#include "freedreno_context.h"
#include "freedreno_state.h"
#include "freedreno_resource.h"
+#include "freedreno_query_hw.h"
#include "freedreno_util.h"
-static enum pc_di_primtype
-mode2primtype(unsigned mode)
-{
- switch (mode) {
- case PIPE_PRIM_POINTS: return DI_PT_POINTLIST;
- case PIPE_PRIM_LINES: return DI_PT_LINELIST;
- case PIPE_PRIM_LINE_STRIP: return DI_PT_LINESTRIP;
- case PIPE_PRIM_TRIANGLES: return DI_PT_TRILIST;
- case PIPE_PRIM_TRIANGLE_STRIP: return DI_PT_TRISTRIP;
- case PIPE_PRIM_TRIANGLE_FAN: return DI_PT_TRIFAN;
- case PIPE_PRIM_QUADS: return DI_PT_QUADLIST;
- case PIPE_PRIM_QUAD_STRIP: return DI_PT_QUADSTRIP;
- case PIPE_PRIM_POLYGON: return DI_PT_POLYGON;
- }
- DBG("unsupported mode: (%s) %d", u_prim_name(mode), mode);
- assert(0);
- return DI_PT_NONE;
-}
-
-static enum pc_di_index_size
-size2indextype(unsigned index_size)
-{
- switch (index_size) {
- case 1: return INDEX_SIZE_8_BIT;
- case 2: return INDEX_SIZE_16_BIT;
- case 4: return INDEX_SIZE_32_BIT;
- }
- DBG("unsupported index size: %d", index_size);
- assert(0);
- return INDEX_SIZE_IGN;
-}
-
-/* this is same for a2xx/a3xx, so split into helper: */
-void
-fd_draw_emit(struct fd_context *ctx, const struct pipe_draw_info *info)
-{
- struct fd_ringbuffer *ring = ctx->ring;
- struct pipe_index_buffer *idx = &ctx->indexbuf;
- struct fd_bo *idx_bo = NULL;
- enum pc_di_index_size idx_type = INDEX_SIZE_IGN;
- enum pc_di_src_sel src_sel;
- uint32_t idx_size, idx_offset;
-
- if (info->indexed) {
- assert(!idx->user_buffer);
-
- idx_bo = fd_resource(idx->buffer)->bo;
- idx_type = size2indextype(idx->index_size);
- idx_size = idx->index_size * info->count;
- idx_offset = idx->offset;
- src_sel = DI_SRC_SEL_DMA;
- } else {
- idx_bo = NULL;
- idx_type = INDEX_SIZE_IGN;
- idx_size = 0;
- idx_offset = 0;
- src_sel = DI_SRC_SEL_AUTO_INDEX;
- }
-
- OUT_PKT3(ring, CP_DRAW_INDX, info->indexed ? 5 : 3);
- OUT_RING(ring, 0x00000000); /* viz query info. */
- OUT_RING(ring, DRAW(mode2primtype(info->mode),
- src_sel, idx_type, IGNORE_VISIBILITY));
- OUT_RING(ring, info->count); /* NumIndices */
- if (info->indexed) {
- OUT_RELOC(ring, idx_bo, idx_offset, 0);
- OUT_RING (ring, idx_size);
- }
-}
-
static void
fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
{
struct fd_context *ctx = fd_context(pctx);
struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
- unsigned buffers;
+ struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
+ unsigned i, buffers = 0;
/* if we supported transform feedback, we'd have to disable this: */
- if (((ctx->scissor.maxx - ctx->scissor.minx) *
- (ctx->scissor.maxy - ctx->scissor.miny)) == 0) {
+ if (((scissor->maxx - scissor->minx) *
+ (scissor->maxy - scissor->miny)) == 0) {
+ return;
+ }
+
+ /* emulate unsupported primitives: */
+ if (!fd_supported_prim(ctx, info->mode)) {
+ util_primconvert_save_index_buffer(ctx->primconvert, &ctx->indexbuf);
+ util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
+ util_primconvert_draw_vbo(ctx->primconvert, info);
return;
}
ctx->needs_flush = true;
- fd_resource(pfb->cbufs[0]->texture)->dirty = true;
+ /*
+ * Figure out the buffers/features we need:
+ */
- /* figure out the buffers we need: */
- buffers = FD_BUFFER_COLOR;
if (fd_depth_enabled(ctx)) {
buffers |= FD_BUFFER_DEPTH;
fd_resource(pfb->zsbuf->texture)->dirty = true;
+ ctx->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
}
+
if (fd_stencil_enabled(ctx)) {
buffers |= FD_BUFFER_STENCIL;
fd_resource(pfb->zsbuf->texture)->dirty = true;
+ ctx->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
}
- /* any buffers that haven't been cleared, we need to restore: */
+ if (fd_logicop_enabled(ctx))
+ ctx->gmem_reason |= FD_GMEM_LOGICOP_ENABLED;
+
+ for (i = 0; i < pfb->nr_cbufs; i++) {
+ struct pipe_resource *surf;
+
+ if (!pfb->cbufs[i])
+ continue;
+
+ surf = pfb->cbufs[i]->texture;
+
+ fd_resource(surf)->dirty = true;
+ buffers |= PIPE_CLEAR_COLOR0 << i;
+
+ if (surf->nr_samples > 1)
+ ctx->gmem_reason |= FD_GMEM_MSAA_ENABLED;
+
+ if (fd_blend_enabled(ctx, i))
+ ctx->gmem_reason |= FD_GMEM_BLEND_ENABLED;
+ }
+
+ ctx->num_draws++;
+
+ ctx->stats.draw_calls++;
+ ctx->stats.prims_emitted +=
+ u_reduced_prims_for_vertices(info->mode, info->count);
+
+ /* any buffers that haven't been cleared yet, we need to restore: */
ctx->restore |= buffers & (FD_BUFFER_ALL & ~ctx->cleared);
/* and any buffers used, need to be resolved: */
ctx->resolve |= buffers;
- ctx->draw(ctx, info);
+ DBG("%x num_draws=%u (%s/%s)", buffers, ctx->num_draws,
+ util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
+ util_format_short_name(pipe_surface_format(pfb->zsbuf)));
+
+ fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_DRAW);
+ ctx->draw_vbo(ctx, info);
+
+ /* if an app (or, well, piglit test) does many thousands of draws
+ * without flush (or anything which implicitly flushes, like
+ * changing render targets), we can exceed the ringbuffer size.
+ * Since we don't currently have a sane way to wrapparound, and
+ * we use the same buffer for both draw and tiling commands, for
+ * now we need to do this hack and trigger flush if we are running
+ * low on remaining space for cmds:
+ */
+ if (((ctx->ring->cur - ctx->ring->start) >
+ (ctx->ring->size/4 - FD_TILING_COMMANDS_DWORDS)) ||
+ (fd_mesa_debug & FD_DBG_FLUSH))
+ fd_context_render(pctx);
}
/* TODO figure out how to make better use of existing state mechanism
{
struct fd_context *ctx = fd_context(pctx);
struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
-
- ctx->cleared |= buffers;
+ struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
+ unsigned cleared_buffers;
+ int i;
+
+ /* for bookkeeping about which buffers have been cleared (and thus
+ * can fully or partially skip mem2gmem) we need to ignore buffers
+ * that have already had a draw, in case apps do silly things like
+ * clear after draw (ie. if you only clear the color buffer, but
+ * something like alpha-test causes side effects from the draw in
+ * the depth buffer, etc)
+ */
+ cleared_buffers = buffers & (FD_BUFFER_ALL & ~ctx->restore);
+
+ /* do we have full-screen scissor? */
+ if (!memcmp(scissor, &ctx->disabled_scissor, sizeof(*scissor))) {
+ ctx->cleared |= cleared_buffers;
+ } else {
+ ctx->partial_cleared |= cleared_buffers;
+ if (cleared_buffers & PIPE_CLEAR_COLOR)
+ ctx->cleared_scissor.color = *scissor;
+ if (cleared_buffers & PIPE_CLEAR_DEPTH)
+ ctx->cleared_scissor.depth = *scissor;
+ if (cleared_buffers & PIPE_CLEAR_STENCIL)
+ ctx->cleared_scissor.stencil = *scissor;
+ }
ctx->resolve |= buffers;
ctx->needs_flush = true;
if (buffers & PIPE_CLEAR_COLOR)
- fd_resource(pfb->cbufs[0]->texture)->dirty = true;
+ for (i = 0; i < pfb->nr_cbufs; i++)
+ if (buffers & (PIPE_CLEAR_COLOR0 << i))
+ fd_resource(pfb->cbufs[i]->texture)->dirty = true;
- if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL))
+ if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
fd_resource(pfb->zsbuf->texture)->dirty = true;
+ ctx->gmem_reason |= FD_GMEM_CLEARS_DEPTH_STENCIL;
+ }
DBG("%x depth=%f, stencil=%u (%s/%s)", buffers, depth, stencil,
- util_format_name(pfb->cbufs[0]->format),
- pfb->zsbuf ? util_format_name(pfb->zsbuf->format) : "none");
+ util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
+ util_format_short_name(pipe_surface_format(pfb->zsbuf)));
+
+ fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_CLEAR);
ctx->clear(ctx, buffers, color, depth, stencil);
ctx->dirty |= FD_DIRTY_ZSA |
+ FD_DIRTY_VIEWPORT |
FD_DIRTY_RASTERIZER |
FD_DIRTY_SAMPLE_MASK |
FD_DIRTY_PROG |