avfilter/scale_cuda: code cleanup

2 files changed, 90 insertions, 222 deletions

diff --git a/libavfilter/vf_scale_cuda.c b/libavfilter/vf_scale_cuda.c
index b287bd8c12..dc565cda89 100644
--- a/libavfilter/vf_scale_cuda.c
+++ b/libavfilter/vf_scale_cuda.c
@@ -48,8 +48,6 @@ static const enum AVPixelFormat supported_formats[] = {
 };
 
 #define DIV_UP(a, b) ( ((a) + (b) - 1) / (b) )
-#define ALIGN_UP(a, b) (((a) + (b) - 1) & ~((b) - 1))
-#define NUM_BUFFERS 2
 #define BLOCKX 32
 #define BLOCKY 16
 
@@ -72,11 +70,6 @@ typedef struct CUDAScaleContext {
     enum AVPixelFormat in_fmt;
     enum AVPixelFormat out_fmt;
 
-    struct {
-        int width;
-        int height;
-    } planes_in[3], planes_out[3];
-
     AVBufferRef *frames_ctx;
     AVFrame     *frame;
 
@@ -110,6 +103,7 @@ typedef struct CUDAScaleContext {
 
     int interp_algo;
     int interp_use_linear;
+    int interp_as_integer;
 } CUDAScaleContext;
 
 static av_cold int cudascale_init(AVFilterContext *ctx)
@@ -153,30 +147,17 @@ static int cudascale_query_formats(AVFilterContext *ctx)
         AV_PIX_FMT_CUDA, AV_PIX_FMT_NONE,
     };
     AVFilterFormats *pix_fmts = ff_make_format_list(pixel_formats);
+    if (!pix_fmts)
+        return AVERROR(ENOMEM);
 
     return ff_set_common_formats(ctx, pix_fmts);
 }
 
-static av_cold int init_stage(CUDAScaleContext *s, AVBufferRef *device_ctx)
+static av_cold int init_hwframe_ctx(CUDAScaleContext *s, AVBufferRef *device_ctx, int width, int height)
 {
     AVBufferRef *out_ref = NULL;
     AVHWFramesContext *out_ctx;
-    int in_sw, in_sh, out_sw, out_sh;
-    int ret, i;
-
-    av_pix_fmt_get_chroma_sub_sample(s->in_fmt,  &in_sw,  &in_sh);
-    av_pix_fmt_get_chroma_sub_sample(s->out_fmt, &out_sw, &out_sh);
-    if (!s->planes_out[0].width) {
-        s->planes_out[0].width  = s->planes_in[0].width;
-        s->planes_out[0].height = s->planes_in[0].height;
-    }
-
-    for (i = 1; i < FF_ARRAY_ELEMS(s->planes_in); i++) {
-        s->planes_in[i].width   = s->planes_in[0].width   >> in_sw;
-        s->planes_in[i].height  = s->planes_in[0].height  >> in_sh;
-        s->planes_out[i].width  = s->planes_out[0].width  >> out_sw;
-        s->planes_out[i].height = s->planes_out[0].height >> out_sh;
-    }
+    int ret;
 
     out_ref = av_hwframe_ctx_alloc(device_ctx);
     if (!out_ref)
@@ -185,8 +166,8 @@ static av_cold int init_stage(CUDAScaleContext *s, AVBufferRef *device_ctx)
 
     out_ctx->format    = AV_PIX_FMT_CUDA;
     out_ctx->sw_format = s->out_fmt;
-    out_ctx->width     = FFALIGN(s->planes_out[0].width,  32);
-    out_ctx->height    = FFALIGN(s->planes_out[0].height, 32);
+    out_ctx->width     = FFALIGN(width,  32);
+    out_ctx->height    = FFALIGN(height, 32);
 
     ret = av_hwframe_ctx_init(out_ref);
     if (ret < 0)
@@ -197,8 +178,8 @@ static av_cold int init_stage(CUDAScaleContext *s, AVBufferRef *device_ctx)
     if (ret < 0)
         goto fail;
 
-    s->frame->width  = s->planes_out[0].width;
-    s->frame->height = s->planes_out[0].height;
+    s->frame->width  = width;
+    s->frame->height = height;
 
     av_buffer_unref(&s->frames_ctx);
     s->frames_ctx = out_ref;
@@ -250,20 +231,20 @@ static av_cold int init_processing_chain(AVFilterContext *ctx, int in_width, int
         return AVERROR(ENOSYS);
     }
 
