Skip to content

Commit 1fe0029

Browse files
authored
vulkan: fuse adds (#15252)
* vulkan: fuse adds Fuse adds that have the same shape, which are common in MoE models. It will currently fuse up to 6 adds, because we assume no more than 8 descriptors per dispatch. But this could be changed. * check runtimeDescriptorArray feature * disable multi_add for Intel due to likely driver bug
1 parent de21927 commit 1fe0029

File tree

6 files changed

+301
-25
lines changed

6 files changed

+301
-25
lines changed

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

Lines changed: 201 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -103,6 +103,8 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
103103
struct ggml_backend_vk_context;
104104

105105
#define MAX_PARAMETER_COUNT 8
106+
// Max number of adds that can be fused without exceeding MAX_PARAMETER_COUNT.
107+
#define MAX_FUSED_ADDS (MAX_PARAMETER_COUNT - 2)
106108

107109
struct vk_pipeline_struct {
108110
std::string name;
@@ -368,6 +370,7 @@ struct vk_device_struct {
368370
bool float_controls_rte_fp16;
369371
bool subgroup_add;
370372
bool subgroup_shuffle;
373+
bool multi_add;
371374

372375
bool integer_dot_product;
373376

@@ -449,6 +452,9 @@ struct vk_device_struct {
449452
vk_pipeline pipeline_div[2][2][2];
450453
vk_pipeline pipeline_div_norepeat[2][2][2];
451454

455+
// indexed by num_additional_fused_ops == num_adds - 1
456+
vk_pipeline pipeline_multi_add[MAX_FUSED_ADDS];
457+
452458
vk_pipeline pipeline_add_id_f32;
453459

454460
vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
@@ -801,6 +807,14 @@ struct vk_op_binary_push_constants {
801807
float param1; float param2; int32_t param3;
802808
};
803809

810+
struct vk_op_multi_add_push_constants {
811+
// shape for dst
812+
uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23;
813+
814+
// strides for srcs+dst
815+
uint32_t nb[8][4];
816+
};
817+
804818
struct vk_op_add_id_push_constants {
805819
uint32_t ne0;
806820
uint32_t ne1;
@@ -2994,6 +3008,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
29943008
CREATE_BINARY(div, _norepeat, {1})
29953009
#undef CREATE_BINARY
29963010

3011+
if (device->multi_add) {
3012+
for (uint32_t i = 0; i < MAX_FUSED_ADDS; ++i) {
3013+
ggml_vk_create_pipeline(device, device->pipeline_multi_add[i], "multi_add_f32_" + std::to_string(i+1), multi_add_f32_len, multi_add_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1);
3014+
}
3015+
}
3016+
29973017
ggml_vk_create_pipeline(device, device->pipeline_add_id_f32, "add_id_f32", add_id_f32_len, add_id_f32_data, "main", 4, sizeof(vk_op_add_id_push_constants), {1, 1, 1}, {}, 1);
29983018

29993019
ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
@@ -3533,6 +3553,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
35333553

35343554
device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
35353555

3556+
device->multi_add = vk12_props.shaderRoundingModeRTEFloat16 &&
3557+
device->properties.limits.maxPushConstantsSize >= sizeof(vk_op_multi_add_push_constants) &&
3558+
vk12_features.runtimeDescriptorArray &&
3559+
device->vendor_id != VK_VENDOR_ID_INTEL &&
3560+
getenv("GGML_VK_DISABLE_MULTI_ADD") == nullptr;
3561+
35363562
if (device->subgroup_size_control) {
35373563
device->subgroup_min_size = subgroup_size_control_props.minSubgroupSize;
35383564
device->subgroup_max_size = subgroup_size_control_props.maxSubgroupSize;
@@ -6887,6 +6913,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
68876913
switch (op) {
68886914
case GGML_OP_ADD:
68896915
{
6916+
if (ctx->num_additional_fused_ops > 0) {
6917+
return ctx->device->pipeline_multi_add[ctx->num_additional_fused_ops];
6918+
}
68906919
auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_norepeat : ctx->device->pipeline_add;
68916920
return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16];
68926921
}
@@ -7743,6 +7772,107 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
77437772
}, dryrun);
77447773
}
77457774

