PDF Kernel (10700): Improved performance on AMD GPU by using shared memory for the scratch buffer and disable inlining to save spilling

Inspired by https://github.com/reger-men/hashcat/blob/6.2.4/OpenCL/m10700-optimized.cl

OpenCL/inc_vendor.h

@@ -130,6 +130,17 @@
 #define DECLSPEC
 #endif
 
+#define INLINE0 __attribute__ ((noinline))
+#define INLINE1 __attribute__ ((inline))
+
+#if defined IS_AMD && defined IS_GPU
+#define INLINE INLINE0
+#elif defined IS_HIP
+#define INLINE INLINE0
+#else
+#define INLINE
+#endif
+
 /**
  * AMD specific
  */

OpenCL/m10700-optimized.cl

@@ -315,7 +315,7 @@ DECLSPEC void make_w_with_offset (ctx_t *ctx, const u32 W_len, const u32 offset,
   }
 }
 
-DECLSPEC u32 do_round (LOCAL_AS u32 *sc, const u32 *pw, const u32 pw_len, ctx_t *ctx, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
+DECLSPEC INLINE u32 do_round (LOCAL_AS u32 *sc, const u32 *pw, const u32 pw_len, ctx_t *ctx, SHM_TYPE u32 *s_te0, SHM_TYPE u32 *s_te1, SHM_TYPE u32 *s_te2, SHM_TYPE u32 *s_te3, SHM_TYPE u32 *s_te4)
 {
   // make scratch buffer
 

docs/changes.txt

@@ -13,7 +13,7 @@
 ## Performance
 ##
 
-- PDF Kernel (10700): Improved performance on AMD GPU by using shared memory for the scratch buffer
+- PDF Kernel (10700): Improved performance on AMD GPU by using shared memory for the scratch buffer and disable inlining to save spilling
 
 ##
 ## Bugs