[FFmpeg-devel] [PATCH] avcodec/dpx: Support for RGB and RGBA 12-bit packed decoding

[Jerome Martinez]

Currently RGB and RGBA 12-bit are supported by DPX decoder only if 
component values are padded (packing "Filled to 32-bit words, method A").
This patch adds decoding of RGB and RGBA 12-bit with no padding (packing 
"Packed into 32-bit words").

As I have no file with line boundaries not aligned on 32-bit, I can not 
have good tests about the stride computing (so code about non aligned 
boundaries is theory) so I preferred to limit risks by decoding only if 
line boundaries are aligned on 32-bit words:
- 8 pixels for RGB (8 pixels x 3 components x 12 bits = 288 bits = 9 x 
32-bit words)
- 2 pixels for RGBA (2 pixels x 4 components x 12 bits = 3 x 32-bit words)

I think Little Endian parsing is fine thanks to the generic code about 
Big vs Little endian but I have no Little Endian test file so I also 
limited the decoding to Big Endian files.

Would be happy to check with cases I was not able to check if someone 
provides files.

I kept "Packing to 16bit required\n" message but interested in any 
suggestion about a better wording due to exceptions.

My test files:
https://github.com/MediaArea/RAWcooked-RegressionTestingFiles/tree/master/Formats/DPX/Flavors/RGB_12_Packed_BE

Regression tests done on 12-bit content "Filled to 32-bit words, method 
A" in:
https://samples.ffmpeg.org/image-samples/dpx_samples.zip

-------------- next part --------------
From bf95371e3964e198e22dc545fc75706dedf9029b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?J=C3=A9r=C3=B4me=20Martinez?= <jerome at mediaarea.net>
Date: Thu, 8 Feb 2018 09:22:08 +0100
Subject: [PATCH] avcodec/dpx: Support for RGB and RGBA 12-bit packed decoding

Limited to widths multiple of 8 (RGB) and 2 (RGBA) due to lack of test files for such corner case
---
 libavcodec/dpx.c | 70 +++++++++++++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 67 insertions(+), 3 deletions(-)

diff --git a/libavcodec/dpx.c b/libavcodec/dpx.c
index 1aa2cbd1c8..aaacd243c9 100644
--- a/libavcodec/dpx.c
+++ b/libavcodec/dpx.c
@@ -65,6 +65,38 @@ static uint16_t read10in32(const uint8_t **ptr, uint32_t * lbuf,
     return *lbuf & 0x3FF;
 }
 
+static uint16_t read12in32(const uint8_t **ptr, uint32_t * lbuf,
+                                  int * n_datum, int is_big)
+{
+    if (*n_datum)
+        (*n_datum)--;
+    else {
+        *lbuf = read32(ptr, is_big);
+        *n_datum = 7;
+    }
+
+    switch (*n_datum){
+    case 7: return *lbuf & 0xFFF;
+    case 6: return (*lbuf >> 12) & 0xFFF;
+    case 5: {
+            uint32_t c = *lbuf >> 24;
+            *lbuf = read32(ptr, is_big);
+            c |= *lbuf << 8;
+            return c & 0xFFF;
+            }
+    case 4: return (*lbuf >> 4) & 0xFFF;
+    case 3: return (*lbuf >> 16) & 0xFFF;
+    case 2: {
+            uint32_t c = *lbuf >> 28;
+            *lbuf = read32(ptr, is_big);
+            c |= *lbuf << 4;
+            return c & 0xFFF;
+            }
+    case 1: return (*lbuf >> 8) & 0xFFF;
+    default: return *lbuf >> 20;
+    }
+}
+
 static int decode_frame(AVCodecContext *avctx,
                         void *data,
                         int *got_frame,
@@ -201,10 +233,29 @@ static int decode_frame(AVCodecContext *avctx,
         break;
     case 12:
         if (!packing) {
-            av_log(avctx, AV_LOG_ERROR, "Packing to 16bit required\n");
-            return -1;
+            int tested = 0;
+            if (endian && descriptor == 50 && (avctx->width%8) == 0) // Little endian and widths not a multiple of 8 need tests
+                tested = 1;
+            if (endian && descriptor == 51 && (avctx->width%2) == 0) // Little endian and widths not a multiple of 2 need tests
+                tested = 1;
+            if (!tested) {
+                av_log(avctx, AV_LOG_ERROR, "Packing to 16bit required\n");
+                return -1;
+            }
+        }
+        stride = avctx->width * elements;
+        if (packing)
+            stride *= 2;
+        else {
+            stride *= 3; // 12 bits are 1.5 byte so multiplied by 3 then divided by 2
+            if (stride % 8) {
+                // Align to 32-bit boundaries (not tested)
+                stride /= 8;
+                stride++;
+                stride *= 8;
+            }
+            stride /= 2;
         }
-        stride = 2 * avctx->width * elements;
         break;
     case 16:
         stride = 2 * avctx->width * elements;
@@ -349,6 +400,7 @@ static int decode_frame(AVCodecContext *avctx,
                                 (uint16_t*)ptr[2],
                                 (uint16_t*)ptr[3]};
             for (y = 0; y < avctx->width; y++) {
+                if (packing) {
                 if (elements >= 3)
                     *dst[2]++ = read16(&buf, endian) >> 4;
                 *dst[0] = read16(&buf, endian) >> 4;
@@ -357,6 +409,18 @@ static int decode_frame(AVCodecContext *avctx,
                     *dst[1]++ = read16(&buf, endian) >> 4;
                 if (elements == 4)
                     *dst[3]++ = read16(&buf, endian) >> 4;
+                }
+                else {
+                    *dst[2]++ = read12in32(&buf, &rgbBuffer,
+                                           &n_datum, endian);
+                    *dst[0]++ = read12in32(&buf, &rgbBuffer,
+                                           &n_datum, endian);
+                    *dst[1]++ = read12in32(&buf, &rgbBuffer,
+                                           &n_datum, endian);
+                    if (elements == 4)
+                        *dst[3]++ = read12in32(&buf, &rgbBuffer,
+                                               &n_datum, endian);
+                }
             }
             for (i = 0; i < elements; i++)
                 ptr[i] += p->linesize[i];
-- 
2.13.3.windows.1