7775+
static void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx, bool dryrun = false) {
7776+
const ggml_tensor *first_node = cgraph->nodes[node_idx];
7777+
const ggml_tensor *dst = cgraph->nodes[node_idx + ctx->num_additional_fused_ops];
7778+
7779+
// Make a list of all the tensors used by the op.
7780+
// Last element of the list is the dest tensor.
7781+
const ggml_tensor *tensors[MAX_PARAMETER_COUNT];
7782+
uint32_t num_srcs = ctx->num_additional_fused_ops + 2;
7783+
uint32_t num_tensors = num_srcs + 1;
7784+
GGML_ASSERT(num_tensors <= MAX_PARAMETER_COUNT);
7785+
7786+
tensors[0] = first_node->src[0];
7787+
tensors[1] = first_node->src[1];
7788+
for (int32_t i = 0; i < ctx->num_additional_fused_ops; ++i) {
7789+
// check whether the previous result is src[0] or src[1]
7790+
if (cgraph->nodes[node_idx + i] == cgraph->nodes[node_idx + i + 1]->src[0]) {
7791+
tensors[i+2] = cgraph->nodes[node_idx + i + 1]->src[1];
7792+
} else {
7793+
tensors[i+2] = cgraph->nodes[node_idx + i + 1]->src[0];
7794+
}
7795+
}
7796+
tensors[num_srcs] = dst;
7797+
7798+
vk_op_multi_add_push_constants pc;
7799+
pc.ne20 = (uint32_t)dst->ne[0];
7800+
pc.ne21 = (uint32_t)dst->ne[1];
7801+
pc.ne22 = (uint32_t)dst->ne[2];
7802+
pc.ne23 = (uint32_t)dst->ne[3];
7803+
7804+
for (uint32_t i = 0; i < num_tensors; ++i) {
7805+
const ggml_tensor *t = tensors[i];
7806+
pc.nb[i][0] = (uint32_t)t->nb[0] / sizeof(float);
7807+
pc.nb[i][1] = (uint32_t)t->nb[1] / sizeof(float);
7808+
pc.nb[i][2] = (uint32_t)t->nb[2] / sizeof(float);
7809+
pc.nb[i][3] = (uint32_t)t->nb[3] / sizeof(float);
7810+
}
7811+
7812+
vk_pipeline pipeline = ctx->device->pipeline_multi_add[ctx->num_additional_fused_ops];
7813+
7814+
if (pipeline == nullptr) {
7815+
std::cerr << "ggml_vulkan: Error: Missing multi_add";
7816+
GGML_ABORT("fatal error");
7817+
}
7818+
7819+
if (dryrun) {
7820+
ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
7821+
return;
7822+
}
7823+
7824+
ggml_backend_vk_buffer_context * buf_ctx[MAX_PARAMETER_COUNT];
7825+
vk_buffer buf[MAX_PARAMETER_COUNT];
7826+
size_t offset[MAX_PARAMETER_COUNT];
7827+
bool uma[MAX_PARAMETER_COUNT];
7828+
7829+
for (uint32_t i = 0; i < num_tensors; ++i) {
7830+
buf_ctx[i] = (ggml_backend_vk_buffer_context *)tensors[i]->buffer->context;
7831+
buf[i] = nullptr;
7832+
offset[i] = 0;
7833+
uma[i] = false;
7834+
7835+
if (ctx->device->uma) {
7836+
ggml_vk_host_get(ctx->device, tensors[i]->data, buf[i], offset[i]);
7837+
uma[i] = buf[i] != nullptr;
7838+
}
7839+
if (!uma[i]) {
7840+
buf[i] = buf_ctx[i]->dev_buffer;
7841+
offset[i] = vk_tensor_offset(tensors[i]) + tensors[i]->view_offs;
7842+
}
7843+
GGML_ASSERT(buf[i] != nullptr);
7844+
}
7845+
// If any remaining descriptors are unused, just point them at src[0]
7846+
for (uint32_t i = num_tensors; i < MAX_PARAMETER_COUNT; ++i) {
7847+
buf[i] = buf[0];
7848+
offset[i] = 0;
7849+
}
7850+
7851+
std::array<uint32_t, 3> elements;
7852+
7853+
uint32_t ne = ggml_nelements(dst);
7854+
if (ne > 262144) {
7855+
elements = { 512, 512, CEIL_DIV(ne, 262144) };
7856+
} else if (ne > 512) {
7857+
elements = { 512, CEIL_DIV(ne, 512), 1 };
7858+
} else {
7859+
elements = { ne, 1, 1 };
7860+
}
7861+
7862+
ggml_vk_sync_buffers(subctx);
7863+
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
7864+
{
7865+
vk_subbuffer{ buf[0], offset[0], VK_WHOLE_SIZE },
7866+
vk_subbuffer{ buf[1], offset[1], VK_WHOLE_SIZE },
7867+
vk_subbuffer{ buf[2], offset[2], VK_WHOLE_SIZE },
7868+
vk_subbuffer{ buf[3], offset[3], VK_WHOLE_SIZE },
7869+
vk_subbuffer{ buf[4], offset[4], VK_WHOLE_SIZE },
7870+
vk_subbuffer{ buf[5], offset[5], VK_WHOLE_SIZE },
7871+
vk_subbuffer{ buf[6], offset[6], VK_WHOLE_SIZE },
7872+
vk_subbuffer{ buf[7], offset[7], VK_WHOLE_SIZE },
7873+
}, pc, elements);
7874+
}
7875+
77467876
static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
77477877
const uint32_t src0_type_size = ggml_type_size(src0->type);
77487878
const uint32_t src1_type_size = ggml_type_size(src1->type);
@@ -9703,8 +9833,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
97039833

