whispercpp 1.3.7 → 1.3.8

data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c

@@ -23,6 +23,7 @@
 #define HTP_ROPE_TYPE_NORMAL 0
 #define HTP_ROPE_TYPE_NEOX   2
 #define HTP_ROPE_TYPE_MROPE  8
+#define HTP_ROPE_TYPE_VISION 24
 #define HTP_ROPE_TYPE_IMROPE 40
 
 #define HTP_ROPE_SPAD_NROWS  16
@@ -70,7 +71,9 @@ struct htp_rope_context {
     struct htp_ops_context * octx;
 
     size_t src0_row_size;
+    size_t src0_row_stride;
     size_t dst_row_size;
+    size_t dst_row_stride;
     size_t src0_row_size_aligned;
     size_t dst_row_size_aligned;
     size_t theta_cache_offset;
@@ -210,6 +213,7 @@ static __attribute__((noinline)) void mrope_cache_init(const float    pos_t,
                              const float    pos_e,
                              const int32_t  sections[4],
                              const bool     is_imrope,
+                             const bool     indep_sects,
                              const float    freq_scale,
                              const float *  freq_factors,
                              float *        corr_dims,
@@ -231,6 +235,14 @@ static __attribute__((noinline)) void mrope_cache_init(const float    pos_t,
         const float ff     = freq_factors ? freq_factors[i0 / 2] : 1.0f;
         const int   sector = (i0 / 2) % sect_dims;
 
+        if (indep_sects) {
+            // Reset theta when crossing into a new section.
+            if      (sector == 0)           { theta_t = pos_t; }
+            else if (sector == sections[0]) { theta_h = pos_h; }
+            else if (sector == sec_w)       { theta_w = pos_w; }
+            else if (sector == sec_e)       { theta_e = pos_e; }
+        }
+
         float theta;
         if (is_imrope) {
             // Interleaved: sector mod 3 selects component
@@ -422,6 +434,17 @@ static void inline rope_neox_f32(struct htp_rope_context * rctx, uint8_t * restr
     }
 }
 
+static void inline rope_vision_f32(struct htp_rope_context * rctx, uint8_t * restrict dst, uint8_t * restrict src,
+                   uint32_t nr, uint32_t ne0, const float * restrict theta_cache) {
+    #pragma unroll(4)
+    for (uint32_t i = 0; i < nr; i++) {
+        float * d = (float *) (dst + i * rctx->dst_row_size_aligned);
+        float * s = (float *) (src + i * rctx->src0_row_size_aligned);
+
+        hvx_rope_neox_f32_aa(d, s, ne0, theta_cache);
+    }
+}
+
 static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
     struct htp_rope_context * rctx = (struct htp_rope_context *) data;
     struct htp_ops_context * octx = rctx->octx;
@@ -447,8 +470,9 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
     uint64_t tt = HAP_perf_get_qtimer_count();
 
     const int32_t mode    = rctx->mode;
-    // MROPE and IMROPE use NEOX-style pairing for the rotation
+    // MROPE, IMROPE and VISION use NEOX-style pairing for the rotation
     const bool    is_neox = (mode & HTP_ROPE_TYPE_NEOX) || (mode & HTP_ROPE_TYPE_MROPE);
+    const bool    is_vision = (mode == HTP_ROPE_TYPE_VISION);
 
     // VTCM setup
     uint8_t * src0_spad_base = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
@@ -496,8 +520,10 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
 
                     const uint8_t * src_addr = (const uint8_t *) src0->data + i3 * nb03 + i2 * nb02 + pi1 * nb01;
                           uint8_t * src_spad = src0_spad_base + pr * rctx->src0_row_size_aligned;
-                    dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src_spad, src_addr),
-                        rctx->src0_row_size_aligned, rctx->src0_row_size, pnr);
+
+                    // Copy only the row payload while striding the DDR source
+                    dma_queue_push(dma_queue, dma_make_ptr(src_spad, src_addr),
+                        rctx->src0_row_size_aligned, rctx->src0_row_stride, rctx->src0_row_size, pnr);
 
                     // FARF(HIGH, "rope-prefetch %u: pr %u i1 %u i2 %u i3 %u src-spad %p src-addr %p pnr %u", ith, pir, pi1, i2, i3, src_spad, src_addr, pnr);
                 }
@@ -516,7 +542,7 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
                             (float) pos[i2 + ne2],
                             (float) pos[i2 + ne2 * 2],
                             (float) pos[i2 + ne2 * 3],
-                            rctx->sections, is_imrope,
+                            rctx->sections, is_imrope, is_vision,
                             rctx->freq_scale, freq_factors, rctx->corr_dims,
                             ne0, rctx->ext_factor, rctx->attn_factor,
                             theta_cache, rctx->theta_scale);
@@ -542,14 +568,19 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
                     // FARF(HIGH, "rope-compute %u: ir %u i1 %u i2 %u i3 %u src-spad %p cnr %u : usec %u", ith, ir, i1, i2, i3, src_spad, cnr,
                     //         (unsigned) HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - rctx->t_start));
 
-                    if (is_neox) {
+                    if (is_vision) {
+                        rope_vision_f32(rctx, dst_spad, src_spad, cnr, ne0, theta_cache);
+                    } else if (is_neox) {
                         rope_neox_f32(rctx, dst_spad, src_spad, cnr, ne0, theta_cache);
                     } else {
                         rope_basic_f32(rctx, dst_spad, src_spad, cnr, ne0, theta_cache);
                     }
 
                     uint8_t * dst_addr = (uint8_t *) dst->data + i3 * nb3 + i2 * nb2 + i1 * nb1;
-                    dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr(dst_addr, dst_spad), rctx->dst_row_size, rctx->dst_row_size_aligned, cnr);
+
+                    // Write only the row payload while striding the DDR dst
+                    dma_queue_push(dma_queue, dma_make_ptr(dst_addr, dst_spad),
+                        rctx->dst_row_stride, rctx->dst_row_size_aligned, rctx->dst_row_size, cnr);
 
                     // Prefetch more rows (if any)
                     if ((cr + HTP_ROPE_SPAD_NROWS) < nrows) {
@@ -558,8 +589,8 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
                         uint32_t pir = ir + HTP_ROPE_SPAD_NROWS;
 
                         const uint8_t * src_addr = (const uint8_t *) src0->data + i3 * nb03 + i2 * nb02 + pi1 * nb01;
-                        dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src_spad, src_addr),
-                            rctx->src0_row_size_aligned, rctx->src0_row_size, pnr);
+                        dma_queue_push(dma_queue, dma_make_ptr(src_spad, src_addr),
+                            rctx->src0_row_size_aligned, rctx->src0_row_stride, rctx->src0_row_size, pnr);
 
                         // FARF(HIGH, "rope-prefetch %u: pr %u i1 %u i2 %u i3 %u src-spad %p src-addr %p pnr %u", ith, pir, pi1, i2, i3, src_spad, src_addr, pnr);
                     }
@@ -598,12 +629,14 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
     const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
     const uint32_t n_threads = MIN(octx->n_threads, src0_nrows);
 
-    const size_t src0_row_size = src0->nb[1];
-    const size_t dst_row_size  = dst->nb[1];
+    const size_t src0_row_size   = src0->ne[0] * sizeof(float);
+    const size_t src0_row_stride = src0->nb[1];
+    const size_t dst_row_size    = dst->ne[0] * sizeof(float);
+    const size_t dst_row_stride  = dst->nb[1];
 
     // Aligned row sizes for VTCM
     const size_t src0_row_size_aligned    = hex_round_up(src0_row_size, VLEN);
-    const size_t dst_row_size_aligned     = hex_round_up(dst_row_size, VLEN);
+    const size_t dst_row_size_aligned     = hex_round_up(dst_row_stride, VLEN);
     const size_t theta_cache_size_aligned = hex_round_up(src0->ne[0] * sizeof(float), 256);
 
     // Calculate spad sizes per thread
@@ -652,8 +685,10 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
 
     rope_corr_dims(rctx.n_dims, rctx.n_ctx_orig, rctx.freq_base, rctx.beta_fast, rctx.beta_slow, rctx.corr_dims);
 
-    rctx.src0_row_size = src0_row_size;
-    rctx.dst_row_size  = dst_row_size;
+    rctx.src0_row_size   = src0_row_size;
+    rctx.src0_row_stride = src0_row_stride;
+    rctx.dst_row_size    = dst_row_size;
+    rctx.dst_row_stride  = dst_row_stride;
     rctx.src0_row_size_aligned = src0_row_size_aligned;
     rctx.dst_row_size_aligned  = dst_row_size_aligned;
     rctx.theta_cache_offset    = theta_cache_size_aligned;

data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c

@@ -183,24 +183,25 @@ static inline void hvx_transpose_32x32_f32(HVX_Vector m[32]) {
 // transposed into VTCM.
 //
 // VTCM layouts (per thread):
-//   src1_T : {d_inner_per_thread, d_conv}   — staged once per launch (small).
-//   src0_T : {d_inner_tile,     ncs}        — staged per d_inner-tile.
+//   src1_T : {d_inner_stride, d_conv}       - staged once per launch (small).
+//   src0_T : {d_inner_tile,     ncs}        - staged per d_inner-tile.
 //
 // d_inner_tile is chosen so that per-thread VTCM stays under the budget.
 // Each thread iterates ceil(d_inner_per_thread d_inner_tile) tiles serially.
 #define HTP_SSM_CONV_VTCM_BUDGET (1u << 20) // 1 MiB per thread
 
-// Scalar transpose: src1 {d_conv, d_inner} (DDR) -> {d_inner_per_thread, d_conv} (VTCM)
+// Scalar transpose: src1 {d_conv, d_inner} (DDR) -> {d_inner_stride, d_conv} (VTCM)
 static inline void transpose_src1(const float * src1_data,
                                   uint32_t      src1_stride_inner,
                                   uint32_t      i1_off,
                                   uint32_t      d_inner_per_thread,
+                                  uint32_t      d_inner_stride,
                                   uint32_t      d_conv,
                                   float *       src1_T) {
     for (uint32_t i = 0; i < d_inner_per_thread; ++i) {
         const float * src_row = src1_data + (i1_off + i) * src1_stride_inner;
         for (uint32_t j = 0; j < d_conv; ++j) {
-            src1_T[j * d_inner_per_thread + i] = src_row[j];
+            src1_T[j * d_inner_stride + i] = src_row[j];
         }
     }
 }
@@ -280,6 +281,7 @@ static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void
     }
 
     const uint32_t d_inner_per_thread = ir1 - ir0;
+    const uint32_t d_inner_stride     = scctx->nrows_per_thread;
     const uint32_t d_inner_tile       = scctx->d_inner_tile;
 
     const float * src0_data = (const float *) src0->data;
@@ -290,8 +292,8 @@ static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void
     float * src0_T = (float *)(octx->src0_spad.data + ith * octx->src0_spad.size_per_thread);
     float * src1_T = (float *)(octx->src1_spad.data + ith * octx->src1_spad.size_per_thread);
 
-    // Stage src1 weights once into VTCM in {d_inner_per_thread, d_conv} layout.
-    transpose_src1(src1_data, src1_stride_inner, ir0, d_inner_per_thread, d_conv, src1_T);
+    // Stage src1 weights once into VTCM in {d_inner_stride, d_conv} layout.
+    transpose_src1(src1_data, src1_stride_inner, ir0, d_inner_per_thread, d_inner_stride, d_conv, src1_T);
 
     const uint32_t C_TILE = VLEN_FP32;
 
@@ -314,7 +316,7 @@ static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void
                     HVX_Vector acc = hvx_vec_splat_f32(0.0f);
                     for (uint32_t j = 0; j < d_conv; ++j) {
                         HVX_Vector x = *(const HVX_Vector *) (src0_T + (t + j) * d_inner_tile + cb);
-                        HVX_Vector w = *(const HVX_Vector *) (src1_T + j * d_inner_per_thread + tile_off + cb);
+                        HVX_Vector w = *(const HVX_Vector *) (src1_T + j * d_inner_stride + tile_off + cb);
                         acc          = Q6_Vqf32_vadd_Vqf32Vqf32(acc, Q6_Vqf32_vmpy_VsfVsf(x, w));
                     }
                     HVX_Vector res = Q6_Vsf_equals_Vqf32(acc);
@@ -362,8 +364,7 @@ int op_ssm_conv_f32(struct htp_ops_context * octx) {
             use_hvx = 1;
         }
 
-        scctx.nrows_per_thread  = (d_inner + n_threads - 1) / n_threads;
-        scctx.nrows_per_thread += (scctx.nrows_per_thread & 1);
+        scctx.nrows_per_thread = hex_round_up((d_inner + n_threads - 1) / n_threads, VLEN_FP32);
 
         const uint32_t d_inner_per_thread = scctx.nrows_per_thread;
         const uint32_t ncs                = src0->ne[0];

data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c

@@ -1,6 +1,9 @@
 #include "worker-pool.h"
+#include "hex-utils.h"
 
 #include <qurt.h>
+#include <qurt_hvx.h>
+
 #include <stdatomic.h>
 #include <stdint.h>
 #include <stdio.h>
@@ -9,7 +12,6 @@
 
 #include "HAP_farf.h"
 
-#define WORKER_THREAD_STACK_SZ  (2 * 16384)
 #define LOWEST_USABLE_QURT_PRIO (254)
 
 struct worker_pool_s;
@@ -42,17 +44,27 @@ static void worker_pool_main(void * context) {
     FARF(HIGH, "worker-pool: thread %u started", me->id);
 
     unsigned int prev_seqn = 0;
+    unsigned int poll_cnt  = WORKER_POOL_POLL_COUNT;
     while (!atomic_load(&pool->killed)) {
         unsigned int seqn = atomic_load(&pool->seqn);
         if (seqn == prev_seqn) {
-            // Nothing to do
+            // drop HVX context while spinning
+            if (poll_cnt > 1 && poll_cnt == WORKER_POOL_POLL_COUNT) {
+                qurt_hvx_unlock();
+            }
+            if (--poll_cnt) {
+                hex_pause();
+                continue;
+            }
             qurt_futex_wait(&pool->seqn, prev_seqn);
+            poll_cnt = WORKER_POOL_POLL_COUNT;
             continue;
         }
 
-        // New job
         prev_seqn = seqn;
+        poll_cnt  = WORKER_POOL_POLL_COUNT;
 
+        // New job
         unsigned int n = atomic_load(&pool->n_jobs);
         unsigned int i = atomic_fetch_add(&pool->next_job, 1);
         if (i >= n) {

data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.h

@@ -24,9 +24,17 @@ typedef struct {
     void *            data;
 } worker_pool_job_t;
 
+#define WORKER_THREAD_STACK_SZ  (2 * 16384)
+
 /// Maximum supported number of worker threads.
 #define MAX_NUM_WORKERS 10
 
+#if __HVX_ARCH__ > 79
+#define WORKER_POOL_POLL_COUNT  2000
+#else
+#define WORKER_POOL_POLL_COUNT  1
+#endif
+
 // Initialize worker pool.
 WORKERPOOL_API AEEResult worker_pool_init(worker_pool_context_t * context, uint32_t n_threads);
 

data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h

@@ -5,10 +5,13 @@
 #include "ggml-backend-impl.h"
 #include "ggml-common.h"
 
+#include <algorithm>
 #include <string>
 #include <vector>
 #include <stdio.h>
 #include "htp-ops.h"
+#include "htp/matmul-ops.h"
+#include "htp/flash-attn-ops.h"
 
 struct htp_opnode {
     ggml_tensor * node = nullptr;
@@ -17,6 +20,13 @@ struct htp_opnode {
 
     htp_op_code opcode = HTP_OP_INVALID;
 
+    std::vector<ggml_tensor *> extra_dsts;
+
+    int32_t kernel_params[HTP_OP_MAX_KERN_PARAMS] = {0};
+
+    htp_opnode(ggml_tensor * node = nullptr, std::vector<ggml_tensor *> fused = {}, htp_op_code opcode = HTP_OP_INVALID, std::vector<ggml_tensor *> extra_dsts = {})
+        : node(node), fused(std::move(fused)), opcode(opcode), extra_dsts(std::move(extra_dsts)) {}
+
     ggml_op op() const {
         return node->op;
     }
@@ -25,6 +35,26 @@ struct htp_opnode {
         return fused.empty() ? node : fused.back();
     }
 
+    void add_fused(ggml_tensor * t, bool extra_dst = false) {
+        fused.push_back(t);
+        if (extra_dst) {
+            extra_dsts.push_back(t);
+        }
+    }
+
+    std::vector<const ggml_tensor *> get_outputs() const {
+        std::vector<const ggml_tensor *> res;
+        if (extra_dsts.empty()) {
+            res.push_back(dst());
+        } else {
+            res.push_back(node);
+            for (const auto * x : extra_dsts) {
+                res.push_back(x);
+            }
+        }
+        return res;
+    }
+
     const ggml_tensor * src0() const {
         return node->src[0];
     }
@@ -37,10 +67,6 @@ struct htp_opnode {
         return ggml_op_is_empty(node->op);
     }
 
-    void add_fused(ggml_tensor * t) {
-        fused.push_back(t);
-    }
-
     bool stackable() const {
         switch (this->op()) {
             case GGML_OP_MUL_MAT:
@@ -131,87 +157,117 @@ struct htp_opformat {
     char types[16 * GGML_MAX_SRC];
     char buffs[64 * GGML_MAX_SRC];
     char names[64 * GGML_MAX_SRC];
+    char kparams[128];
 
-    int format_tensor_dims(char * str, const struct ggml_tensor * t) {
+    int format_tensor_dims(char * str, size_t max_size, const struct ggml_tensor * t) {
         if (!t) {
-            return sprintf(str, "NONE");
+            return snprintf(str, max_size, "NONE");
         }
         if (t->ne[2] == 1 && t->ne[3] == 1) {
-            return sprintf(str, "%d:%d", (int) t->ne[0], (int) t->ne[1]);
+            return snprintf(str, max_size, "%d:%d", (int) t->ne[0], (int) t->ne[1]);
         } else {
-            return sprintf(str, "%d:%d:%d:%d", (int) t->ne[0], (int) t->ne[1], (int) t->ne[2], (int) t->ne[3]);
+            return snprintf(str, max_size, "%d:%d:%d:%d", (int) t->ne[0], (int) t->ne[1], (int) t->ne[2], (int) t->ne[3]);
         }
     }
 
-    void format_op_dims(char * str, const htp_opnode & node) {
+    void format_op_dims(char * str, size_t max_size, const htp_opnode & node) {
         char * p = str;
+        char * p_end = str + max_size;
         auto inputs = node.get_inputs();
 
         if (!inputs.empty()) {
-            p += format_tensor_dims(p, inputs[0]);
+            p += std::min((size_t)format_tensor_dims(p, p_end - p, inputs[0]), (size_t)(p_end - p));
 
             for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += format_tensor_dims(p, inputs[i]);
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)format_tensor_dims(p, p_end - p, inputs[i]), (size_t)(p_end - p));
+                }
             }
 
-            p += sprintf(p, " -> ");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
         }
 
         char self[64];
-        format_tensor_dims(self, node.dst());
-        p += sprintf(p, "%s", self);
+        format_tensor_dims(self, sizeof(self), node.dst());
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", self), (size_t)(p_end - p));
+        }
     }
 
-    int format_tensor_strides(char * str, const struct ggml_tensor * t) {
+    int format_tensor_strides(char * str, size_t max_size, const struct ggml_tensor * t) {
         if (!t) {
-            return sprintf(str, "NONE");
+            return snprintf(str, max_size, "NONE");
         }
         const char * c = ggml_is_contiguous(t) ? "" : "!";
 
         if (t->ne[2] == 1 && t->ne[3] == 1) {
-            return sprintf(str, "%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], c);
+            return snprintf(str, max_size, "%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], c);
         } else {
-            return sprintf(str, "%zu:%zu:%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], (size_t) t->nb[2], (size_t) t->nb[3], c);
+            return snprintf(str, max_size, "%zu:%zu:%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], (size_t) t->nb[2], (size_t) t->nb[3], c);
         }
     }
 
-    void format_op_strides(char * str, const htp_opnode & node) {
+    void format_op_strides(char * str, size_t max_size, const htp_opnode & node) {
         char * p = str;
+        char * p_end = str + max_size;
         auto inputs = node.get_inputs();
 
         if (!inputs.empty()) {
-            p += format_tensor_strides(p, inputs[0]);
+            p += std::min((size_t)format_tensor_strides(p, p_end - p, inputs[0]), (size_t)(p_end - p));
 
             for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += format_tensor_strides(p, inputs[i]);
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)format_tensor_strides(p, p_end - p, inputs[i]), (size_t)(p_end - p));
+                }
             }
 
