#include "util/u_inlines.h"
#include "util/u_memory.h"
#include "util/u_blitter.h"
+#include "util/u_upload_mgr.h"
#include "indices/u_primconvert.h"
#include "pipe/p_screen.h"
uint32_t resolve_uncleared = vc4->resolve & ~vc4->cleared;
uint32_t width = vc4->framebuffer.width;
uint32_t height = vc4->framebuffer.height;
- uint32_t xtiles = align(width, 64) / 64;
- uint32_t ytiles = align(height, 64) / 64;
+ uint32_t stride_in_tiles = align(width, 64) / 64;
+
+ assert(vc4->draw_min_x != ~0 && vc4->draw_min_y != ~0);
+ uint32_t min_x_tile = vc4->draw_min_x / 64;
+ uint32_t min_y_tile = vc4->draw_min_y / 64;
+ uint32_t max_x_tile = (vc4->draw_max_x - 1) / 64;
+ uint32_t max_y_tile = (vc4->draw_max_y - 1) / 64;
+ uint32_t xtiles = max_x_tile - min_x_tile + 1;
+ uint32_t ytiles = max_y_tile - min_y_tile + 1;
#if 0
fprintf(stderr, "RCL: resolve 0x%x clear 0x%x resolve uncleared 0x%x\n",
resolve_uncleared);
#endif
- cl_u8(&vc4->rcl, VC4_PACKET_CLEAR_COLORS);
- cl_u32(&vc4->rcl, vc4->clear_color[0]);
- cl_u32(&vc4->rcl, vc4->clear_color[1]);
- cl_u32(&vc4->rcl, vc4->clear_depth);
- cl_u8(&vc4->rcl, vc4->clear_stencil);
+ uint32_t reloc_size = 9;
+ uint32_t clear_size = 14;
+ uint32_t config_size = 11 + reloc_size;
+ uint32_t loadstore_size = 7 + reloc_size;
+ uint32_t tilecoords_size = 3;
+ uint32_t branch_size = 5 + reloc_size;
+ uint32_t color_store_size = 1;
+ uint32_t semaphore_size = 1;
+ cl_ensure_space(&vc4->rcl,
+ clear_size +
+ config_size +
+ loadstore_size +
+ semaphore_size +
+ xtiles * ytiles * (loadstore_size * 4 +
+ tilecoords_size * 3 +
+ branch_size +
+ color_store_size));
+
+ if (vc4->cleared) {
+ cl_u8(&vc4->rcl, VC4_PACKET_CLEAR_COLORS);
+ cl_u32(&vc4->rcl, vc4->clear_color[0]);
+ cl_u32(&vc4->rcl, vc4->clear_color[1]);
+ cl_u32(&vc4->rcl, vc4->clear_depth);
+ cl_u8(&vc4->rcl, vc4->clear_stencil);
+ }
/* The rendering mode config determines the pointer that's used for
* VC4_PACKET_STORE_MS_TILE_BUFFER address computations. The kernel
VC4_RENDER_CONFIG_MEMORY_FORMAT_SHIFT) |
(vc4_rt_format_is_565(render_surf->base.format) ?
VC4_RENDER_CONFIG_FORMAT_BGR565 :
- VC4_RENDER_CONFIG_FORMAT_RGBA8888) |
- VC4_RENDER_CONFIG_EARLY_Z_COVERAGE_DISABLE));
+ VC4_RENDER_CONFIG_FORMAT_RGBA8888)));
/* The tile buffer normally gets cleared when the previous tile is
* stored. If the clear values changed between frames, then the tile
uint32_t depth_hindex = ztex ? vc4_gem_hindex(vc4, ztex->bo) : 0;
uint32_t tile_alloc_hindex = vc4_gem_hindex(vc4, vc4->tile_alloc);
- for (int y = 0; y < ytiles; y++) {
- for (int x = 0; x < xtiles; x++) {
- bool end_of_frame = (x == xtiles - 1 &&
- y == ytiles - 1);
+ for (int y = min_y_tile; y <= max_y_tile; y++) {
+ for (int x = min_x_tile; x <= max_x_tile; x++) {
+ bool end_of_frame = (x == max_x_tile &&
+ y == max_y_tile);
bool coords_emitted = false;
/* Note that the load doesn't actually occur until the
/* Wait for the binner before jumping to the first
* tile's lists.
*/
- if (x == 0 && y == 0)
+ if (x == min_x_tile && y == min_y_tile)
cl_u8(&vc4->rcl, VC4_PACKET_WAIT_ON_SEMAPHORE);
cl_start_reloc(&vc4->rcl, 1);
cl_u8(&vc4->rcl, VC4_PACKET_BRANCH_TO_SUB_LIST);
cl_reloc_hindex(&vc4->rcl, tile_alloc_hindex,
- (y * xtiles + x) * 32);
+ (y * stride_in_tiles + x) * 32);
if (vc4->resolve & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
vc4_tile_coordinates(vc4, x, y, &coords_emitted);
if (!vc4->needs_flush)
return;
+ /* The RCL setup would choke if the draw bounds cause no drawing, so
+ * just drop the drawing if that's the case.
+ */
+ if (vc4->draw_max_x <= vc4->draw_min_x ||
+ vc4->draw_max_y <= vc4->draw_min_y) {
+ vc4_job_reset(vc4);
+ return;
+ }
+
/* Increment the semaphore indicating that binning is done and
* unblocking the render thread. Note that this doesn't act until the
* FLUSH completes.
