*
**************************************************************************/
+#include <limits.h>
#include "util/u_memory.h"
+#include "util/u_math.h"
+#include "util/u_cpu_detect.h"
+#include "util/u_surface.h"
-#include "lp_state.h"
+#include "lp_scene_queue.h"
+#include "lp_debug.h"
+#include "lp_fence.h"
#include "lp_rast.h"
#include "lp_rast_priv.h"
#include "lp_tile_soa.h"
#include "lp_bld_debug.h"
-
-#define RAST_DEBUG debug_printf
-
-struct lp_rasterizer *lp_rast_create( struct pipe_screen *screen )
-{
- struct lp_rasterizer *rast;
-
- rast = CALLOC_STRUCT(lp_rasterizer);
- if(!rast)
- return NULL;
-
- rast->screen = screen;
- rast->tile.color = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
- rast->tile.depth = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
-
- return rast;
-}
+#include "lp_scene.h"
/**
* Begin the rasterization phase.
* Map the framebuffer surfaces. Initialize the 'rast' state.
*/
-boolean lp_rast_begin( struct lp_rasterizer *rast,
- struct pipe_surface *cbuf,
- struct pipe_surface *zsbuf,
- boolean write_color,
- boolean write_zstencil,
- unsigned width,
- unsigned height )
+static boolean
+lp_rast_begin( struct lp_rasterizer *rast,
+ const struct pipe_framebuffer_state *fb,
+ boolean write_color,
+ boolean write_zstencil )
{
struct pipe_screen *screen = rast->screen;
+ struct pipe_surface *cbuf, *zsbuf;
+ int i;
- RAST_DEBUG("%s %dx%d\n", __FUNCTION__, width, height);
+ LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
- pipe_surface_reference(&rast->state.cbuf, cbuf);
- pipe_surface_reference(&rast->state.zsbuf, zsbuf);
+ util_copy_framebuffer_state(&rast->state.fb, fb);
- rast->width = width;
- rast->height = height;
rast->state.write_zstencil = write_zstencil;
rast->state.write_color = write_color;
- rast->check_for_clipped_tiles = (width % TILE_SIZE != 0 ||
- height % TILE_SIZE != 0);
-
- if (cbuf) {
- rast->cbuf_transfer = screen->get_tex_transfer(rast->screen,
- cbuf->texture,
- cbuf->face,
- cbuf->level,
- cbuf->zslice,
- PIPE_TRANSFER_READ_WRITE,
- 0, 0, width, height);
- if (!rast->cbuf_transfer)
- return FALSE;
+ rast->check_for_clipped_tiles = (fb->width % TILE_SIZE != 0 ||
+ fb->height % TILE_SIZE != 0);
- rast->cbuf_map = screen->transfer_map(rast->screen,
- rast->cbuf_transfer);
- if (!rast->cbuf_map)
- return FALSE;
+
+ for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
+ cbuf = rast->state.fb.cbufs[i];
+ if (cbuf) {
+ rast->cbuf_transfer[i] = screen->get_tex_transfer(rast->screen,
+ cbuf->texture,
+ cbuf->face,
+ cbuf->level,
+ cbuf->zslice,
+ PIPE_TRANSFER_READ_WRITE,
+ 0, 0,
+ cbuf->width,
+ cbuf->height);
+ if (!rast->cbuf_transfer[i])
+ goto fail;
+
+ rast->cbuf_map[i] = screen->transfer_map(rast->screen,
+ rast->cbuf_transfer[i]);
+ if (!rast->cbuf_map[i])
+ goto fail;
+ }
}
+ zsbuf = rast->state.fb.zsbuf;
if (zsbuf) {
rast->zsbuf_transfer = screen->get_tex_transfer(rast->screen,
- zsbuf->texture,
- zsbuf->face,
- zsbuf->level,
- zsbuf->zslice,
- PIPE_TRANSFER_READ_WRITE,
- 0, 0, width, height);
+ zsbuf->texture,
+ zsbuf->face,
+ zsbuf->level,
+ zsbuf->zslice,
+ PIPE_TRANSFER_READ_WRITE,
+ 0, 0,
+ zsbuf->width,
+ zsbuf->height);
if (!rast->zsbuf_transfer)
- return FALSE;
+ goto fail;
rast->zsbuf_map = screen->transfer_map(rast->screen,
rast->zsbuf_transfer);
if (!rast->zsbuf_map)
- return FALSE;
+ goto fail;
}
return TRUE;
+
+fail:
+ /* Unmap and release transfers?
