* IN THE SOFTWARE.
*/
-#include <stdio.h>
+#include "util/u_prim.h"
+#include "util/u_format.h"
+#include "util/u_pack_color.h"
+#include "indices/u_primconvert.h"
#include "vc4_context.h"
#include "vc4_resource.h"
-static struct vc4_bo *
-get_vbo(struct vc4_context *vc4, uint32_t width, uint32_t height)
-{
- struct {
- float x, y, z, w;
- } verts[] = {
- { -1, -1, 1, 1 },
- { 1, -1, 1, 1 },
- { -1, 1, 1, 1 },
- };
-
- return vc4_bo_alloc_mem(vc4->screen, verts, sizeof(verts), "verts");
-}
-static struct vc4_bo *
-get_ibo(struct vc4_context *vc4)
-{
- static const uint8_t indices[] = { 0, 1, 2 };
-
- return vc4_bo_alloc_mem(vc4->screen, indices, sizeof(indices), "indices");
-}
-
static void
-vc4_rcl_tile_calls(struct vc4_context *vc4,
- uint32_t xtiles, uint32_t ytiles,
- struct vc4_bo *tile_alloc)
+vc4_get_draw_cl_space(struct vc4_context *vc4)
{
- for (int x = 0; x < xtiles; x++) {
- for (int y = 0; y < ytiles; y++) {
- cl_u8(&vc4->rcl, VC4_PACKET_TILE_COORDINATES);
- cl_u8(&vc4->rcl, x);
- cl_u8(&vc4->rcl, y);
-
- cl_start_reloc(&vc4->rcl, 1);
- cl_u8(&vc4->rcl, VC4_PACKET_BRANCH_TO_SUB_LIST);
- cl_reloc(vc4, &vc4->rcl, tile_alloc,
- (y * xtiles + x) * 32);
-
- if (x == xtiles - 1 && y == ytiles - 1) {
- cl_u8(&vc4->rcl,
- VC4_PACKET_STORE_MS_TILE_BUFFER_AND_EOF);
- } else {
- cl_u8(&vc4->rcl,
- VC4_PACKET_STORE_MS_TILE_BUFFER);
- }
- }
- }
+ /* Binner gets our packet state -- vc4_emit.c contents,
+ * and the primitive itself.
+ */
+ cl_ensure_space(&vc4->bcl, 256);
+
+ /* Nothing for rcl -- that's covered by vc4_context.c */
+
+ /* shader_rec gets up to 12 dwords of reloc handles plus a maximally
+ * sized shader_rec (104 bytes base for 8 vattrs plus 32 bytes of
+ * vattr stride).
+ */
+ cl_ensure_space(&vc4->shader_rec, 12 * sizeof(uint32_t) + 104 + 8 * 32);
+
+ /* Uniforms are covered by vc4_write_uniforms(). */
+
+ /* There could be up to 16 textures per stage, plus misc other
+ * pointers.
+ */
+ cl_ensure_space(&vc4->bo_handles, (2 * 16 + 20) * sizeof(uint32_t));
+ cl_ensure_space(&vc4->bo_pointers,
+ (2 * 16 + 20) * sizeof(struct vc4_bo *));
}
+/**
+ * Does the initial bining command list setup for drawing to a given FBO.
