h264: miscellaneous micro-optimisations

[nihav.git] / nihav-itu / src / codecs / h264 / decoder_st.rs

diff --git a/nihav-itu/src/codecs/h264/decoder_st.rs b/nihav-itu/src/codecs/h264/decoder_st.rs
index 85deab9ed0fea8f9a613cff39d6f9616152991e0..7e8b83a8187777cc0a62a1ce51e2aaf250d90fa0 100644 (file)
--- a/nihav-itu/src/codecs/h264/decoder_st.rs
+++ b/nihav-itu/src/codecs/h264/decoder_st.rs
@@ -222,7 +222,7 @@ println!("PAFF?");
                     self.has_pic = self.decode_slice_cavlc(&mut br, &slice_hdr, full_size)?;
                 } else {
                     br.align();
-                    let start = (br.tell() / 8) as usize;
+                    let start = br.tell() / 8;
                     let csrc = &src[start..];
                     validate!(csrc.len() >= 2);
                     let mut cabac = CABAC::new(csrc, slice_hdr.slice_type, slice_hdr.slice_qp, slice_hdr.cabac_init_idc as usize)?;
@@ -362,25 +362,29 @@ println!("PAFF?");
                 mb_info.coeffs[i][0] = mb_info.coeffs[24][i];
             }
         }
-        if !mb_info.transform_size_8x8 {
-            let quant_dc = !mb_info.mb_type.is_intra16x16();
-            for i in 0..16 {
-                if mb_info.coded[i] {
-                    if !tx_bypass {
-                        idct(&mut mb_info.coeffs[i], qp_y, quant_dc);
+        if !tx_bypass {
+            if !mb_info.transform_size_8x8 {
+                let quant_dc = !mb_info.mb_type.is_intra16x16();
+                for (coded, coeffs) in mb_info.coded[..16].iter_mut().zip(mb_info.coeffs[..16].iter_mut()) {
+                    if *coded {
+                        idct(coeffs, qp_y, quant_dc);
+                    } else if has_dc {
+                        idct_dc(coeffs, qp_y, quant_dc);
+                        *coded = true;
                     }
-                } else if has_dc {
-                    if !tx_bypass {
-                        idct_dc(&mut mb_info.coeffs[i], qp_y, quant_dc);
+                }
+            } else {
+                for i in 0..4 {
+                    if mb_info.coded[(i & 1) * 2 + (i & 2) * 4] {
+                        dequant8x8(&mut mb_info.coeffs8x8[i].coeffs, &pps.scaling_list_8x8[!mb_info.mb_type.is_intra() as usize]);
+                        idct8x8(&mut mb_info.coeffs8x8[i].coeffs, qp_y);
                     }
-                    mb_info.coded[i] = true;
                 }
             }
-        } else {
-            for i in 0..4 {
-                if mb_info.coded[(i & 1) * 2 + (i & 2) * 4] && !tx_bypass {
-                    dequant8x8(&mut mb_info.coeffs8x8[i].coeffs, &pps.scaling_list_8x8[!mb_info.mb_type.is_intra() as usize]);
-                    idct8x8(&mut mb_info.coeffs8x8[i].coeffs, qp_y);
+        } else if !mb_info.transform_size_8x8 {
+            for i in 0..16 {
+                if !mb_info.coded[i] && has_dc {
+                    mb_info.coded[i] = true;
                 }
             }
         }
@@ -482,7 +486,7 @@ _ => {},
             17, 18, 20, 24, 19, 21, 26, 28, 23, 27, 29, 30, 22, 25, 38, 41
         ];
 
-        let mut mb_idx = slice_hdr.first_mb_in_slice as usize;
+        let mut mb_idx = slice_hdr.first_mb_in_slice;
         let mut mb_info = CurrentMBInfo { qp_y: slice_hdr.slice_qp, ..Default::default() };
         let skip_type = if slice_hdr.slice_type.is_p() { MBType::PSkip } else { MBType::BSkip };
         while br.tell() < full_size && mb_idx < self.num_mbs {
@@ -576,7 +580,7 @@ _ => {},
         Ok(mb_idx == self.num_mbs)
     }
     fn decode_slice_cabac(&mut self, cabac: &mut CABAC, slice_hdr: &SliceHeader) -> DecoderResult<bool> {
-        let mut mb_idx = slice_hdr.first_mb_in_slice as usize;
+        let mut mb_idx = slice_hdr.first_mb_in_slice;
         let mut prev_mb_skipped = false;
         let skip_type = if slice_hdr.slice_type.is_p() { MBType::PSkip } else { MBType::BSkip };
         let mut last_qp_diff = false;
@@ -763,7 +767,7 @@ impl NADecoder for H264Decoder {
             }
 
             let num_bufs = if !self.sps.is_empty() {
-                    self.sps[0].num_ref_frames as usize + 1
+                    self.sps[0].num_ref_frames + 1
                 } else {
                     3
                 }.max(16 + 1);