+ */
+ return FALSE;
}
* Finish the rasterization phase.
* Unmap framebuffer surfaces.
*/
-void lp_rast_end( struct lp_rasterizer *rast )
+static void
+lp_rast_end( struct lp_rasterizer *rast )
{
struct pipe_screen *screen = rast->screen;
+ unsigned i;
+
+ for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
+ if (rast->cbuf_map[i])
+ screen->transfer_unmap(screen, rast->cbuf_transfer[i]);
- if (rast->cbuf_map)
- screen->transfer_unmap(screen, rast->cbuf_transfer);
+ if (rast->cbuf_transfer[i])
+ screen->tex_transfer_destroy(rast->cbuf_transfer[i]);
+
+ rast->cbuf_transfer[i] = NULL;
+ rast->cbuf_map[i] = NULL;
+ }
if (rast->zsbuf_map)
screen->transfer_unmap(screen, rast->zsbuf_transfer);
- if (rast->cbuf_transfer)
- screen->tex_transfer_destroy(rast->cbuf_transfer);
-
if (rast->zsbuf_transfer)
screen->tex_transfer_destroy(rast->zsbuf_transfer);
- rast->cbuf_transfer = NULL;
rast->zsbuf_transfer = NULL;
- rast->cbuf_map = NULL;
rast->zsbuf_map = NULL;
}
* \param x window X position of the tile, in pixels
* \param y window Y position of the tile, in pixels
*/
-void lp_rast_start_tile( struct lp_rasterizer *rast,
- unsigned x,
- unsigned y )
+static void
+lp_rast_start_tile( struct lp_rasterizer *rast,
+ unsigned thread_index,
+ unsigned x, unsigned y )
{
- RAST_DEBUG("%s %d,%d\n", __FUNCTION__, x, y);
+ LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, x, y);
- rast->x = x;
- rast->y = y;
+ rast->tasks[thread_index].x = x;
+ rast->tasks[thread_index].y = y;
}
* This is a bin command called during bin processing.
*/
void lp_rast_clear_color( struct lp_rasterizer *rast,
+ unsigned thread_index,
const union lp_rast_cmd_arg arg )
{
const uint8_t *clear_color = arg.clear_color;
-
- RAST_DEBUG("%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__,
+ uint8_t **color_tile = rast->tasks[thread_index].tile.color;
+ unsigned i;
+
+ LP_DBG(DEBUG_RAST, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__,
clear_color[0],
clear_color[1],
clear_color[2],
if (clear_color[0] == clear_color[1] &&
clear_color[1] == clear_color[2] &&
clear_color[2] == clear_color[3]) {
- memset(rast->tile.color, clear_color[0], TILE_SIZE * TILE_SIZE * 4);
+ for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
+ memset(color_tile[i], clear_color[0], TILE_SIZE * TILE_SIZE * 4);
+ }
}
else {
unsigned x, y, chan;
- for (y = 0; y < TILE_SIZE; y++)
- for (x = 0; x < TILE_SIZE; x++)
- for (chan = 0; chan < 4; ++chan)
- TILE_PIXEL(rast->tile.color, x, y, chan) = clear_color[chan];
+ for (i = 0; i < rast->state.fb.nr_cbufs; i++)
+ for (y = 0; y < TILE_SIZE; y++)
+ for (x = 0; x < TILE_SIZE; x++)
+ for (chan = 0; chan < 4; ++chan)
+ TILE_PIXEL(color_tile[i], x, y, chan) = clear_color[chan];
}
}
* This is a bin command called during bin processing.