97049834
break;
97059835
case GGML_OP_ADD:
9706-
ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
9707-
9836+
if (ctx->num_additional_fused_ops) {
9837+
ggml_vk_multi_add(ctx, compute_ctx, cgraph, node_idx, dryrun);
9838+
} else {
9839+
ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
9840+
}
97089841
break;
97099842
case GGML_OP_SUB:
97109843
ggml_vk_sub(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -10586,6 +10719,58 @@ static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, st
1058610719
return true;
1058710720
}
1058810721

10722+
static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx) {
10723+
10724+
const ggml_tensor *first_node = cgraph->nodes[node_idx];
10725+
if (first_node->op != GGML_OP_ADD) {
10726+
return 0;
10727+
}
10728+
10729+
if (!ctx->device->multi_add) {
10730+
return 0;
10731+
}
10732+
10733+
int32_t num_adds = 1;
10734+
while (node_idx + num_adds < cgraph->n_nodes &&
10735+
cgraph->nodes[node_idx + num_adds]->op == GGML_OP_ADD &&
10736+
num_adds < MAX_FUSED_ADDS) {
10737+
num_adds++;
10738+
}
10739+
10740+
// The shader currently requires same shapes (but different strides are allowed),
10741+
// everything f32, and no misalignment
10742+
for (int32_t i = 0; i < num_adds; ++i) {
10743+
const ggml_tensor *next_node = cgraph->nodes[node_idx + i];
10744+
if (!ggml_are_same_shape(first_node, next_node->src[0]) ||
10745+
!ggml_are_same_shape(first_node, next_node->src[1]) ||
10746+
next_node->type != GGML_TYPE_F32 ||
10747+
next_node->src[0]->type != GGML_TYPE_F32 ||
10748+
next_node->src[1]->type != GGML_TYPE_F32 ||
10749+
get_misalign_bytes(ctx, next_node) ||
10750+
get_misalign_bytes(ctx, next_node->src[0]) ||
10751+
get_misalign_bytes(ctx, next_node->src[1])) {
10752+
num_adds = i;
10753+
}
10754+
}
10755+
10756+
// Verify we can fuse these
10757+
ggml_op adds[MAX_FUSED_ADDS];
10758+
for (int32_t i = 0; i < num_adds; ++i) {
10759+
adds[i] = GGML_OP_ADD;
10760+
}
10761+
10762+
// decrease num_adds if they can't all be fused
10763+
while (num_adds > 1 && !ggml_can_fuse(cgraph, node_idx, adds, num_adds)) {
10764+
num_adds--;
10765+
}
10766+
10767+
// a single add is not "fused", so just return zero
10768+
if (num_adds == 1) {
10769+
return 0;
10770+
}
10771+
return num_adds;
10772+
}
10773+
1058910774
static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
1059010775
VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
1059110776
ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
@@ -10599,8 +10784,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
1059910784

1060010785
uint64_t total_mat_mul_bytes = 0;
1060110786
for (int i = 0; i < cgraph->n_nodes; i++) {
10602-
if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
10603-
ctx->num_additional_fused_ops = 1;
10787+
if (!ctx->device->disable_fusion) {
10788+
uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
10789+
if (num_adds) {
10790+
ctx->num_additional_fused_ops = num_adds - 1;
10791+
} else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
10792+
ctx->num_additional_fused_ops = 1;
10793+
}
1060410794
}
1060510795
ggml_vk_build_graph(ctx, cgraph, i, nullptr, 0, true, false, false, false);
1060610796
if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
@@ -10675,8 +10865,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
1067510865
mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
1067610866
}
1067710867

