avutil/hwcontext_cuda: use generic size and pointer assignment functions
authorTimo Rothenpieler <timo@rothenpieler.org>
Wed, 9 May 2018 19:18:15 +0000 (21:18 +0200)
committerTimo Rothenpieler <timo@rothenpieler.org>
Wed, 9 May 2018 22:34:22 +0000 (00:34 +0200)
libavutil/hwcontext_cuda.c

index b90c29a..cb5d15c 100644 (file)
@@ -24,6 +24,7 @@
 #include "mem.h"
 #include "pixdesc.h"
 #include "pixfmt.h"
+#include "imgutils.h"
 
 #define CUDA_FRAME_ALIGNMENT 256
 
@@ -117,7 +118,6 @@ fail:
 static int cuda_frames_init(AVHWFramesContext *ctx)
 {
     CUDAFramesContext *priv = ctx->internal->priv;
-    int aligned_width = FFALIGN(ctx->width, CUDA_FRAME_ALIGNMENT);
     int i;
 
     for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++) {
@@ -133,29 +133,9 @@ static int cuda_frames_init(AVHWFramesContext *ctx)
     av_pix_fmt_get_chroma_sub_sample(ctx->sw_format, &priv->shift_width, &priv->shift_height);
 
     if (!ctx->pool) {
-        int size;
-
-        switch (ctx->sw_format) {
-        case AV_PIX_FMT_NV12:
-        case AV_PIX_FMT_YUV420P:
-            size = aligned_width * ctx->height * 3 / 2;
-            break;
-        case AV_PIX_FMT_YUV444P:
-        case AV_PIX_FMT_P010:
-        case AV_PIX_FMT_P016:
-            size = aligned_width * ctx->height * 3;
-            break;
-        case AV_PIX_FMT_YUV444P16:
-            size = aligned_width * ctx->height * 6;
-            break;
-        case AV_PIX_FMT_0RGB32:
-        case AV_PIX_FMT_0BGR32:
-            size = aligned_width * ctx->height * 4;
-            break;
-        default:
-            av_log(ctx, AV_LOG_ERROR, "BUG: Pixel format missing from size calculation.");
-            return AVERROR_BUG;
-        }
+        int size = av_image_get_buffer_size(ctx->sw_format, ctx->width, ctx->height, CUDA_FRAME_ALIGNMENT);
+        if (size < 0)
+            return size;
 
         ctx->internal->pool_internal = av_buffer_pool_init2(size, ctx, cuda_pool_alloc, NULL);
         if (!ctx->internal->pool_internal)
@@ -167,54 +147,22 @@ static int cuda_frames_init(AVHWFramesContext *ctx)
 
 static int cuda_get_buffer(AVHWFramesContext *ctx, AVFrame *frame)
 {
-    int aligned_width;
-    int width_in_bytes = ctx->width;
-
-    if (ctx->sw_format == AV_PIX_FMT_P010 ||
-        ctx->sw_format == AV_PIX_FMT_P016 ||
-        ctx->sw_format == AV_PIX_FMT_YUV444P16) {
-       width_in_bytes *= 2;
-    }
-    aligned_width = FFALIGN(width_in_bytes, CUDA_FRAME_ALIGNMENT);
+    int res;
 
     frame->buf[0] = av_buffer_pool_get(ctx->pool);
     if (!frame->buf[0])
         return AVERROR(ENOMEM);
 
-    switch (ctx->sw_format) {
-    case AV_PIX_FMT_NV12:
-    case AV_PIX_FMT_P010:
-    case AV_PIX_FMT_P016:
-        frame->data[0]     = frame->buf[0]->data;
-        frame->data[1]     = frame->data[0] + aligned_width * ctx->height;
-        frame->linesize[0] = aligned_width;
-        frame->linesize[1] = aligned_width;
-        break;
-    case AV_PIX_FMT_YUV420P:
-        frame->data[0]     = frame->buf[0]->data;
-        frame->data[2]     = frame->data[0] + aligned_width * ctx->height;
-        frame->data[1]     = frame->data[2] + aligned_width * ctx->height / 4;
-        frame->linesize[0] = aligned_width;
-        frame->linesize[1] = aligned_width / 2;
-        frame->linesize[2] = aligned_width / 2;
-        break;
-    case AV_PIX_FMT_YUV444P:
-    case AV_PIX_FMT_YUV444P16:
-        frame->data[0]     = frame->buf[0]->data;
-        frame->data[1]     = frame->data[0] + aligned_width * ctx->height;
-        frame->data[2]     = frame->data[1] + aligned_width * ctx->height;
-        frame->linesize[0] = aligned_width;
-        frame->linesize[1] = aligned_width;
-        frame->linesize[2] = aligned_width;
-        break;
-    case AV_PIX_FMT_0BGR32:
-    case AV_PIX_FMT_0RGB32:
-        frame->data[0]     = frame->buf[0]->data;
-        frame->linesize[0] = aligned_width * 4;
-        break;
-    default:
-        av_frame_unref(frame);
-        return AVERROR_BUG;
+    res = av_image_fill_arrays(frame->data, frame->linesize, frame->buf[0]->data,
+                               ctx->sw_format, ctx->width, ctx->height, CUDA_FRAME_ALIGNMENT);
+    if (res < 0)
+        return res;
+
+    // YUV420P is a special case.
+    // Nvenc expects the U/V planes in swapped order from how ffmpeg expects them.
+    if (ctx->sw_format == AV_PIX_FMT_YUV420P) {
+        FFSWAP(uint8_t*, frame->data[1], frame->data[2]);
+        FFSWAP(int, frame->linesize[1], frame->linesize[2]);
     }
 
     frame->format = AV_PIX_FMT_CUDA;