*/
void lp_rast_clear_zstencil( struct lp_rasterizer *rast,
+ unsigned thread_index,
const union lp_rast_cmd_arg arg)
{
unsigned i, j;
+ uint32_t *depth_tile = rast->tasks[thread_index].tile.depth;
- RAST_DEBUG("%s 0x%x\n", __FUNCTION__, arg.clear_zstencil);
+ LP_DBG(DEBUG_RAST, "%s 0x%x\n", __FUNCTION__, arg.clear_zstencil);
for (i = 0; i < TILE_SIZE; i++)
for (j = 0; j < TILE_SIZE; j++)
- rast->tile.depth[i*TILE_SIZE + j] = arg.clear_zstencil;
+ depth_tile[i*TILE_SIZE + j] = arg.clear_zstencil;
}
* This is a bin command called during bin processing.
*/
void lp_rast_load_color( struct lp_rasterizer *rast,
+ unsigned thread_index,
const union lp_rast_cmd_arg arg)
{
- RAST_DEBUG("%s\n", __FUNCTION__);
-
- /* call u_tile func to load colors from surface */
+ struct lp_rasterizer_task *task = &rast->tasks[thread_index];
+ const unsigned x = task->x;
+ const unsigned y = task->y;
+ unsigned i;
+
+ LP_DBG(DEBUG_RAST, "%s at %u, %u\n", __FUNCTION__, x, y);
+
+ for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
+ struct pipe_transfer *transfer = rast->cbuf_transfer[i];
+ int w = TILE_SIZE;
+ int h = TILE_SIZE;
+
+ if (x >= transfer->width)
+ continue;
+
+ if (y >= transfer->height)
+ continue;
+ /* XXX: require tile-size aligned render target dimensions:
+ */
+ if (x + w > transfer->width)
+ w -= x + w - transfer->width;
+
+ if (y + h > transfer->height)
+ h -= y + h - transfer->height;
+
+ assert(w >= 0);
+ assert(h >= 0);
+ assert(w <= TILE_SIZE);
+ assert(h <= TILE_SIZE);
+
+ lp_tile_read_4ub(transfer->texture->format,
+ rast->tasks[thread_index].tile.color[i],
+ rast->cbuf_map[i],
+ transfer->stride,
+ x, y,
+ w, h);
+ }
}
* This is a bin command called during bin processing.
*/
void lp_rast_load_zstencil( struct lp_rasterizer *rast,
+ unsigned thread_index,
const union lp_rast_cmd_arg arg )
{
- RAST_DEBUG("%s\n", __FUNCTION__);
+ LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
/* call u_tile func to load depth (and stencil?) from surface */
}
void lp_rast_set_state( struct lp_rasterizer *rast,
+ unsigned thread_index,
const union lp_rast_cmd_arg arg )
{
const struct lp_rast_state *state = arg.set_state;
- RAST_DEBUG("%s\n", __FUNCTION__);
+ LP_DBG(DEBUG_RAST, "%s %p\n", __FUNCTION__, (void *) state);
- /* XXX to do */
+ /* just set the current state pointer for this rasterizer */
+ rast->tasks[thread_index].current_state = state;
}
* This is a bin command called during bin processing.
*/
void lp_rast_shade_tile( struct lp_rasterizer *rast,
+ unsigned thread_index,
const union lp_rast_cmd_arg arg )
{
+ /* Set c1,c2,c3 to large values so the in/out test always passes */
+ const int32_t c1 = INT_MIN, c2 = INT_MIN, c3 = INT_MIN;
const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
- const unsigned mask = ~0;
+ const unsigned tile_x = rast->tasks[thread_index].x;
+ const unsigned tile_y = rast->tasks[thread_index].y;
unsigned x, y;
- RAST_DEBUG("%s\n", __FUNCTION__);
+ LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
/* Use the existing preference for 4x4 (four quads) shading:
*/
for (y = 0; y < TILE_SIZE; y += 4)
for (x = 0; x < TILE_SIZE; x += 4)
- lp_rast_shade_quads( rast, inputs, rast->x + x, rast->y + y, mask);
+ lp_rast_shade_quads( rast,
+ thread_index,
+ inputs,
+ tile_x + x,
+ tile_y + y,
+ c1, c2, c3);
}
* This is a bin command called during bin processing.