-    if (in_width == out_width && in_height == out_height)
-        s->passthrough = 1;
-
     s->in_fmt = in_format;
     s->out_fmt = out_format;
 
-    s->planes_in[0].width   = in_width;
-    s->planes_in[0].height  = in_height;
-    s->planes_out[0].width  = out_width;
-    s->planes_out[0].height = out_height;
+    if (s->passthrough && in_width == out_width && in_height == out_height && in_format == out_format) {
+        s->frames_ctx = av_buffer_ref(ctx->inputs[0]->hw_frames_ctx);
+        if (!s->frames_ctx)
+            return AVERROR(ENOMEM);
+    } else {
+        s->passthrough = 0;
 
-    ret = init_stage(s, in_frames_ctx->device_ref);
-    if (ret < 0)
-        return ret;
+        ret = init_hwframe_ctx(s, in_frames_ctx->device_ref, out_width, out_height);
+        if (ret < 0)
+            return ret;
+    }
 
     ctx->outputs[0]->hw_frames_ctx = av_buffer_ref(s->frames_ctx);
     if (!ctx->outputs[0]->hw_frames_ctx)
@@ -296,12 +277,14 @@ static av_cold int cudascale_config_props(AVFilterLink *outlink)
         scaler_ptx = vf_scale_cuda_ptx;
         function_infix = "_Bilinear";
         s->interp_use_linear = 1;
+        s->interp_as_integer = 1;
         break;
     case INTERP_ALGO_DEFAULT:
     case INTERP_ALGO_BICUBIC:
         scaler_ptx = vf_scale_cuda_bicubic_ptx;
         function_infix = "_Bicubic";
         s->interp_use_linear = 0;
+        s->interp_as_integer = 0;
         break;
     default:
         av_log(ctx, AV_LOG_ERROR, "Unknown interpolation algorithm\n");
@@ -372,8 +355,8 @@ static av_cold int cudascale_config_props(AVFilterLink *outlink)
     if (ret < 0)
         return ret;
 
-    av_log(ctx, AV_LOG_VERBOSE, "w:%d h:%d -> w:%d h:%d\n",
-           inlink->w, inlink->h, outlink->w, outlink->h);
+    av_log(ctx, AV_LOG_VERBOSE, "w:%d h:%d -> w:%d h:%d%s\n",
+           inlink->w, inlink->h, outlink->w, outlink->h, s->passthrough ? " (passthrough)" : "");
 
     if (inlink->sample_aspect_ratio.num) {
         outlink->sample_aspect_ratio = av_mul_q((AVRational){outlink->h*inlink->w,
@@ -398,14 +381,14 @@ static int call_resize_kernel(AVFilterContext *ctx, CUfunction func, int channel
     CudaFunctions *cu = s->hwctx->internal->cuda_dl;
     CUdeviceptr dst_devptr = (CUdeviceptr)dst_dptr;
     CUtexObject tex = 0;
-    void *args_uchar[] = { &tex, &dst_devptr, &dst_width, &dst_height, &dst_pitch, &src_width, &src_height, &src_pitch, &bit_depth };
+    void *args_uchar[] = { &tex, &dst_devptr, &dst_width, &dst_height, &dst_pitch, &src_width, &src_height, &bit_depth };
     int ret;
 
     CUDA_TEXTURE_DESC tex_desc = {
         .filterMode = s->interp_use_linear ?
                       CU_TR_FILTER_MODE_LINEAR :
                       CU_TR_FILTER_MODE_POINT,
-        .flags = CU_TRSF_READ_AS_INTEGER,
+        .flags = s->interp_as_integer ? CU_TRSF_READ_AS_INTEGER : 0,
     };
 
