k_KQ_1];
316+
ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/warp_size], &KV_tmp_f [i_KQ*(kq_nbatch + cpy_ne) + k_KQ_1]);
290317
#endif // FAST_FP16_AVAILABLE
291318
}
292319
#pragma unroll
293320
for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
294321
const int j_KQ = j_KQ_0 + threadIdx.y;
295322

296323
#ifdef FAST_FP16_AVAILABLE
297-
Q_k[j_KQ_0/nwarps] = Q_tmp[j_KQ][k_KQ_0/2 + k_KQ_1];
324+
ggml_cuda_memcpy_1<cpy_nb>(&Q_k[j_KQ_0/nwarps], &Q_tmp[j_KQ][k_KQ_0/2 + k_KQ_1]);
298325
#else
299-
Q_k[j_KQ_0/nwarps] = Q_tmp[j_KQ][k_KQ_0 + k_KQ_1];
326+
ggml_cuda_memcpy_1<cpy_nb>(&Q_k[j_KQ_0/nwarps], &Q_tmp[j_KQ][k_KQ_0 + k_KQ_1]);
300327
#endif // FAST_FP16_AVAILABLE
301328
}
302329

303330
#pragma unroll
304331
for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
305332
#pragma unroll
306333
for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
307-
ggml_cuda_mad(sum[i_KQ_0/warp_size][j_KQ_0/nwarps], K_k[i_KQ_0/warp_size], Q_k[j_KQ_0/nwarps]);
334+
#pragma unroll
335+
for (int k = 0; k < cpy_ne; ++k) {
336+
ggml_cuda_mad(sum[i_KQ_0/warp_size][j_KQ_0/nwarps], K_k[i_KQ_0/warp_size][k], Q_k[j_KQ_0/nwarps][k]);
337+
}
308338
}
309339
}
310340
}
@@ -345,14 +375,54 @@ static __global__ void flash_attn_tile(
345375
kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
346376

347377
float kqsum_add = 0.0f;
378+
if (kq_stride % (4*warp_size) == 0 && cpy_ne % 4 == 0) {
348379
#pragma unroll
349-
for (int i0 = 0; i0 < kq_stride; i0 += warp_size) {
350-
const int i = i0 + threadIdx.x;
380+
for (int i0 = 0; i0 < kq_stride; i0 += 4*warp_size) {
381+
const int i = i0 + 4*threadIdx.x;
351382

352-
const float diff = KQ[j][i] - kqmax[j0/nwarps];
353-
const float val = expf(diff);
354-
kqsum_add += val;
355-
KQ[j][i] = val;
383+
float4 val = *(const float4 *) &KQ[j][i];
384+
val.x = expf(val.x - kqmax[j0/nwarps]);
385+
val.y = expf(val.y - kqmax[j0/nwarps]);
386+
val.z = expf(val.z - kqmax[j0/nwarps]);
387+
val.w = expf(val.w - kqmax[j0/nwarps]);
388+
kqsum_add += val.x + val.y + val.z + val.w;
389+
390+
#ifdef FAST_FP16_AVAILABLE
391+
const half2 tmp[2] = {make_half2(val.x, val.y), make_half2(val.z, val.w)};
392+
ggml_cuda_memcpy_1<sizeof(tmp)>(&KQ[j][i/2], &tmp);
393+
#else
394+
ggml_cuda_memcpy_1<sizeof(val)>(&KQ[j][i], &val);
395+
#endif // FAST_FP16_AVAILABLE
396+
}
397+
} else if (kq_stride % (2*warp_size) == 0 && cpy_ne % 2 == 0) {
398+
#pragma unroll
399+
for (int i0 = 0; i0 < kq_stride; i0 += 2*warp_size) {
400+
const int i = i0 + 2*threadIdx.x;
401+
402+
float2 val = *(const float2 *) &KQ[j][i];
403+
val.x = expf(val.x - kqmax[j0/nwarps]);
404+
val.y = expf(val.y - kqmax[j0/nwarps]);
405+
kqsum_add += val.x + val.y;
406+
#ifdef FAST_FP16_AVAILABLE
407+
const half2 tmp = make_half2(val.x, val.y);
408+
ggml_cuda_memcpy_1<sizeof(tmp)>(&KQ[j][i/2], &tmp);
409+
#else
410+
ggml_cuda_memcpy_1<sizeof(val)>(&KQ[j][i], &val);
411+
#endif // FAST_FP16_AVAILABLE
412+
}
413+
} else {
414+
for (int i0 = 0; i0 < kq_stride; i0 += warp_size) {
415+
const int i = i0 + threadIdx.x;
416+
417+
const float diff = KQ[j][i] - kqmax[j0/nwarps];
418+
const float val = expf(diff);
419+
kqsum_add += val;
420+
#ifdef FAST_FP16_AVAILABLE
421+
((half *) KQ[j])[i] = val;
422+
#else
423+
KQ[j][i] = val;
424+
#endif // FAST_FP16_AVAILABLE
425+
}
356426
}
357427
kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + kqsum_add;
358428

@@ -419,8 +489,7 @@ static __global__ void flash_attn_tile(
419489
const int j = j0 + threadIdx.y;
420490

421491
#ifdef FAST_FP16_AVAILABLE
422-
const float tmp = KQ[j][k0 + k1];
423-
KQ_k[j0/nwarps] = make_half2(tmp, tmp);
492+
KQ_k[j0/nwarps] = __half2half2(((const half *)KQ[j])[k0 + k1]);
424493
#else
425494
KQ_k[j0/nwarps] = KQ[j][k0 + k1];
426495
#endif // FAST_FP16_AVAILABLE

‎ggml/src/ggml-cuda/vendors/hip.h‎

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -162,6 +162,14 @@
162162
#define GCN
163163
#endif
164164

165+
#if defined(__gfx900__) || defined(__gfx906__)
166+
#define GCN5
167+
#endif
168+
169+
#if defined(__gfx803__)
170+
#define GCN4
171+
#endif
172+
165173
#if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx942__)
166174
#define CDNA // For the entire family
167175
#endif

0 commit comments

Comments
 (0)