[g3dvl] use a vertex element instead of the instance id

5 files changed, 85 insertions, 21 deletions

diff --git a/src/gallium/auxiliary/vl/vl_mpeg12_decoder.c b/src/gallium/auxiliary/vl/vl_mpeg12_decoder.c
index 35bf79261d3..201f3abab13 100644
--- a/src/gallium/auxiliary/vl/vl_mpeg12_decoder.c
+++ b/src/gallium/auxiliary/vl/vl_mpeg12_decoder.c
@@ -94,7 +94,7 @@ init_zscan_buffer(struct vl_mpeg12_buffer *buffer)
    formats[0] = formats[1] = formats[2] = dec->zscan_source_format;
    buffer->zscan_source = vl_video_buffer_init(dec->base.context, dec->pipe,
                                                dec->blocks_per_line * BLOCK_WIDTH * BLOCK_HEIGHT,
-                                               align(dec->max_blocks, dec->blocks_per_line) / dec->blocks_per_line,
+                                               align(dec->num_blocks, dec->blocks_per_line) / dec->blocks_per_line,
                                                1, PIPE_VIDEO_CHROMA_FORMAT_444,
                                                formats, PIPE_USAGE_STATIC);
    if (!buffer->zscan_source)
@@ -563,12 +563,14 @@ vl_mpeg12_decoder_flush_buffer(struct pipe_video_decode_buffer *buffer,
       }
    }
 
+   vb[2] = dec->block_num;
+
    dec->pipe->bind_vertex_elements_state(dec->pipe, dec->ves_ycbcr);
    for (i = 0; i < VL_MAX_PLANES; ++i) {
       if (!num_ycbcr_blocks[i]) continue;
 
       vb[1] = vl_vb_get_ycbcr(&buf->vertex_stream, i);
-      dec->pipe->set_vertex_buffers(dec->pipe, 2, vb);
+      dec->pipe->set_vertex_buffers(dec->pipe, 3, vb);
 
       vl_zscan_render(&buf->zscan[i] , num_ycbcr_blocks[i]);
 
@@ -585,7 +587,7 @@ vl_mpeg12_decoder_flush_buffer(struct pipe_video_decode_buffer *buffer,
          if (!num_ycbcr_blocks[i]) continue;
 
          vb[1] = vl_vb_get_ycbcr(&buf->vertex_stream, component);
-         dec->pipe->set_vertex_buffers(dec->pipe, 2, vb);
+         dec->pipe->set_vertex_buffers(dec->pipe, 3, vb);
 
          if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
             vl_idct_prepare_stage2(component == 0 ? &dec->idct_y : &dec->idct_c, &buf->idct[component]);
@@ -680,14 +682,10 @@ find_format_config(struct vl_mpeg12_decoder *dec, const struct format_config con
 static bool
 init_zscan(struct vl_mpeg12_decoder *dec, const struct format_config* format_config)
 {
-   const unsigned block_size_pixels = BLOCK_WIDTH * BLOCK_HEIGHT;
    unsigned num_channels;
 
    assert(dec);
 
-   dec->blocks_per_line = MAX2(util_next_power_of_two(dec->base.width) / block_size_pixels, 4);
-   dec->max_blocks = (dec->base.width * dec->base.height) / block_size_pixels;
-
    dec->zscan_source_format = format_config->zscan_source_format;
    dec->zscan_linear = vl_zscan_layout(dec->pipe, vl_zscan_linear, dec->blocks_per_line);
    dec->zscan_normal = vl_zscan_layout(dec->pipe, vl_zscan_normal, dec->blocks_per_line);
@@ -696,11 +694,11 @@ init_zscan(struct vl_mpeg12_decoder *dec, const struct format_config* format_con
    num_channels = dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT ? 4 : 1;
 
    if (!vl_zscan_init(&dec->zscan_y, dec->pipe, dec->base.width, dec->base.height,
-                      dec->blocks_per_line, dec->max_blocks, num_channels))
+                      dec->blocks_per_line, dec->num_blocks, num_channels))
       return false;
 
    if (!vl_zscan_init(&dec->zscan_c, dec->pipe, dec->chroma_width, dec->chroma_height,
-                      dec->blocks_per_line, dec->max_blocks, num_channels))
+                      dec->blocks_per_line, dec->num_blocks, num_channels))
       return false;
 
    return true;
@@ -838,6 +836,7 @@ vl_create_mpeg12_decoder(struct pipe_video_context *context,
                          enum pipe_video_chroma_format chroma_format,
                          unsigned width, unsigned height)
 {
+   const unsigned block_size_pixels = BLOCK_WIDTH * BLOCK_HEIGHT;
    const struct format_config *format_config;
    struct vl_mpeg12_decoder *dec;
 