     CUDA_RESOURCE_DESC res_desc = {
@@ -448,12 +431,12 @@ static int scalecuda_resize(AVFilterContext *ctx,
                            out->data[0], out->width, out->height, out->linesize[0],
                            1, 8);
         call_resize_kernel(ctx, s->cu_func_uchar, 1,
-                           in->data[1], in->width / 2, in->height / 2, in->linesize[0] / 2,
-                           out->data[1], out->width / 2, out->height / 2, out->linesize[0] / 2,
+                           in->data[1], in->width / 2, in->height / 2, in->linesize[1],
+                           out->data[1], out->width / 2, out->height / 2, out->linesize[1],
                            1, 8);
         call_resize_kernel(ctx, s->cu_func_uchar, 1,
-                           in->data[2], in->width / 2, in->height / 2, in->linesize[0] / 2,
-                           out->data[2], out->width / 2, out->height / 2, out->linesize[0] / 2,
+                           in->data[2], in->width / 2, in->height / 2, in->linesize[2],
+                           out->data[2], out->width / 2, out->height / 2, out->linesize[2],
                            1, 8);
         break;
     case AV_PIX_FMT_YUV444P:
@@ -462,12 +445,12 @@ static int scalecuda_resize(AVFilterContext *ctx,
                            out->data[0], out->width, out->height, out->linesize[0],
                            1, 8);
         call_resize_kernel(ctx, s->cu_func_uchar, 1,
-                           in->data[1], in->width, in->height, in->linesize[0],
-                           out->data[1], out->width, out->height, out->linesize[0],
+                           in->data[1], in->width, in->height, in->linesize[1],
+                           out->data[1], out->width, out->height, out->linesize[1],
                            1, 8);
         call_resize_kernel(ctx, s->cu_func_uchar, 1,
-                           in->data[2], in->width, in->height, in->linesize[0],
-                           out->data[2], out->width, out->height, out->linesize[0],
+                           in->data[2], in->width, in->height, in->linesize[2],
+                           out->data[2], out->width, out->height, out->linesize[2],
                            1, 8);
         break;
     case AV_PIX_FMT_YUV444P16:
@@ -524,6 +507,7 @@ static int scalecuda_resize(AVFilterContext *ctx,
 static int cudascale_scale(AVFilterContext *ctx, AVFrame *out, AVFrame *in)
 {
     CUDAScaleContext *s = ctx->priv;
+    AVFilterLink *outlink = ctx->outputs[0];
     AVFrame *src = in;
     int ret;
 
@@ -539,8 +523,8 @@ static int cudascale_scale(AVFilterContext *ctx, AVFrame *out, AVFrame *in)
     av_frame_move_ref(out, s->frame);
     av_frame_move_ref(s->frame, s->tmp_frame);
 
-    s->frame->width  = s->planes_out[0].width;
-    s->frame->height = s->planes_out[0].height;
+    s->frame->width  = outlink->w;
+    s->frame->height = outlink->h;
 
     ret = av_frame_copy_props(out, in);
     if (ret < 0)
@@ -560,6 +544,9 @@ static int cudascale_filter_frame(AVFilterLink *link, AVFrame *in)
     CUcontext dummy;
     int ret = 0;
 
+    if (s->passthrough)
+        return ff_filter_frame(outlink, in);
+
     out = av_frame_alloc();
     if (!out) {
         ret = AVERROR(ENOMEM);
@@ -589,18 +576,28 @@ fail:
     return ret;
 }
 
