代码¶
https://github.com/facebookincubator/oculus-linux-kernel
mesa 中freedreno代码¶
@startuml
package ringbuffer <<Rectangle>> {
class msm_ringbuffer
class fd_ringbuffer
}
class msm_ringbuffer {
struct fd_ringbuffer base;
union u
---
fd_pipe * pipe;
fd_bo * reloc_bos;//array
set * ring_set;
---
fd_submit * submit
msm_cmd* cmds;//array
---
msm_cmd *cmd;//cur
fd_bo * ring_bo;
}
class fd_ringbuffer {
struct fd_ringbuffer_funcs * funcs
---
.funcs = ring_funcs
}
fd_ringbuffer <|-- msm_ringbuffer
interface ring_funcs {
.grow = msg_ringbuffer_grow,
.emit_reloc = msg_ringbuffer_emit_reloc,
.emit_reloc_ring = msg_ringbuffer_emit_reloc_ring,
.cmd_count = msg_ringbuffer_cmd_count,
.destroy = msg_ringbuffer_destroy,
}
fd_ringbuffer::ring_funcs .> ring_funcs
package pipe <<Rectangle>> {
class fd_pipe
class msm_pipe
}
class fd_pipe {
fd_pipe_id id;
uint32_t gpu_id;
int32_t refcnt;
struct fd_pipe_funcs * funcs;
---
.funcs = legacy_funcs
fd_device *dev;
}
interface fd_pipe_funcs {
.ringbuffer_new_object = msm_ringbuffer_new_object,
.submit_new = msm_submit_new,
.get_param = msm_pipe_get_param,
.wait = msm_pipe_wait,
.destroy = msm_pipe_destroy,
}
fd_pipe::legacy_funcs .> fd_pipe_funcs
class msm_pipe {
fd_pipe base;
---
uint32_t pipe;
uint32_t gpu_id;
uint64_t gmem_base;
uint32_t gmem;
uint32_t chip_id;
uint32_t queue_id;
struct slab_parent_pool ring_pool;
}
fd_pipe <|-- msm_pipe
class fd_submit_funcs {
*new_ringbuffer
* **flush**
*destroy
}
class fd_submit {
fd_pipe *pipe;
fd_submit_funcs *funcs;
}
msm_ringbuffer::submit .> fd_submit
fd_submit::funcs .> fd_submit_funcs
package device <<Rectangle>> {
class fd_device
class msm_device
}
class fd_device {
int fd
fd_version version
int32_t refcnt
fd_device_funcs * funcs
fd_bo_cache bo_cache
fd_bo_cache ring_cache;
}
class msm_device {
fd_device base;
fd_bo_cache ring_cache;
}
fd_device <|-- msm_device
interface fd_device_funcs {
.bo_new_handle = msm_bo_new_handle,
.bo_from_handle = msm_bo_from_handle,
.pipe_new = msm_pipe_new,
.destroy = msm_device_destroy,
}
fd_pipe::dev ..> fd_device
msm_ringbuffer::pipe .> fd_pipe
fd_device::funcs .> fd_device_funcs
package context <<Rectangle>> {
class pipe_context
class fd_context
}
pipe_context <|-- fd_context
class pipe_context {
pipe_screen * screen
void * priv
void * draw
u_upload_mgr *stream_uploader
u_upload_mgr *const_uploader
-- func pointers --
*destroy
*draw_vbo
*set_vertex_buffers
*set_stream_out_targets
*blit
*clear
*create_surface
*surface_destroy
*buffer_subdata
*texture_subdata
*resource_commit
*launch_grid
a very long list
}
class fd_context {
pipe_context base;
list_head node;
-- uplink --
fd_device *dev;
fd_screen * screen;
fd_pipe * pipe;
-- other contexts --
blitter_context *blitter
primconvert_context * primconvert
-- current batch --
fd_batch * batch
-- function pointers --
emit_tile_init
emit_tile_prep
emit_tile_mem2gmem
emit_tile_rendergrep
emit_tile
emit_title_gmem2mem
emit_tile_fini
emit_sysmem_prep
emit_sysmem_fini
draw_vbo
clear
blit
ts_to_ns
launch_grid
*query*
-- end --
}
fd_context::dev .> fd_device
class cso_context {
pipe_context *pipe;
}
cso_context::pipe -left-> pipe_context
class st_context {
st_context_iface iface;
gl_context *ctx;
pipe_context *pipe;
draw_context *draw;
cso_context *cso_context
}
st_context::pipe -left-> pipe_context
class fd_batch {
unsigned num_bins_per_pipe;
unsigned prim_strm_bits;
unsigned draw_strm_bits;
--draw pass cmdstream--
fd_ringbuffer *draw;
--binning pass cmdstream--
fd_ringbuffer *binning;
--tiling/gmem (IB0) cmdstream--
fd_ringbuffer *gmem;
--epilogue cmdstream--
fd_ringbuffer *epilogue;
// TODO maybe more generically
//split out clear
//and clear_binning rings?
