Superminaren/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-24 13:37:01 +02:00
Code Issues Packages Projects Releases Wiki Activity

hexagon: add support for GEGLU_QUICK (#29114)

Browse Source
This commit is contained in:
Aparna M P
2026-09-19 09:48:07 -07:00
committed by GitHub
parent 7d4b92bb9b
commit 851cb34f21
4 changed files with 71 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-hexagon/ggml-hexagon.cpp
+2
View File
@@ -5749,6 +5749,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
case GGML_GLU_OP_SWIGLU_OAI: return HTP_OP_GLU_SWIGLU_OAI;
case GGML_GLU_OP_SWIGLU_CLAMP: return HTP_OP_GLU_SWIGLU_CLAMP;
case GGML_GLU_OP_GEGLU: return HTP_OP_GLU_GEGLU;
case GGML_GLU_OP_GEGLU_QUICK: return HTP_OP_GLU_GEGLU_QUICK;
default: break;
}
break;
@@ -6769,6 +6770,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
case GGML_GLU_OP_SWIGLU_OAI:
case GGML_GLU_OP_SWIGLU_CLAMP:
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_GEGLU_QUICK:
supp = ggml_hexagon_supported_activations(sess, op);
break;
default:
ggml/src/ggml-hexagon/htp/act-ops.c
+67 -1
View File
@@ -312,6 +312,45 @@ static inline void hvx_geglu_f32_aa(uint8_t * restrict dst, const uint8_t * rest
}
}
static inline void hvx_geglu_quick_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
assert((unsigned long) dst % 128 == 0);
assert((unsigned long) src0 % 128 == 0);
assert((unsigned long) src1 % 128 == 0);
HVX_Vector * restrict vdst = (HVX_Vector *) dst;
const HVX_Vector * restrict vsrc0 = (const HVX_Vector *) src0;
const HVX_Vector * restrict vsrc1 = (const HVX_Vector *) src1;
const uint32_t epv = 128 / sizeof(float);
const uint32_t nvec = n / epv;
const uint32_t nloe = n % epv;
const HVX_Vector v_scale = hvx_vec_splat_f32(1.702f);
const HVX_Vector v_one = hvx_vec_splat_f32(1.0f);
const HVX_Vector v_max_exp = hvx_vec_splat_f32(87.0f);
const HVX_Vector v_min_exp = hvx_vec_splat_f32(-87.0f);
uint32_t i = 0;
_Pragma("unroll(4)")
for (; i < nvec; i++) {
HVX_Vector x = vsrc0[i];
HVX_Vector g = vsrc1[i];
HVX_Vector scaled_x = hvx_vec_mul_f32_f32(x, v_scale);
HVX_Vector sigmoid_x = hvx_vec_fast_sigmoid_f32_guard_2it(scaled_x, v_one, v_max_exp, v_min_exp);
vdst[i] = hvx_vec_mul_f32_f32(hvx_vec_mul_f32_f32(x, sigmoid_x), g);
}
if (nloe) {
HVX_Vector x = vsrc0[i];
HVX_Vector g = vsrc1[i];
HVX_Vector scaled_x = hvx_vec_mul_f32_f32(x, v_scale);
HVX_Vector sigmoid_x = hvx_vec_fast_sigmoid_f32_guard_2it(scaled_x, v_one, v_max_exp, v_min_exp);
HVX_Vector result = hvx_vec_mul_f32_f32(hvx_vec_mul_f32_f32(x, sigmoid_x), g);
hvx_vec_store_a((void *) &vdst[i], nloe * sizeof(float), result);
}
}
// geglu(x, g) = gelu(x) * g
static void geglu_f32(const float * restrict src0,
const float * restrict src1,
@@ -329,6 +368,23 @@ static void geglu_f32(const float * restrict src0,
}
}
// geglu_quick(x, g) = x * sigmoid(1.702 * x) * g
static void geglu_quick_f32(const float * restrict src0,
const float * restrict src1,
