arm: Implement a NEON version of 422 h264_h_loop_filter_chroma

Previously, the 420 version was used even for 422.

This fixes occasional checkasm failures.

Signed-off-by: Martin Storsjö <martin@martin.st>

libavcodec/arm/h264dsp_init_arm.c

@@ -33,6 +33,8 @@ void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, int stride, int alpha,
                                        int beta, int8_t *tc0);
 void ff_h264_h_loop_filter_chroma_neon(uint8_t *pix, int stride, int alpha,
                                        int beta, int8_t *tc0);
+void ff_h264_h_loop_filter_chroma422_neon(uint8_t *pix, int stride, int alpha,
+                                          int beta, int8_t *tc0);
 
 void ff_weight_h264_pixels_16_neon(uint8_t *dst, int stride, int height,
                                    int log2_den, int weight, int offset);
@@ -76,7 +78,11 @@ static av_cold void h264dsp_init_neon(H264DSPContext *c, const int bit_depth,
         c->h264_v_loop_filter_luma   = ff_h264_v_loop_filter_luma_neon;
         c->h264_h_loop_filter_luma   = ff_h264_h_loop_filter_luma_neon;
         c->h264_v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_neon;
-        c->h264_h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_neon;
+
+        if (chroma_format_idc <= 1)
+            c->h264_h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_neon;
+        else
+            c->h264_h_loop_filter_chroma = ff_h264_h_loop_filter_chroma422_neon;
 
         c->weight_h264_pixels_tab[0] = ff_weight_h264_pixels_16_neon;
         c->weight_h264_pixels_tab[1] = ff_weight_h264_pixels_8_neon;

libavcodec/arm/h264dsp_neon.S

@@ -237,6 +237,7 @@ function ff_h264_h_loop_filter_chroma_neon, export=1
         h264_loop_filter_start
 
         sub             r0,  r0,  #2
+h_loop_filter_chroma420:
         vld1.32         {d18[0]}, [r0], r1
         vld1.32         {d16[0]}, [r0], r1
         vld1.32         {d0[0]},  [r0], r1
@@ -271,6 +272,24 @@ function ff_h264_h_loop_filter_chroma_neon, export=1
         bx              lr
 endfunc
 
+function ff_h264_h_loop_filter_chroma422_neon, export=1
+        h264_loop_filter_start
+        push            {r4, lr}
+        add             r4,  r0,  r1
+        add             r1,  r1,  r1
+        sub             r0,  r0,  #2
+
+        bl              h_loop_filter_chroma420
+
+        ldr             r12, [sp, #8]
+        ldr             r12, [r12]
+        vmov.32         d24[0], r12
+        sub             r0,  r4,  #2
+
+        bl              h_loop_filter_chroma420
+        pop             {r4, pc}
+endfunc
+
 @ Biweighted prediction
 
 .macro  biweight_16     macs, macd