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

hexagon: add f32 ssm_conv op (#20122)

Browse Source 
* hexagon: add ssm_conv op

* hexagon: hvx kernel is functional

* hexagon: improvements to ssm-conv hvx kernel

* hexagon: added dma to ssm-conv hvx kernel

* hexagon: ssm-conv dynamically compute gather scratchpad

* hex-ssm-conv: add local context and fix various issues (spad indexing, etc)

---------

Co-authored-by: Max Krasnyansky <maxk@qti.qualcomm.com>
This commit is contained in:
Todor Boinovski
2026-03-06 09:59:26 -08:00
committed by GitHub
co-authored by Max Krasnyansky
parent e68f2fb894
commit 34df42f7be
7 changed files with 456 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-hexagon/htp/main.c
+49
View File
@@ -757,6 +757,47 @@ static void proc_sum_rows_req(struct htp_context * ctx, struct htp_general_req *
send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
}
static void proc_ssm_conv_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
// We've written to the output buffer, we'd also need to flush it
rsp_bufs[0].fd = bufs[2].fd;
rsp_bufs[0].ptr = bufs[2].ptr;
rsp_bufs[0].offset = bufs[2].offset;
rsp_bufs[0].size = bufs[2].size;
rsp_bufs[0].flags = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER | // Flush HTP
DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT); // Invalidate CPU
// Setup OP context
struct htp_ops_context octx = { 0 };
octx.ctx = ctx;
octx.src0 = req->src0;
octx.src1 = req->src1;
octx.dst = req->dst;
octx.flags = req->flags;
octx.op = req->op;
memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
// Update data pointers
octx.src0.data = (uint32_t) bufs[0].ptr;
octx.src1.data = (uint32_t) bufs[1].ptr;
octx.dst.data = (uint32_t) bufs[2].ptr;
octx.n_threads = ctx->n_threads;
struct profile_data prof;
profile_start(&prof);
uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
if (vtcm_acquire(ctx) == AEE_SUCCESS) {
rsp_status = op_ssm_conv(&octx);
vtcm_release(ctx);
}
profile_stop(&prof);
send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
}
static void proc_activations_req(struct htp_context * ctx,
struct htp_general_req * req,
struct dspqueue_buffer * bufs,
@@ -1142,6 +1183,14 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
proc_argsort_req(ctx, &req, bufs);
break;
case HTP_OP_SSM_CONV:
if (n_bufs != 3) {
FARF(ERROR, "Bad ssm-conv-req buffer list");
continue;
}
proc_ssm_conv_req(ctx, &req, bufs);
break;
default:
FARF(ERROR, "Unknown Op %u", req.op);
break;