/**
* Emit context states and 3DPRIMITIVE.
*/
bool
ilo_3d_pipeline_emit_draw(struct ilo_3d_pipeline *p,
const struct ilo_context *ilo,
int *prim_generated, int *prim_emitted)
{
bool success;
if (ilo->dirty & ILO_DIRTY_SO &&
ilo->so.enabled && !ilo->so.append_bitmask) {
/*
* We keep track of the SVBI in the driver, so that we can restore it
* when the HW context is invalidated (by another process). The value
* needs to be reset when stream output is enabled and the targets are
* changed.
*/
p->state.so_num_vertices = 0;
/* on GEN7+, we need SOL_RESET to reset the SO write offsets */
if (p->dev->gen >= ILO_GEN(7))
ilo_cp_set_one_off_flags(p->cp, INTEL_EXEC_GEN7_SOL_RESET);
}
while (true) {
struct ilo_builder_snapshot snapshot;
/* we will rewind if aperture check below fails */
ilo_builder_batch_snapshot(&p->cp->builder, &snapshot);
handle_invalid_batch_bo(p, false);
/* draw! */
p->emit_draw(p, ilo);
if (ilo_builder_validate(&ilo->cp->builder, 0, NULL)) {
success = true;
break;
}
/* rewind */
ilo_builder_batch_restore(&p->cp->builder, &snapshot);
if (ilo_cp_empty(p->cp)) {
success = false;
break;
}
else {
/* flush and try again */
ilo_cp_flush(p->cp, "out of aperture");
}
}
if (success) {
const int num_verts =
u_vertices_per_prim(u_reduced_prim(ilo->draw->mode));
const int max_emit =
(p->state.so_max_vertices - p->state.so_num_vertices) / num_verts;
const int generated =
u_reduced_prims_for_vertices(ilo->draw->mode, ilo->draw->count);
const int emitted = MIN2(generated, max_emit);
p->state.so_num_vertices += emitted * num_verts;
if (prim_generated)
*prim_generated = generated;
if (prim_emitted)
*prim_emitted = emitted;
}
p->invalidate_flags = 0x0;
return success;
}
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;
struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
unsigned i, prims, buffers = 0;
/* if we supported transform feedback, we'd have to disable this: */
if (((scissor->maxx - scissor->minx) *
(scissor->maxy - scissor->miny)) == 0) {
return;
}
/* TODO: push down the region versions into the tiles */
if (!fd_render_condition_check(pctx))
return;
/* emulate unsupported primitives: */
if (!fd_supported_prim(ctx, info->mode)) {
if (ctx->streamout.num_targets > 0)
debug_error("stream-out with emulated prims");
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;
/*
* Figure out the buffers/features we need:
*/
if (fd_depth_enabled(ctx)) {
buffers |= FD_BUFFER_DEPTH;
resource_written(ctx, pfb->zsbuf->texture);
ctx->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
}
if (fd_stencil_enabled(ctx)) {
buffers |= FD_BUFFER_STENCIL;
resource_written(ctx, pfb->zsbuf->texture);
ctx->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
}
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;
resource_written(ctx, surf);
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;
}
/* Skip over buffer 0, that is sent along with the command stream */
for (i = 1; i < PIPE_MAX_CONSTANT_BUFFERS; i++) {
resource_read(ctx, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
resource_read(ctx, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
}
/* Mark VBOs as being read */
for (i = 0; i < ctx->vtx.vertexbuf.count; i++) {
assert(!ctx->vtx.vertexbuf.vb[i].user_buffer);
resource_read(ctx, ctx->vtx.vertexbuf.vb[i].buffer);
}
/* Mark index buffer as being read */
resource_read(ctx, ctx->indexbuf.buffer);
/* Mark textures as being read */
for (i = 0; i < ctx->verttex.num_textures; i++)
if (ctx->verttex.textures[i])
resource_read(ctx, ctx->verttex.textures[i]->texture);
for (i = 0; i < ctx->fragtex.num_textures; i++)
if (ctx->fragtex.textures[i])
resource_read(ctx, ctx->fragtex.textures[i]->texture);
/* Mark streamout buffers as being written.. */
for (i = 0; i < ctx->streamout.num_targets; i++)
if (ctx->streamout.targets[i])
resource_written(ctx, ctx->streamout.targets[i]->buffer);
ctx->num_draws++;
prims = u_reduced_prims_for_vertices(info->mode, info->count);
ctx->stats.draw_calls++;
/* TODO prims_emitted should be clipped when the stream-out buffer is
* not large enough. See max_tf_vtx().. probably need to move that
* into common code. Although a bit more annoying since a2xx doesn't
* use ir3 so no common way to get at the pipe_stream_output_info
* which is needed for this calculation.
*/
if (ctx->streamout.num_targets > 0)
ctx->stats.prims_emitted += prims;
ctx->stats.prims_generated += prims;
/* 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;
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);
for (i = 0; i < ctx->streamout.num_targets; i++)
ctx->streamout.offsets[i] += info->count;
if (fd_mesa_debug & FD_DBG_DDRAW)
ctx->dirty = 0xffffffff;
/* 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);
}
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;
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 (((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;
/*
* Figure out the buffers/features we need:
*/
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;
}
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 |= FD_BUFFER_COLOR;
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, 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);
}
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;
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 (((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;
/*
* Figure out the buffers/features we need:
*/
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;
}
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 |= FD_BUFFER_COLOR;
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;
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);
}
