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_hw_2116_workaround(pctx);
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) {
uint32_t offset = vc4->indexbuf.offset;
uint32_t index_size = vc4->indexbuf.index_size;
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 {
if (vc4->indexbuf.user_buffer) {
prsc = NULL;
u_upload_data(vc4->uploader, 0,
info->count * index_size, 4,
vc4->indexbuf.user_buffer,
&offset, &prsc);
} else {
prsc = vc4->indexbuf.buffer;
}
}
struct vc4_resource *rsc = vc4_resource(prsc);
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(&bcl, info->count);
cl_reloc(vc4, &vc4->bcl, &bcl, rsc->bo, offset);
cl_u32(&bcl, vc4->max_index);
if (vc4->indexbuf.index_size == 4 || vc4->indexbuf.user_buffer)
pipe_resource_reference(&prsc, NULL);
} else {
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;
}
if (vc4->zsa && vc4->zsa->base.stencil[0].enabled)
vc4->resolve |= PIPE_CLEAR_STENCIL;
vc4->resolve |= PIPE_CLEAR_COLOR0;
vc4->shader_rec_count++;
if (vc4_debug & VC4_DEBUG_ALWAYS_FLUSH)
vc4_flush(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
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);
return;
}
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, 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, (1 << num_elements_emit) - 1); /* vs attribute array bitfield */
cl_u8(&vc4->shader_rec, 16 * num_elements_emit); /* vs total attribute size */
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, (1 << num_elements_emit) - 1); /* cs attribute array bitfield */
cl_u8(&vc4->shader_rec, 16 * num_elements_emit); /* cs total attribute size */
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;
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, i * 16); /* VS VPM offset */
cl_u8(&vc4->shader_rec, i * 16); /* CS VPM offset */
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;
index_size = 2;
}
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++;
if (vc4_debug & VC4_DEBUG_ALWAYS_FLUSH)
vc4_flush(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, 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 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);
}
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_predraw_check_textures(pctx, &vc4->verttex);
vc4_predraw_check_textures(pctx, &vc4->fragtex);
vc4_hw_2116_workaround(pctx, info->count);
struct vc4_job *job = vc4_get_job_for_fbo(vc4);
vc4_get_draw_cl_space(job, info->count);
if (vc4->prim_mode != info->mode) {
vc4->prim_mode = info->mode;
vc4->dirty |= VC4_DIRTY_PRIM_MODE;
}
vc4_start_draw(vc4);
if (!vc4_update_compiled_shaders(vc4, info->mode)) {
debug_warn_once("shader compile failed, skipping draw call.\n");
return;
}
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, 0);
}
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(&job->bcl);
if (info->indexed) {
uint32_t offset = vc4->indexbuf.offset;
uint32_t index_size = vc4->indexbuf.index_size;
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 {
if (vc4->indexbuf.user_buffer) {
prsc = NULL;
u_upload_data(vc4->uploader, 0,
info->count * index_size, 4,
vc4->indexbuf.user_buffer,
&offset, &prsc);
} else {
prsc = vc4->indexbuf.buffer;
}
}
struct vc4_resource *rsc = vc4_resource(prsc);
cl_start_reloc(&job->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(&bcl, info->count);
cl_reloc(job, &job->bcl, &bcl, rsc->bo, offset);
cl_u32(&bcl, vc4->max_index);
job->draw_calls_queued++;
if (vc4->indexbuf.index_size == 4 || vc4->indexbuf.user_buffer)
pipe_resource_reference(&prsc, NULL);
} else {
uint32_t count = info->count;
uint32_t start = info->start;
uint32_t extra_index_bias = 0;
while (count) {
uint32_t this_count = count;
uint32_t step = count;
static const uint32_t max_verts = 65535;
/* GFXH-515 / SW-5891: The binner emits 16 bit indices
* for drawarrays, which means that if start + count >
* 64k it would truncate the top bits. Work around
* this by emitting a limited number of primitives at
* a time and reemitting the shader state pointing
* farther down the vertex attribute arrays.
*
* To do this properly for line loops or trifans, we'd
* need to make a new VB containing the first vertex
* plus whatever remainder.
*/
if (extra_index_bias) {
cl_end(&job->bcl, bcl);
vc4_emit_gl_shader_state(vc4, info,
extra_index_bias);
bcl = cl_start(&job->bcl);
}
if (start + count > max_verts) {
switch (info->mode) {
case PIPE_PRIM_POINTS:
this_count = step = max_verts;
break;
case PIPE_PRIM_LINES:
this_count = step = max_verts - (max_verts % 2);
break;
case PIPE_PRIM_LINE_STRIP:
this_count = max_verts;
step = max_verts - 1;
break;
case PIPE_PRIM_LINE_LOOP:
this_count = max_verts;
step = max_verts - 1;
debug_warn_once("unhandled line loop "
"looping behavior with "
">65535 verts\n");
break;
case PIPE_PRIM_TRIANGLES:
this_count = step = max_verts - (max_verts % 3);
break;
case PIPE_PRIM_TRIANGLE_STRIP:
this_count = max_verts;
step = max_verts - 2;
break;
default:
debug_warn_once("unhandled primitive "
"max vert count, truncating\n");
this_count = step = max_verts;
}
}
cl_u8(&bcl, VC4_PACKET_GL_ARRAY_PRIMITIVE);
cl_u8(&bcl, info->mode);
cl_u32(&bcl, this_count);
cl_u32(&bcl, start);
job->draw_calls_queued++;
count -= step;
extra_index_bias += start + step;
start = 0;
}
}
cl_end(&job->bcl, bcl);
/* We shouldn't have tripped the HW_2116 bug with the GFXH-515
* workaround.
*/
assert(job->draw_calls_queued <= VC4_HW_2116_COUNT);
if (vc4->zsa && vc4->framebuffer.zsbuf) {
struct vc4_resource *rsc =
vc4_resource(vc4->framebuffer.zsbuf->texture);
if (vc4->zsa->base.depth.enabled) {
job->resolve |= PIPE_CLEAR_DEPTH;
rsc->initialized_buffers = PIPE_CLEAR_DEPTH;
}
if (vc4->zsa->base.stencil[0].enabled) {
job->resolve |= PIPE_CLEAR_STENCIL;
rsc->initialized_buffers |= PIPE_CLEAR_STENCIL;
}
}
job->resolve |= PIPE_CLEAR_COLOR0;
if (vc4_debug & VC4_DEBUG_ALWAYS_FLUSH)
vc4_flush(pctx);
}