-            p += sprintf(p, " -> ");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
         }
 
         char self[64];
-        format_tensor_strides(self, node.dst());
-        p += sprintf(p, "%s", self);
+        format_tensor_strides(self, sizeof(self), node.dst());
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", self), (size_t)(p_end - p));
+        }
     }
 
-    void format_op_types(char * str, const htp_opnode & node) {
+    void format_op_types(char * str, size_t max_size, const htp_opnode & node) {
         char * p = str;
+        char * p_end = str + max_size;
         auto inputs = node.get_inputs();
 
         if (!inputs.empty()) {
-            p += sprintf(p, "%s", inputs[0] ? ggml_type_name(inputs[0]->type) : "NONE");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[0] ? ggml_type_name(inputs[0]->type) : "NONE"), (size_t)(p_end - p));
+            }
 
             for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += sprintf(p, "%s", inputs[i] ? ggml_type_name(inputs[i]->type) : "NONE");
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[i] ? ggml_type_name(inputs[i]->type) : "NONE"), (size_t)(p_end - p));
+                }
             }
 
-            p += sprintf(p, " -> ");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
         }
 
-        p += sprintf(p, "%s", ggml_type_name(node.dst()->type));
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", ggml_type_name(node.dst()->type)), (size_t)(p_end - p));
+        }
     }
 
     const char * tensor_buff_name(const struct ggml_tensor * t) {
@@ -221,51 +277,113 @@ struct htp_opformat {
         return "NONE";
     }
 
-    void format_op_buffs(char * str, const htp_opnode & node) {
+    void format_op_buffs(char * str, size_t max_size, const htp_opnode & node) {
         char * p = str;
+        char * p_end = str + max_size;
         auto inputs = node.get_inputs();
 
         if (!inputs.empty()) {
-            p += sprintf(p, "%s", tensor_buff_name(inputs[0]));
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, "%s", tensor_buff_name(inputs[0])), (size_t)(p_end - p));
+            }
 
             for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += sprintf(p, "%s", tensor_buff_name(inputs[i]));
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, "%s", tensor_buff_name(inputs[i])), (size_t)(p_end - p));
+                }
             }
 
-            p += sprintf(p, " -> ");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
         }
 
-        p += sprintf(p, "%s", tensor_buff_name(node.dst()));
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", tensor_buff_name(node.dst())), (size_t)(p_end - p));
+        }
     }
 
-    void format_op_names(char * str, const htp_opnode & node) {
+    void format_op_names(char * str, size_t max_size, const htp_opnode & node) {
         char * p = str;
+        char * p_end = str + max_size;
         auto inputs = node.get_inputs();
 
         if (!inputs.empty()) {
-            p += sprintf(p, "%s", inputs[0] ? inputs[0]->name : "NONE");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[0] ? inputs[0]->name : "NONE"), (size_t)(p_end - p));
+            }
 
             for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += sprintf(p, "%s", inputs[i] ? inputs[i]->name : "NONE");
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[i] ? inputs[i]->name : "NONE"), (size_t)(p_end - p));
+                }
             }
 
-            p += sprintf(p, " -> ");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
         }
 