+static AVFrame *cudascale_get_video_buffer(AVFilterLink *inlink, int w, int h)
+{
+    CUDAScaleContext *s = inlink->dst->priv;
+
+    return s->passthrough ?
+        ff_null_get_video_buffer   (inlink, w, h) :
+        ff_default_get_video_buffer(inlink, w, h);
+}
+
 #define OFFSET(x) offsetof(CUDAScaleContext, x)
 #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM)
 static const AVOption options[] = {
-    { "w",      "Output video width",  OFFSET(w_expr),     AV_OPT_TYPE_STRING, { .str = "iw"   }, .flags = FLAGS },
-    { "h",      "Output video height", OFFSET(h_expr),     AV_OPT_TYPE_STRING, { .str = "ih"   }, .flags = FLAGS },
+    { "w",      "Output video width",  OFFSET(w_expr),     AV_OPT_TYPE_STRING, { .str = "iw" }, .flags = FLAGS },
+    { "h",      "Output video height", OFFSET(h_expr),     AV_OPT_TYPE_STRING, { .str = "ih" }, .flags = FLAGS },
     { "interp_algo", "Interpolation algorithm used for resizing", OFFSET(interp_algo), AV_OPT_TYPE_INT, { .i64 = INTERP_ALGO_DEFAULT }, 0, INTERP_ALGO_COUNT - 1, FLAGS, "interp_algo" },
-        { "bilinear",    "bilinear",     0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_BILINEAR }, 0, 0, FLAGS, "interp_algo" },
-        { "bicubic",     "bicubic",      0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_BICUBIC  }, 0, 0, FLAGS, "interp_algo" },
-    { "force_original_aspect_ratio", "decrease or increase w/h if necessary to keep the original AR", OFFSET(force_original_aspect_ratio), AV_OPT_TYPE_INT, { .i64 = 0}, 0, 2, FLAGS, "force_oar" },
-    { "disable",  NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 0 }, 0, 0, FLAGS, "force_oar" },
-    { "decrease", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 1 }, 0, 0, FLAGS, "force_oar" },
-    { "increase", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 2 }, 0, 0, FLAGS, "force_oar" },
+        { "bilinear", "bilinear", 0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_BILINEAR }, 0, 0, FLAGS, "interp_algo" },
+        { "bicubic",  "bicubic",  0, AV_OPT_TYPE_CONST, { .i64 = INTERP_ALGO_BICUBIC  }, 0, 0, FLAGS, "interp_algo" },
+    { "passthrough", "Do not process frames at all if parameters match", OFFSET(passthrough), AV_OPT_TYPE_BOOL, { .i64 = 1 }, 0, 1, FLAGS },
+    { "force_original_aspect_ratio", "decrease or increase w/h if necessary to keep the original AR", OFFSET(force_original_aspect_ratio), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 2, FLAGS, "force_oar" },
+        { "disable",  NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 0 }, 0, 0, FLAGS, "force_oar" },
+        { "decrease", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 1 }, 0, 0, FLAGS, "force_oar" },
+        { "increase", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = 2 }, 0, 0, FLAGS, "force_oar" },
     { "force_divisible_by", "enforce that the output resolution is divisible by a defined integer when force_original_aspect_ratio is used", OFFSET(force_divisible_by), AV_OPT_TYPE_INT, { .i64 = 1}, 1, 256, FLAGS },
     { NULL },
 };
@@ -617,6 +614,7 @@ static const AVFilterPad cudascale_inputs[] = {
         .name        = "default",
         .type        = AVMEDIA_TYPE_VIDEO,
         .filter_frame = cudascale_filter_frame,
+        .get_video_buffer = cudascale_get_video_buffer,
     },
     { NULL }
 };
diff --git a/libavfilter/vf_scale_cuda.cu b/libavfilter/vf_scale_cuda.cu
index 3f3f40546d..24b1151215 100644
--- a/libavfilter/vf_scale_cuda.cu
+++ b/libavfilter/vf_scale_cuda.cu
@@ -20,12 +20,14 @@
  * DEALINGS IN THE SOFTWARE.
  */
 
-extern "C" {
+#include "cuda/vector_helpers.cuh"
 
-__global__ void Subsample_Bilinear_uchar(cudaTextureObject_t uchar_tex,
-                                    unsigned char *dst,
-                                    int dst_width, int dst_height, int dst_pitch,
-                                    int src_width, int src_height)
+template<typename T>
+__device__ inline void Subsample_Bilinear(cudaTextureObject_t tex,
+                                          T *dst,
+                                          int dst_width, int dst_height, int dst_pitch,
+                                          int src_width, int src_height,
+                                          int bit_depth)
 {
     int xo = blockIdx.x * blockDim.x + threadIdx.x;
     int yo = blockIdx.y * blockDim.y + threadIdx.y;
@@ -42,170 +44,38 @@ __global__ void Subsample_Bilinear_uchar(cudaTextureObject_t uchar_tex,
         // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
         float dx = wh / (0.5f + wh);
         float dy = wv / (0.5f + wv);
-        int y0 = tex2D<unsigned char>(uchar_tex, xi-dx, yi-dy);
-        int y1 = tex2D<unsigned char>(uchar_tex, xi+dx, yi-dy);
-        int y2 = tex2D<unsigned char>(uchar_tex, xi-dx, yi+dy);
-        int y3 = tex2D<unsigned char>(uchar_tex, xi+dx, yi+dy);
-        dst[yo*dst_pitch+xo] = (unsigned char)((y0+y1+y2+y3+2) >> 2);
-    }
-}
-
-__global__ void Subsample_Bilinear_uchar2(cudaTextureObject_t uchar2_tex,
-                                    uchar2 *dst,
-                                    int dst_width, int dst_height, int dst_pitch2,
-                                    int src_width, int src_height)
-{
-    int xo = blockIdx.x * blockDim.x + threadIdx.x;
-    int yo = blockIdx.y * blockDim.y + threadIdx.y;
-
-    if (yo < dst_height && xo < dst_width)
-    {
-        float hscale = (float)src_width / (float)dst_width;
-        float vscale = (float)src_height / (float)dst_height;
-        float xi = (xo + 0.5f) * hscale;
-        float yi = (yo + 0.5f) * vscale;
-        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
-        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
-        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
-        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
-        float dx = wh / (0.5f + wh);
-        float dy = wv / (0.5f + wv);
-        uchar2 c0 = tex2D<uchar2>(uchar2_tex, xi-dx, yi-dy);
-        uchar2 c1 = tex2D<uchar2>(uchar2_tex, xi+dx, yi-dy);
-        uchar2 c2 = tex2D<uchar2>(uchar2_tex, xi-dx, yi+dy);
-        uchar2 c3 = tex2D<uchar2>(uchar2_tex, xi+dx, yi+dy);
-        int2 uv;
-        uv.x = ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
-        uv.y = ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
-        dst[yo*dst_pitch2+xo] = make_uchar2((unsigned char)uv.x, (unsigned char)uv.y);
-    }
-}
-
-__global__ void Subsample_Bilinear_uchar4(cudaTextureObject_t uchar4_tex,
-                                    uchar4 *dst,
-                                    int dst_width, int dst_height, int dst_pitch,
-                                    int src_width, int src_height)
-{
-    int xo = blockIdx.x * blockDim.x + threadIdx.x;
-    int yo = blockIdx.y * blockDim.y + threadIdx.y;
 
