nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 34 Packages Projects Releases Wiki Activity

hexagon: Add ARGSORT, DIV, SQR, SQRT, SUM_ROWS, GEGLU (#19406)

Browse Source 
* hexagon: add ARGSORT op

Co-authored-by: Yarden Tal <yardent@qti.qualcomm.com>

* hexagon: argsort reject tensors with huge rows for now

* Adding support for DIV,SQR,SQRT,SUM_ROWS ops in hexagon backend

* hexagon : Add GEGLU op

* hexagon: fix editor config check

* hexagon: rewrite and optimize binary ops ADD/SUB/MUL/DIV/ADD_ID to use DMA

---------

Co-authored-by: Yarden Tal <yardent@qti.qualcomm.com>
Co-authored-by: Manohara Hosakoppa Krishnamurthy <mhosakop@qti.qualcomm.com>
This commit is contained in:
Max Krasnyansky
2026-02-10 23:21:12 -08:00
committed by GitHub
co-authored by Yarden Tal Manohara Hosakoppa Krishnamurthy
parent 8ee538ce73
commit 73cd5e1b97
17 changed files with 1925 additions and 407 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-hexagon/ggml-hexagon.cpp
+106 -7
View File
@@ -1935,11 +1935,6 @@ static bool ggml_hexagon_supported_binary(const struct ggml_hexagon_session * se
return false;
}
// TODO: add support for non-contigiuos tensors
if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1) || !ggml_is_contiguous(dst)) {
return false;
}
return true;
}
@@ -1991,6 +1986,25 @@ static bool ggml_hexagon_supported_unary(const struct ggml_hexagon_session * ses
return true;
}
static bool ggml_hexagon_supported_sum_rows(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;
if (!hex_supported_src0_type(src0->type)) {
return false;
}
if (!hex_supported_dst_type(dst->type)) {
return false;
}
// TODO: add support for non-contigiuos tensors
if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(dst)) {
return false;
}
return true;
}
static bool ggml_hexagon_supported_activations(const struct ggml_hexagon_session * sess,
const struct ggml_tensor * op) {
const struct ggml_tensor * src0 = op->src[0];
@@ -2111,6 +2125,26 @@ static bool ggml_hexagon_supported_get_rows(const struct ggml_hexagon_session *
return true;
}
static bool ggml_hexagon_supported_argsort(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
const struct ggml_tensor * src0 = op->src[0]; // values
const struct ggml_tensor * dst = op; // indices
if (src0->type != GGML_TYPE_F32) {
return false;
}
if (dst->type != GGML_TYPE_I32) {
return false;
}
if (src0->ne[0] > (16*1024)) {
// reject tensors with huge rows for now
return false;
}
return true;
}
static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
const int32_t * op_params = &op->op_params[0];
@@ -2278,6 +2312,9 @@ static inline size_t init_binary_req(htp_general_req * req, dspqueue_buffer * bu
case GGML_OP_SUB:
req->op = HTP_OP_SUB;
break;
case GGML_OP_DIV:
req->op = HTP_OP_DIV;
break;
default:
GGML_ABORT("ggml-hex: binary : unsupported op: %d\n", t->op);
break;
@@ -2316,6 +2353,17 @@ static inline size_t init_get_rows_req(htp_general_req * req, dspqueue_buffer *
return n_bufs;
}
static inline size_t init_argsort_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
req->op = HTP_OP_ARGSORT;
memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
size_t n_bufs = 0;
n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
n_bufs += htp_req_buff_init(&req->dst, &bufs[n_bufs], t, DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
return n_bufs;
}
template <bool _is_src0_constant>
static inline size_t init_binary_id_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
switch (t->op) {
@@ -2370,6 +2418,16 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
supported = true;
break;
case GGML_OP_SQR:
req->op = HTP_OP_SQR;
