Skip to content

Commit c3827e6

Browse files
ggerganovggerganov
committed
metal : add comments
ggml-ci
1 parent 14d8c54 commit c3827e6

File tree

1 file changed

+33
-25
lines changed

1 file changed

+33
-25
lines changed

‎ggml/src/ggml-metal/ggml-metal-common.cpp‎

Lines changed: 33 additions & 25 deletions
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,8 @@
55

66
#include <vector>
77

8+
// represents a memory range (i.e. an interval from a starting address p0 to an ending address p1 in a given buffer pb)
9+
// the type indicates whether it is a source range (i.e. ops read data from it) or a destination range (i.e. ops write data to it)
810
struct ggml_mem_range {
911
uint64_t pb; // buffer id
1012

@@ -37,8 +39,8 @@ void ggml_mem_ranges_reset(ggml_mem_ranges * mrs) {
3739
mrs->ranges.clear();
3840
}
3941

40-
static bool ggml_mem_ranges_add(ggml_mem_ranges * mrs, ggml_mem_range mrp) {
41-
mrs->ranges.push_back(mrp);
42+
static bool ggml_mem_ranges_add(ggml_mem_ranges * mrs, ggml_mem_range mr) {
43+
mrs->ranges.push_back(mr);
4244

4345
return true;
4446
}
@@ -49,32 +51,32 @@ static ggml_mem_range ggml_mem_range_from_tensor(const ggml_tensor * tensor, ggm
4951

5052
GGML_ASSERT(!tensor->view_src);
5153

52-
ggml_mem_range mrp;
54+
ggml_mem_range mr;
5355

5456
if (tensor->buffer) {
5557
// when the tensor is allocated, use the actual memory address range in the buffer
5658
//
57-
// take the actual allocated size
59+
// take the actual allocated size with ggml_backend_buft_get_alloc_size()
5860
// this can be larger than the tensor size if the buffer type allocates extra memory
5961
// ref: https://github.com/ggml-org/llama.cpp/pull/15966
60-
mrp = {
62+
mr = {
6163
/*.pb =*/ (uint64_t) tensor->buffer,
6264
/*.p0 =*/ (uint64_t) tensor->data,
6365
/*.p1 =*/ (uint64_t) tensor->data + ggml_backend_buft_get_alloc_size(tensor->buffer->buft, tensor),
6466
/*.pt =*/ pt,
6567
};
6668
} else {
67-
// otherwise, the tensor ptr is used as an unique id of the memory ranges
69+
// otherwise, the pointer address is used as an unique id of the memory ranges
6870
// that the tensor will be using when it is allocated
69-
mrp = {
71+
mr = {
7072
/*.pb =*/ (uint64_t) tensor,
7173
/*.p0 =*/ 0, //
7274
/*.p1 =*/ 1024, // [0, 1024) is a dummy range, not used
7375
/*.pt =*/ pt,
7476
};
7577
};
7678

77-
return mrp;
79+
return mr;
7880
}
7981

8082
static ggml_mem_range ggml_mem_range_from_tensor_src(const ggml_tensor * tensor) {
@@ -88,25 +90,25 @@ static ggml_mem_range ggml_mem_range_from_tensor_dst(const ggml_tensor * tensor)
8890
static bool ggml_mem_ranges_add_src(ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
8991
GGML_ASSERT(tensor);
9092

91-
ggml_mem_range mrp = ggml_mem_range_from_tensor_src(tensor);
93+
ggml_mem_range mr = ggml_mem_range_from_tensor_src(tensor);
9294

9395
if (mrs->debug > 2) {
94-
GGML_LOG_DEBUG("%s: add src range buf=%lld, [%lld, %lld)\n", __func__, mrp.pb, mrp.p0, mrp.p1);
96+
GGML_LOG_DEBUG("%s: add src range buf=%lld, [%lld, %lld)\n", __func__, mr.pb, mr.p0, mr.p1);
9597
}
9698

97-
return ggml_mem_ranges_add(mrs, mrp);
99+
return ggml_mem_ranges_add(mrs, mr);
98100
}
99101

100102
static bool ggml_mem_ranges_add_dst(ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
101103
GGML_ASSERT(tensor);
102104

103-
ggml_mem_range mrp = ggml_mem_range_from_tensor_dst(tensor);
105+
ggml_mem_range mr = ggml_mem_range_from_tensor_dst(tensor);
104106

105107
if (mrs->debug > 2) {
106-
GGML_LOG_DEBUG("%s: add dst range buf=%lld, [%lld, %lld)\n", __func__, mrp.pb, mrp.p0, mrp.p1);
108+
GGML_LOG_DEBUG("%s: add dst range buf=%lld, [%lld, %lld)\n", __func__, mr.pb, mr.p0, mr.p1);
107109
}
108110