@@ -861,12 +860,16 @@ vl_create_mpeg12_decoder(struct pipe_video_context *context,
 
    dec->pipe = pipe;
 
+   dec->blocks_per_line = MAX2(util_next_power_of_two(dec->base.width) / block_size_pixels, 4);
+   dec->num_blocks = (dec->base.width * dec->base.height) / block_size_pixels;
+
    dec->quads = vl_vb_upload_quads(dec->pipe);
    dec->pos = vl_vb_upload_pos(
       dec->pipe,
       dec->base.width / MACROBLOCK_WIDTH,
       dec->base.height / MACROBLOCK_HEIGHT
    );
+   dec->block_num = vl_vb_upload_block_num(dec->pipe, dec->num_blocks);
 
    dec->ves_ycbcr = vl_vb_get_ves_ycbcr(dec->pipe);
    dec->ves_mv = vl_vb_get_ves_mv(dec->pipe);
diff --git a/src/gallium/auxiliary/vl/vl_mpeg12_decoder.h b/src/gallium/auxiliary/vl/vl_mpeg12_decoder.h
index 4fc19ff9524..474ae2d5d29 100644
--- a/src/gallium/auxiliary/vl/vl_mpeg12_decoder.h
+++ b/src/gallium/auxiliary/vl/vl_mpeg12_decoder.h
@@ -49,12 +49,13 @@ struct vl_mpeg12_decoder
    unsigned chroma_width, chroma_height;
 
    unsigned blocks_per_line;
-   unsigned max_blocks;
+   unsigned num_blocks;
 
    enum pipe_format zscan_source_format;
 
    struct pipe_vertex_buffer quads;
    struct pipe_vertex_buffer pos;
+   struct pipe_vertex_buffer block_num;
 
    void *ves_ycbcr;
    void *ves_mv;
diff --git a/src/gallium/auxiliary/vl/vl_vertex_buffers.c b/src/gallium/auxiliary/vl/vl_vertex_buffers.c
index c45f96cec51..c0f1449bf80 100644
--- a/src/gallium/auxiliary/vl/vl_vertex_buffers.c
+++ b/src/gallium/auxiliary/vl/vl_vertex_buffers.c
@@ -125,6 +125,49 @@ vl_vb_upload_pos(struct pipe_context *pipe, unsigned width, unsigned height)
    return pos;
 }
 
+struct pipe_vertex_buffer
+vl_vb_upload_block_num(struct pipe_context *pipe, unsigned num_blocks)
+{
+   struct pipe_vertex_buffer buf;
+   struct pipe_transfer *buf_transfer;
+   struct vertex2s *v;
+   unsigned i;
+
+   assert(pipe);
+
+   /* create buffer */
+   buf.stride = sizeof(struct vertex2s);
+   buf.buffer_offset = 0;
+   buf.buffer = pipe_buffer_create
+   (
+      pipe->screen,
+      PIPE_BIND_VERTEX_BUFFER,
+      PIPE_USAGE_STATIC,
+      sizeof(struct vertex2s) * num_blocks
+   );
+
+   if(!buf.buffer)
+      return buf;
+
+   /* and fill it */
+   v = pipe_buffer_map
+   (
+      pipe,
+      buf.buffer,
+      PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
+      &buf_transfer
+   );
+
+   for ( i = 0; i < num_blocks; ++i, ++v) {
+      v->x = i;
+      v->y = i;
+   }
+
+   pipe_buffer_unmap(pipe, buf_transfer);
+
+   return buf;
+}
+
 static struct pipe_vertex_element
 vl_vb_get_quad_vertex_element(void)
 {
@@ -170,7 +213,12 @@ vl_vb_get_ves_ycbcr(struct pipe_context *pipe)
 
    vl_vb_element_helper(&vertex_elems[VS_I_VPOS], 1, 1);
 
-   return pipe->create_vertex_elements_state(pipe, 2, vertex_elems);
+   /* block num element */
+   vertex_elems[VS_I_BLOCK_NUM].src_format = PIPE_FORMAT_R16G16_SSCALED;
+
+   vl_vb_element_helper(&vertex_elems[VS_I_BLOCK_NUM], 1, 2);
+
+   return pipe->create_vertex_elements_state(pipe, 3, vertex_elems);
 }
 
 void *
@@ -249,6 +297,12 @@ error_ycbcr:
    return false;
 }
 
+unsigned
+vl_vb_attributes_per_plock(struct vl_vertex_buffer *buffer)
+{
+   return 1;
+}
+
 struct pipe_vertex_buffer
 vl_vb_get_ycbcr(struct vl_vertex_buffer *buffer, int component)
 {
diff --git a/src/gallium/auxiliary/vl/vl_vertex_buffers.h b/src/gallium/auxiliary/vl/vl_vertex_buffers.h
index 7e727cbca12..74845a42b69 100644
--- a/src/gallium/auxiliary/vl/vl_vertex_buffers.h
+++ b/src/gallium/auxiliary/vl/vl_vertex_buffers.h
@@ -41,13 +41,15 @@
 /* inputs to the vertex shaders */
 enum VS_INPUT
 {
-   VS_I_RECT,
-   VS_I_VPOS,
+   VS_I_RECT = 0,
+   VS_I_VPOS = 1,
 
