nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 33 Packages Projects Releases Wiki Activity

hexagon: general DMA and Binary Op fixes for large strides (#20918)

Browse Source 
* hex-dma: make chained dma the default to handle newer models

This also includes some new instrumentation that we can remove later.

* hexagon: add uint32 dump helper

* hexagon: use single-page VTCM allocation to avoid issues with large gather ops in ssm-conv

ssm-conv uses HVX gather instruction and that instruction cannot handle cases where the base+offset
spans page boundaries.

* hexagon: update ssm-conv to make base-addr compute a bit easier to read

* hex-dma: use 1d mode for reshaping, it supports sizes up to 24-bits (>16MB)

* hex-bin: fix incorrect stride logic

* hexagon: make sure repack buffs are dumped for verbose > 2

* hex-bin: consistently use dma_queue_push even for dummy dst transactions

* hex-dma: start using 2d-wide mode on v75 and up

The removes the need to deal with the 16-bit limitaion for the strides.

* hex-bin: cleanup kernel selection logic

* hex-bin: cleanup binary op core and fix transposed tensor handling

* snapdragon: update run-bench to use larger ubatch and fa-on
This commit is contained in:
Max Krasnyansky
2026-03-23 15:33:49 -07:00
committed by GitHub
parent 1fb2290a51
commit 7cadbfce10
10 changed files with 369 additions and 330 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-hexagon/ggml-hexagon.cpp
+6 -6
View File
@@ -461,7 +461,7 @@ static void repack_row_q4x4x2(uint8_t * y, const block_q4_0 * x, int64_t k) {
d[7] = x[i * 8 + 7].d;
}
if (opt_verbose > 1) {
if (opt_verbose > 2) {
for (int i = 0; i < nb; i++) {
dump_packed_block_q4x4x2(y, i, k);
}
@@ -480,7 +480,7 @@ static void unpack_row_q4x4x2(block_q4_0 * x, const uint8_t * y, int64_t k) {
const uint8_t * y_q = y + 0; // quants first
const uint8_t * y_d = y + qrow_size; // then scales
if (opt_verbose > 1) {
if (opt_verbose > 2) {
for (int i = 0; i < nb; i++) {
dump_packed_block_q4x4x2(y, i, k);
}
@@ -796,7 +796,7 @@ static void repack_row_q8x4x2(uint8_t * y, const block_q8_0 * x, int64_t k) {
d[7] = x[i * 8 + 7].d;
}
if (opt_verbose > 1) {
if (opt_verbose > 2) {
for (int i = 0; i < nb; i++) {
dump_packed_block_q8x4x2(y, i, k);
}
@@ -814,7 +814,7 @@ static void unpack_row_q8x4x2(block_q8_0 * x, const uint8_t * y, int64_t k) {
const uint8_t * y_q = y + 0; // quants first
const uint8_t * y_d = y + qrow_size; // then scales
if (opt_verbose > 1) {
if (opt_verbose > 2) {
for (int i = 0; i < nb; i++) {
dump_packed_block_q8x4x2(y, i, k);
}
@@ -1149,7 +1149,7 @@ static void repack_row_mxfp4x4x2(uint8_t * y, const block_mxfp4 * x, int64_t k)
e[7] = x[i * 8 + 7].e;
}
if (opt_verbose > 1) {
if (opt_verbose > 2) {
for (int i = 0; i < nb; i++) {
dump_packed_block_mxfp4x4x2(y, i, k);
}
@@ -1168,7 +1168,7 @@ static void unpack_row_mxfp4x4x2(block_mxfp4 * x, const uint8_t * y, int64_t k)
const uint8_t * y_q = y + 0; // quants first
const uint8_t * y_e = y + qrow_size; // then scales
if (opt_verbose > 1) {
if (opt_verbose > 2) {
for (int i = 0; i < nb; i++) {
dump_packed_block_mxfp4x4x2(y, i, k);
}