Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml-hexagon: gelu operation (#17921)

Browse Source 
* feat: inital support for gelu using sigmoid approximation

* snapshot: faster gelu using polynomial approximation

* test: disable l2-block prefetch in polynomail approximation

* Revert "test: disable l2-block prefetch in polynomail approximation"

This reverts commit 72339994d45b2bed887e79994403c378d90b62b5.

* Revert "snapshot: faster gelu using polynomial approximation"

This reverts commit 2a787a61d11f9e63e5943a2e6d134b2f0c402ace.

* debug: temporarily disable unnecessary log message for debug purpose

* Feat: optiized unaligned sigmoid_f32

* Feat: larger l2prefetch block

* feat: apply unaligned-load optimization on mul and mul_scalar

* Revert "debug: temporarily disable unnecessary log message for debug purpose"

This reverts commit 84f2f23aa9f17e2fa826db969cd825d0ab192995.

* refactor: cleanup commented unused code

* chore: reformat code with clang-formatter to pass cli test

* Revert "chore: reformat code with clang-formatter to pass cli test"

This reverts commit 952877ec24732b12010c7fa7ed3fc8de4b74e718.

* fix: fix loop overflow

* chore: fix formating ci error
This commit is contained in:
Shouyu
2025-12-17 10:39:32 -08:00
committed by GitHub
parent 4301e27319
commit 4470a0764a
6 changed files with 256 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-hexagon/htp/hvx-utils.h
+57
View File
@@ -265,12 +265,16 @@ static inline void hvx_bcast_fp32_a(uint8_t * restrict dst, float elem, uint32_t
}
}
/* Return whether 'n' elements from vector are in the one chunk of 'chunk_size'. */
static __attribute__((always_inline)) int32_t is_in_one_chunk(void * addr, uint32_t n, uint32_t chunk_size) {
uint32_t left_off = (size_t) addr & (chunk_size - 1);
uint32_t right_off = left_off + n;
return right_off <= chunk_size;
}
static void hvx_vec_dump_fp16_n(char * pref, HVX_Vector v, uint32_t n) {
HVX_VectorAlias u = { .v = v };
@@ -994,6 +998,59 @@ static inline void hvx_fast_sigmoid_f32(const uint8_t * restrict src, uint8_t *
}
}
static inline void hvx_sigmoid_f32(const uint8_t * restrict src, uint8_t * restrict dst, const int num_elems){
int step_of_1 = num_elems >> 5; // divby 32, because 32 float = 128 bytes per HVX vector
int leftover = num_elems - (step_of_1 * VLEN_FP32);
int32_t leftover_size = leftover * sizeof(float);
static const float kMinExp = -87.f; // 0
static const float kMaxExp = 87.f; // 1
const HVX_Vector one = hvx_vec_splat_fp32(1.f);
const HVX_Vector max_exp = hvx_vec_splat_fp32(kMaxExp);
const HVX_Vector min_exp = hvx_vec_splat_fp32(kMinExp);
const float *input = (float *)src;
float *output = (float *)dst;
HVX_Vector * input_v_ptr = (HVX_Vector *) input;
HVX_UVector * output_v_ptr = (HVX_UVector *) output;
HVX_Vector slinep;
HVX_Vector slinec;
HVX_Vector sline;
slinep = *input_v_ptr++;
#pragma unroll(4)
for (int i = step_of_1 - 1; i > 0; i--) {
slinec = *input_v_ptr++;
sline = Q6_V_valign_VVR(slinec, slinep, (size_t) input);
*((HVX_UVector *) (output_v_ptr++)) = hvx_vec_fast_sigmoid_fp32_guard(sline, one, max_exp, min_exp);
/* Prepare slinep for next iteration */
slinep = slinec;
}
if (step_of_1 > 0) {
slinec = htp_is_aligned(input_v_ptr, 128) && leftover == 0 ? slinep : *input_v_ptr++;
sline = Q6_V_valign_VVR(slinec, slinep, (size_t) input);
*((HVX_UVector *) (output_v_ptr++)) = hvx_vec_fast_sigmoid_fp32_guard(sline, one, max_exp, min_exp);
;
slinep = slinec;
}
if (leftover > 0) {
slinec = (is_in_one_chunk(input_v_ptr, leftover_size, 128) ? slinep : *input_v_ptr++);
sline = Q6_V_valign_VVR(slinec, slinep, (size_t) input);
HVX_Vector sout = hvx_vec_fast_sigmoid_fp32_guard(sline, one, max_exp, min_exp);
hvx_vec_store_u(output_v_ptr, leftover_size, sout);
}
}
float hvx_sum_of_squares_f32(const uint8_t * restrict src, const int num_elems);
void hvx_mul_f32(const uint8_t * restrict src0,
const uint8_t * restrict src1,