*/
void lp_rast_shade_quads( struct lp_rasterizer *rast,
+ unsigned thread_index,
const struct lp_rast_shader_inputs *inputs,
unsigned x, unsigned y,
- unsigned mask)
+ int32_t c1, int32_t c2, int32_t c3)
{
-#if 1
- const struct lp_rast_state *state = inputs->state;
- struct lp_rast_tile *tile = &rast->tile;
- void *color;
+ const struct lp_rast_state *state = rast->tasks[thread_index].current_state;
+ struct lp_rast_tile *tile = &rast->tasks[thread_index].tile;
+ uint8_t *color[PIPE_MAX_COLOR_BUFS];
void *depth;
- uint32_t ALIGN16_ATTRIB masks[2][2][2][2];
+ unsigned i;
unsigned ix, iy;
int block_offset;
+#ifdef DEBUG
+ assert(state);
+
/* Sanity checks */
assert(x % TILE_VECTOR_WIDTH == 0);
assert(y % TILE_VECTOR_HEIGHT == 0);
- /* mask: the rasterizer wants to treat pixels in 4x4 blocks, but
- * the pixel shader wants to swizzle them into 4 2x2 quads.
- *
- * Additionally, the pixel shader wants masks as full dword ~0,
- * while the rasterizer wants to pack per-pixel bits tightly.
- */
-#if 0
- unsigned qx, qy;
- for (qy = 0; qy < 2; ++qy)
- for (qx = 0; qx < 2; ++qx)
- for (iy = 0; iy < 2; ++iy)
- for (ix = 0; ix < 2; ++ix)
- masks[qy][qx][iy][ix] = mask & (1 << (qy*8+iy*4+qx*2+ix)) ? ~0 : 0;
-#else
- masks[0][0][0][0] = mask & (1 << (0*8+0*4+0*2+0)) ? ~0 : 0;
- masks[0][0][0][1] = mask & (1 << (0*8+0*4+0*2+1)) ? ~0 : 0;
- masks[0][0][1][0] = mask & (1 << (0*8+1*4+0*2+0)) ? ~0 : 0;
- masks[0][0][1][1] = mask & (1 << (0*8+1*4+0*2+1)) ? ~0 : 0;
- masks[0][1][0][0] = mask & (1 << (0*8+0*4+1*2+0)) ? ~0 : 0;
- masks[0][1][0][1] = mask & (1 << (0*8+0*4+1*2+1)) ? ~0 : 0;
- masks[0][1][1][0] = mask & (1 << (0*8+1*4+1*2+0)) ? ~0 : 0;
- masks[0][1][1][1] = mask & (1 << (0*8+1*4+1*2+1)) ? ~0 : 0;
-
- masks[1][0][0][0] = mask & (1 << (1*8+0*4+0*2+0)) ? ~0 : 0;
- masks[1][0][0][1] = mask & (1 << (1*8+0*4+0*2+1)) ? ~0 : 0;
- masks[1][0][1][0] = mask & (1 << (1*8+1*4+0*2+0)) ? ~0 : 0;
- masks[1][0][1][1] = mask & (1 << (1*8+1*4+0*2+1)) ? ~0 : 0;
- masks[1][1][0][0] = mask & (1 << (1*8+0*4+1*2+0)) ? ~0 : 0;
- masks[1][1][0][1] = mask & (1 << (1*8+0*4+1*2+1)) ? ~0 : 0;
- masks[1][1][1][0] = mask & (1 << (1*8+1*4+1*2+0)) ? ~0 : 0;
- masks[1][1][1][1] = mask & (1 << (1*8+1*4+1*2+1)) ? ~0 : 0;
+ assert((x % 4) == 0);
+ assert((y % 4) == 0);
#endif
- assert((x % 2) == 0);
- assert((y % 2) == 0);
-
ix = x % TILE_SIZE;
iy = y % TILE_SIZE;
block_offset = ((iy/4)*(16*16) + (ix/4)*16);
/* color buffer */
- color = tile->color + 4 * block_offset;
+ for (i = 0; i < rast->state.fb.nr_cbufs; i++)
+ color[i] = tile->color[i] + 4 * block_offset;
/* depth buffer */
depth = tile->depth + block_offset;
- /* XXX: This will most likely fail on 32bit x86 without -mstackrealign */
- assert(lp_check_alignment(masks, 16));
- assert(lp_check_alignment(depth, 16));
- assert(lp_check_alignment(color, 16));
+
+#ifdef DEBUG
+ assert(lp_check_alignment(tile->depth, 16));
+ assert(lp_check_alignment(tile->color[0], 16));
assert(lp_check_alignment(state->jit_context.blend_color, 16));
+ assert(lp_check_alignment(inputs->step[0], 16));
+ assert(lp_check_alignment(inputs->step[1], 16));
+ assert(lp_check_alignment(inputs->step[2], 16));
+#endif
+
/* run shader */
state->jit_function( &state->jit_context,
x, y,
inputs->a0,
inputs->dadx,
inputs->dady,
- &masks[0][0][0][0],
color,
- depth);
-#else
- struct lp_rast_tile *tile = &rast->tile;
- unsigned chan_index;
- unsigned q, ix, iy;
-
- x %= TILE_SIZE;
- y %= TILE_SIZE;
-
- /* mask */
- for (q = 0; q < 4; ++q)
- for(iy = 0; iy < 2; ++iy)
- for(ix = 0; ix < 2; ++ix)
- if(masks[q] & (1 << (iy*2 + ix)))
- for (chan_index = 0; chan_index < NUM_CHANNELS; ++chan_index)
- TILE_PIXEL(tile->color, x + q*2 + ix, y + iy, chan_index) = 0xff;
-
-#endif
+ depth,
+ c1, c2, c3,
+ inputs->step[0], inputs->step[1], inputs->step[2]);
}
/**
* Write the rasterizer's color tile to the framebuffer.
*/
-static void lp_rast_store_color( struct lp_rasterizer *rast )
+static void lp_rast_store_color( struct lp_rasterizer *rast,
+ unsigned thread_index)
{
- const unsigned x = rast->x;
- const unsigned y = rast->y;
- unsigned w = TILE_SIZE;
- unsigned h = TILE_SIZE;
-
- if (x + w > rast->width)
- w -= x + w - rast->width;
-
- if (y + h > rast->height)
- h -= y + h - rast->height;
-
- RAST_DEBUG("%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h);
-
- lp_tile_write_4ub(rast->cbuf_transfer->format,
- rast->tile.color,
- rast->cbuf_map,
- rast->cbuf_transfer->stride,
- x, y,
- w, h);
+ const unsigned x = rast->tasks[thread_index].x;
+ const unsigned y = rast->tasks[thread_index].y;
+ unsigned i;
+
+ for (i = 0; i < rast->state.fb.nr_cbufs; i++) {
+ struct pipe_transfer *transfer = rast->cbuf_transfer[i];
+ int w = TILE_SIZE;
+ int h = TILE_SIZE;
+
+ if (x >= transfer->width)
+ continue;
+
+ if (y >= transfer->height)
+ continue;
+
+ /* XXX: require tile-size aligned render target dimensions:
+ */
+ if (x + w > transfer->width)
+ w -= x + w - transfer->width;
+
+ if (y + h > transfer->height)
+ h -= y + h - transfer->height;
+
+ assert(w >= 0);
+ assert(h >= 0);
+ assert(w <= TILE_SIZE);
+ assert(h <= TILE_SIZE);
+
+ LP_DBG(DEBUG_RAST, "%s [%u] %d,%d %dx%d\n", __FUNCTION__,
+ thread_index, x, y, w, h);
+
+ lp_tile_write_4ub(transfer->texture->format,
+ rast->tasks[thread_index].tile.color[i],
+ rast->cbuf_map[i],
+ transfer->stride,
+ x, y,
+ w, h);
+ }
}
/**
* Write the rasterizer's z/stencil tile to the framebuffer.