-    if (yo < dst_height && xo < dst_width)
-    {
-        float hscale = (float)src_width / (float)dst_width;
-        float vscale = (float)src_height / (float)dst_height;
-        float xi = (xo + 0.5f) * hscale;
-        float yi = (yo + 0.5f) * vscale;
-        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
-        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
-        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
-        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
-        float dx = wh / (0.5f + wh);
-        float dy = wv / (0.5f + wv);
-        uchar4 c0 = tex2D<uchar4>(uchar4_tex, xi-dx, yi-dy);
-        uchar4 c1 = tex2D<uchar4>(uchar4_tex, xi+dx, yi-dy);
-        uchar4 c2 = tex2D<uchar4>(uchar4_tex, xi-dx, yi+dy);
-        uchar4 c3 = tex2D<uchar4>(uchar4_tex, xi+dx, yi+dy);
-        int4 res;
-        res.x =  ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
-        res.y =  ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
-        res.z =  ((int)c0.z+(int)c1.z+(int)c2.z+(int)c3.z+2) >> 2;
-        res.w =  ((int)c0.w+(int)c1.w+(int)c2.w+(int)c3.w+2) >> 2;
-        dst[yo*dst_pitch+xo] = make_uchar4(
-            (unsigned char)res.x, (unsigned char)res.y, (unsigned char)res.z, (unsigned char)res.w);
+        intT r = { 0 };
+        vec_set_scalar(r, 2);
+        r += tex2D<T>(tex, xi - dx, yi - dy);
+        r += tex2D<T>(tex, xi + dx, yi - dy);
+        r += tex2D<T>(tex, xi - dx, yi + dy);
+        r += tex2D<T>(tex, xi + dx, yi + dy);
+        vec_set(dst[yo*dst_pitch+xo], r >> 2);
     }
 }
 
-__global__ void Subsample_Bilinear_ushort(cudaTextureObject_t ushort_tex,
-                                    unsigned short *dst,
-                                    int dst_width, int dst_height, int dst_pitch,
-                                    int src_width, int src_height)
-{
-    int xo = blockIdx.x * blockDim.x + threadIdx.x;
-    int yo = blockIdx.y * blockDim.y + threadIdx.y;
-
-    if (yo < dst_height && xo < dst_width)
-    {
-        float hscale = (float)src_width / (float)dst_width;
-        float vscale = (float)src_height / (float)dst_height;
-        float xi = (xo + 0.5f) * hscale;
-        float yi = (yo + 0.5f) * vscale;
-        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
-        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
-        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
-        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
-        float dx = wh / (0.5f + wh);
-        float dy = wv / (0.5f + wv);
-        int y0 = tex2D<unsigned short>(ushort_tex, xi-dx, yi-dy);
-        int y1 = tex2D<unsigned short>(ushort_tex, xi+dx, yi-dy);
-        int y2 = tex2D<unsigned short>(ushort_tex, xi-dx, yi+dy);
-        int y3 = tex2D<unsigned short>(ushort_tex, xi+dx, yi+dy);
-        dst[yo*dst_pitch+xo] = (unsigned short)((y0+y1+y2+y3+2) >> 2);
-    }
-}
-
-__global__ void Subsample_Bilinear_ushort2(cudaTextureObject_t ushort2_tex,
-                                    ushort2 *dst,
-                                    int dst_width, int dst_height, int dst_pitch2,
-                                    int src_width, int src_height)
-{
-    int xo = blockIdx.x * blockDim.x + threadIdx.x;
-    int yo = blockIdx.y * blockDim.y + threadIdx.y;
+extern "C" {
 
