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

hexagon: support for OP_CPY, host buffers now optional, hvx-utils refactoring and optimizations (#18822)

Browse Source 
* hexagon: disable repack buffers if host buffers are disabled, improved handling of env vars

* hexagon: add support for OP_CPY fp16/fp32 -> fp16/fp32

Factore out all hvx_copy functions into hvx-copy.h header and reduced code duplication.
Update HTP ops infra to support OP_CPY

* hexagon: cleanup and refactor hex/hvx/htp headers and helper libs

hex is basically all scalar/core platform stuff (L2, DMA, basic utils)
hvx is all hvx related utils, helpers, etc
htp is higher level stuff like Ops, etc

hvx-utils library got a nice round of cleanup and refactoring to reduce duplication

use hvx_vec_store_a where possible

* hexagon: refactor HVX sigmoid functions to hvx-sigmoid.h

Moved sigmoid and tanh vector functions from hvx-utils.h to a new header
hvx-sigmoid.h. Implemented aligned and unaligned variants for sigmoid
array processing using a macro pattern similar to hvx-copy.h. Updated
act-ops.c to use the new aligned variant hvx_sigmoid_f32_aa. Removed
unused hvx-sigmoid.c.

* hexagon: factor out hvx-sqrt.h

* hexagon: mintor update to hvx-utils.h

* hexagon: remove spurios log

* hexagon: factor out and optimize hvx_add/sub/mul

* hexagon: remove _opt variants of add/sub/mul as they simply fully aligned versions

* hexagon: refactor reduction functions to hvx-reduce.h

Moved `hvx_self_max_f32` and `hvx_self_sum_f32` from `hvx-utils.h`/`.c` to `hvx-reduce.h`.
Renamed them to `hvx_reduce_max_f32` and `hvx_reduce_sum_f32`.
Added aligned (`_a`) and unaligned (`_u`) variants and used macros to unify logic.
Updated `softmax-ops.c` to use the new functions.

* hexagon: refactor the rest of arithmetic functions to hvx-arith.h

Moved `hvx_sum_of_squares_f32`, `hvx_min_scalar_f32`, and `hvx_clamp_scalar_f32` from `hvx-utils.c/h` to `hvx-arith.h`. Implemented aligned/unaligned variants (`_aa`, `_au`, etc.) and used macros to reduce code duplication. Updated `hvx_min_scalar_f32` and `hvx_clamp_scalar_f32` to use `dst, src, ..., n` argument order. Updated call sites in `act-ops.c`.

Refactor Hexagon HVX arithmetic functions (min, clamp) to hvx-arith.h

Moved `hvx_min_scalar_f32` and `hvx_clamp_scalar_f32` from `hvx-utils.c/h` to `hvx-arith.h`. Implemented aligned/unaligned variants (`_aa`, `_au`, etc.) and used macros to reduce code duplication. Updated these functions to use `dst, src, ..., n` argument order and updated call sites in `act-ops.c`. `hvx_sum_of_squares_f32` remains in `hvx-utils.c` as requested.

* hexagon: refactor hvx_sum_of_squares_f32

- Modify `hvx_sum_of_squares_f32` in `ggml/src/ggml-hexagon/htp/hvx-reduce.h` to use `dst, src` signature.
- Implement `_a` (aligned) and `_u` (unaligned) variants for `hvx_sum_of_squares_f32`.
- Update `hvx_reduce_loop_body` macro to support both returning and storing results via `finalize_op`.
- Update existing reduction functions in `hvx-reduce.h` to use the updated macro.
- Update `rms_norm_htp_f32` in `ggml/src/ggml-hexagon/htp/unary-ops.c` to match the new signature.

* hexagon: use hvx_splat instead of memset

* hexagon: consistent use of f32/f16 in all function names to match the rest of GGML