fd_ringbuffer *lrz_clear;
fd_ringbuffer *tile_setup;
fd_ringbuffer *tile_fini;
}
fd_context::batch .left.> fd_batch
fd_batch .left.> fd_ringbuffer
@enduml
1 draw的调用流程¶
@startuml
!theme bluegray
app -> _mesa_: _mesa_DrawElements
_mesa_ -> st: Driver->st_draw_vbo
st -> cso: cso_draw_vbo
cso -> pipe: draw_vbo(fd6_draw_vbo)
pipe -> fd_draw: draw_emit
@enduml
1.1 mesa adreno code¶
1.1.1 类图¶
class pipe_screen {
context_create: fun*
... fd6_context_create
}
class fd_screen
fd_screen -up-|> pipe_screen
class pipe_context {
screen : pipe_screen *
---
flush: func *
draw_vbo: func *
}
class fd_context {
base : pipe_context
screen : fd_screen *
}
class fd6_context {
base : fd_context
}
fd6_context -up-|> fd_context
fd_context -up-|> pipe_context
pipe_context .right.> pipe_screen : ref
fd_context .right.> fd_screen : ref
class st_context_iface {
flush : func *
}
class st_context {
iface: st_context_iface
ctx : gl_context*
---
pipe: pipe_context*
---
draw: draw_context*
}
st_context::pipe .right.> pipe_context
st_context -up-|> st_context_iface
class dd_function_table {
Flush : func *
}
class gl_context {
Driver : dd_function_table
---
st : st_context *
}
gl_context::st .right.> st_context
gl_context::Driver .up.|> dd_function_table
1.1.2 时序图(sequence)¶
整个flush的调用sequence如下图。下图即说明了各个类的调用seqeunce，也说明了flush的最终目标。 flush的最终目标就是真正的tiled rendering，向kmd发送ringbuffer，包括一个binning pass和多个render pass。
app-> context++: _mesa_flush
context -> st_context++: Driver.Flush=\nst_glFlush(gl_context*)
context--
st_context -> st_context: st_flush(*st)
st_context -> pipe_context++: st->pipe->flush(pipe*)
st_context--
pipe_context -> freedreno_context++ : .flush=\nfd_context_flush(pipe)
pipe_context--
freedreno_context -> fd_batch++: fd_batch_flush\n(fd_batch*)
freedreno_context--
freedreno_context--
fd_batch -> fd_batch++: batch_flush(fd_batch*)
fd_batch -> freedreno_gmem++: fd_gmem_render_tiles\n(fd_batch*)
fd_batch--
fd_batch--
freedreno_gmem -> freedreno_gmem++: render_tiles\n(fd_batch*,gmem)
freedreno_gmem -> fd_context++: emit_tile_init
fd_context -> fd_context: fd6_emit_title_init
fd_context -> fd6_gmem: emit_binning_pass
fd_context--
loop foreach tile
freedreno_gmem -> fd_context: emit_tile_prep
freedreno_gmem -> fd_context: emit_tile_mem2gmem(unresolve)
freedreno_gmem -> fd_context: emit_tile_renderprep
freedreno_gmem -> fd_context: query_prepare_tile
alt if has emit_tile
freedreno_gmem -> fd_context: **emit_tile**
else
freedreno_gmem -> fd_context: emit_ib
end alt
freedreno_gmem -> fd_context: emit_tile_gmem2mem
end loop
freedreno_gmem -> fd_context: emit_tile_fini