-        p += sprintf(p, "%s", node.dst()->name);
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", node.dst()->name), (size_t)(p_end - p));
+        }
+    }
+    void format_kernel_params(char * str, size_t max_size, const htp_opnode & node) {
+        if (node.opcode == HTP_OP_MUL_MAT || node.opcode == HTP_OP_MUL_MAT_ID ||
+            node.opcode == HTP_OP_MUL_MAT_QKV || node.opcode == HTP_OP_MUL_MAT_FFN ||
+            node.opcode == HTP_OP_MUL_MAT_ADD) {
+            const auto * kparams = (const struct htp_mm_kernel_params *) node.kernel_params;
+            const char * path = "unknown";
+            int32_t type = kparams->kernel_type;
+            if (type == HTP_MM_KERNEL_HMX_2D || type == HTP_MM_KERNEL_HMX_F16_BATCHED) {
+                path = "hmx-tiled";
+            } else if (type == HTP_MM_KERNEL_HVX_F16_F16_VTCM || type == HTP_MM_KERNEL_HVX_F32_F32_VTCM ||
+                       type == HTP_MM_KERNEL_HVX_QUANT_ROW    || type == HTP_MM_KERNEL_HVX_QUANT_BLOCK) {
+                path = "hvx-tiled";
+            } else if (type == HTP_MM_KERNEL_HVX_F16_F16_DDR  || type == HTP_MM_KERNEL_HVX_F16_F32_DDR ||
+                       type == HTP_MM_KERNEL_HVX_F32_F32_DDR  || type == HTP_MM_KERNEL_HVX_F32_F16_DDR ||
+                       type == HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT) {
+                path = "hvx-flat";
+            }
+            snprintf(str, max_size, "%s vtcm %d", path, (int) kparams->vtcm_size);
+        } else if (node.opcode == HTP_OP_FLASH_ATTN_EXT) {
+            const auto * kparams = (const struct htp_fa_kernel_params *) node.kernel_params;
+            const char * path = "unknown";
+            int32_t type = kparams->kernel_type;
+            if (type == HTP_FA_KERNEL_HMX) {
+                path = kparams->u.hmx.pipeline ? "hmx-pipe" : "hmx-seq";
+            } else if (type == HTP_FA_KERNEL_HVX) {
+                path = "hvx";
+            }
+            snprintf(str, max_size, "%s vtcm %d", path, (int) kparams->vtcm_size);
+        } else {
+            snprintf(str, max_size, "----");
+        }
     }
 
     void format(const htp_opnode & node) {
-        format_op_dims(dims, node);
-        format_op_strides(strides, node);
-        format_op_types(types, node);
-        format_op_buffs(buffs, node);
-        format_op_names(names, node);
+        format_op_dims(dims, sizeof(dims), node);
+        format_op_strides(strides, sizeof(strides), node);
+        format_op_types(types, sizeof(types), node);
+        format_op_buffs(buffs, sizeof(buffs), node);
+        format_op_names(names, sizeof(names), node);
+        format_kernel_params(kparams, sizeof(kparams), node);
     }
 
-    htp_opformat() {}
+    htp_opformat() {
+        strides[0] = '\0';
+        dims[0]    = '\0';
+        types[0]   = '\0';
+        buffs[0]   = '\0';
+        names[0]   = '\0';
+        kparams[0] = '\0';
+    }
     htp_opformat(const htp_opnode & node) { format(node); }
 };
 

data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf

@@ -14,8 +14,6 @@ Drivers_Dir = 13
 1 = %DiskId%
 
 [SourceDisksFiles]
-libggml-htp-v68.so = 1
-libggml-htp-v69.so = 1
 libggml-htp-v73.so = 1
 libggml-htp-v75.so = 1
 libggml-htp-v79.so = 1
@@ -28,8 +26,6 @@ ExcludeFromSelect = *
 CopyFiles=Drivers_Dir
 
 [Drivers_Dir]
-libggml-htp-v68.so,,,0x10 ;COPYFLG_NO_OVERWRITE
-libggml-htp-v69.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v73.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v75.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v79.so,,,0x10 ;COPYFLG_NO_OVERWRITE

data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt

@@ -130,6 +130,9 @@ if (GGML_HIP_EXPORT_METRICS)
     set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -Rpass-analysis=kernel-resource-usage --save-temps")
 endif()
 
+# Fast math for HIP, like CUDA's -use_fast_math. Not -ffast-math: that implies -ffinite-math-only, which breaks ggml's INFINITY masking and produces NaNs.
+set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -funsafe-math-optimizations")
+
 if (NOT GGML_CUDA_FA)
     add_compile_definitions(GGML_CUDA_NO_FA)
 endif()
@@ -155,3 +158,5 @@ if (GGML_HIP_RCCL)
 endif()
 
 target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas)
+
+target_compile_options(ggml-hip PRIVATE "$<$<COMPILE_LANGUAGE:HIP>:-ffast-math;-fno-finite-math-only>")