-    if (yo < dst_height && xo < dst_width)
-    {
-        float hscale = (float)src_width / (float)dst_width;
-        float vscale = (float)src_height / (float)dst_height;
-        float xi = (xo + 0.5f) * hscale;
-        float yi = (yo + 0.5f) * vscale;
-        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
-        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
-        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
-        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
-        float dx = wh / (0.5f + wh);
-        float dy = wv / (0.5f + wv);
-        ushort2 c0 = tex2D<ushort2>(ushort2_tex, xi-dx, yi-dy);
-        ushort2 c1 = tex2D<ushort2>(ushort2_tex, xi+dx, yi-dy);
-        ushort2 c2 = tex2D<ushort2>(ushort2_tex, xi-dx, yi+dy);
-        ushort2 c3 = tex2D<ushort2>(ushort2_tex, xi+dx, yi+dy);
-        int2 uv;
-        uv.x = ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
-        uv.y = ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
-        dst[yo*dst_pitch2+xo] = make_ushort2((unsigned short)uv.x, (unsigned short)uv.y);
+#define BILINEAR_KERNEL(T) \
+    __global__ void Subsample_Bilinear_ ## T(cudaTextureObject_t src_tex,                  \
+                                             T *dst,                                       \
+                                             int dst_width, int dst_height, int dst_pitch, \
+                                             int src_width, int src_height,                \
+                                             int bit_depth)                                \
+    {                                                                                      \
+        Subsample_Bilinear<T>(src_tex, dst,                                                \
+                              dst_width, dst_height, dst_pitch,                            \
+                              src_width, src_height,                                       \
+                              bit_depth);                                                  \
     }
-}
 
-__global__ void Subsample_Bilinear_ushort4(cudaTextureObject_t ushort4_tex,
-                                    ushort4 *dst,
-                                    int dst_width, int dst_height, int dst_pitch,
-                                    int src_width, int src_height)
-{
-    int xo = blockIdx.x * blockDim.x + threadIdx.x;
-    int yo = blockIdx.y * blockDim.y + threadIdx.y;
+BILINEAR_KERNEL(uchar)
+BILINEAR_KERNEL(uchar2)
+BILINEAR_KERNEL(uchar4)
 
-    if (yo < dst_height && xo < dst_width)
-    {
-        float hscale = (float)src_width / (float)dst_width;
-        float vscale = (float)src_height / (float)dst_height;
-        float xi = (xo + 0.5f) * hscale;
-        float yi = (yo + 0.5f) * vscale;
-        // 3-tap filter weights are {wh,1.0,wh} and {wv,1.0,wv}
-        float wh = min(max(0.5f * (hscale - 1.0f), 0.0f), 1.0f);
-        float wv = min(max(0.5f * (vscale - 1.0f), 0.0f), 1.0f);
-        // Convert weights to two bilinear weights -> {wh,1.0,wh} -> {wh,0.5,0} + {0,0.5,wh}
-        float dx = wh / (0.5f + wh);
-        float dy = wv / (0.5f + wv);
-        ushort4 c0 = tex2D<ushort4>(ushort4_tex, xi-dx, yi-dy);
-        ushort4 c1 = tex2D<ushort4>(ushort4_tex, xi+dx, yi-dy);
-        ushort4 c2 = tex2D<ushort4>(ushort4_tex, xi-dx, yi+dy);
-        ushort4 c3 = tex2D<ushort4>(ushort4_tex, xi+dx, yi+dy);
-        int4 res;
-        res.x =  ((int)c0.x+(int)c1.x+(int)c2.x+(int)c3.x+2) >> 2;
-        res.y =  ((int)c0.y+(int)c1.y+(int)c2.y+(int)c3.y+2) >> 2;
-        res.z =  ((int)c0.z+(int)c1.z+(int)c2.z+(int)c3.z+2) >> 2;
-        res.w =  ((int)c0.w+(int)c1.w+(int)c2.w+(int)c3.w+2) >> 2;
-        dst[yo*dst_pitch+xo] = make_ushort4(
-            (unsigned short)res.x, (unsigned short)res.y, (unsigned short)res.z, (unsigned short)res.w);
-    }
-}
+BILINEAR_KERNEL(ushort)
+BILINEAR_KERNEL(ushort2)
+BILINEAR_KERNEL(ushort4)
 
 }