*/
-static void lp_rast_store_zstencil( struct lp_rasterizer *rast )
+static void lp_rast_store_zstencil( struct lp_rasterizer *rast,
+ unsigned thread_index )
{
- const unsigned x = rast->x;
- const unsigned y = rast->y;
+ const unsigned x = rast->tasks[thread_index].x;
+ const unsigned y = rast->tasks[thread_index].y;
unsigned w = TILE_SIZE;
unsigned h = TILE_SIZE;
- if (x + w > rast->width)
- w -= x + w - rast->width;
+ if (x + w > rast->state.fb.width)
+ w -= x + w - rast->state.fb.width;
- if (y + h > rast->height)
- h -= y + h - rast->height;
+ if (y + h > rast->state.fb.height)
+ h -= y + h - rast->state.fb.height;
- RAST_DEBUG("%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h);
+ LP_DBG(DEBUG_RAST, "%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h);
- assert(rast->zsbuf_transfer->format == PIPE_FORMAT_Z32_UNORM);
- lp_tile_write_z32(rast->tile.depth,
+ assert(rast->zsbuf_transfer->texture->format == PIPE_FORMAT_Z32_UNORM);
+ lp_tile_write_z32(rast->tasks[thread_index].tile.depth,
rast->zsbuf_map,
rast->zsbuf_transfer->stride,
x, y, w, h);
/**
* Write the rasterizer's tiles to the framebuffer.
*/
-void lp_rast_end_tile( struct lp_rasterizer *rast )
+static void
+lp_rast_end_tile( struct lp_rasterizer *rast,
+ unsigned thread_index )
{
- RAST_DEBUG("%s\n", __FUNCTION__);
+ LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
if (rast->state.write_color)
- lp_rast_store_color(rast);
+ lp_rast_store_color(rast, thread_index);
if (rast->state.write_zstencil)
- lp_rast_store_zstencil(rast);
+ lp_rast_store_zstencil(rast, thread_index);
+}
+
+
+/**
+ * Signal on a fence. This is called during bin execution/rasterization.
+ * Called per thread.
+ */
+void lp_rast_fence( struct lp_rasterizer *rast,
+ unsigned thread_index,
+ const union lp_rast_cmd_arg arg )
+{
+ struct lp_fence *fence = arg.fence;
+
+ pipe_mutex_lock( fence->mutex );
+
+ fence->count++;
+ assert(fence->count <= fence->rank);
+
+ LP_DBG(DEBUG_RAST, "%s count=%u rank=%u\n", __FUNCTION__,
+ fence->count, fence->rank);
+
+ pipe_condvar_signal( fence->signalled );
+
+ pipe_mutex_unlock( fence->mutex );
+}
+
+
+/**
+ * When all the threads are done rasterizing a scene, one thread will
+ * call this function to reset the scene and put it onto the empty queue.
+ */
+static void
+release_scene( struct lp_rasterizer *rast,
+ struct lp_scene *scene )
+{
+ util_unreference_framebuffer_state( &scene->fb );
+
+ lp_scene_reset( scene );
+ lp_scene_enqueue( rast->empty_scenes, scene );
+ rast->curr_scene = NULL;
+}
+
+
+/**
+ * Rasterize commands for a single bin.
+ * \param x, y position of the bin's tile in the framebuffer
+ * Must be called between lp_rast_begin() and lp_rast_end().
+ * Called per thread.
+ */
+static void
+rasterize_bin( struct lp_rasterizer *rast,
+ unsigned thread_index,
+ const struct cmd_bin *bin,
+ int x, int y)
+{
+ const struct cmd_block_list *commands = &bin->commands;
+ struct cmd_block *block;
+ unsigned k;
+
+ lp_rast_start_tile( rast, thread_index, x, y );
+
+ /* simply execute each of the commands in the block list */
+ for (block = commands->head; block; block = block->next) {
+ for (k = 0; k < block->count; k++) {
+ block->cmd[k]( rast, thread_index, block->arg[k] );
+ }
+ }
+
+ lp_rast_end_tile( rast, thread_index );
+}
+
+/* An empty bin is one that just loads the contents of the tile and
+ * stores them again unchanged. This typically happens when bins have
+ * been flushed for some reason in the middle of a frame, or when
+ * incremental updates are being made to a render target.
+ *
+ * Try to avoid doing pointless work in this case.
+ */
+static boolean
+is_empty_bin( const struct cmd_bin *bin )
+{
+ const struct cmd_block *head = bin->commands.head;
+ int i;
+
+ /* We emit at most two load-tile commands at the start of the first
+ * command block. If there are more than two commands in the
+ * block, we know that the bin is non-empty.
