* IN THE SOFTWARE.
*/
+#include "util/u_prim.h"
#include "util/u_format.h"
#include "util/u_pack_color.h"
#include "indices/u_primconvert.h"
uint32_t height = vc4->framebuffer.height;
uint32_t tilew = align(width, 64) / 64;
uint32_t tileh = align(height, 64) / 64;
-
- /* Tile alloc memory setup: We use an initial alloc size of 32b. The
- * hardware then aligns that to 256b (we use 4096, because all of our
- * BO allocations align to that anyway), then for some reason the
- * simulator wants an extra page available, even if you have overflow
- * memory set up.
- *
- * XXX: The binner only does 28-bit addressing math, so the tile alloc
- * and tile state should be in the same BO and that BO needs to not
- * cross a 256MB boundary, somehow.
- */
- uint32_t tile_alloc_size = 32 * tilew * tileh;
- tile_alloc_size = align(tile_alloc_size, 4096);
- tile_alloc_size += 4096;
- uint32_t tile_state_size = 48 * tilew * tileh;
- if (!vc4->tile_alloc || vc4->tile_alloc->size < tile_alloc_size) {
- vc4_bo_unreference(&vc4->tile_alloc);
- vc4->tile_alloc = vc4_bo_alloc(vc4->screen, tile_alloc_size,
- "tile_alloc");
- }
- if (!vc4->tile_state || vc4->tile_state->size < tile_state_size) {
- vc4_bo_unreference(&vc4->tile_state);
- vc4->tile_state = vc4_bo_alloc(vc4->screen, tile_state_size,
- "tile_state");
- }
+ struct vc4_cl_out *bcl = cl_start(&vc4->bcl);
// Tile state data is 48 bytes per tile, I think it can be thrown away
// as soon as binning is finished.
- cl_start_reloc(&vc4->bcl, 2);
- cl_u8(&vc4->bcl, VC4_PACKET_TILE_BINNING_MODE_CONFIG);
- cl_reloc(vc4, &vc4->bcl, vc4->tile_alloc, 0);
- cl_u32(&vc4->bcl, vc4->tile_alloc->size);
- cl_reloc(vc4, &vc4->bcl, vc4->tile_state, 0);
- cl_u8(&vc4->bcl, tilew);
- cl_u8(&vc4->bcl, tileh);
- cl_u8(&vc4->bcl,
- VC4_BIN_CONFIG_AUTO_INIT_TSDA |
- VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_32 |
- VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_32);
+ cl_u8(&bcl, VC4_PACKET_TILE_BINNING_MODE_CONFIG);
+ cl_u32(&bcl, 0); /* tile alloc addr, filled by kernel */
+ cl_u32(&bcl, 0); /* tile alloc size, filled by kernel */
+ cl_u32(&bcl, 0); /* tile state addr, filled by kernel */
+ cl_u8(&bcl, tilew);
+ cl_u8(&bcl, tileh);
+ cl_u8(&bcl, 0); /* flags, filled by kernel. */
/* START_TILE_BINNING resets the statechange counters in the hardware,
* which are what is used when a primitive is binned to a tile to
* figure out what new state packets need to be written to that tile's
* command list.
*/
- cl_u8(&vc4->bcl, VC4_PACKET_START_TILE_BINNING);
+ cl_u8(&bcl, VC4_PACKET_START_TILE_BINNING);
/* Reset the current compressed primitives format. This gets modified
* by VC4_PACKET_GL_INDEXED_PRIMITIVE and
* VC4_PACKET_GL_ARRAY_PRIMITIVE, so it needs to be reset at the start
* of every tile.
