*
**************************************************************************/
+#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_bin_queue.h"
+#include "lp_scene_queue.h"
#include "lp_debug.h"
-#include "lp_state.h"
+#include "lp_fence.h"
+#include "lp_perf.h"
#include "lp_rast.h"
#include "lp_rast_priv.h"
#include "lp_tile_soa.h"
-#include "lp_bld_debug.h"
-#include "lp_bin.h"
+#include "gallivm/lp_bld_debug.h"
+#include "lp_scene.h"
-
-/**
- * Called by rasterization threads to get the next chunk of work.
- * We use a lock to make sure that all the threads get the same bins.
- */
-static struct lp_bins *
-get_next_full_bin( struct lp_rasterizer *rast )
-{
- pipe_mutex_lock( rast->get_bin_mutex );
- if (!rast->curr_bins) {
- /* this will wait until there's something in the queue */
- rast->curr_bins = lp_bins_dequeue( rast->full_bins );
- rast->release_count = 0;
-
- lp_bin_iter_begin( rast->curr_bins );
- }
- pipe_mutex_unlock( rast->get_bin_mutex );
- return rast->curr_bins;
-}
-
-
-/**
- * Called by rasterization threads after they've finished with
- * the current bin. When all threads have called this, we reset
- * the bin and put it into the 'empty bins' queue.
- */
-static void
-release_current_bin( struct lp_rasterizer *rast )
-{
- pipe_mutex_lock( rast->get_bin_mutex );
- rast->release_count++;
- if (rast->release_count == rast->num_threads) {
- assert(rast->curr_bins);
- lp_reset_bins( rast->curr_bins );
- lp_bins_enqueue( rast->empty_bins, rast->curr_bins );
- rast->curr_bins = NULL;
- }
- pipe_mutex_unlock( rast->get_bin_mutex );
-}
-
-
-
-/**
- * Begin the rasterization phase.
- * Map the framebuffer surfaces. Initialize the 'rast' state.
+/* Begin rasterizing a scene:
*/
static boolean
lp_rast_begin( struct lp_rasterizer *rast,
- const struct pipe_framebuffer_state *fb,
- boolean write_color,
- boolean write_zstencil )
+ struct lp_scene *scene )
{
- struct pipe_screen *screen = rast->screen;
- struct pipe_surface *cbuf, *zsbuf;
+ const struct pipe_framebuffer_state *fb = &scene->fb;
+ boolean write_color = fb->nr_cbufs != 0;
+ boolean write_zstencil = fb->zsbuf != NULL;
+ int i;
- LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+ rast->curr_scene = scene;
- util_copy_framebuffer_state(&rast->state.fb, fb);
+ LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+ rast->state.nr_cbufs = scene->fb.nr_cbufs;
rast->state.write_zstencil = write_zstencil;
rast->state.write_color = write_color;
-
- rast->check_for_clipped_tiles = (fb->width % TILE_SIZE != 0 ||
- fb->height % TILE_SIZE != 0);
-
- /* XXX support multiple color buffers here */
- cbuf = rast->state.fb.cbufs[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,
- fb->width, fb->height);
- if (!rast->cbuf_transfer)
- return FALSE;
-
- rast->cbuf_map = screen->transfer_map(rast->screen,
- rast->cbuf_transfer);
- if (!rast->cbuf_map)
- return FALSE;
+
+ for (i = 0; i < rast->state.nr_cbufs; i++) {
+ struct pipe_surface *cbuf = scene->fb.cbufs[i];
+ rast->cbuf[i].map = scene->cbuf_map[i];
+ rast->cbuf[i].format = cbuf->texture->format;
+ rast->cbuf[i].width = cbuf->width;
+ rast->cbuf[i].height = cbuf->height;
+ rast->cbuf[i].stride = llvmpipe_texture_stride(cbuf->texture, cbuf->level);
}
- 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,
- fb->width, fb->height);
- if (!rast->zsbuf_transfer)
- return FALSE;
-
- rast->zsbuf_map = screen->transfer_map(rast->screen,
- rast->zsbuf_transfer);
- if (!rast->zsbuf_map)
- return FALSE;
+ if (write_zstencil) {
+ struct pipe_surface *zsbuf = scene->fb.zsbuf;
+ rast->zsbuf.map = scene->zsbuf_map;
+ rast->zsbuf.stride = llvmpipe_texture_stride(zsbuf->texture, zsbuf->level);
+ rast->zsbuf.blocksize =
+ util_format_get_blocksize(zsbuf->texture->format);
}
+ lp_scene_bin_iter_begin( scene );
+
return TRUE;
}
-/**
- * Finish the rasterization phase.
