nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 30 Packages Projects Releases Wiki Activity

Hexagon: DAIG op (#22195)

Browse Source 
* hexagon: Add DIAG op

* hexagon: add HVX support and DMA double buffering

* hexagon: fix fatal error

* hexagon: remove as many pragma(s) as possible
This commit is contained in:
Shreya Jain
2026-04-21 14:16:04 -07:00
committed by GitHub
parent 2248799a58
commit 5a4cd6741f
6 changed files with 250 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-hexagon/ggml-hexagon.cpp
+28
View File
@@ -2596,6 +2596,29 @@ static bool ggml_hexagon_supported_cumsum(const struct ggml_hexagon_session * se
return true;
}
static bool ggml_hexagon_supported_diag(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * dst = op;
// diag only supports F32 currently
if (src0->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
return false;
}
// Input must have ne[1] == 1 (vector input)
if (src0->ne[1] != 1) {
return false;
}
// Output must be square in first two dimensions
if (dst->ne[0] != dst->ne[1] || dst->ne[0] != src0->ne[0]) {
return false;
}
GGML_UNUSED(sess);
return true;
}
static const char * ggml_backend_hexagon_name(ggml_backend_t backend) {
auto sess = static_cast<ggml_hexagon_session *>(backend->context);
return sess->c_name();
@@ -2632,6 +2655,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
case GGML_OP_ROPE: return HTP_OP_ROPE;
case GGML_OP_REPEAT: return HTP_OP_REPEAT;
case GGML_OP_CUMSUM: return HTP_OP_CUMSUM;
case GGML_OP_DIAG: return HTP_OP_DIAG;
case GGML_OP_UNARY:
switch (ggml_get_unary_op(t)) {
@@ -3159,6 +3183,10 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
supp = ggml_hexagon_supported_cumsum(sess, op);
break;
case GGML_OP_DIAG:
supp = ggml_hexagon_supported_diag(sess, op);
break;
default:
break;
}