Skip to content

LLM inference in C/C++

C++ ggml

ggml: optimize concat op by replacing per-element memcpy with row-level memcpy (#24575)

Kartik Sirohi369e1cd6140b
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 2b5f68444b..b869f4bddd 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -1896,7 +1896,6 @@ void ggml_compute_forward_repeat_back(
}
// ggml_compute_forward_concat
-
static void ggml_compute_forward_concat_any(
const ggml_compute_params * params,
ggml_tensor * dst) {
@@ -1904,8 +1903,6 @@ static void ggml_compute_forward_concat_any(
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
- const size_t len = ggml_type_size(src0->type);
-
const int ith = params->ith;
const int nth = params->nth;
@@ -1914,31 +1911,38 @@ static void ggml_compute_forward_concat_any(
const int32_t dim = ggml_get_op_params_i32(dst, 0);
GGML_ASSERT(dim >= 0 && dim < 4);
+ GGML_ASSERT(ggml_is_contiguous_rows(src0));
+ GGML_ASSERT(ggml_is_contiguous_rows(src1));
int64_t o[4] = {0, 0, 0, 0};
+
if (dim == 0) {
+ GGML_ASSERT(src0->ne[0] % ggml_blck_size(src0->type) == 0);
+ GGML_ASSERT(src1->ne[0] % ggml_blck_size(src1->type) == 0);
+
o[dim] = src0->ne[dim]/ggml_blck_size(src0->type);
} else {
o[dim] = src0->ne[dim];
}
- const char * x;
-
- // TODO: smarter multi-theading
- for (int i3 = 0; i3 < ne3; i3++) {
- for (int i2 = ith; i2 < ne2; i2 += nth) {
- for (int i1 = 0; i1 < ne1; i1++) {
- for (int i0 = 0; i0 < ne0/ggml_blck_size(dst->type); i0++) {
- if (i0 < ne00/ggml_blck_size(src0->type) && i1 < ne01 && i2 < ne02 && i3 < ne03) {
- x = (const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03;
- } else {
- x = (const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13;
- }
-
- char * y = (char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3;
+ // Region 1: copy rows from src0
+ for (int i3 = 0; i3 < ne03; i3++) {
+ for (int i2 = ith; i2 < ne02; i2 += nth) {
+ for (int i1 = 0; i1 < ne01; i1++) {
+ const char * x = (const char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03;
+ char * y = ( char *) dst->data + i1*nb1 + i2*nb2 + i3*nb3;
+ memcpy(y, x, ggml_row_size(src0->type, ne00));
+ }
+ }
+ }
- memcpy(y, x, len);
- }
+ // Region 2: copy rows from src1, offset into dst by o[]
+ for (int i3 = 0; i3 < ne13; i3++) {
+ for (int i2 = ith; i2 < ne12; i2 += nth) {
+ for (int i1 = 0; i1 < ne11; i1++) {
+ const char * x = (const char *) src1->data + i1*nb11 + i2*nb12 + i3*nb13;
+ char * y = ( char *) dst->data + (i1 + o[1])*nb1 + (i2 + o[2])*nb2 + (i3 + o[3])*nb3 + o[0]*nb0;
+ memcpy(y, x, ggml_row_size(src1->type, ne10));
}
}
}
@@ -2078,14 +2082,6 @@ void ggml_compute_forward_concat(
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
- const ggml_tensor * src1 = dst->src[1];
-
- if (ggml_is_quantized(src0->type)) {
- GGML_ASSERT(ggml_is_contiguous_rows(src0));
- GGML_ASSERT(ggml_is_contiguous_rows(src1));
- GGML_ASSERT(src0->ne[0] % ggml_blck_size(src0->type) == 0);
- GGML_ASSERT(src1->ne[0] % ggml_blck_size(src1->type) == 0);
- }
switch (src0->type) {
case GGML_TYPE_F16: