avcodec/ac3: Implement ac3_exponent_min for aarch64 NEON

Signed-off-by: Geoff Hill <geoff@geoffhill.org>
Signed-off-by: Martin Storsjö <martin@martin.st>

libavcodec/aarch64/ac3dsp_init_aarch64.c

@@ -25,6 +25,7 @@
 #include "libavcodec/ac3dsp.h"
 #include "config.h"
 
+void ff_ac3_exponent_min_neon(uint8_t *exp, int num_reuse_blocks, int nb_coefs);
 void ff_float_to_fixed24_neon(int32_t *dst, const float *src, size_t len);
 
 av_cold void ff_ac3dsp_init_aarch64(AC3DSPContext *c)
@@ -32,5 +33,6 @@ av_cold void ff_ac3dsp_init_aarch64(AC3DSPContext *c)
     int cpu_flags = av_get_cpu_flags();
     if (!have_neon(cpu_flags)) return;
 
+    c->ac3_exponent_min = ff_ac3_exponent_min_neon;
     c->float_to_fixed24 = ff_float_to_fixed24_neon;
 }

libavcodec/aarch64/ac3dsp_neon.S

@@ -21,6 +21,22 @@
 
 #include "libavutil/aarch64/asm.S"
 
+function ff_ac3_exponent_min_neon, export=1
+        cbz             w1, 3f
+1:      ld1             {v0.16b}, [x0]
+        mov             w3, w1
+        add             x4, x0, #256
+2:      ld1             {v1.16b}, [x4]
+        umin            v0.16b, v0.16b, v1.16b
+        add             x4, x4, #256
+        subs            w3, w3, #1
+        b.gt            2b
+        st1             {v0.16b}, [x0], #16
+        subs            w2, w2, #16
+        b.gt            1b
+3:      ret
+endfunc
+
 function ff_float_to_fixed24_neon, export=1
 1:      ld1             {v0.4s, v1.4s}, [x1], #32
         fcvtzs          v0.4s, v0.4s, #24

tests/checkasm/ac3dsp.c

@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences (ISCAS).
+ * Copyright (c) 2024 Geoff Hill <geoff@geoffhill.org>
  *
  * This file is part of FFmpeg.
  *
@@ -27,6 +28,14 @@
 
 #include "checkasm.h"
 
+#define randomize_exp(buf, len)        \
+    do {                               \
+        int i;                         \
+        for (i = 0; i < len; i++) {    \
+            buf[i] = (uint8_t)rnd();   \
+        }                              \
+    } while (0)
+
 #define randomize_float(buf, len)                               \
     do {                                                        \
         int i;                                                  \
@@ -36,6 +45,38 @@
         }                                                       \
     } while (0)
 
+static void check_ac3_exponent_min(AC3DSPContext *c) {
+#define MAX_COEFS 256
+#define MAX_CTXT 6
+#define EXP_SIZE (MAX_CTXT * MAX_COEFS)
+
+    LOCAL_ALIGNED_16(uint8_t, src, [EXP_SIZE]);
+    LOCAL_ALIGNED_16(uint8_t, v1, [EXP_SIZE]);
+    LOCAL_ALIGNED_16(uint8_t, v2, [EXP_SIZE]);
+    int n;
+
+    declare_func(void, uint8_t *, int, int);
+
+    for (n = 0; n < MAX_CTXT; ++n) {
+        if (check_func(c->ac3_exponent_min, "ac3_exponent_min_reuse%d", n)) {
+            randomize_exp(src, EXP_SIZE);
+
+            memcpy(v1, src, EXP_SIZE);
+            memcpy(v2, src, EXP_SIZE);
+
+            call_ref(v1, n, MAX_COEFS);
+            call_new(v2, n, MAX_COEFS);
+
+            if (memcmp(v1, v2, EXP_SIZE) != 0)
+                fail();
+
+            bench_new(v2, n, MAX_COEFS);
+        }
+    }
+
+    report("ac3_exponent_min");
+}
+
 static void check_float_to_fixed24(AC3DSPContext *c) {
 #define BUF_SIZE 1024
     LOCAL_ALIGNED_32(float, src, [BUF_SIZE]);
@@ -66,5 +107,6 @@ void checkasm_check_ac3dsp(void)
     AC3DSPContext c;
     ff_ac3dsp_init(&c);
 
+    check_ac3_exponent_min(&c);
     check_float_to_fixed24(&c);
 }