Skip to content

Commit 7f459c9

Browse files
0cc4m0cc4m
authored
vulkan: use fewer FA rows for small cache runs (#18280)
1 parent cf2ffc0 commit 7f459c9

1 file changed

Lines changed: 29 additions & 27 deletions

File tree

‎ggml/src/ggml-vulkan/ggml-vulkan.cpp‎

Lines changed: 29 additions & 27 deletions
Original file line numberDiff line numberDiff line change
@@ -379,18 +379,18 @@ enum FaCodePath {
379379
};
380380

381381
struct vk_fa_pipeline_state {
382-
vk_fa_pipeline_state(uint32_t HSK, uint32_t HSV, bool small_rows, FaCodePath path, bool aligned, bool f32acc)
383-
: HSK(HSK), HSV(HSV), small_rows(small_rows), path(path), aligned(aligned), f32acc(f32acc) {}
382+
vk_fa_pipeline_state(uint32_t HSK, uint32_t HSV, bool small_rows, bool small_cache, FaCodePath path, bool aligned, bool f32acc)
383+
: HSK(HSK), HSV(HSV), small_rows(small_rows), small_cache(small_cache), path(path), aligned(aligned), f32acc(f32acc) {}
384384

385385
uint32_t HSK, HSV;
386-
bool small_rows;
386+
bool small_rows, small_cache;
387387
FaCodePath path;
388388
bool aligned;
389389
bool f32acc;
390390

391391
bool operator<(const vk_fa_pipeline_state &b) const {
392-
return std::tie(HSK, HSV, small_rows, path, aligned, f32acc) <
393-
std::tie(b.HSK, b.HSV, b.small_rows, b.path, b.aligned, b.f32acc);
392+
return std::tie(HSK, HSV, small_rows, small_cache, path, aligned, f32acc) <
393+
std::tie(b.HSK, b.HSV, b.small_rows, b.small_cache, b.path, b.aligned, b.f32acc);
394394
}
395395
};
396396

@@ -2582,10 +2582,10 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
25822582
static constexpr uint32_t flash_attention_num_small_rows = 32;
25832583
static constexpr uint32_t scalar_flash_attention_num_small_rows = 1;
25842584

2585-
static uint32_t get_fa_scalar_num_large_rows(uint32_t hsk, uint32_t hsv) {
2585+
static uint32_t get_fa_scalar_num_large_rows(uint32_t hsk, uint32_t hsv, bool small_cache) {
25862586
if (hsv >= 192) {
25872587
return 2;
2588-
} else if ((hsv | hsk) & 8) {
2588+
} else if ((hsv | hsk) & 8 || small_cache) {
25892589
return 4;
25902590
} else {
25912591
return 8;
@@ -2607,9 +2607,8 @@ static uint32_t get_fa_num_small_rows(FaCodePath path) {
26072607
}
26082608
}
26092609

2610-
static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) {
2610+
static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows, bool small_cache) {
26112611
GGML_UNUSED(clamp);
2612-
GGML_UNUSED(hsv);
26132612

26142613
if (path == FA_SCALAR) {
26152614
if (small_rows) {
@@ -2618,9 +2617,9 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint3
26182617
if ((hsv | hsk) & 8) {
26192618
// HSV/HSK not being a multiple of 16 makes D_split smaller, which makes cols_per_iter
26202619
// larger, and Bc needs to be >= cols_per_thread. 64 is large enough, 32 is not.
2621-
return {get_fa_scalar_num_large_rows(hsk, hsv), 64};
2620+
return {get_fa_scalar_num_large_rows(hsk, hsv, small_cache), 64};
26222621
} else {
2623-
return {get_fa_scalar_num_large_rows(hsk, hsv), 32};
2622+
return {get_fa_scalar_num_large_rows(hsk, hsv, small_cache), 32};
26242623
}
26252624
}
26262625
}
@@ -2649,8 +2648,8 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint3
26492648
return {64, 64};
26502649
}
26512650

2652-
static uint32_t fa_align(FaCodePath path, uint32_t hsk, uint32_t hsv, ggml_type type, bool small_rows) {
2653-
return fa_rows_cols(path, hsk, hsv, 0, type, small_rows)[1];
2651+
static uint32_t fa_align(FaCodePath path, uint32_t hsk, uint32_t hsv, ggml_type type, bool small_rows, bool small_cache) {
2652+
return fa_rows_cols(path, hsk, hsv, 0, type, small_rows, small_cache)[1];
26542653
}
26552654

26562655
static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vector<uint32_t>& warptile, bool mul_mat_id, ggml_type src0_type) {
@@ -2992,11 +2991,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
29922991
align, disable_robustness, require_full_subgroups, required_subgroup_size);
29932992
};
29942993

2995-
auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
2996-
return {fa_rows_cols(path, hsk, hsv, clamp, type, small_rows)[0], 1, 1};
2994+
auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows, bool small_cache) -> std::array<uint32_t, 3> {
2995+
return {fa_rows_cols(path, hsk, hsv, clamp, type, small_rows, small_cache)[0], 1, 1};
29972996
};
29982997

