include llama.cpp since there were some edits

llama.cpp/ggml/src/ggml-hexagon/htp/allreduce-ops.h

@@ -0,0 +1,63 @@
+#ifndef ALLREDUCE_OPS_H
+#define ALLREDUCE_OPS_H
+
+#include <stdint.h>
+#include <stddef.h>
+#include <stdbool.h>
+
+#define HTP_ALLREDUCE_MAX_RANKS 4
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum htp_allreduce_kernel_type {
+    HTP_ALLREDUCE_KERNEL_UNSUPPORTED = 0,
+    HTP_ALLREDUCE_KERNEL_DMA_1D,
+    HTP_ALLREDUCE_KERNEL_DMA_2D,
+};
+
+enum htp_allreduce_mode {
+    HTP_ALLREDUCE_FULL           = 0, // all ranks reduce full tensor
+    HTP_ALLREDUCE_SHARDED_FANOUT = 1, // each rank reduces 1/N shard and fans out to all buffers
+};
+
+static inline size_t htp_allreduce_vtcm_buffer_count(
+    uint32_t n_ranks,
+    uint32_t n_threads,
+    bool has_add,
+    bool is_row_bcast
+) {
+    return (size_t) (n_ranks + 1) * n_threads + (has_add ? (is_row_bcast ? 1 : n_threads) : 0);
+}
+
+struct htp_allreduce_kernel_params {
+    uint8_t kernel_type;
+    int32_t rank;
+    int32_t n_ranks;
+    int32_t n_threads;
+    int32_t block_elems;          // 1D: block_elems, 2D: block_rows
+    int32_t elems_per_thread;     // 1D: nelem_per_thread, 2D: nrows_per_thread
+    int32_t vtcm_size_per_thread;
+    int32_t vtcm_size;
+    int32_t ne0;
+    int32_t ne1;
+    int32_t row_size_aligned;
+    int32_t rank_elem_start;
+    int32_t rank_nelem;
+    int32_t n_dsts;
+    int32_t is_row_bcast;
+    int32_t mode;                 // enum htp_allreduce_mode: FULL or SHARDED_FANOUT
+};
+
+#if defined(__cplusplus)
+static_assert(sizeof(struct htp_allreduce_kernel_params) <= 128, "htp_allreduce_kernel_params is too large for kernel_params blob");
+#else
+_Static_assert(sizeof(struct htp_allreduce_kernel_params) <= 128, "htp_allreduce_kernel_params is too large for kernel_params blob");
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* ALLREDUCE_OPS_H */