supported = true;
break;
case GGML_OP_SQRT:
req->op = HTP_OP_SQRT;
supported = true;
break;
case GGML_OP_UNARY:
if (ggml_get_unary_op(t) == GGML_UNARY_OP_SILU) {
req->op = HTP_OP_UNARY_SILU;
@@ -2387,6 +2445,9 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
} else if (ggml_get_glu_op(t) == GGML_GLU_OP_SWIGLU_OAI) {
req->op = HTP_OP_GLU_SWIGLU_OAI;
supported = true;
} else if (ggml_get_glu_op(t) == GGML_GLU_OP_GEGLU) {
req->op = HTP_OP_GLU_GEGLU;
supported = true;
}
break;
@@ -2411,6 +2472,17 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
return n_bufs;
}
static inline size_t init_sum_rows_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
req->op = HTP_OP_SUM_ROWS;
size_t n_bufs = 0;
n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
n_bufs += htp_req_buff_init(&req->dst, &bufs[n_bufs], t, DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
return n_bufs;
}
static inline size_t init_rope_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
req->op = HTP_OP_ROPE;
@@ -2519,6 +2591,7 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
case GGML_OP_MUL:
case GGML_OP_ADD:
case GGML_OP_SUB:
case GGML_OP_DIV:
ggml_hexagon_dispatch_op<init_binary_req<false>>(sess, node, flags);
break;
case GGML_OP_ADD_ID:
@@ -2528,6 +2601,13 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
case GGML_OP_SCALE:
ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
break;
case GGML_OP_SQR:
case GGML_OP_SQRT:
ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
break;
case GGML_OP_SUM_ROWS:
ggml_hexagon_dispatch_op<init_sum_rows_req>(sess, node, flags);
break;
case GGML_OP_UNARY:
if ((ggml_get_unary_op(node) == GGML_UNARY_OP_SILU) ||
(ggml_get_unary_op(node) == GGML_UNARY_OP_GELU)) {
@@ -2536,7 +2616,8 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
break;
case GGML_OP_GLU:
if ((ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU) ||
(ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU_OAI)) {
(ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU_OAI) ||
(ggml_get_glu_op(node) == GGML_GLU_OP_GEGLU)) {
ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
}
break;
@@ -2564,6 +2645,10 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
ggml_hexagon_dispatch_op<init_cpy_req>(sess, node, flags);
break;
case GGML_OP_ARGSORT:
ggml_hexagon_dispatch_op<init_argsort_req>(sess, node, flags);
break;
default:
GGML_ABORT("\nggml-hex: graph-compute %s is not supported\n", ggml_op_desc(node));
}
@@ -2916,6 +3001,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
case GGML_OP_MUL:
case GGML_OP_ADD:
case GGML_OP_SUB:
case GGML_OP_DIV:
supp = ggml_hexagon_supported_binary(sess, op);
break;
@@ -2928,6 +3014,15 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
supp = ggml_hexagon_supported_unary(sess, op);
break;
case GGML_OP_SQR:
case GGML_OP_SQRT:
supp = ggml_hexagon_supported_unary(sess, op);
break;
case GGML_OP_SUM_ROWS:
supp = ggml_hexagon_supported_sum_rows(sess, op);
break;
case GGML_OP_SOFT_MAX:
supp = ggml_hexagon_supported_softmax(sess, op);
break;
@@ -2943,7 +3038,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
case GGML_OP_GLU:
{
const auto glu_op = ggml_get_glu_op(op);
if ((glu_op == GGML_GLU_OP_SWIGLU) || (glu_op == GGML_GLU_OP_SWIGLU_OAI)) {
if ((glu_op == GGML_GLU_OP_SWIGLU) || (glu_op == GGML_GLU_OP_SWIGLU_OAI) || (glu_op == GGML_GLU_OP_GEGLU)) {
supp = ggml_hexagon_supported_activations(sess, op);
}
break;
@@ -2968,6 +3063,10 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
supp = ggml_hexagon_supported_cpy(sess, op);
break;
case GGML_OP_ARGSORT:
supp = ggml_hexagon_supported_argsort(sess, op);
break;
default:
break;
}