- * Unmap framebuffer surfaces.
- */
static void
lp_rast_end( struct lp_rasterizer *rast )
{
- struct pipe_screen *screen = rast->screen;
-
- if (rast->cbuf_map)
- screen->transfer_unmap(screen, rast->cbuf_transfer);
+ int i;
- if (rast->zsbuf_map)
- screen->transfer_unmap(screen, rast->zsbuf_transfer);
+ lp_scene_reset( rast->curr_scene );
- if (rast->cbuf_transfer)
- screen->tex_transfer_destroy(rast->cbuf_transfer);
+ for (i = 0; i < rast->state.nr_cbufs; i++)
+ rast->cbuf[i].map = NULL;
- 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;
+ rast->zsbuf.map = NULL;
+ rast->curr_scene = NULL;
}
-
/**
* Begining rasterization of a tile.
* \param x window X position of the tile, in pixels
* \param y window Y position of the tile, in pixels
*/
static void
-lp_rast_start_tile( struct lp_rasterizer *rast,
- unsigned thread_index,
- unsigned x, unsigned y )
+lp_rast_start_tile(struct lp_rasterizer_task *task,
+ unsigned x, unsigned y)
{
LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, x, y);
- rast->tasks[thread_index].x = x;
- rast->tasks[thread_index].y = y;
+ task->x = x;
+ task->y = y;
}
* Clear the rasterizer's current color tile.
* 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 )
+void
+lp_rast_clear_color(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
{
+ struct lp_rasterizer *rast = task->rast;
const uint8_t *clear_color = arg.clear_color;
- uint8_t *color_tile = rast->tasks[thread_index].tile.color;
-
+ uint8_t **color_tile = task->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],
if (clear_color[0] == clear_color[1] &&
clear_color[1] == clear_color[2] &&
clear_color[2] == clear_color[3]) {
- memset(color_tile, clear_color[0], TILE_SIZE * TILE_SIZE * 4);
+ /* clear to grayscale value {x, x, x, x} */
+ for (i = 0; i < rast->state.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(color_tile, x, y, chan) = clear_color[chan];
+ /* Non-gray color.
+ * Note: if the swizzled tile layout changes (see TILE_PIXEL) this code
+ * will need to change. It'll be pretty obvious when clearing no longer
+ * works.
+ */
+ const unsigned chunk = TILE_SIZE / 4;
+ for (i = 0; i < rast->state.nr_cbufs; i++) {
+ uint8_t *c = color_tile[i];
+ unsigned j;
+ for (j = 0; j < 4 * TILE_SIZE; j++) {
+ memset(c, clear_color[0], chunk);
+ c += chunk;
+ memset(c, clear_color[1], chunk);
+ c += chunk;
+ memset(c, clear_color[2], chunk);
+ c += chunk;
+ memset(c, clear_color[3], chunk);
+ c += chunk;
+ }
+ assert(c - color_tile[i] == TILE_SIZE * TILE_SIZE * 4);
+ }
}
+
+ LP_COUNT(nr_color_tile_clear);
}
* Clear the rasterizer's current z/stencil tile.
* 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)
+void
+lp_rast_clear_zstencil(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
{
+ struct lp_rasterizer *rast = task->rast;
+ const unsigned tile_x = task->x;
+ const unsigned tile_y = task->y;
+ const unsigned height = TILE_SIZE / TILE_VECTOR_HEIGHT;
+ const unsigned width = TILE_SIZE * TILE_VECTOR_HEIGHT;
+ unsigned block_size = rast->zsbuf.blocksize;
+ uint8_t *dst;
+ unsigned dst_stride = rast->zsbuf.stride * TILE_VECTOR_HEIGHT;
unsigned i, j;
- uint32_t *depth_tile = rast->tasks[thread_index].tile.depth;
-
+
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++)
- depth_tile[i*TILE_SIZE + j] = arg.clear_zstencil;
+ /*assert(rast->zsbuf.map);*/
+ if (!rast->zsbuf.map)
+ return;
+
+ LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+ /*
+ * Clear the aera of the swizzled depth/depth buffer matching this tile, in
+ * stripes of TILE_VECTOR_HEIGHT x TILE_SIZE at a time.
+ *
+ * The swizzled depth format is such that the depths for
+ * TILE_VECTOR_HEIGHT x TILE_VECTOR_WIDTH pixels have consecutive offsets.
+ */
+
+ dst = lp_rast_depth_pointer(rast, tile_x, tile_y);
+
+ switch (block_size) {
+ case 1:
+ memset(dst, (uint8_t) arg.clear_zstencil, height * width);
+ break;
+ case 2:
+ for (i = 0; i < height; i++) {
+ uint16_t *row = (uint16_t *)dst;
+ for (j = 0; j < width; j++)
+ *row++ = (uint16_t) arg.clear_zstencil;
+ dst += dst_stride;
+ }
+ break;
+ case 4:
+ for (i = 0; i < height; i++) {
+ uint32_t *row = (uint32_t *)dst;
+ for (j = 0; j < width; j++)
+ *row++ = arg.clear_zstencil;
+ dst += dst_stride;
+ }
+ break;
+ default:
+ assert(0);
+ break;
+ }
}
* Load tile color from the framebuffer surface.
* 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)
+void
+lp_rast_load_color(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
{
- LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+ struct lp_rasterizer *rast = task->rast;
+ const unsigned x = task->x, y = task->y;
+ unsigned i;
- /* call u_tile func to load colors from surface */
-}
+ LP_DBG(DEBUG_RAST, "%s at %u, %u\n", __FUNCTION__, x, y);
+ for (i = 0; i < rast->state.nr_cbufs; i++) {
+ if (x >= rast->cbuf[i].width || y >= rast->cbuf[i].height)
+ continue;
-/**
- * Load tile z/stencil from the framebuffer surface.
- * 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 )
-{
- LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+ lp_tile_read_4ub(rast->cbuf[i].format,
+ task->tile.color[i],
+ rast->cbuf[i].map,
+ rast->cbuf[i].stride,
+ x, y,
+ TILE_SIZE, TILE_SIZE);
- /* call u_tile func to load depth (and stencil?) from surface */
+ LP_COUNT(nr_color_tile_load);
+ }
}
-void lp_rast_set_state( struct lp_rasterizer *rast,
- unsigned thread_index,
- const union lp_rast_cmd_arg arg )
+void
+lp_rast_set_state(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
{
const struct lp_rast_state *state = arg.set_state;
LP_DBG(DEBUG_RAST, "%s %p\n", __FUNCTION__, (void *) state);
/* just set the current state pointer for this rasterizer */
- rast->tasks[thread_index].current_state = state;
+ task->current_state = state;
}
-/* Within a tile:
- */
-
/**
* Run the shader on all blocks in a tile. This is used when a tile is
* completely contained inside a triangle.
* 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 )
+void
+lp_rast_shade_tile(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
{
+ struct lp_rasterizer *rast = task->rast;
+ const struct lp_rast_state *state = task->current_state;
+ struct lp_rast_tile *tile = &task->tile;
const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
- const unsigned tile_x = rast->tasks[thread_index].x;
- const unsigned tile_y = rast->tasks[thread_index].y;
- const unsigned mask = ~0;
+ const unsigned tile_x = task->x, tile_y = task->y;
unsigned x, y;
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,
- thread_index,
- inputs,
- tile_x + x,
- tile_y + y,
- mask);
+ /* render the whole 64x64 tile in 4x4 chunks */
+ for (y = 0; y < TILE_SIZE; y += 4){
+ for (x = 0; x < TILE_SIZE; x += 4) {
+ uint8_t *color[PIPE_MAX_COLOR_BUFS];
+ uint32_t *depth;
+ unsigned block_offset, i;
+
+ /* offset of the 16x16 pixel block within the tile */
+ block_offset = ((y / 4) * (16 * 16) + (x / 4) * 16);
+
+ /* color buffer */
+ for (i = 0; i < rast->state.nr_cbufs; i++)
+ color[i] = tile->color[i] + 4 * block_offset;
+
+ /* depth buffer */
+ depth = lp_rast_depth_pointer(rast, tile_x + x, tile_y + y);
+
+ /* run shader */
+ state->jit_function[RAST_WHOLE]( &state->jit_context,
+ tile_x + x, tile_y + y,
+ inputs->facing,
+ inputs->a0,
+ inputs->dadx,
+ inputs->dady,
+ color,
+ depth,
+ INT_MIN, INT_MIN, INT_MIN,
+ NULL, NULL, NULL );
+ }
+ }
}
* Compute shading for a 4x4 block of pixels.
* This is a bin command called during bin processing.
*/
-void lp_rast_shade_quads( struct lp_rasterizer *rast,
- unsigned thread_index,
+void lp_rast_shade_quads( struct lp_rasterizer_task *task,
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 = rast->tasks[thread_index].current_state;
- struct lp_rast_tile *tile = &rast->tasks[thread_index].tile;
- void *color;
+ const struct lp_rast_state *state = task->current_state;
+ struct lp_rasterizer *rast = task->rast;
+ struct lp_rast_tile *tile = &task->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;
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;
-#endif
-
- assert((x % 2) == 0);
- assert((y % 2) == 0);
+ assert((x % 4) == 0);
+ assert((y % 4) == 0);
ix = x % TILE_SIZE;
iy = y % TILE_SIZE;
/* offset of the 16x16 pixel block within the tile */
- block_offset = ((iy/4)*(16*16) + (ix/4)*16);
+ block_offset = ((iy / 4) * (16 * 16) + (ix / 4) * 16);
/* color buffer */
- color = tile->color + 4 * block_offset;
+ for (i = 0; i < rast->state.nr_cbufs; i++)
+ color[i] = tile->color[i] + 4 * block_offset;
/* depth buffer */
- depth = tile->depth + block_offset;
+ depth = lp_rast_depth_pointer(rast, x, y);
- /* 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));
+ assert(lp_check_alignment(tile->color[0], 16));
assert(lp_check_alignment(state->jit_context.blend_color, 16));
- /* 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;
+ assert(lp_check_alignment(inputs->step[0], 16));
+ assert(lp_check_alignment(inputs->step[1], 16));
+ assert(lp_check_alignment(inputs->step[2], 16));
- /* 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
+ /* run shader */
+ state->jit_function[RAST_EDGE_TEST]( &state->jit_context,
+ x, y,
+ inputs->facing,
+ inputs->a0,
+ inputs->dadx,
+ inputs->dady,
+ color,
+ depth,
+ c1, c2, c3,
+ inputs->step[0],
+ inputs->step[1],
+ inputs->step[2]);
}
-/* End of tile:
- */
-
-
/**
- * Write the rasterizer's color tile to the framebuffer.
+ * Set top row and left column of the tile's pixels to white. For debugging.
*/
-static void lp_rast_store_color( struct lp_rasterizer *rast,
- unsigned thread_index)
+static void
+outline_tile(uint8_t *tile)
{
- const unsigned x = rast->tasks[thread_index].x;
- const unsigned y = rast->tasks[thread_index].y;
- int w = TILE_SIZE;
- int h = TILE_SIZE;
-
- if (x + w > rast->state.fb.width)
- w -= x + w - rast->state.fb.width;
-
- if (y + h > rast->state.fb.height)
- h -= y + h - rast->state.fb.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(rast->cbuf_transfer->format,
- rast->tasks[thread_index].tile.color,
- rast->cbuf_map,
- rast->cbuf_transfer->stride,
- x, y,
- w, h);
+ const uint8_t val = 0xff;
+ unsigned i;
+
+ for (i = 0; i < TILE_SIZE; i++) {
+ TILE_PIXEL(tile, i, 0, 0) = val;
+ TILE_PIXEL(tile, i, 0, 1) = val;
+ TILE_PIXEL(tile, i, 0, 2) = val;
+ TILE_PIXEL(tile, i, 0, 3) = val;
+
+ TILE_PIXEL(tile, 0, i, 0) = val;
+ TILE_PIXEL(tile, 0, i, 1) = val;
+ TILE_PIXEL(tile, 0, i, 2) = val;
+ TILE_PIXEL(tile, 0, i, 3) = val;
+ }
}
+/**
+ * Draw grid of gray lines at 16-pixel intervals across the tile to
+ * show the sub-tile boundaries. For debugging.
+ */
static void
-lp_tile_write_z32(const uint32_t *src, uint8_t *dst, unsigned dst_stride,
- unsigned x0, unsigned y0, unsigned w, unsigned h)
+outline_subtiles(uint8_t *tile)
{
- unsigned x, y;
- uint8_t *dst_row = dst + y0*dst_stride;
- for (y = 0; y < h; ++y) {
- uint32_t *dst_pixel = (uint32_t *)(dst_row + x0*4);
- for (x = 0; x < w; ++x) {
- *dst_pixel++ = *src++;
+ const uint8_t val = 0x80;
+ const unsigned step = 16;
+ unsigned i, j;
+
+ for (i = 0; i < TILE_SIZE; i += step) {
+ for (j = 0; j < TILE_SIZE; j++) {
+ TILE_PIXEL(tile, i, j, 0) = val;
+ TILE_PIXEL(tile, i, j, 1) = val;
+ TILE_PIXEL(tile, i, j, 2) = val;
+ TILE_PIXEL(tile, i, j, 3) = val;
+
+ TILE_PIXEL(tile, j, i, 0) = val;
+ TILE_PIXEL(tile, j, i, 1) = val;
+ TILE_PIXEL(tile, j, i, 2) = val;
+ TILE_PIXEL(tile, j, i, 3) = val;
}
- dst_row += dst_stride;
}
+
+ outline_tile(tile);
}
+
+
/**
- * Write the rasterizer's z/stencil tile to the framebuffer.
+ * Write the rasterizer's color tile to the framebuffer.
*/
-static void lp_rast_store_zstencil( struct lp_rasterizer *rast,
- unsigned thread_index )
+static void
+lp_rast_store_color(struct lp_rasterizer_task *task)
{
- const unsigned x = rast->tasks[thread_index].x;
- const unsigned y = rast->tasks[thread_index].y;
- unsigned w = TILE_SIZE;
- unsigned h = TILE_SIZE;
+ struct lp_rasterizer *rast = task->rast;
+ const unsigned x = task->x, y = task->y;
+ unsigned i;
+
+ for (i = 0; i < rast->state.nr_cbufs; i++) {
+ if (x >= rast->cbuf[i].width)
+ continue;
+
+ if (y >= rast->cbuf[i].height)
+ continue;
- if (x + w > rast->state.fb.width)
- w -= x + w - rast->state.fb.width;
+ LP_DBG(DEBUG_RAST, "%s [%u] %d,%d\n", __FUNCTION__,
+ task->thread_index, x, y);
- if (y + h > rast->state.fb.height)
- h -= y + h - rast->state.fb.height;
+ if (LP_DEBUG & DEBUG_SHOW_SUBTILES)
+ outline_subtiles(task->tile.color[i]);
+ else if (LP_DEBUG & DEBUG_SHOW_TILES)
+ outline_tile(task->tile.color[i]);
- LP_DBG(DEBUG_RAST, "%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h);
+ lp_tile_write_4ub(rast->cbuf[i].format,
+ task->tile.color[i],
+ rast->cbuf[i].map,
+ rast->cbuf[i].stride,
+ x, y,
+ TILE_SIZE, TILE_SIZE);
- assert(rast->zsbuf_transfer->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);
+ LP_COUNT(nr_color_tile_store);
+ }
}
+
/**
- * Write the rasterizer's tiles to the framebuffer.
+ * Signal on a fence. This is called during bin execution/rasterization.
+ * Called per thread.
*/
-static void
-lp_rast_end_tile( struct lp_rasterizer *rast,
- unsigned thread_index )
+void
+lp_rast_fence(struct lp_rasterizer_task *task,
+ const union lp_rast_cmd_arg arg)
{
- LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+ struct lp_fence *fence = arg.fence;
+ lp_fence_signal(fence);
+}
- if (rast->state.write_color)
- lp_rast_store_color(rast, thread_index);
- if (rast->state.write_zstencil)
- lp_rast_store_zstencil(rast, thread_index);
-}
/**
* Called per thread.
*/
static void
-rasterize_bin( struct lp_rasterizer *rast,
- unsigned thread_index,
- const struct cmd_bin *bin,
- int x, int y)
+rasterize_bin(struct lp_rasterizer_task *task,
+ 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 );
+ lp_rast_start_tile( task, x * TILE_SIZE, y * TILE_SIZE );
/* 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] );
+ block->cmd[k]( task, block->arg[k] );
+ }
+ }
+
+ /* Write the rasterizer's tiles to the framebuffer.
+ */
+ if (task->rast->state.write_color)
+ lp_rast_store_color(task);
+
+ /* Free data for this bin.
+ */
+ lp_scene_bin_reset( task->rast->curr_scene, x, y);
+}
+
+
+#define RAST(x) { lp_rast_##x, #x }
+
+static struct {
+ lp_rast_cmd cmd;
+ const char *name;
+} cmd_names[] =
+{
+ RAST(load_color),
+ RAST(clear_color),
+ RAST(clear_zstencil),
+ RAST(triangle),
+ RAST(shade_tile),
+ RAST(set_state),
+ RAST(fence),
+};
+
+static void
+debug_bin( const struct cmd_bin *bin )
+{
+ const struct cmd_block *head = bin->commands.head;
+ int i, j;
+
+ for (i = 0; i < head->count; i++) {
+ debug_printf("%d: ", i);
+ for (j = 0; j < Elements(cmd_names); j++) {
+ if (head->cmd[i] == cmd_names[j].cmd) {
+ debug_printf("%s\n", cmd_names[j].name);
+ break;
+ }
}
+ if (j == Elements(cmd_names))
+ debug_printf("...other\n");
+ }
+
+}
+
+/* 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;
+
+ if (0)
+ debug_bin(bin);
+
+ /* We emit at most two load-tile commands at the start of the first
+ * command block. In addition we seem to emit a couple of
+ * set-state commands even in empty bins.
+ *
+ * As a heuristic, if a bin has more than 4 commands, consider it
+ * non-empty.
+ */
+ if (head->next != NULL ||
+ head->count > 4) {
+ return FALSE;
}
- lp_rast_end_tile( rast, thread_index );
+ for (i = 0; i < head->count; i++)
+ if (head->cmd[i] != lp_rast_load_color &&
+ head->cmd[i] != lp_rast_set_state) {
+ return FALSE;
+ }
+
+ return TRUE;
}
+
/**
- * Rasterize/execute all bins.
+ * Rasterize/execute all bins within a scene.
* Called per thread.
*/
static void
-rasterize_bins( struct lp_rasterizer *rast,
- unsigned thread_index,
- struct lp_bins *bins,
- bool write_depth )
+rasterize_scene(struct lp_rasterizer_task *task,
+ struct lp_scene *scene)
{
- /* loop over tile bins, rasterize each */
+ /* loop over scene bins, rasterize each */
#if 0
{
unsigned i, j;
- for (i = 0; i < bins->tiles_x; i++) {
- for (j = 0; j < bins->tiles_y; j++) {
- struct cmd_bin *bin = lp_get_bin(bins, i, j);
- rasterize_bin( rast, thread_index,
- bin, i * TILE_SIZE, j * TILE_SIZE );
+ for (i = 0; i < scene->tiles_x; i++) {
+ for (j = 0; j < scene->tiles_y; j++) {
+ struct cmd_bin *bin = lp_scene_get_bin(scene, i, j);
+ rasterize_bin(task, bin, i, j);
}
}
}
struct cmd_bin *bin;
int x, y;
- assert(bins);
- while ((bin = lp_bin_iter_next(bins, &x, &y))) {
- rasterize_bin( rast, thread_index, bin, x * TILE_SIZE, y * TILE_SIZE);
+ assert(scene);
+ while ((bin = lp_scene_bin_iter_next(scene, &x, &y))) {
+ if (!is_empty_bin( bin ))
+ rasterize_bin(task, bin, x, y);
}
}
#endif
* Called by setup module when it has something for us to render.
*/
void
-lp_rasterize_bins( struct lp_rasterizer *rast,
- struct lp_bins *bins,
- const struct pipe_framebuffer_state *fb,
- bool write_depth )
+lp_rast_queue_scene( struct lp_rasterizer *rast,
+ struct lp_scene *scene)
{
- boolean debug = false;
-
LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
- if (debug) {
- unsigned x, y;
- printf("rasterize bins:\n");
- printf(" data size: %u\n", lp_bin_data_size(bins));
- for (y = 0; y < bins->tiles_y; y++) {
- for (x = 0; x < bins->tiles_x; x++) {
- printf(" bin %u, %u size: %u\n", x, y,
- lp_bin_cmd_size(bins, x, y));
- }
- }
- }
-
- lp_rast_begin( rast, fb,
- fb->cbufs[0]!= NULL,
- fb->zsbuf != NULL && write_depth );
-
if (rast->num_threads == 0) {
/* no threading */
- lp_bin_iter_begin( bins );
- rasterize_bins( rast, 0, bins, write_depth );
- /* reset bins and put into the empty queue */
- lp_reset_bins( bins );
- lp_bins_enqueue( rast->empty_bins, bins);
+ lp_rast_begin( rast, scene );
+
+ rasterize_scene( &rast->tasks[0], scene );
+
+ lp_scene_reset( scene );
+ rast->curr_scene = NULL;
}
else {
/* threaded rendering! */
unsigned i;
- lp_bins_enqueue( rast->full_bins, bins );
-
- /* XXX need to move/fix these */
- rast->fb = fb;
- rast->write_depth = write_depth;
-
- /*lp_bin_iter_begin( bins );*/
+ 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);
}
+ }
+
+ LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
+}
+
+
+void
+lp_rast_finish( struct lp_rasterizer *rast )
+{
+ if (rast->num_threads == 0) {
+ /* nothing to do */
+ }
+ else {
+ int i;
/* wait for work to complete */
for (i = 0; i < rast->num_threads; i++) {
pipe_semaphore_wait(&rast->tasks[i].work_done);
}
}
-
- lp_rast_end( rast );
-
- LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
}
* 2. do work
* 3. signal that we're done
*/
-static void *
-thread_func( void *init_data )
+static PIPE_THREAD_ROUTINE( thread_func, init_data )
{
struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
struct lp_rasterizer *rast = task->rast;
boolean debug = false;
while (1) {
- struct lp_bins *bins;
-
/* wait for work */
if (debug)
debug_printf("thread %d waiting for work\n", task->thread_index);