freedreno_gmem--
freedreno_gmem -> freedreno_gmem: flush_ring(fd_batch*)
freedreno_gmem--
fd_batch是在freedreno_context阶段新创建的临时对象。 fd_batch包含一次提交所需要的各种资源（纹理、buffer之类的）和相应的ringbuffer。 ringbuffer包括 binning, draw, gmem，这三个是最主要的，分别用于binning stage、render stage和resolve stage。另外还有lrz_clear、tile_setup、tile_fini等非主要的ringbuffer。
ringbuffer中draw是必须的。binning可能没有，即使有，它也是从draw创建出来的，因为它所需的所有命令和资源draw中都有，所以binning这个ringbuffer自己的内容很少，它会把draw接进来作为一个indirect buffer(IB)，再结合自己独有的一些状态设置，就可以提交给显卡做binning了。
1.1.3 ringbuffer及指令录制¶
ringbuffer本身真的只是一个ringbuffer:
class fd_ringbuffer {
uint32_t *cur, *end, *start;
-- buffer的指针--
fd_ringbuffer_funcs funcs
-- buffer的操作--
}
上一节讲到fd_batch及其draw, binning, gmem三个ringbuffer是在flush的时候创建的。其实说的不太对。它是在调用drawcall之前就要创建的。兴许这个flush在eglMakeCurrent的时候也会调用，所以可以保证draw call的时候有batch及ringbuffer存在。
ringbuffer录制的大部分工作发生在drawcall里，主要是fd6_draw_vbo函数。fd6_draw_vbo会创建一个emit结构体，这个结构体基本是由ctx中绑定的各种当前对象（当前program及其vs,fs,gs，顶点缓冲区等）初始化的。然后fd6_draw_vbo调用fd6_emit_state函数，把这些状态写入到ringbuffer里。 所有fd6_emit_开头的函数都是往ringbuffer写入数据用的，它的第一个参数一般是一个ringbuffer。fd6_emit_state里各种DIRTY_XXX判断，是为了减少不必要的状态切换，如果某些状态没有变化，就不通过指令去更新它。
static bool
fd6_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info,
unsigned index_offset)
{
struct fd6_context *fd6_ctx = fd6_context(ctx);
struct ir3_shader *gs = ctx->prog.gs;
struct fd6_emit emit = {
.ctx = ctx,
.vtx = &ctx->vtx,
.info = info,
.key = {
.vs = ctx->prog.vs,
.gs = ctx->prog.gs,
.fs = ctx->prog.fs,
.key = {
.color_two_side = ctx->rasterizer->light_twoside,
.vclamp_color = ctx->rasterizer->clamp_vertex_color,
.fclamp_color = ctx->rasterizer->clamp_fragment_color,
.rasterflat = ctx->rasterizer->flatshade,
.ucp_enables = ctx->rasterizer->clip_plane_enable,
.has_per_samp = (fd6_ctx->fsaturate || fd6_ctx->vsaturate),
.vsaturate_s = fd6_ctx->vsaturate_s,
.vsaturate_t = fd6_ctx->vsaturate_t,
.vsaturate_r = fd6_ctx->vsaturate_r,
.fsaturate_s = fd6_ctx->fsaturate_s,
.fsaturate_t = fd6_ctx->fsaturate_t,
.fsaturate_r = fd6_ctx->fsaturate_r,
.layer_zero = !gs || !(gs->nir->info.outputs_written & VARYING_BIT_LAYER),
.vsamples = ctx->tex[PIPE_SHADER_VERTEX].samples,
.fsamples = ctx->tex[PIPE_SHADER_FRAGMENT].samples,
.sample_shading = (ctx->min_samples > 1),
.msaa = (ctx->framebuffer.samples > 1),
},
},
.rasterflat = ctx->rasterizer->flatshade,
.sprite_coord_enable = ctx->rasterizer->sprite_coord_enable,
.sprite_coord_mode = ctx->rasterizer->sprite_coord_mode,
.primitive_restart = info->primitive_restart && info->index_size,
};
if (!(ctx->prog.vs && ctx->prog.fs))
return false;
if (info->mode == PIPE_PRIM_PATCHES) {
emit.key.hs = ctx->prog.hs;
emit.key.ds = ctx->prog.ds;
if (!(ctx->prog.hs && ctx->prog.ds))
return false;
shader_info *ds_info = &emit.key.ds->nir->info;
emit.key.key.tessellation = ir3_tess_mode(ds_info->tess.primitive_mode);
}
if (emit.key.gs)
emit.key.key.has_gs = true;
if (!(emit.key.hs || emit.key.ds || emit.key.gs || info->indirect))
fd6_vsc_update_sizes(ctx->batch, info);
fixup_shader_state(ctx, &emit.key.key);
if (!(ctx->dirty & FD_DIRTY_PROG)) {
emit.prog = fd6_ctx->prog;
} else {
fd6_ctx->prog = fd6_emit_get_prog(&emit);
}
/* bail if compile failed: */
if (!fd6_ctx->prog)
return NULL;
fixup_draw_state(ctx, &emit);
emit.dirty = ctx->dirty; /* *after* fixup_shader_state() */
emit.bs = fd6_emit_get_prog(&emit)->bs;
emit.vs = fd6_emit_get_prog(&emit)->vs;
emit.hs = fd6_emit_get_prog(&emit)->hs;
emit.ds = fd6_emit_get_prog(&emit)->ds;
emit.gs = fd6_emit_get_prog(&emit)->gs;
emit.fs = fd6_emit_get_prog(&emit)->fs;
ctx->stats.vs_regs += ir3_shader_halfregs(emit.vs);
ctx->stats.hs_regs += COND(emit.hs, ir3_shader_halfregs(emit.hs));
ctx->stats.ds_regs += COND(emit.ds, ir3_shader_halfregs(emit.ds));
ctx->stats.gs_regs += COND(emit.gs, ir3_shader_halfregs(emit.gs));
ctx->stats.fs_regs += ir3_shader_halfregs(emit.fs);
struct fd_ringbuffer *ring = ctx->batch->draw;
struct CP_DRAW_INDX_OFFSET_0 draw0 = {
.prim_type = ctx->primtypes[info->mode],
.vis_cull = USE_VISIBILITY,
.gs_enable = !!emit.key.gs,
};
if (info->index_size) {
draw0.source_select = DI_SRC_SEL_DMA;
draw0.index_size = fd4_size2indextype(info->index_size);
} else {
draw0.source_select = DI_SRC_SEL_AUTO_INDEX;
}
if (info->mode == PIPE_PRIM_PATCHES) {
shader_info *ds_info = &emit.ds->shader->nir->info;
uint32_t factor_stride;
switch (ds_info->tess.primitive_mode) {
case GL_ISOLINES:
draw0.patch_type = TESS_ISOLINES;
factor_stride = 12;
break;
case GL_TRIANGLES:
draw0.patch_type = TESS_TRIANGLES;
factor_stride = 20;
break;
case GL_QUADS:
draw0.patch_type = TESS_QUADS;
factor_stride = 28;
break;
default:
unreachable("bad tessmode");
}
draw0.prim_type = DI_PT_PATCHES0 + info->vertices_per_patch;
draw0.tess_enable = true;
ctx->batch->tessellation = true;
ctx->batch->tessparam_size = MAX2(ctx->batch->tessparam_size,
emit.hs->output_size * 4 * info->count);
ctx->batch->tessfactor_size = MAX2(ctx->batch->tessfactor_size,
factor_stride * info->count);
if (!ctx->batch->tess_addrs_constobj) {
/* Reserve space for the bo address - we'll write them later in
* setup_tess_buffers(). We need 2 bo address, but indirect
* constant upload needs at least 4 vec4s.
*/
unsigned size = 4 * 16;
ctx->batch->tess_addrs_constobj = fd_submit_new_ringbuffer(
ctx->batch->submit, size, FD_RINGBUFFER_STREAMING);
ctx->batch->tess_addrs_constobj->cur += size;
}
}
uint32_t index_start = info->index_size ? info->index_bias : info->start;
if (ctx->last.dirty || (ctx->last.index_start != index_start)) {
OUT_PKT4(ring, REG_A6XX_VFD_INDEX_OFFSET, 1);
OUT_RING(ring, index_start); /* VFD_INDEX_OFFSET */
ctx->last.index_start = index_start;
}
if (ctx->last.dirty || (ctx->last.instance_start != info->start_instance)) {
OUT_PKT4(ring, REG_A6XX_VFD_INSTANCE_START_OFFSET, 1);
OUT_RING(ring, info->start_instance); /* VFD_INSTANCE_START_OFFSET */
ctx->last.instance_start = info->start_instance;
}
uint32_t restart_index = info->primitive_restart ? info->restart_index : 0xffffffff;
if (ctx->last.dirty || (ctx->last.restart_index != restart_index)) {
OUT_PKT4(ring, REG_A6XX_PC_RESTART_INDEX, 1);
OUT_RING(ring, restart_index); /* PC_RESTART_INDEX */
ctx->last.restart_index = restart_index;
}
fd6_emit_state(ring, &emit);
/* for debug after a lock up, write a unique counter value
* to scratch7 for each draw, to make it easier to match up
* register dumps to cmdstream. The combination of IB
* (scratch6) and DRAW is enough to "triangulate" the
* particular draw that caused lockup.
*/
emit_marker6(ring, 7);
if (info->indirect) {
draw_emit_indirect(ring, &draw0, info, index_offset);
} else {
draw_emit(ring, &draw0, info, index_offset);
}
emit_marker6(ring, 7);
fd_reset_wfi(ctx->batch);
if (emit.streamout_mask) {
struct fd_ringbuffer *ring = ctx->batch->draw;
for (unsigned i = 0; i < PIPE_MAX_SO_BUFFERS; i++) {
if (emit.streamout_mask & (1 << i)) {
fd6_event_write(ctx->batch, ring, FLUSH_SO_0 + i, false);
}
}
}
fd_context_all_clean(ctx);
return true;
}
另一个函数fd6_emit_state:
void
fd6_emit_state(struct fd_ringbuffer *ring, struct fd6_emit *emit)
{
struct fd_context *ctx = emit->ctx;
struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
const struct fd6_program_state *prog = fd6_emit_get_prog(emit);
const struct ir3_shader_variant *vs = emit->vs;
const struct ir3_shader_variant *hs = emit->hs;
const struct ir3_shader_variant *ds = emit->ds;
const struct ir3_shader_variant *gs = emit->gs;
const struct ir3_shader_variant *fs = emit->fs;
const enum fd_dirty_3d_state dirty = emit->dirty;
bool needs_border = false;
emit_marker6(ring, 5);
/* NOTE: we track fb_read differently than _BLEND_ENABLED since
* we might at some point decide to do sysmem in some cases when