+ */
+ if (head->next != NULL ||
+ head->count > 2)
+ return FALSE;
+
+ for (i = 0; i < head->count; i++)
+ if (head->cmd[i] != lp_rast_load_color &&
+ head->cmd[i] != lp_rast_load_zstencil)
+ return FALSE;
+
+ return TRUE;
+}
+
+
+
+/**
+ * Rasterize/execute all bins within a scene.
+ * Called per thread.
+ */
+static void
+rasterize_scene( struct lp_rasterizer *rast,
+ unsigned thread_index,
+ struct lp_scene *scene,
+ bool write_depth )
+{
+ /* loop over scene bins, rasterize each */
+#if 0
+ {
+ unsigned i, j;
+ for (i = 0; i < scene->tiles_x; i++) {
+ for (j = 0; j < scene->tiles_y; j++) {
+ struct cmd_bin *bin = lp_get_bin(scene, i, j);
+ rasterize_bin( rast, thread_index,
+ bin, i * TILE_SIZE, j * TILE_SIZE );
+ }
+ }
+ }
+#else
+ {
+ struct cmd_bin *bin;
+ int x, y;
+
+ assert(scene);
+ while ((bin = lp_scene_bin_iter_next(scene, &x, &y))) {
+ if (!is_empty_bin( bin ))
+ rasterize_bin( rast, thread_index, bin, x * TILE_SIZE, y * TILE_SIZE);
+ }
+ }
+#endif
+}
+
+
+/**
+ * Called by setup module when it has something for us to render.
+ */
+void
+lp_rasterize_scene( struct lp_rasterizer *rast,
+ struct lp_scene *scene,
+ const struct pipe_framebuffer_state *fb,
+ bool write_depth )
+{
+ boolean debug = false;
+
+ LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
+
+ if (debug) {
+ unsigned x, y;
+ printf("rasterize scene:\n");
+ printf(" data size: %u\n", lp_scene_data_size(scene));
+ for (y = 0; y < scene->tiles_y; y++) {
+ for (x = 0; x < scene->tiles_x; x++) {
+ printf(" bin %u, %u size: %u\n", x, y,
+ lp_scene_bin_size(scene, x, y));
+ }
+ }
+ }
+
+ /* save framebuffer state in the bin */
+ util_copy_framebuffer_state(&scene->fb, fb);
+ scene->write_depth = write_depth;
+
+ if (rast->num_threads == 0) {
+ /* no threading */
+
+ lp_rast_begin( rast, fb,
+ fb->nr_cbufs != 0, /* always write color if cbufs present */
+ fb->zsbuf != NULL && write_depth );
+
+ lp_scene_bin_iter_begin( scene );
+ rasterize_scene( rast, 0, scene, write_depth );
+
+ release_scene( rast, scene );
+
+ lp_rast_end( rast );
+ }
+ else {
+ /* threaded rendering! */
+ unsigned i;
+
+ lp_scene_enqueue( rast->full_scenes, scene );
+
+ /* signal the threads that there's work to do */
+ for (i = 0; i < rast->num_threads; i++) {
+ pipe_semaphore_signal(&rast->tasks[i].work_ready);
+ }
+
+ /* wait for work to complete */
+ for (i = 0; i < rast->num_threads; i++) {
+ pipe_semaphore_wait(&rast->tasks[i].work_done);
+ }
+ }
+
+ LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
+}
+
+
+/**
+ * This is the thread's main entrypoint.
+ * It's a simple loop:
+ * 1. wait for work
+ * 2. do work
+ * 3. signal that we're done
+ */
+static void *
+thread_func( void *init_data )
+{
+ struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
+ struct lp_rasterizer *rast = task->rast;
+ boolean debug = false;
+
+ while (1) {
+ /* wait for work */
+ if (debug)
+ debug_printf("thread %d waiting for work\n", task->thread_index);
+ pipe_semaphore_wait(&task->work_ready);
+
+ if (task->thread_index == 0) {
+ /* thread[0]:
+ * - get next scene to rasterize
+ * - map the framebuffer surfaces
+ */
+ const struct pipe_framebuffer_state *fb;
+ boolean write_depth;
+
+ rast->curr_scene = lp_scene_dequeue( rast->full_scenes );
+
+ lp_scene_bin_iter_begin( rast->curr_scene );
+
+ fb = &rast->curr_scene->fb;
+ write_depth = rast->curr_scene->write_depth;
+
+ lp_rast_begin( rast, fb,
+ fb->nr_cbufs != 0,
+ fb->zsbuf != NULL && write_depth );
+ }
+
+ /* Wait for all threads to get here so that threads[1+] don't
+ * get a null rast->curr_scene pointer.