10678-
if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
10679-
ctx->num_additional_fused_ops = 1;
10868+
if (!ctx->device->disable_fusion) {
10869+
uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
10870+
if (num_adds) {
10871+
ctx->num_additional_fused_ops = num_adds - 1;
10872+
} else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
10873+
ctx->num_additional_fused_ops = 1;
10874+
}
1068010875
}
1068110876

1068210877
// Signal the almost_ready fence when the graph is mostly complete (< 20% remaining)

ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp

Lines changed: 2 additions & 17 deletions
Original file line numberDiff line numberDiff line change
@@ -2,6 +2,7 @@
22
#extension GL_EXT_control_flow_attributes : require
33

44
#include "rte.comp"
5+
#include "utils.comp"
56

67
layout (push_constant) uniform parameter
78
{
@@ -28,25 +29,9 @@ uint get_aoffset() { return p.misalign_offsets >> 16; }
2829
uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; }
2930
uint get_doffset() { return p.misalign_offsets & 0xFF; }
3031

31-
// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
32-
uint fastmod(uint a, uint b) {
33-
if ((b & (b-1)) == 0) {
34-
return a & (b-1);
35-
}
36-
return a % b;
37-
}
38-
39-
uint fastdiv(uint a, uint b) {
40-
return (a < b) ? 0 : (a / b);
41-
}
4232

4333
void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) {
44-
i03 = fastdiv(idx, (p.ne02*p.ne01*p.ne00));
45-
const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
46-
i02 = fastdiv((idx - i03_offset), (p.ne01*p.ne00));
47-
const uint i02_offset = i02*p.ne01*p.ne00;
48-
i01 = (idx - i03_offset - i02_offset) / p.ne00;
49-
i00 = idx - i03_offset - i02_offset - i01*p.ne00;
34+
get_indices(idx, i00, i01, i02, i03, p.ne00, p.ne01, p.ne02, p.ne03);
5035
}
5136

5237
uint src0_idx(uint i00, uint i01, uint i02, uint i03) {
Lines changed: 68 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,68 @@
1+
#version 450
2+
3+
#extension GL_EXT_shader_16bit_storage : require
4+
#extension GL_EXT_nonuniform_qualifier : enable
5+
#extension GL_EXT_control_flow_attributes : require
6+
7+
#include "rte.comp"
8+
#include "types.comp"
9+
#include "utils.comp"
10+
11+
layout (push_constant) uniform parameter2
12+
{
13+
// shape for dst
14+
uint ne20; uint ne21; uint ne22; uint ne23;
15+
16+
// strides for srcs+dst
17+
uint nb[8][4];
18+
} p;
19+
20+
layout (binding = 0) readonly buffer A {A_TYPE data_a[];} a[];
21+
layout (binding = 0) writeonly buffer D {D_TYPE data_d[];} d[];
22+
23+
layout(constant_id = 0) const uint num_srcs = 2;
24+
25+
uint src_idx(uint s, uint i00, uint i01, uint i02, uint i03) {
26+
return i03*p.nb[s][3] + i02*p.nb[s][2] + i01*p.nb[s][1] + i00*p.nb[s][0];
27+
}
28+
29+
uint dst_idx(uint i00, uint i01, uint i02, uint i03) {
30+
uint nb20 = p.nb[num_srcs][0];
31+
uint nb21 = p.nb[num_srcs][1];
32+
uint nb22 = p.nb[num_srcs][2];
33+
uint nb23 = p.nb[num_srcs][3];
34+
return i03*nb23 + i02*nb22 + i01*nb21 + i00*nb20;
35+
}
36+
37+
uint get_idx() {
38+
return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
39+
}
40+
41+
const uint num_threads = 256;
42+
43+
layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
44+
45+
void main() {
46+
uint idx = get_idx();
47+
48+
uint ne = p.ne20 * p.ne21 * p.ne22 * p.ne23;
49+
50+
// num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
51+
const uint num_iter = 2;
52+
53+
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
54+
if (idx >= ne) {
55+
continue;
56+
}
57+
uint i00, i01, i02, i03;
58+
get_indices(idx, i00, i01, i02, i03, p.ne20, p.ne21, p.ne22, p.ne23);
59+
60+
FLOAT_TYPE sum = FLOAT_TYPE(0);
61+
[[unroll]] for (uint s = 0; s < num_srcs; ++s) {
62+
sum += FLOAT_TYPE(a[s].data_a[src_idx(s, i00, i01, i02, i03)]);
63+
}
64+
d[num_srcs].data_d[dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
65+
66+
idx += num_threads;
67+
}
68+
}

0 commit comments

Comments
 (0)