*/
- cl_u8(&vc4->bcl, VC4_PACKET_PRIMITIVE_LIST_FORMAT);
- cl_u8(&vc4->bcl, (VC4_PRIMITIVE_LIST_FORMAT_16_INDEX |
- VC4_PRIMITIVE_LIST_FORMAT_TYPE_TRIANGLES));
+ cl_u8(&bcl, VC4_PACKET_PRIMITIVE_LIST_FORMAT);
+ cl_u8(&bcl, (VC4_PRIMITIVE_LIST_FORMAT_16_INDEX |
+ VC4_PRIMITIVE_LIST_FORMAT_TYPE_TRIANGLES));
vc4->needs_flush = true;
vc4->draw_call_queued = true;
+ vc4->draw_width = width;
+ vc4->draw_height = height;
+
+ cl_end(&vc4->bcl, bcl);
}
static void
-vc4_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
+vc4_update_shadow_textures(struct pipe_context *pctx,
+ struct vc4_texture_stateobj *stage_tex)
{
- struct vc4_context *vc4 = vc4_context(pctx);
-
- if (info->mode >= PIPE_PRIM_QUADS) {
- util_primconvert_save_index_buffer(vc4->primconvert, &vc4->indexbuf);
- util_primconvert_save_rasterizer_state(vc4->primconvert, &vc4->rasterizer->base);
- util_primconvert_draw_vbo(vc4->primconvert, info);
- return;
+ for (int i = 0; i < stage_tex->num_textures; i++) {
+ struct pipe_sampler_view *view = stage_tex->textures[i];
+ if (!view)
+ continue;
+ struct vc4_resource *rsc = vc4_resource(view->texture);
+ if (rsc->shadow_parent)
+ vc4_update_shadow_baselevel_texture(pctx, view);
}
+}
- vc4_get_draw_cl_space(vc4);
-
+static void
+vc4_emit_gl_shader_state(struct vc4_context *vc4, const struct pipe_draw_info *info)
+{
+ /* VC4_DIRTY_VTXSTATE */
struct vc4_vertex_stateobj *vtx = vc4->vtx;
+ /* VC4_DIRTY_VTXBUF */
struct vc4_vertexbuf_stateobj *vertexbuf = &vc4->vertexbuf;
- if (vc4->prim_mode != info->mode) {
- vc4->prim_mode = info->mode;
- vc4->dirty |= VC4_DIRTY_PRIM_MODE;
- }
-
- vc4_start_draw(vc4);
- vc4_update_compiled_shaders(vc4, info->mode);
-
- vc4_emit_state(pctx);
- vc4->dirty = 0;
-
- vc4_write_uniforms(vc4, vc4->prog.fs,
- &vc4->constbuf[PIPE_SHADER_FRAGMENT],
- &vc4->fragtex);
- vc4_write_uniforms(vc4, vc4->prog.vs,
- &vc4->constbuf[PIPE_SHADER_VERTEX],
- &vc4->verttex);
- vc4_write_uniforms(vc4, vc4->prog.cs,
- &vc4->constbuf[PIPE_SHADER_VERTEX],
- &vc4->verttex);
-
/* The simulator throws a fit if VS or CS don't read an attribute, so
* we emit a dummy read.
*/
uint32_t num_elements_emit = MAX2(vtx->num_elements, 1);
/* Emit the shader record. */
- cl_start_shader_reloc(&vc4->shader_rec, 3 + num_elements_emit);
- cl_u16(&vc4->shader_rec,
+ struct vc4_cl_out *shader_rec =
+ cl_start_shader_reloc(&vc4->shader_rec, 3 + num_elements_emit);
+ /* VC4_DIRTY_PRIM_MODE | VC4_DIRTY_RASTERIZER */
+ cl_u16(&shader_rec,
VC4_SHADER_FLAG_ENABLE_CLIPPING |
+ VC4_SHADER_FLAG_FS_SINGLE_THREAD |
((info->mode == PIPE_PRIM_POINTS &&
vc4->rasterizer->base.point_size_per_vertex) ?