109-
return ggml_mem_ranges_add(mrs, mrp);
111+
return ggml_mem_ranges_add(mrs, mr);
110112
}
111113

112114
bool ggml_mem_ranges_add(ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
@@ -119,24 +121,26 @@ bool ggml_mem_ranges_add(ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
119121
return ggml_mem_ranges_add_dst(mrs, tensor);
120122
}
121123

122-
static bool ggml_mem_ranges_check(const ggml_mem_ranges * mrs, ggml_mem_range mrp) {
124+
static bool ggml_mem_ranges_check(const ggml_mem_ranges * mrs, ggml_mem_range mr) {
123125
for (size_t i = 0; i < mrs->ranges.size(); i++) {
124126
const auto & cmp = mrs->ranges[i];
125127

126-
if (mrp.pb != cmp.pb) {
128+
// two memory ranges cannot intersect if they are in different buffers
129+
if (mr.pb != cmp.pb) {
127130
continue;
128131
}
129132

130-
if (mrp.pt == MEM_RANGE_TYPE_SRC && cmp.pt == MEM_RANGE_TYPE_SRC) {
133+
// intersecting source ranges are allowed
134+
if (mr.pt == MEM_RANGE_TYPE_SRC && cmp.pt == MEM_RANGE_TYPE_SRC) {
131135
continue;
132136
}
133137

134-
if (mrp.p0 < cmp.p1 && mrp.p1 >= cmp.p0) {
138+
if (mr.p0 < cmp.p1 && mr.p1 >= cmp.p0) {
135139
if (mrs->debug > 2) {
136140
GGML_LOG_DEBUG("%s: the %s range buf=%lld, [%lld, %lld) overlaps with a previous %s range buf=%lld, [%lld, %lld)\n",
137141
__func__,
138-
mrp.pt == MEM_RANGE_TYPE_SRC ? "src" : "dst",
139-
mrp.pb, mrp.p0, mrp.p1,
142+
mr.pt == MEM_RANGE_TYPE_SRC ? "src" : "dst",
143+
mr.pb, mr.p0, mr.p1,
140144
cmp.pt == MEM_RANGE_TYPE_SRC ? "src" : "dst",
141145
cmp.pb, cmp.p0, cmp.p1);
142146
}
@@ -151,19 +155,19 @@ static bool ggml_mem_ranges_check(const ggml_mem_ranges * mrs, ggml_mem_range mr
151155
static bool ggml_mem_ranges_check_src(const ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
152156
GGML_ASSERT(tensor);
153157

154-
ggml_mem_range mrp = ggml_mem_range_from_tensor_src(tensor);
158+
ggml_mem_range mr = ggml_mem_range_from_tensor_src(tensor);
155159

156-
const bool res = ggml_mem_ranges_check(mrs, mrp);
160+
const bool res = ggml_mem_ranges_check(mrs, mr);
157161

158162
return res;
159163
}
160164

161165
static bool ggml_mem_ranges_check_dst(const ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
162166
GGML_ASSERT(tensor);
163167

164-
ggml_mem_range mrp = ggml_mem_range_from_tensor_dst(tensor);
168+
ggml_mem_range mr = ggml_mem_range_from_tensor_dst(tensor);
165169

166-
const bool res = ggml_mem_ranges_check(mrs, mrp);
170+
const bool res = ggml_mem_ranges_check(mrs, mr);
167171

168172
return res;
169173
}
@@ -227,6 +231,7 @@ static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node
227231
}
228232
}
229233

234+
// keep track of the sources of the fused nodes as well
230235
for (const auto * fused : node.fused) {
231236
for (int i = 0; i < GGML_MAX_SRC; i++) {
232237
if (fused->src[i]) {
@@ -295,7 +300,10 @@ static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node
295300

296301
std::vector<bool> used(n, false);
297302

303+
// the memory ranges for the set of currently concurrent nodes
298304
ggml_mem_ranges * mrs0 = ggml_mem_ranges_init(0);
305+
306+
// the memory ranges for the set of nodes that haven't been processed yet, when looking forward for a node to reorder
299307
ggml_mem_ranges * mrs1 = ggml_mem_ranges_init(0);
300308

301309
for (int i0 = 0; i0 < n; i0++) {
@@ -424,8 +432,8 @@ void ggml_metal_graph_optimize(ggml_cgraph * gf) {
424432
nodes.push_back(std::move(node));
425433
}
426434

427-
// reorder to improve concurrency
428435
#if 1
436+
// reorder to improve concurrency
429437
const auto order = ggml_metal_graph_optimize_reorder(nodes);
430438
#else
431439
std::vector<int> order(nodes.size());

0 commit comments

Comments
 (0)