+ */
static void
-vc4_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
+vc4_start_draw(struct vc4_context *vc4)
{
- struct vc4_context *vc4 = vc4_context(pctx);
+ if (vc4->needs_flush)
+ return;
+
+ vc4_get_draw_cl_space(vc4);
+
uint32_t width = vc4->framebuffer.width;
uint32_t height = vc4->framebuffer.height;
uint32_t tilew = align(width, 64) / 64;
uint32_t tileh = align(height, 64) / 64;
- struct vc4_bo *tile_alloc = vc4_bo_alloc(vc4->screen,
- 32 * tilew * tileh, "tilea");
- struct vc4_bo *tile_state = vc4_bo_alloc(vc4->screen,
- 48 * tilew * tileh, "tilestate");
- struct vc4_bo *ibo = get_ibo(vc4);
-
- struct vc4_bo *vbo = get_vbo(vc4, width, height);
- static const uint32_t fs_uni[] = { 0 };
- uint32_t vs_uni[] = {
- fui(vc4->framebuffer.width * 16.0f / 2.0f),
- fui(vc4->framebuffer.height * 16.0f / 2.0f),
- };
- uint32_t cs_uni[] = {
- fui(vc4->framebuffer.width * 16.0f / 2.0f),
- fui(vc4->framebuffer.height * 16.0f / 2.0f),
- };
- struct vc4_bo *fs_ubo = vc4_bo_alloc_mem(vc4->screen, fs_uni,
- sizeof(fs_uni), "fs_ubo");
- struct vc4_bo *vs_ubo = vc4_bo_alloc_mem(vc4->screen, vs_uni,
- sizeof(vs_uni), "vs_ubo");
- struct vc4_bo *cs_ubo = vc4_bo_alloc_mem(vc4->screen, cs_uni,
- sizeof(cs_uni), "cs_ubo");
- vc4->needs_flush = true;
+ /* 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");
+ }
// 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, tile_alloc, 0);
- cl_u32(&vc4->bcl, 0x8000); /* tile allocation memory size */
- cl_reloc(vc4, &vc4->bcl, tile_state, 0);
+ 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);
-
+ 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);
+
+ /* 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);
+ /* 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, 0x12); // 16 bit triangle
+ cl_u8(&vc4->bcl, (VC4_PRIMITIVE_LIST_FORMAT_16_INDEX |
+ VC4_PRIMITIVE_LIST_FORMAT_TYPE_TRIANGLES));
- vc4_emit_state(pctx);
+ vc4->needs_flush = true;
+ vc4->draw_call_queued = true;
+}
- /* the actual draw call. */
- uint32_t nr_attributes = 1;
- cl_u8(&vc4->bcl, VC4_PACKET_GL_SHADER_STATE);
-#ifndef USE_VC4_SIMULATOR
- cl_u32(&vc4->bcl, nr_attributes & 0x7); /* offset into shader_rec */
-#else
- cl_u32(&vc4->bcl, simpenrose_hw_addr(vc4->shader_rec.next) |
- (nr_attributes & 0x7));
-#endif
-
- cl_start_reloc(&vc4->bcl, 1);
- cl_u8(&vc4->bcl, VC4_PACKET_GL_INDEXED_PRIMITIVE);
- cl_u8(&vc4->bcl, 0x04); // 8bit index, trinagles
- cl_u32(&vc4->bcl, 3); // Length
- cl_reloc(vc4, &vc4->bcl, ibo, 0);
- cl_u32(&vc4->bcl, 2); // Maximum index
-
- cl_u8(&vc4->bcl, VC4_PACKET_FLUSH_ALL);
- cl_u8(&vc4->bcl, VC4_PACKET_NOP);
- cl_u8(&vc4->bcl, VC4_PACKET_HALT);
-
-// Shader Record
-
- cl_start_shader_reloc(&vc4->shader_rec, 7);
- cl_u16(&vc4->shader_rec, VC4_SHADER_FLAG_ENABLE_CLIPPING);
+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;
+ }
+
+ vc4_get_draw_cl_space(vc4);
+
+ struct vc4_vertex_stateobj *vtx = vc4->vtx;
+ 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,
+ VC4_SHADER_FLAG_ENABLE_CLIPPING |
+ ((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, 0); /* fs num varyings */
+ cl_u8(&vc4->shader_rec, vc4->prog.fs->num_inputs);
cl_reloc(vc4, &vc4->shader_rec, vc4->prog.fs->bo, 0);
- cl_reloc(vc4, &vc4->shader_rec, fs_ubo, 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, 1); /* vs attribute array bitfield */
- cl_u8(&vc4->shader_rec, 16); /* vs total attribute size */
+ 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_reloc(vc4, &vc4->shader_rec, vs_ubo, 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, 1); /* cs attribute array bitfield */
- cl_u8(&vc4->shader_rec, 16); /* vs total attribute size */
- cl_reloc(vc4, &vc4->shader_rec, vc4->prog.vs->bo,
- vc4->prog.vs->coord_shader_offset);
- cl_reloc(vc4, &vc4->shader_rec, cs_ubo, 0);
-
- cl_reloc(vc4, &vc4->shader_rec, vbo, 0);
- cl_u8(&vc4->shader_rec, 15); /* bytes - 1 in the attribute*/
- cl_u8(&vc4->shader_rec, 16); /* attribute stride */
- cl_u8(&vc4->shader_rec, 0); /* VS VPM offset */
- cl_u8(&vc4->shader_rec, 0); /* CS VPM offset */
+ 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 */
+
+ 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;
+ 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);
+
+ if (vb->stride > 0) {
+ max_index = MIN2(max_index,
+ (vb_size - elem_size) / vb->stride);
+ }
+ }
+
+ 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 */
+ vc4_bo_unreference(&bo);
+ }
+
+ /* the actual draw call. */
+ cl_u8(&vc4->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);
+
+ /* Note that the primitive type fields match with OpenGL/gallium
+ * definitions, up to but not including QUADS.