VC4_SHADER_FLAG_VS_POINT_SIZE : 0));
- cl_u8(&vc4->shader_rec, 0); /* fs num uniforms (unused) */
- cl_u8(&vc4->shader_rec, vc4->prog.fs->num_inputs);
- cl_reloc(vc4, &vc4->shader_rec, vc4->prog.fs->bo, 0);
- cl_u32(&vc4->shader_rec, 0); /* UBO offset written by kernel */
-
- cl_u16(&vc4->shader_rec, 0); /* vs num uniforms */
- cl_u8(&vc4->shader_rec, vc4->prog.vs->vattrs_live);
- cl_u8(&vc4->shader_rec, vc4->prog.vs->vattr_offsets[8]);
- cl_reloc(vc4, &vc4->shader_rec, vc4->prog.vs->bo, 0);
- cl_u32(&vc4->shader_rec, 0); /* UBO offset written by kernel */
-
- cl_u16(&vc4->shader_rec, 0); /* cs num uniforms */
- cl_u8(&vc4->shader_rec, vc4->prog.cs->vattrs_live);
- cl_u8(&vc4->shader_rec, vc4->prog.cs->vattr_offsets[8]);
- cl_reloc(vc4, &vc4->shader_rec, vc4->prog.cs->bo, 0);
- cl_u32(&vc4->shader_rec, 0); /* UBO offset written by kernel */
+
+ /* VC4_DIRTY_COMPILED_FS */
+ cl_u8(&shader_rec, 0); /* fs num uniforms (unused) */
+ cl_u8(&shader_rec, vc4->prog.fs->num_inputs);
+ cl_reloc(vc4, &vc4->shader_rec, &shader_rec, vc4->prog.fs->bo, 0);
+ cl_u32(&shader_rec, 0); /* UBO offset written by kernel */
+
+ /* VC4_DIRTY_COMPILED_VS */
+ cl_u16(&shader_rec, 0); /* vs num uniforms */
+ cl_u8(&shader_rec, vc4->prog.vs->vattrs_live);
+ cl_u8(&shader_rec, vc4->prog.vs->vattr_offsets[8]);
+ cl_reloc(vc4, &vc4->shader_rec, &shader_rec, vc4->prog.vs->bo, 0);
+ cl_u32(&shader_rec, 0); /* UBO offset written by kernel */
+
+ /* VC4_DIRTY_COMPILED_CS */
+ cl_u16(&shader_rec, 0); /* cs num uniforms */
+ cl_u8(&shader_rec, vc4->prog.cs->vattrs_live);
+ cl_u8(&shader_rec, vc4->prog.cs->vattr_offsets[8]);
+ cl_reloc(vc4, &vc4->shader_rec, &shader_rec, vc4->prog.cs->bo, 0);
+ cl_u32(&shader_rec, 0); /* UBO offset written by kernel */
uint32_t max_index = 0xffff;
- uint32_t vpm_offset = 0;
for (int i = 0; i < vtx->num_elements; i++) {
struct pipe_vertex_element *elem = &vtx->pipe[i];
struct pipe_vertex_buffer *vb =
&vertexbuf->vb[elem->vertex_buffer_index];
struct vc4_resource *rsc = vc4_resource(vb->buffer);
- uint32_t offset = vb->buffer_offset + elem->src_offset;
+ /* not vc4->dirty tracked: vc4->last_index_bias */
+ uint32_t offset = (vb->buffer_offset +
+ elem->src_offset +
+ vb->stride * info->index_bias);
uint32_t vb_size = rsc->bo->size - offset;
uint32_t elem_size =
util_format_get_blocksize(elem->src_format);
- cl_reloc(vc4, &vc4->shader_rec, rsc->bo, offset);
- cl_u8(&vc4->shader_rec, elem_size - 1);
- cl_u8(&vc4->shader_rec, vb->stride);
- cl_u8(&vc4->shader_rec, vc4->prog.vs->vattr_offsets[i]);
- cl_u8(&vc4->shader_rec, vc4->prog.cs->vattr_offsets[i]);
-
- vpm_offset += align(elem_size, 4);
+ cl_reloc(vc4, &vc4->shader_rec, &shader_rec, rsc->bo, offset);
+ cl_u8(&shader_rec, elem_size - 1);
+ cl_u8(&shader_rec, vb->stride);
+ cl_u8(&shader_rec, vc4->prog.vs->vattr_offsets[i]);
+ cl_u8(&shader_rec, vc4->prog.cs->vattr_offsets[i]);
if (vb->stride > 0) {
max_index = MIN2(max_index,