* hexagon: fix hvx_copy_f16_f32 on v75 and older

* hexagon: update readme to include GGML_HEXAGON_EXPERIMENTAL

* scripts: update snapdragon/adb scripts to enable host param
This commit is contained in:
Max Krasnyansky
2026-01-14 21:46:12 -08:00
committed by GitHub
parent 36f0132464
commit cff777f226
46 changed files with 2940 additions and 3153 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-hexagon/htp/binary-ops.c
+20 -37
View File
@@ -2,36 +2,25 @@
#pragma clang diagnostic ignored "-Wunused-function"
#pragma clang diagnostic ignored "-Wunused-but-set-variable"
#ifdef HTP_DEBUG
# define FARF_HIGH 1
#endif
#include <HAP_farf.h>
#include <HAP_mem.h>
#include <HAP_perf.h>
#include <HAP_ps.h>
#include <hexagon_protos.h>
#include <hexagon_types.h>
#include <math.h>
#include <qurt_thread.h>
#include <string.h>
#include "hex-dma.h"
#include "hvx-utils.h"
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "htp-ctx.h"
#include "htp-dma.h"
#include "htp-msg.h"
#include "htp-ops.h"
#include "hvx-utils.h"
#include "ops-utils.h"
typedef void (*hvx_elemwise_f32_func)(const uint8_t * src0,
const uint8_t * src1,
uint8_t * data_dst,
const int num_elems);
typedef void (*hvx_elemwise_f32_func)(uint8_t * data_dst, const uint8_t * src0, const uint8_t * src1, const uint32_t num_elems);
static hvx_elemwise_f32_func func_table_HVX[] = { hvx_mul_f32, hvx_add_f32, hvx_sub_f32 };
static hvx_elemwise_f32_func func_table_HVX_opt[] = { hvx_mul_f32_opt, hvx_add_f32_opt, hvx_sub_f32_opt };
static hvx_elemwise_f32_func func_table_HVX_opt[] = { hvx_mul_f32_aa, hvx_add_f32_aa, hvx_sub_f32_aa };
#define htp_binary_preamble \
const struct htp_tensor * src0 = &octx->src0; \
@@ -98,9 +87,8 @@ static void binary_job_f32_per_thread(struct htp_ops_context * octx,
int is_aligned = 1;
int opt_path = 0;
if ((0 == htp_is_aligned((void *) src0->data, VLEN)) || (0 == htp_is_aligned((void *) src1->data, VLEN)) ||
(0 == htp_is_aligned((void *) dst->data, VLEN))) {
FARF(HIGH, "binary-f32: unaligned addresses in elementwise op, possibly slower execution\n");
if ((0 == hex_is_aligned((void *) src0->data, VLEN)) || (0 == hex_is_aligned((void *) src1->data, VLEN)) ||
(0 == hex_is_aligned((void *) dst->data, VLEN))) {
is_aligned = 0;
}
if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) {
@@ -130,24 +118,24 @@ static void binary_job_f32_per_thread(struct htp_ops_context * octx,
const uint8_t * restrict src1_ptr = data_src1 + i13 * nb13 + i12 * nb12 + i11 * src1_row_size;
if (ir + 1 < src0_end_row) {
htp_l2fetch(src0_ptr + ne00, 1, src0_row_size, src0_row_size);
hex_l2fetch(src0_ptr + ne00, src0_row_size, src0_row_size, 1);
if (src1_row_size == src0_row_size) {
htp_l2fetch(src1_ptr, 1, src1_row_size, src1_row_size);
hex_l2fetch(src1_ptr, src1_row_size, src1_row_size, 1);
}
}
const uint32_t nr0 = ne00 / ne10;
if (nr0 > 1) {
if ((1 == is_aligned) && (nr0 == ne00)) {
hvx_bcast_fp32_a(spad_data_th, *(float *) src1_ptr, nr0);
hvx_splat_f32_a(spad_data_th, *(float *) src1_ptr, nr0);
} else {
for (uint32_t r = 0; r < nr0; r++) {
memcpy(spad_data_th + r * nb11, (const uint8_t *) src1_ptr, nb11);
}
}
func_HVX((const uint8_t *) src0_ptr, (const uint8_t *) spad_data_th, (uint8_t *) dst_ptr, ne00);
func_HVX((uint8_t *) dst_ptr, (const uint8_t *) src0_ptr, (const uint8_t *) spad_data_th, ne00);
} else {
func_HVX((const uint8_t *) src0_ptr, (const uint8_t *) src1_ptr, (uint8_t *) dst_ptr, ne00);
func_HVX((uint8_t *) dst_ptr, (const uint8_t *) src0_ptr, (const uint8_t *) src1_ptr, ne00);
}
src0_ptr += src0_row_size;
@@ -185,11 +173,6 @@ static void binary_add_id_job_f32_per_thread(struct htp_ops_context * octx,
uint64_t t1, t2;
t1 = HAP_perf_get_qtimer_count();
if ((0 == htp_is_aligned((void *) src0->data, VLEN)) || (0 == htp_is_aligned((void *) src1->data, VLEN)) ||
(0 == htp_is_aligned((void *) dst->data, VLEN))) {
FARF(HIGH, "add-id-f32: unaligned addresses, possibly slower execution\n");
}
const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
uint8_t * restrict data_dst = (uint8_t *) dst->data;
@@ -210,9 +193,9 @@ static void binary_add_id_job_f32_per_thread(struct htp_ops_context * octx,
const float * restrict src1_ptr = (const float *) (data_src1 + 0 + 0 + i11 * nb11);
if (ir + 1 < src0_end_row) {
htp_l2fetch(src0_ptr + ne00, 1, src0_row_size, src0_row_size);
hex_l2fetch(src0_ptr + ne00, src0_row_size, src0_row_size, 1);
if (src1_row_size == src0_row_size) {
htp_l2fetch(src1_ptr + ne10, 1, src1_row_size, src1_row_size);
hex_l2fetch(src1_ptr + ne10, src1_row_size, src1_row_size, 1);
}
}
@@ -221,9 +204,9 @@ static void binary_add_id_job_f32_per_thread(struct htp_ops_context * octx,
for (uint32_t r = 0; r < nr0; r++) {
memcpy(spad_data + r * nb10, (const uint8_t *) src1_ptr, nb10);
}
func_HVX((const uint8_t *) src0_ptr, (const uint8_t *) spad_data, (uint8_t *) dst_ptr, ne00);
func_HVX((uint8_t *) dst_ptr, (const uint8_t *) src0_ptr, (const uint8_t *) spad_data, ne00);
} else {
func_HVX((const uint8_t *) src0_ptr, (const uint8_t *) src1_ptr, (uint8_t *) dst_ptr, ne00);
func_HVX((uint8_t *) dst_ptr, (const uint8_t *) src0_ptr, (const uint8_t *) src1_ptr, ne00);
}
}
@@ -299,9 +282,9 @@ static int execute_op_binary_f32(struct htp_ops_context * octx) {
const size_t dst_row_size = dst->nb[1];
// VTCM scratchpads for all tensors
octx->dst_spad.size = htp_round_up(dst_row_size, 128) * n_threads;
octx->src0_spad.size = htp_round_up(src0_row_size, 128) * n_threads;
octx->src1_spad.size = htp_round_up(src1_row_size, 128) * n_threads;
octx->dst_spad.size = hex_round_up(dst_row_size, 128) * n_threads;
octx->src0_spad.size = hex_round_up(src0_row_size, 128) * n_threads;
octx->src1_spad.size = hex_round_up(src1_row_size, 128) * n_threads;
size_t spad_size = octx->src0_spad.size + octx->src1_spad.size + octx->dst_spad.size;