* blend is enabled:
*/
if (fs->fb_read)
ctx->batch->gmem_reason |= FD_GMEM_FB_READ;
if (emit->dirty & FD_DIRTY_VTXSTATE) {
struct fd6_vertex_stateobj *vtx = fd6_vertex_stateobj(ctx->vtx.vtx);
fd6_emit_add_group(emit, vtx->stateobj, FD6_GROUP_VTXSTATE, ENABLE_ALL);
}
if (emit->dirty & FD_DIRTY_VTXBUF) {
struct fd_ringbuffer *state;
state = build_vbo_state(emit);
fd6_emit_take_group(emit, state, FD6_GROUP_VBO, ENABLE_ALL);
}
if (dirty & FD_DIRTY_ZSA) {
struct fd6_zsa_stateobj *zsa = fd6_zsa_stateobj(ctx->zsa);
if (util_format_is_pure_integer(pipe_surface_format(pfb->cbufs[0])))
fd6_emit_add_group(emit, zsa->stateobj_no_alpha, FD6_GROUP_ZSA, ENABLE_ALL);
else
fd6_emit_add_group(emit, zsa->stateobj, FD6_GROUP_ZSA, ENABLE_ALL);
}
if (dirty & (FD_DIRTY_ZSA | FD_DIRTY_BLEND | FD_DIRTY_PROG)) {
struct fd_ringbuffer *state;
state = build_lrz(emit, false);
if (state) {
fd6_emit_take_group(emit, state, FD6_GROUP_LRZ, ENABLE_DRAW);
}
state = build_lrz(emit, true);
if (state) {
fd6_emit_take_group(emit, state,
FD6_GROUP_LRZ_BINNING, CP_SET_DRAW_STATE__0_BINNING);
}
}
if (dirty & FD_DIRTY_STENCIL_REF) {
struct pipe_stencil_ref *sr = &ctx->stencil_ref;
OUT_PKT4(ring, REG_A6XX_RB_STENCILREF, 1);
OUT_RING(ring, A6XX_RB_STENCILREF_REF(sr->ref_value[0]) |
A6XX_RB_STENCILREF_BFREF(sr->ref_value[1]));
}
/* NOTE: scissor enabled bit is part of rasterizer state, but
* fd_rasterizer_state_bind() will mark scissor dirty if needed:
*/
if (dirty & FD_DIRTY_SCISSOR) {
struct fd_ringbuffer *ring = fd_submit_new_ringbuffer(
emit->ctx->batch->submit, 3*4, FD_RINGBUFFER_STREAMING);
struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
OUT_REG(ring,
A6XX_GRAS_SC_SCREEN_SCISSOR_TL(0,
.x = scissor->minx,
.y = scissor->miny
),
A6XX_GRAS_SC_SCREEN_SCISSOR_BR(0,
.x = MAX2(scissor->maxx, 1) - 1,
.y = MAX2(scissor->maxy, 1) - 1
)
);
fd6_emit_take_group(emit, ring, FD6_GROUP_SCISSOR, ENABLE_ALL);
ctx->batch->max_scissor.minx = MIN2(ctx->batch->max_scissor.minx, scissor->minx);
ctx->batch->max_scissor.miny = MIN2(ctx->batch->max_scissor.miny, scissor->miny);
ctx->batch->max_scissor.maxx = MAX2(ctx->batch->max_scissor.maxx, scissor->maxx);
ctx->batch->max_scissor.maxy = MAX2(ctx->batch->max_scissor.maxy, scissor->maxy);
}
if (dirty & FD_DIRTY_VIEWPORT) {
struct pipe_scissor_state *scissor = &ctx->viewport_scissor;
OUT_REG(ring,
A6XX_GRAS_CL_VPORT_XOFFSET(0, ctx->viewport.translate[0]),
A6XX_GRAS_CL_VPORT_XSCALE(0, ctx->viewport.scale[0]),
A6XX_GRAS_CL_VPORT_YOFFSET(0, ctx->viewport.translate[1]),
A6XX_GRAS_CL_VPORT_YSCALE(0, ctx->viewport.scale[1]),
A6XX_GRAS_CL_VPORT_ZOFFSET(0, ctx->viewport.translate[2]),
A6XX_GRAS_CL_VPORT_ZSCALE(0, ctx->viewport.scale[2])
);
OUT_REG(ring,
A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL(0,
.x = scissor->minx,
.y = scissor->miny
),
A6XX_GRAS_SC_VIEWPORT_SCISSOR_BR(0,
.x = MAX2(scissor->maxx, 1) - 1,
.y = MAX2(scissor->maxy, 1) - 1
)
);
unsigned guardband_x =
fd_calc_guardband(ctx->viewport.translate[0], ctx->viewport.scale[0],
false);
unsigned guardband_y =
fd_calc_guardband(ctx->viewport.translate[1], ctx->viewport.scale[1],
false);
OUT_REG(ring, A6XX_GRAS_CL_GUARDBAND_CLIP_ADJ(
.horz = guardband_x,
.vert = guardband_y
)
);
}
if (dirty & FD_DIRTY_PROG) {
fd6_emit_add_group(emit, prog->config_stateobj, FD6_GROUP_PROG_CONFIG, ENABLE_ALL);
fd6_emit_add_group(emit, prog->stateobj, FD6_GROUP_PROG, ENABLE_DRAW);
fd6_emit_add_group(emit, prog->binning_stateobj,
FD6_GROUP_PROG_BINNING, CP_SET_DRAW_STATE__0_BINNING);
/* emit remaining streaming program state, ie. what depends on
* other emit state, so cannot be pre-baked.
*/
struct fd_ringbuffer *streaming = fd6_program_interp_state(emit);
fd6_emit_take_group(emit, streaming, FD6_GROUP_PROG_INTERP, ENABLE_DRAW);
}
if (dirty & FD_DIRTY_RASTERIZER) {
struct fd_ringbuffer *stateobj =
fd6_rasterizer_state(ctx, emit->primitive_restart);
fd6_emit_add_group(emit, stateobj,
FD6_GROUP_RASTERIZER, ENABLE_ALL);
}
if (dirty & (FD_DIRTY_FRAMEBUFFER | FD_DIRTY_RASTERIZER_DISCARD | FD_DIRTY_PROG)) {
struct fd_ringbuffer *ring = fd_submit_new_ringbuffer(
emit->ctx->batch->submit, 5 * 4, FD_RINGBUFFER_STREAMING);
unsigned nr = pfb->nr_cbufs;
if (ctx->rasterizer->rasterizer_discard)
nr = 0;
OUT_PKT4(ring, REG_A6XX_RB_FS_OUTPUT_CNTL0, 2);
OUT_RING(ring, COND(fs->writes_pos, A6XX_RB_FS_OUTPUT_CNTL0_FRAG_WRITES_Z) |
COND(fs->writes_smask && pfb->samples > 1,
A6XX_RB_FS_OUTPUT_CNTL0_FRAG_WRITES_SAMPMASK));
OUT_RING(ring, A6XX_RB_FS_OUTPUT_CNTL1_MRT(nr));
OUT_PKT4(ring, REG_A6XX_SP_FS_OUTPUT_CNTL1, 1);
OUT_RING(ring, A6XX_SP_FS_OUTPUT_CNTL1_MRT(nr));
fd6_emit_take_group(emit, ring, FD6_GROUP_PROG_FB_RAST, ENABLE_DRAW);
}
fd6_emit_consts(emit);
struct ir3_stream_output_info *info = &fd6_last_shader(prog)->shader->stream_output;
if (info->num_outputs)
fd6_emit_streamout(ring, emit, info);
if (dirty & (FD_DIRTY_BLEND | FD_DIRTY_SAMPLE_MASK)) {
struct fd6_blend_variant *blend = fd6_blend_variant(ctx->blend,