if (vtx->num_elements == 0) {
assert(num_elements_emit == 1);
struct vc4_bo *bo = vc4_bo_alloc(vc4->screen, 4096, "scratch VBO");
- cl_reloc(vc4, &vc4->shader_rec, bo, 0);
- cl_u8(&vc4->shader_rec, 16 - 1); /* element size */
- cl_u8(&vc4->shader_rec, 0); /* stride */
- cl_u8(&vc4->shader_rec, 0); /* VS VPM offset */
- cl_u8(&vc4->shader_rec, 0); /* CS VPM offset */
+ cl_reloc(vc4, &vc4->shader_rec, &shader_rec, bo, 0);
+ cl_u8(&shader_rec, 16 - 1); /* element size */
+ cl_u8(&shader_rec, 0); /* stride */
+ cl_u8(&shader_rec, 0); /* VS VPM offset */
+ cl_u8(&shader_rec, 0); /* CS VPM offset */
vc4_bo_unreference(&bo);
}
+ cl_end(&vc4->shader_rec, shader_rec);
+ struct vc4_cl_out *bcl = cl_start(&vc4->bcl);
/* the actual draw call. */
- cl_u8(&vc4->bcl, VC4_PACKET_GL_SHADER_STATE);
+ cl_u8(&bcl, VC4_PACKET_GL_SHADER_STATE);
assert(vtx->num_elements <= 8);
/* Note that number of attributes == 0 in the packet means 8
* attributes. This field also contains the offset into shader_rec.
*/
- cl_u32(&vc4->bcl, num_elements_emit & 0x7);
+ cl_u32(&bcl, num_elements_emit & 0x7);
+ cl_end(&vc4->bcl, bcl);
+
+ vc4_write_uniforms(vc4, vc4->prog.fs,
+ &vc4->constbuf[PIPE_SHADER_FRAGMENT],
+ &vc4->fragtex);
+ vc4_write_uniforms(vc4, vc4->prog.vs,
+ &vc4->constbuf[PIPE_SHADER_VERTEX],
+ &vc4->verttex);
+ vc4_write_uniforms(vc4, vc4->prog.cs,
+ &vc4->constbuf[PIPE_SHADER_VERTEX],
+ &vc4->verttex);
+
+ vc4->last_index_bias = info->index_bias;
+ vc4->max_index = max_index;
+}
+
+static void
+vc4_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
+{
+ struct vc4_context *vc4 = vc4_context(pctx);
+
+ if (info->mode >= PIPE_PRIM_QUADS) {
+ util_primconvert_save_index_buffer(vc4->primconvert, &vc4->indexbuf);
+ util_primconvert_save_rasterizer_state(vc4->primconvert, &vc4->rasterizer->base);
+ util_primconvert_draw_vbo(vc4->primconvert, info);
+ perf_debug("Fallback conversion for %d %s vertices\n",
+ info->count, u_prim_name(info->mode));
+ return;
+ }
+
+ /* Before setting up the draw, do any fixup blits necessary. */
+ vc4_update_shadow_textures(pctx, &vc4->verttex);
+ vc4_update_shadow_textures(pctx, &vc4->fragtex);
+
+ vc4_get_draw_cl_space(vc4);
+
+ if (vc4->prim_mode != info->mode) {
+ vc4->prim_mode = info->mode;
+ vc4->dirty |= VC4_DIRTY_PRIM_MODE;
+ }
+
+ vc4_start_draw(vc4);
+ vc4_update_compiled_shaders(vc4, info->mode);
+
+ vc4_emit_state(pctx);
+
+ if ((vc4->dirty & (VC4_DIRTY_VTXBUF |
+ VC4_DIRTY_VTXSTATE |
+ VC4_DIRTY_PRIM_MODE |
+ VC4_DIRTY_RASTERIZER |
+ VC4_DIRTY_COMPILED_CS |
+ VC4_DIRTY_COMPILED_VS |
+ VC4_DIRTY_COMPILED_FS |
+ vc4->prog.cs->uniform_dirty_bits |
+ vc4->prog.vs->uniform_dirty_bits |
+ vc4->prog.fs->uniform_dirty_bits)) ||
+ vc4->last_index_bias != info->index_bias) {
+ vc4_emit_gl_shader_state(vc4, info);
+ }
+
+ vc4->dirty = 0;
/* Note that the primitive type fields match with OpenGL/gallium
* definitions, up to but not including QUADS.
