From 18b0f345212a5950bddebe00edc3ec5f54781456 Mon Sep 17 00:00:00 2001 From: Brian Date: Wed, 9 Jan 2008 16:12:25 -0700 Subject: Cell: start using DMA tags in a more sensible way, move waits() later when possible. --- src/mesa/pipe/cell/spu/main.c | 52 +++++++++++++++++++++++++++---------------- src/mesa/pipe/cell/spu/main.h | 12 ++++++++++ src/mesa/pipe/cell/spu/tri.c | 7 +++++- 3 files changed, 51 insertions(+), 20 deletions(-) (limited to 'src/mesa/pipe/cell/spu') diff --git a/src/mesa/pipe/cell/spu/main.c b/src/mesa/pipe/cell/spu/main.c index 8e7f8be1eb..7b63e85ae2 100644 --- a/src/mesa/pipe/cell/spu/main.c +++ b/src/mesa/pipe/cell/spu/main.c @@ -45,7 +45,7 @@ helpful headers: /opt/ibm/cell-sdk/prototype/sysroot/usr/include/libmisc.h */ -static boolean Debug = TRUE; +static boolean Debug = FALSE; volatile struct cell_init_info init; @@ -59,9 +59,10 @@ int DefaultTag; void -wait_on_mask(unsigned tag) +wait_on_mask(unsigned tagMask) { - mfc_write_tag_mask( tag ); + mfc_write_tag_mask( tagMask ); + /* wait for completion of _any_ DMAs specified by tagMask */ mfc_read_tag_status_any(); } @@ -125,7 +126,6 @@ clear_surface(const struct cell_command_clear_surface *clear) { uint num_tiles = fb.width_tiles * fb.height_tiles; uint i, j; - int tag = init.id; if (clear->surface == 0) { for (i = 0; i < TILE_SIZE; i++) @@ -147,13 +147,15 @@ clear_surface(const struct cell_command_clear_surface *clear) uint tx = i % fb.width_tiles; uint ty = i / fb.width_tiles; if (clear->surface == 0) - put_tile(&fb, tx, ty, (uint *) ctile, tag, 0); + put_tile(&fb, tx, ty, (uint *) ctile, TAG_SURFACE_CLEAR, 0); else - put_tile(&fb, tx, ty, (uint *) ztile, tag, 1); + put_tile(&fb, tx, ty, (uint *) ztile, TAG_SURFACE_CLEAR, 1); /* XXX we don't want this here, but it fixes bad tile results */ - wait_on_mask(1 << tag); } +#if 0 + wait_on_mask(1 << TAG_SURFACE_CLEAR); +#endif } @@ -198,7 +200,6 @@ static void render(const struct cell_command_render *render) { struct cell_prim_buffer prim_buffer ALIGN16_ATTRIB; - int tag = init.id /**DefaultTag**/; uint i, j, vertex_bytes; /* @@ -217,16 +218,18 @@ render(const struct cell_command_render *render) mfc_get(&prim_buffer, /* dest */ (unsigned int) render->vertex_data, /* src */ vertex_bytes, /* size */ - tag, + TAG_VERTEX_BUFFER, 0, /* tid */ 0 /* rid */); - wait_on_mask( 1 << tag ); /* XXX temporary */ + wait_on_mask(1 << TAG_VERTEX_BUFFER); /* find tiles which intersect the prim bounding box */ uint txmin, tymin, box_width_tiles, box_num_tiles; tile_bounding_box(render, &txmin, &tymin, &box_num_tiles, &box_width_tiles); + /* make sure any pending clears have completed */ + wait_on_mask(1 << TAG_SURFACE_CLEAR); /* loop over tiles */ for (i = init.id; i < box_num_tiles; i += init.num_spus) { @@ -236,12 +239,12 @@ render(const struct cell_command_render *render) assert(tx < fb.width_tiles); assert(ty < fb.height_tiles); - get_tile(&fb, tx, ty, (uint *) ctile, tag, 0); - wait_on_mask(1 << tag); /* XXX temporary */ - + /* Start fetching color/z tiles. We'll wait for completion when + * we need read/write to them later in triangle rasterization. + */ + get_tile(&fb, tx, ty, (uint *) ctile, TAG_READ_TILE_COLOR, 0); if (fb.depth_format == PIPE_FORMAT_Z16_UNORM) { - get_tile(&fb, tx, ty, (uint *) ztile, tag+1, 1); - wait_on_mask(1 << (tag+1)); /* XXX temporary */ + get_tile(&fb, tx, ty, (uint *) ztile, TAG_READ_TILE_Z, 1); } assert(render->prim_type == PIPE_PRIM_TRIANGLES); @@ -274,12 +277,23 @@ render(const struct cell_command_render *render) tri_draw(&prim, tx, ty); } - put_tile(&fb, tx, ty, (uint *) ctile, tag, 0); - wait_on_mask(1 << tag); /* XXX temp */ + /* in case nothing was drawn, wait now for completion */ + /* XXX temporary */ + wait_on_mask(1 << TAG_READ_TILE_COLOR); + if (fb.depth_format == PIPE_FORMAT_Z16_UNORM) { + wait_on_mask(1 << TAG_READ_TILE_Z); /* XXX temporary */ + } + + /* XXX IF we wrote anything into the tile... */ + + put_tile(&fb, tx, ty, (uint *) ctile, TAG_WRITE_TILE_COLOR, 0); + if (fb.depth_format == PIPE_FORMAT_Z16_UNORM) { + put_tile(&fb, tx, ty, (uint *) ztile, TAG_WRITE_TILE_Z, 1); + } + wait_on_mask(1 << TAG_WRITE_TILE_COLOR); /* XXX temp */ if (fb.depth_format == PIPE_FORMAT_Z16_UNORM) { - put_tile(&fb, tx, ty, (uint *) ztile, tag+1, 1); - wait_on_mask(1 << (tag+1)); /* XXX temporary */ + wait_on_mask(1 << TAG_WRITE_TILE_Z); /* XXX temporary */ } } } diff --git a/src/mesa/pipe/cell/spu/main.h b/src/mesa/pipe/cell/spu/main.h index 5e0516d0c1..656d28ea0e 100644 --- a/src/mesa/pipe/cell/spu/main.h +++ b/src/mesa/pipe/cell/spu/main.h @@ -55,6 +55,18 @@ extern ushort ztile[TILE_SIZE][TILE_SIZE] ALIGN16_ATTRIB; extern int DefaultTag; +/* DMA TAGS */ + +#define TAG_SURFACE_CLEAR 10 +#define TAG_VERTEX_BUFFER 11 +#define TAG_READ_TILE_COLOR 12 +#define TAG_READ_TILE_Z 13 +#define TAG_WRITE_TILE_COLOR 14 +#define TAG_WRITE_TILE_Z 15 + + + + void wait_on_mask(unsigned tag); diff --git a/src/mesa/pipe/cell/spu/tri.c b/src/mesa/pipe/cell/spu/tri.c index 5d4790fe6c..b7dfd6ab8c 100644 --- a/src/mesa/pipe/cell/spu/tri.c +++ b/src/mesa/pipe/cell/spu/tri.c @@ -283,11 +283,13 @@ emit_quad( struct setup_stage *setup, int x, int y, unsigned mask ) eval_coeff(setup, 1, (float) x, (float) y, colors); - if (fb.depth_format == PIPE_FORMAT_Z16_UNORM) { float zvals[4]; eval_z(setup, (float) x, (float) y, zvals); + wait_on_mask(1 << TAG_READ_TILE_Z); /* XXX temporary */ + + if (mask & MASK_TOP_LEFT) { z = (uint) (zvals[0] * 65535.0); if (z < ztile[iy][ix]) @@ -321,6 +323,9 @@ emit_quad( struct setup_stage *setup, int x, int y, unsigned mask ) } } + if (mask) + wait_on_mask(1 << TAG_READ_TILE_COLOR); + if (mask & MASK_TOP_LEFT) ctile[iy][ix] = pack_color(colors[QUAD_TOP_LEFT]); if (mask & MASK_TOP_RIGHT) -- cgit v1.2.3