float * restrict dst,
const uint32_t num_rows,
const struct htp_act_context * actx) {
htp_glu_op_preamble;
for (uint32_t ib = 0; ib < num_rows; ib++) {
const uint8_t * restrict src0_ptr = (const uint8_t *) src0 + (ib * src0_row_size_aligned);
const uint8_t * restrict src1_ptr = (const uint8_t *) src1 + (ib * src1_row_size_aligned);
uint8_t * restrict dst_ptr = (uint8_t *) dst + (ib * dst_row_size_aligned);
hvx_geglu_quick_f32_aa(dst_ptr, src0_ptr, src1_ptr, nc);
}
}
#define DEFINE_GLU_PER_THREAD(NAME, OP_STR, CORE_EXPR) \
static void glu_##NAME##_f32_per_thread(unsigned int nth, unsigned int ith, void * data) { \
struct htp_act_context * actx = (struct htp_act_context *) data; \
@@ -433,6 +489,7 @@ DEFINE_GLU_PER_THREAD(swiglu, "swiglu-f32", swiglu_f32(src0_spad, src1_spad, dst
DEFINE_GLU_PER_THREAD(swiglu_oai, "swiglu-oai-f32", swiglu_oai_f32(src0_spad, src1_spad, dst_spad, block_size, actx))
DEFINE_GLU_PER_THREAD(swiglu_clamp, "swiglu-clamp-f32", swiglu_clamp_f32(src0_spad, src1_spad, dst_spad, block_size, actx))
DEFINE_GLU_PER_THREAD(geglu, "geglu-f32", geglu_f32(src0_spad, src1_spad, dst_spad, block_size, actx))
DEFINE_GLU_PER_THREAD(geglu_quick, "geglu-quick-f32", geglu_quick_f32(src0_spad, src1_spad, dst_spad, block_size, actx))
static int execute_op_activations_f32(struct htp_ops_context * octx) {
const struct htp_tensor * src0 = octx->src[0];
@@ -467,6 +524,11 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
act_op_func = (worker_callback_t)glu_geglu_f32_per_thread;
op_type = "geglu-f32";
break;
case HTP_OP_GLU_GEGLU_QUICK:
act_op_func = (worker_callback_t)glu_geglu_quick_f32_per_thread;
op_type = "geglu-quick-f32";
break;
default:
FARF(ERROR, "Unsupported activations Op %u\n", octx->op);
return HTP_STATUS_NO_SUPPORT;
@@ -570,7 +632,11 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
const uint8_t * data_src0 = (const uint8_t *) src0->data;
const uint8_t * data_src1 = src1 ? (const uint8_t *) src1->data : NULL;
if (!src1 && (octx->op == HTP_OP_GLU_SWIGLU || octx->op == HTP_OP_GLU_SWIGLU_OAI || octx->op == HTP_OP_GLU_SWIGLU_CLAMP || octx->op == HTP_OP_GLU_GEGLU)) {
if (!src1 && (octx->op == HTP_OP_GLU_SWIGLU ||
octx->op == HTP_OP_GLU_SWIGLU_OAI ||
octx->op == HTP_OP_GLU_SWIGLU_CLAMP ||
octx->op == HTP_OP_GLU_GEGLU ||
octx->op == HTP_OP_GLU_GEGLU_QUICK)) {
const int32_t swapped = octx->op_params[1];
data_src1 = data_src0;
actx.src1_row_size = actx.src0_row_size;
ggml/src/ggml-hexagon/htp/htp-ops.h
+1
View File
@@ -71,6 +71,7 @@ enum htp_op_code {
HTP_OP_GLU_SWIGLU,
HTP_OP_GLU_SWIGLU_OAI,
HTP_OP_GLU_GEGLU,
HTP_OP_GLU_GEGLU_QUICK,
HTP_OP_SOFTMAX,
HTP_OP_ADD_ID,
HTP_OP_ROPE,
ggml/src/ggml-hexagon/htp/main.c
+1
View File
@@ -828,6 +828,7 @@ static int execute_op(struct htp_ops_context * octx) {
case HTP_OP_GLU_SWIGLU_OAI:
case HTP_OP_GLU_SWIGLU_CLAMP:
case HTP_OP_GLU_GEGLU:
case HTP_OP_GLU_GEGLU_QUICK:
return op_activations(octx);
case HTP_OP_SOFTMAX: