aboutsummaryrefslogtreecommitdiff
path: root/libffwd/cuda.h
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-20 20:10:56 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-21 05:57:19 +0200
commit6b969d25ce136eaf695222bd1a10bff1ecac0220 (patch)
treec23afb50c28b9deaddae5c6526a74ee11cf2d865 /libffwd/cuda.h
parentd4bdb8dae753462f621507d5fc3582e0e53da35c (diff)
downloadffwd-6b969d25ce136eaf695222bd1a10bff1ecac0220.tar.gz
ffwd-6b969d25ce136eaf695222bd1a10bff1ecac0220.zip
all: split Makefile and reorg library/tools
Diffstat (limited to 'libffwd/cuda.h')
-rw-r--r--libffwd/cuda.h80
1 files changed, 80 insertions, 0 deletions
diff --git a/libffwd/cuda.h b/libffwd/cuda.h
new file mode 100644
index 0000000..8be1dd4
--- /dev/null
+++ b/libffwd/cuda.h
@@ -0,0 +1,80 @@
+/*
+ * ffwd_cuda.h - CUDA/cuBLAS backend
+ */
+
+#ifndef FFWD_CUDA_H
+#define FFWD_CUDA_H
+
+/* Public-API export annotation. The libraries are built with
+ * -fvisibility=hidden, so only declarations carrying FFWD_API are exported
+ * from the shared library; everything else stays internal. */
+#ifndef FFWD_API
+# if defined(__GNUC__)
+# define FFWD_API __attribute__((visibility("default")))
+# else
+# define FFWD_API
+# endif
+#endif
+
+#include "ffwd.h"
+
+typedef struct ffwd_cuda_ctx ffwd_cuda_ctx_t;
+typedef struct ffwd_cuda_late_ctx ffwd_cuda_late_ctx_t;
+
+FFWD_API int ffwd_cuda_set_fast_gemm(const char *mode);
+FFWD_API int ffwd_cuda_set_weights_bf16(int on);
+
+FFWD_API ffwd_cuda_ctx_t *ffwd_cuda_load(const char *model_dir);
+FFWD_API void ffwd_cuda_free(ffwd_cuda_ctx_t *ctx);
+
+/*
+ * Late-interaction CUDA path. Encoding (transformer forward + projection) runs
+ * on the GPU; MaxSim scoring stays on the host, where the grouped-GEMM scorer is
+ * much faster than a device graph for this small op. The encoders therefore copy
+ * the projected token vectors back to host memory.
+ */
+FFWD_API ffwd_cuda_late_ctx_t *ffwd_cuda_late_load(const char *model_dir);
+FFWD_API void ffwd_cuda_late_free(ffwd_cuda_late_ctx_t *ctx);
+FFWD_API const ffwd_config_t *ffwd_cuda_late_config(const ffwd_cuda_late_ctx_t *ctx);
+FFWD_API int ffwd_cuda_late_token_dim(const ffwd_cuda_late_ctx_t *ctx);
+
+/* Encode token_ids into out_vectors[n_tokens, token_dim] (host). */
+FFWD_API int ffwd_cuda_late_encode_tokens(ffwd_cuda_late_ctx_t *ctx,
+ const int *token_ids,
+ int n_tokens,
+ int normalize,
+ float *out_vectors);
+
+/*
+ * Encode n_docs documents in one packed forward and pack each document's kept
+ * token vectors back-to-back into out_vectors[total_keep, token_dim] (host) in
+ * document order. out_offsets[n_docs + 1] receives the prefix sum of kept counts
+ * (the layout ffwd_late_maxsim_batch consumes). Returns 0 on success.
+ */
+FFWD_API int ffwd_cuda_late_encode_docs(ffwd_cuda_late_ctx_t *ctx,
+ const int *const *doc_ids,
+ const int *n_tokens,
+ const int *const *keep,
+ const int *n_keep,
+ int n_docs,
+ int normalize,
+ float *out_vectors,
+ int *out_offsets);
+
+FFWD_API const ffwd_config_t *ffwd_cuda_config(const ffwd_cuda_ctx_t *ctx);
+
+FFWD_API int ffwd_cuda_encode_into(ffwd_cuda_ctx_t *ctx,
+ const int *token_ids,
+ int n_tokens,
+ float *out_embedding);
+FFWD_API float *ffwd_cuda_encode(ffwd_cuda_ctx_t *ctx, const int *token_ids, int n_tokens);
+FFWD_API int ffwd_cuda_encode_batch(ffwd_cuda_ctx_t *ctx,
+ const ffwd_input_t *inputs,
+ int batch,
+ float *out_embeddings);
+FFWD_API int ffwd_cuda_encode_spans_batch(ffwd_cuda_ctx_t *ctx,
+ const ffwd_context_input_t *inputs,
+ int batch,
+ float *out_embeddings);
+
+#endif /* FFWD_CUDA_H */