*/
+ struct vc4_cl_out *bcl = cl_start(&vc4->bcl);
if (info->indexed) {
- struct vc4_resource *rsc = vc4_resource(vc4->indexbuf.buffer);
uint32_t offset = vc4->indexbuf.offset;
uint32_t index_size = vc4->indexbuf.index_size;
- if (rsc->shadow_parent) {
- vc4_update_shadow_index_buffer(pctx, &vc4->indexbuf);
- offset = 0;
+ struct pipe_resource *prsc;
+ if (vc4->indexbuf.index_size == 4) {
+ prsc = vc4_get_shadow_index_buffer(pctx, &vc4->indexbuf,
+ info->count, &offset);
+ index_size = 2;
+ } else {
+ prsc = vc4->indexbuf.buffer;
}
+ struct vc4_resource *rsc = vc4_resource(prsc);
- cl_start_reloc(&vc4->bcl, 1);
- cl_u8(&vc4->bcl, VC4_PACKET_GL_INDEXED_PRIMITIVE);
- cl_u8(&vc4->bcl,
+ cl_start_reloc(&vc4->bcl, &bcl, 1);
+ cl_u8(&bcl, VC4_PACKET_GL_INDEXED_PRIMITIVE);
+ cl_u8(&bcl,
info->mode |
(index_size == 2 ?
VC4_INDEX_BUFFER_U16:
VC4_INDEX_BUFFER_U8));
- cl_u32(&vc4->bcl, info->count);
- cl_reloc(vc4, &vc4->bcl, rsc->bo, offset);
- cl_u32(&vc4->bcl, max_index);
+ cl_u32(&bcl, info->count);
+ cl_reloc(vc4, &vc4->bcl, &bcl, rsc->bo, offset);
+ cl_u32(&bcl, vc4->max_index);
+
+ if (vc4->indexbuf.index_size == 4)
+ pipe_resource_reference(&prsc, NULL);
} else {
- cl_u8(&vc4->bcl, VC4_PACKET_GL_ARRAY_PRIMITIVE);
- cl_u8(&vc4->bcl, info->mode);
- cl_u32(&vc4->bcl, info->count);
- cl_u32(&vc4->bcl, info->start);
+ cl_u8(&bcl, VC4_PACKET_GL_ARRAY_PRIMITIVE);
+ cl_u8(&bcl, info->mode);
+ cl_u32(&bcl, info->count);
+ cl_u32(&bcl, info->start);
}
+ cl_end(&vc4->bcl, bcl);
if (vc4->zsa && vc4->zsa->base.depth.enabled) {
vc4->resolve |= PIPE_CLEAR_DEPTH;
{
union util_color uc;
util_pack_color(rgba, format, &uc);
- return uc.ui[0];
+ if (util_format_get_blocksize(format) == 2)
+ return uc.us;
+ else
+ return uc.ui[0];
}
static void
/* We can't flag new buffers for clearing once we've queued draws. We
* could avoid this by using the 3d engine to clear.
*/
- if (vc4->draw_call_queued)
+ if (vc4->draw_call_queued) {
+ perf_debug("Flushing rendering to process new clear.");
vc4_flush(pctx);
+ }
if (buffers & PIPE_CLEAR_COLOR0) {
vc4->clear_color[0] = vc4->clear_color[1] =