*/
+ cl_ensure_space(&vc4->bcl, 8);
cl_u8(&vc4->bcl, VC4_PACKET_INCREMENT_SEMAPHORE);
/* The FLUSH caps all of our bin lists with a VC4_PACKET_RETURN. */
cl_u8(&vc4->bcl, VC4_PACKET_FLUSH);
vc4_setup_rcl(vc4);
- if (vc4_debug & VC4_DEBUG_CL) {
- fprintf(stderr, "BCL:\n");
- vc4_dump_cl(vc4->bcl.base, vc4->bcl.end - vc4->bcl.base, false);
- fprintf(stderr, "RCL:\n");
- vc4_dump_cl(vc4->rcl.base, vc4->rcl.end - vc4->rcl.base, true);
- }
-
- struct drm_vc4_submit_cl submit;
- memset(&submit, 0, sizeof(submit));
-
- submit.bo_handles = vc4->bo_handles.base;
- submit.bo_handle_count = (vc4->bo_handles.next -
- vc4->bo_handles.base) / 4;
- submit.bin_cl = vc4->bcl.base;
- submit.bin_cl_size = vc4->bcl.next - vc4->bcl.base;
- submit.render_cl = vc4->rcl.base;
- submit.render_cl_size = vc4->rcl.next - vc4->rcl.base;
- submit.shader_rec = vc4->shader_rec.base;
- submit.shader_rec_size = vc4->shader_rec.next - vc4->shader_rec.base;
- submit.shader_rec_count = vc4->shader_rec_count;
- submit.uniforms = vc4->uniforms.base;
- submit.uniforms_size = vc4->uniforms.next - vc4->uniforms.base;
-
- if (!(vc4_debug & VC4_DEBUG_NORAST)) {
- int ret;
-
-#ifndef USE_VC4_SIMULATOR
- ret = drmIoctl(vc4->fd, DRM_IOCTL_VC4_SUBMIT_CL, &submit);
-#else
- ret = vc4_simulator_flush(vc4, &submit);
-#endif
- if (ret) {
- fprintf(stderr, "VC4 submit failed\n");
- abort();
- }
- }
-
- vc4->last_emit_seqno = submit.seqno;
-
- if (vc4_debug & VC4_DEBUG_ALWAYS_SYNC) {
- if (!vc4_wait_seqno(vc4->screen, vc4->last_emit_seqno,
- PIPE_TIMEOUT_INFINITE)) {
- fprintf(stderr, "Wait failed.\n");
- abort();
- }
- }
-
- vc4_reset_cl(&vc4->bcl);
- vc4_reset_cl(&vc4->rcl);
- vc4_reset_cl(&vc4->shader_rec);
- vc4_reset_cl(&vc4->uniforms);
- vc4_reset_cl(&vc4->bo_handles);
- struct vc4_bo **referenced_bos = vc4->bo_pointers.base;
- for (int i = 0; i < submit.bo_handle_count; i++)
- vc4_bo_unreference(&referenced_bos[i]);
- vc4_reset_cl(&vc4->bo_pointers);
- vc4->shader_rec_count = 0;
-
- vc4->needs_flush = false;
- vc4->draw_call_queued = false;
-
- /* We have no hardware context saved between our draw calls, so we
- * need to flag the next draw as needing all state emitted. Emitting
- * all state at the start of our draws is also what ensures that we
- * return to the state we need after a previous tile has finished.
- */
- vc4->dirty = ~0;
- vc4->resolve = 0;
- vc4->cleared = 0;
+ vc4_job_submit(vc4);
}
static void
return false;
}
+static void
+vc4_invalidate_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
+{
+ struct vc4_context *vc4 = vc4_context(pctx);
+ struct pipe_surface *zsurf = vc4->framebuffer.zsbuf;
+
+ if (zsurf && zsurf->texture == prsc)
+ vc4->resolve &= ~(PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL);
+}
+
static void
vc4_context_destroy(struct pipe_context *pctx)
{
if (vc4->primconvert)
util_primconvert_destroy(vc4->primconvert);
+ if (vc4->uploader)
+ u_upload_destroy(vc4->uploader);
+
util_slab_destroy(&vc4->transfer_pool);
pipe_surface_reference(&vc4->framebuffer.cbufs[0], NULL);
pctx->priv = priv;
pctx->destroy = vc4_context_destroy;
pctx->flush = vc4_pipe_flush;
+ pctx->invalidate_resource = vc4_invalidate_resource;
vc4_draw_init(pctx);
vc4_state_init(pctx);
vc4_query_init(pctx);
vc4_resource_context_init(pctx);
- vc4_init_cl(vc4, &vc4->bcl);
- vc4_init_cl(vc4, &vc4->rcl);
- vc4_init_cl(vc4, &vc4->shader_rec);
- vc4_init_cl(vc4, &vc4->uniforms);
- vc4_init_cl(vc4, &vc4->bo_handles);
- vc4_init_cl(vc4, &vc4->bo_pointers);
+ vc4_job_init(vc4);
- vc4->dirty = ~0;
vc4->fd = screen->fd;
util_slab_create(&vc4->transfer_pool, sizeof(struct vc4_transfer),
if (!vc4->primconvert)
goto fail;
+ vc4->uploader = u_upload_create(pctx, 16 * 1024, 4,
+ PIPE_BIND_INDEX_BUFFER);
+
vc4_debug |= saved_shaderdb_flag;
return &vc4->base;