-   VS_I_MV_TOP,
-   VS_I_MV_BOTTOM,
+   VS_I_BLOCK_NUM = 2,
 
-   NUM_VS_INPUTS
+   VS_I_MV_TOP = 2,
+   VS_I_MV_BOTTOM = 3,
+
+   NUM_VS_INPUTS = 4
 };
 
 struct vl_vertex_buffer
@@ -71,6 +73,8 @@ struct pipe_vertex_buffer vl_vb_upload_quads(struct pipe_context *pipe);
 
 struct pipe_vertex_buffer vl_vb_upload_pos(struct pipe_context *pipe, unsigned width, unsigned height);
 
+struct pipe_vertex_buffer vl_vb_upload_block_num(struct pipe_context *pipe, unsigned num_blocks);
+
 void *vl_vb_get_ves_ycbcr(struct pipe_context *pipe);
 
 void *vl_vb_get_ves_mv(struct pipe_context *pipe);
@@ -79,6 +83,8 @@ bool vl_vb_init(struct vl_vertex_buffer *buffer,
                 struct pipe_context *pipe,
                 unsigned width, unsigned height);
 
+unsigned vl_vb_attributes_per_plock(struct vl_vertex_buffer *buffer);
+
 void vl_vb_map(struct vl_vertex_buffer *buffer, struct pipe_context *pipe);
 
 struct pipe_vertex_buffer vl_vb_get_ycbcr(struct vl_vertex_buffer *buffer, int component);
diff --git a/src/gallium/auxiliary/vl/vl_zscan.c b/src/gallium/auxiliary/vl/vl_zscan.c
index 4af3962209f..2b52d610f8f 100644
--- a/src/gallium/auxiliary/vl/vl_zscan.c
+++ b/src/gallium/auxiliary/vl/vl_zscan.c
@@ -92,8 +92,8 @@ create_vert_shader(struct vl_zscan *zscan)
 {
    struct ureg_program *shader;
 
-   struct ureg_src scale, instance;
-   struct ureg_src vrect, vpos;
+   struct ureg_src scale;
+   struct ureg_src vrect, vpos, block_num;
 
    struct ureg_dst tmp;
    struct ureg_dst o_vpos, o_vtex[zscan->num_channels];
@@ -108,11 +108,11 @@ create_vert_shader(struct vl_zscan *zscan)
       (float)BLOCK_WIDTH / zscan->buffer_width,
       (float)BLOCK_HEIGHT / zscan->buffer_height);
 
-   instance = ureg_DECL_system_value(shader, 0, TGSI_SEMANTIC_INSTANCEID, 0);
-
    vrect = ureg_DECL_vs_input(shader, VS_I_RECT);
    vpos = ureg_DECL_vs_input(shader, VS_I_VPOS);
 
+   block_num = ureg_DECL_system_value(shader, 0, TGSI_SEMANTIC_INSTANCEID, 0);
+
    tmp = ureg_DECL_temporary(shader);
 
    o_vpos = ureg_DECL_output(shader, TGSI_SEMANTIC_POSITION, VS_O_VPOS);
@@ -136,7 +136,7 @@ create_vert_shader(struct vl_zscan *zscan)
    ureg_MUL(shader, ureg_writemask(o_vpos, TGSI_WRITEMASK_XY), ureg_src(tmp), scale);
    ureg_MOV(shader, ureg_writemask(o_vpos, TGSI_WRITEMASK_ZW), ureg_imm1f(shader, 1.0f));
 
-   ureg_MUL(shader, ureg_writemask(tmp, TGSI_WRITEMASK_XW), ureg_scalar(instance, TGSI_SWIZZLE_X),
+   ureg_MUL(shader, ureg_writemask(tmp, TGSI_WRITEMASK_XW), ureg_scalar(block_num, TGSI_SWIZZLE_X),
             ureg_imm1f(shader, 1.0f / zscan->blocks_per_line));
 
    ureg_FRC(shader, ureg_writemask(tmp, TGSI_WRITEMASK_Y), ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_X));