+ */
+ pipe_barrier_wait( &rast->barrier );
+
+ /* do work */
+ if (debug)
+ debug_printf("thread %d doing work\n", task->thread_index);
+ rasterize_scene(rast,
+ task->thread_index,
+ rast->curr_scene,
+ rast->curr_scene->write_depth);
+
+ /* wait for all threads to finish with this scene */
+ pipe_barrier_wait( &rast->barrier );
+
+ if (task->thread_index == 0) {
+ /* thread[0]:
+ * - release the scene object
+ * - unmap the framebuffer surfaces
+ */
+ release_scene( rast, rast->curr_scene );
+ lp_rast_end( rast );
+ }
+
+ /* signal done with work */
+ if (debug)
+ debug_printf("thread %d done working\n", task->thread_index);
+ pipe_semaphore_signal(&task->work_done);
+ }
+
+ return NULL;
+}
+
+
+/**
+ * Initialize semaphores and spawn the threads.
+ */
+static void
+create_rast_threads(struct lp_rasterizer *rast)
+{
+ unsigned i;
+
+ rast->num_threads = util_cpu_caps.nr_cpus;
+ rast->num_threads = debug_get_num_option("LP_NUM_THREADS", rast->num_threads);
+ rast->num_threads = MIN2(rast->num_threads, MAX_THREADS);
+
+ /* NOTE: if num_threads is zero, we won't use any threads */
+ for (i = 0; i < rast->num_threads; i++) {
+ pipe_semaphore_init(&rast->tasks[i].work_ready, 0);
+ pipe_semaphore_init(&rast->tasks[i].work_done, 0);
+ rast->threads[i] = pipe_thread_create(thread_func,
+ (void *) &rast->tasks[i]);
+ }
+}
+
+
+
+/**
+ * Create new lp_rasterizer.
+ * \param empty the queue to put empty scenes on after we've finished
+ * processing them.
+ */
+struct lp_rasterizer *
+lp_rast_create( struct pipe_screen *screen, struct lp_scene_queue *empty )
+{
+ struct lp_rasterizer *rast;
+ unsigned i, cbuf;
+
+ rast = CALLOC_STRUCT(lp_rasterizer);
+ if(!rast)
+ return NULL;
+
+ rast->screen = screen;
+
+ rast->empty_scenes = empty;
+ rast->full_scenes = lp_scene_queue_create();
+
+ for (i = 0; i < Elements(rast->tasks); i++) {
+ for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ )
+ rast->tasks[i].tile.color[cbuf] = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
+
+ rast->tasks[i].tile.depth = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 );
+ rast->tasks[i].rast = rast;
+ rast->tasks[i].thread_index = i;
+ }
+
+ create_rast_threads(rast);
+
+ /* for synchronizing rasterization threads */
+ pipe_barrier_init( &rast->barrier, rast->num_threads );
+
+ return rast;
}
*/
void lp_rast_destroy( struct lp_rasterizer *rast )
{
- pipe_surface_reference(&rast->state.cbuf, NULL);
- pipe_surface_reference(&rast->state.zsbuf, NULL);
- align_free(rast->tile.depth);
- align_free(rast->tile.color);
+ unsigned i, cbuf;
+
+ util_unreference_framebuffer_state(&rast->state.fb);
+
+ for (i = 0; i < Elements(rast->tasks); i++) {
+ align_free(rast->tasks[i].tile.depth);
+ for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ )
+ align_free(rast->tasks[i].tile.color[cbuf]);
+ }
+
+ /* for synchronizing rasterization threads */
+ pipe_barrier_destroy( &rast->barrier );
+
FREE(rast);
}
+
+/** Return number of rasterization threads */
+unsigned
+lp_rast_get_num_threads( struct lp_rasterizer *rast )
+{
+ return rast->num_threads;
+}