2999-
auto const &fa_spec_constants = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
2998+
auto const &fa_spec_constants = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows, bool small_cache) -> std::vector<uint32_t> {
30002999
// For large number of rows, 128 invocations seems to work best.
30013000
// For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we
30023001
// can't use 256 for D==80.
@@ -3006,7 +3005,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
30063005
uint32_t wg_size = (path == FA_SCALAR || path == FA_COOPMAT1)
30073006
? scalar_flash_attention_workgroup_size
30083007
: ((small_rows && (D % 32) == 0) ? 256 : 128);
3009-
auto rows_cols = fa_rows_cols(path, hsk, hsv, clamp, type, small_rows);
3008+
auto rows_cols = fa_rows_cols(path, hsk, hsv, clamp, type, small_rows, small_cache);
30103009

30113010
// D_split can't be larger than a subgroup because we use subgroupShuffle to reduce it.
30123011
// D_split can't be larger than the LSB of D divided by 4 due to vectorization in the shader.
@@ -3021,21 +3020,22 @@ static void ggml_vk_load_shaders(vk_device& device) {
30213020
uint32_t HSK = fa.first.HSK; \
30223021
uint32_t HSV = fa.first.HSV; \
30233022
bool small_rows = fa.first.small_rows; \
3023+
bool small_cache = fa.first.small_cache; \
30243024
FaCodePath path = fa.first.path; \
30253025
bool aligned = fa.first.aligned; \
30263026
bool f32acc = fa.first.f32acc; \
30273027
if (path == FAPATH) { \
30283028
if (aligned) { \
30293029
if (f32acc) { \
3030-
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_align(FAPATH,HSK,HSV,TYPE,small_rows), true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \
3030+
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \
30313031
} else { \
3032-
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_align(FAPATH,HSK,HSV,TYPE,small_rows), true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \
3032+
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \
30333033
} \
30343034
} else { \
30353035
if (f32acc) { \
3036-
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows), 1, true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \
3036+
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1, true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \
30373037
} else { \
3038-
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows), 1, true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \
3038+
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1, true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \
30393039
} \
30403040
} \
30413041
} \
@@ -8008,11 +8008,11 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx
80088008
}
80098009
}
80108010

8011-
static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv) {
8011+
static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv, bool small_cache) {
80128012
// Needs to be kept up to date on shader changes
80138013
GGML_UNUSED(hsv);
80148014
const uint32_t wg_size = scalar_flash_attention_workgroup_size;
8015-
const uint32_t Br = get_fa_scalar_num_large_rows(hsk, hsv);
8015+
const uint32_t Br = get_fa_scalar_num_large_rows(hsk, hsv, small_cache);
80168016
const uint32_t Bc = scalar_flash_attention_Bc;
80178017

80188018
const uint32_t tmpsh = wg_size * sizeof(float);
@@ -8136,14 +8136,16 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
81368136
uint32_t workgroups_y = (uint32_t)neq2;
81378137
uint32_t workgroups_z = (uint32_t)neq3;
81388138

8139+
const bool small_cache = nek1 < 1024;
8140+
81398141
// For scalar/coopmat1 FA, we can use the "large" size to accommodate qga.
81408142
// For coopmat2 FA, we always use the small size (which is still pretty large for gqa).
81418143
uint32_t max_gqa;
81428144
switch (path) {
81438145
case FA_SCALAR:
81448146
case FA_COOPMAT1:
81458147
// We may switch from coopmat1 to scalar, so use the scalar limit for both
8146-
max_gqa = get_fa_scalar_num_large_rows(HSK, HSV);
8148+
max_gqa = get_fa_scalar_num_large_rows(HSK, HSV, small_cache);
81478149
break;
81488150
case FA_COOPMAT2:
81498151
max_gqa = get_fa_num_small_rows(FA_COOPMAT2);
@@ -8177,7 +8179,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
81778179

81788180
// with large hsk/hsv, scalar path may need to use small_rows to fit in shared memory
81798181
if (path == FA_SCALAR &&
8180-
!ggml_vk_flash_attn_scalar_shmem_support(ctx->device, HSK, HSV)) {
8182+
!ggml_vk_flash_attn_scalar_shmem_support(ctx->device, HSK, HSV, small_cache)) {
81818183
small_rows = true;
81828184
}
81838185

@@ -8193,7 +8195,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
81938195
v_stride /= 4;
81948196
}
81958197

8196-
uint32_t alignment = fa_align(path, HSK, HSV, k->type, small_rows);
8198+
uint32_t alignment = fa_align(path, HSK, HSV, k->type, small_rows, small_cache);
81978199
bool aligned = (KV % alignment) == 0 &&
81988200
// the "aligned" shader variant will forcibly align strides, for performance
81998201
(q_stride & 7) == 0 && (k_stride & 7) == 0 && (v_stride & 7) == 0;
@@ -8205,7 +8207,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
82058207

82068208
bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32;
82078209

8208-
vk_fa_pipeline_state fa_pipeline_state(HSK, HSV, small_rows, path, aligned, f32acc);
8210+
vk_fa_pipeline_state fa_pipeline_state(HSK, HSV, small_rows, small_cache, path, aligned, f32acc);
82098211

82108212
vk_pipeline pipeline = nullptr;
82118213

0 commit comments

Comments
 (0)