+ */
+ 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;
+ }
+
+ cl_start_reloc(&vc4->bcl, 1);
+ cl_u8(&vc4->bcl, VC4_PACKET_GL_INDEXED_PRIMITIVE);
+ cl_u8(&vc4->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);
+ } 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);
+ }
+
+ if (vc4->zsa && vc4->zsa->base.depth.enabled) {
+ vc4->resolve |= PIPE_CLEAR_DEPTH;
+ }
+ if (vc4->zsa && vc4->zsa->base.stencil[0].enabled)
+ vc4->resolve |= PIPE_CLEAR_STENCIL;
+ vc4->resolve |= PIPE_CLEAR_COLOR0;
vc4->shader_rec_count++;
- cl_u8(&vc4->rcl, VC4_PACKET_CLEAR_COLORS);
- cl_u32(&vc4->rcl, 0xff000000); // Opaque Black
- cl_u32(&vc4->rcl, 0xff000000); // 32 bit clear colours need to be repeated twice
- cl_u32(&vc4->rcl, 0);
- cl_u8(&vc4->rcl, 0);
-
- struct vc4_surface *csurf = vc4_surface(vc4->framebuffer.cbufs[0]);
- struct vc4_resource *ctex = vc4_resource(csurf->base.texture);
-
- cl_start_reloc(&vc4->rcl, 1);
- cl_u8(&vc4->rcl, VC4_PACKET_TILE_RENDERING_MODE_CONFIG);
- cl_reloc(vc4, &vc4->rcl, ctex->bo, csurf->offset);
- cl_u16(&vc4->rcl, width);
- cl_u16(&vc4->rcl, height);
- cl_u8(&vc4->rcl, (VC4_RENDER_CONFIG_MEMORY_FORMAT_LINEAR |
- VC4_RENDER_CONFIG_FORMAT_RGBA8888));
- cl_u8(&vc4->rcl, 0);
-
- // Do a store of the first tile to force the tile buffer to be cleared
- /* XXX: I think these two packets may be unnecessary. */
- cl_u8(&vc4->rcl, VC4_PACKET_TILE_COORDINATES);
- cl_u8(&vc4->rcl, 0);
- cl_u8(&vc4->rcl, 0);
-
- cl_u8(&vc4->rcl, VC4_PACKET_STORE_TILE_BUFFER_GENERAL);
- cl_u16(&vc4->rcl, 0); // Store nothing (just clear)
- cl_u32(&vc4->rcl, 0); // no address is needed
-
- vc4_rcl_tile_calls(vc4, tilew, tileh, tile_alloc);
-
- vc4_flush(pctx);
+ if (vc4_debug & VC4_DEBUG_ALWAYS_FLUSH)
+ vc4_flush(pctx);
+}
+
+static uint32_t
+pack_rgba(enum pipe_format format, const float *rgba)
+{
+ union util_color uc;
+ util_pack_color(rgba, format, &uc);
+ if (util_format_get_blocksize(format) == 2)
+ return uc.us;
+ else
+ return uc.ui[0];
}
static void
{
struct vc4_context *vc4 = vc4_context(pctx);
- vc4->needs_flush = true;
+ /* 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) {
+ perf_debug("Flushing rendering to process new clear.");
+ vc4_flush(pctx);
+ }
+
+ if (buffers & PIPE_CLEAR_COLOR0) {
+ vc4->clear_color[0] = vc4->clear_color[1] =
+ pack_rgba(vc4->framebuffer.cbufs[0]->format,
+ color->f);
+ }
+
+ if (buffers & PIPE_CLEAR_DEPTH) {
+ /* Though the depth buffer is stored with Z in the high 24,
+ * for this field we just need to store it in the low 24.
+ */
+ vc4->clear_depth = util_pack_z(PIPE_FORMAT_Z24X8_UNORM, depth);
+ }
+
+ if (buffers & PIPE_CLEAR_STENCIL)
+ vc4->clear_stencil = stencil;
+
+ vc4->draw_min_x = 0;
+ vc4->draw_min_y = 0;
+ vc4->draw_max_x = vc4->framebuffer.width;
+ vc4->draw_max_y = vc4->framebuffer.height;
+ vc4->cleared |= buffers;
+ vc4->resolve |= buffers;
+
+ vc4_start_draw(vc4);
}
static void