pipe_semaphore_wait(&task->work_ready);
- bins = get_next_full_bin( rast );
- assert(bins);
+ if (rast->exit_flag)
+ break;
+
+ if (task->thread_index == 0) {
+ /* thread[0]:
+ * - get next scene to rasterize
+ * - map the framebuffer surfaces
+ */
+ lp_rast_begin( rast,
+ lp_scene_dequeue( rast->full_scenes, TRUE ) );
+ }
+
+ /* 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_bins(rast, task->thread_index,
- bins, rast->write_depth);
+
+ rasterize_scene(task,
+ rast->curr_scene);
- release_current_bin( rast );
+ /* wait for all threads to finish with this scene */
+ pipe_barrier_wait( &rast->barrier );
+
+ /* XXX: shouldn't be necessary:
+ */
+ if (task->thread_index == 0) {
+ 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);
}
{
unsigned i;
+#ifdef PIPE_OS_WINDOWS
+ /* Multithreading not supported on windows until conditions and barriers are
+ * properly implemented. */
+ rast->num_threads = 0;
+#else
+#ifdef PIPE_OS_EMBEDDED
+ rast->num_threads = 0;
+#else
rast->num_threads = util_cpu_caps.nr_cpus;
+#endif
rast->num_threads = debug_get_num_option("LP_NUM_THREADS", rast->num_threads);
rast->num_threads = MIN2(rast->num_threads, MAX_THREADS);
+#endif
/* NOTE: if num_threads is zero, we won't use any threads */
for (i = 0; i < rast->num_threads; i++) {
/**
* Create new lp_rasterizer.
- * \param empty the queue to put empty bins on after we've finished
+ * \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_bins_queue *empty )
+lp_rast_create( void )
{
struct lp_rasterizer *rast;
- unsigned i;
+ unsigned i, cbuf;
rast = CALLOC_STRUCT(lp_rasterizer);
if(!rast)
return NULL;
- rast->screen = screen;
-
- rast->empty_bins = empty;
- rast->full_bins = lp_bins_queue_create();
+ rast->full_scenes = lp_scene_queue_create();
for (i = 0; i < Elements(rast->tasks); i++) {
- rast->tasks[i].tile.color = 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;
+ struct lp_rasterizer_task *task = &rast->tasks[i];
+
+ for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ )
+ task->tile.color[cbuf] = align_malloc(TILE_SIZE * TILE_SIZE * 4, 16);
+
+ task->rast = rast;
+ task->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 )
{
- unsigned i;
-
- util_unreference_framebuffer_state(&rast->state.fb);
+ unsigned i, cbuf;
for (i = 0; i < Elements(rast->tasks); i++) {
- align_free(rast->tasks[i].tile.depth);
- align_free(rast->tasks[i].tile.color);
+ for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ )
+ align_free(rast->tasks[i].tile.color[cbuf]);
}
+ /* Set exit_flag and signal each thread's work_ready semaphore.
+ * Each thread will be woken up, notice that the exit_flag is set and
+ * break out of its main loop. The thread will then exit.
+ */
+ rast->exit_flag = TRUE;
+ for (i = 0; i < rast->num_threads; i++) {
+ pipe_semaphore_signal(&rast->tasks[i].work_ready);
+ }
+
+ /* Wait for threads to terminate before cleaning up per-thread data */
+ for (i = 0; i < rast->num_threads; i++) {
+ pipe_thread_wait(rast->threads[i]);
+ }
+
+ /* Clean up per-thread data */
+ for (i = 0; i < rast->num_threads; i++) {
+ pipe_semaphore_destroy(&rast->tasks[i].work_ready);
+ pipe_semaphore_destroy(&rast->tasks[i].work_done);
+ }
+
+ /* 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;
+}