pfb->samples, ctx->sample_mask);
fd6_emit_add_group(emit, blend->stateobj, FD6_GROUP_BLEND, ENABLE_DRAW);
}
if (dirty & FD_DIRTY_BLEND_COLOR) {
struct pipe_blend_color *bcolor = &ctx->blend_color;
struct fd_ringbuffer *ring = fd_submit_new_ringbuffer(
emit->ctx->batch->submit, 5*4, FD_RINGBUFFER_STREAMING);
OUT_REG(ring,
A6XX_RB_BLEND_RED_F32(bcolor->color[0]),
A6XX_RB_BLEND_GREEN_F32(bcolor->color[1]),
A6XX_RB_BLEND_BLUE_F32(bcolor->color[2]),
A6XX_RB_BLEND_ALPHA_F32(bcolor->color[3])
);
fd6_emit_take_group(emit, ring, FD6_GROUP_BLEND_COLOR, ENABLE_DRAW);
}
needs_border |= fd6_emit_combined_textures(ring, emit, PIPE_SHADER_VERTEX, vs);
if (hs) {
needs_border |= fd6_emit_combined_textures(ring, emit, PIPE_SHADER_TESS_CTRL, hs);
needs_border |= fd6_emit_combined_textures(ring, emit, PIPE_SHADER_TESS_EVAL, ds);
}
if (gs) {
needs_border |= fd6_emit_combined_textures(ring, emit, PIPE_SHADER_GEOMETRY, gs);
}
needs_border |= fd6_emit_combined_textures(ring, emit, PIPE_SHADER_FRAGMENT, fs);
if (needs_border)
emit_border_color(ctx, ring);
if (hs) {
debug_assert(ir3_shader_nibo(hs) == 0);
debug_assert(ir3_shader_nibo(ds) == 0);
}
if (gs) {
debug_assert(ir3_shader_nibo(gs) == 0);
}
#define DIRTY_IBO (FD_DIRTY_SHADER_SSBO | FD_DIRTY_SHADER_IMAGE | \
FD_DIRTY_SHADER_PROG)
if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & DIRTY_IBO) {
struct fd_ringbuffer *state =
fd6_build_ibo_state(ctx, fs, PIPE_SHADER_FRAGMENT);
struct fd_ringbuffer *obj = fd_submit_new_ringbuffer(
ctx->batch->submit, 0x100, FD_RINGBUFFER_STREAMING);
OUT_PKT7(obj, CP_LOAD_STATE6, 3);
OUT_RING(obj, CP_LOAD_STATE6_0_DST_OFF(0) |
CP_LOAD_STATE6_0_STATE_TYPE(ST6_SHADER) |
CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
CP_LOAD_STATE6_0_STATE_BLOCK(SB6_IBO) |
CP_LOAD_STATE6_0_NUM_UNIT(ir3_shader_nibo(fs)));
OUT_RB(obj, state);
OUT_PKT4(obj, REG_A6XX_SP_IBO_LO, 2);
OUT_RB(obj, state);
/* TODO if we used CP_SET_DRAW_STATE for compute shaders, we could
* de-duplicate this from program->config_stateobj
*/
OUT_PKT4(obj, REG_A6XX_SP_IBO_COUNT, 1);
OUT_RING(obj, ir3_shader_nibo(fs));
fd6_emit_ibo_consts(emit, fs, PIPE_SHADER_FRAGMENT, ring);
fd6_emit_take_group(emit, obj, FD6_GROUP_IBO, ENABLE_DRAW);
fd_ringbuffer_del(state);
}
if (emit->num_groups > 0) {
OUT_PKT7(ring, CP_SET_DRAW_STATE, 3 * emit->num_groups);
for (unsigned i = 0; i < emit->num_groups; i++) {
struct fd6_state_group *g = &emit->groups[i];
unsigned n = g->stateobj ?
fd_ringbuffer_size(g->stateobj) / 4 : 0;
debug_assert((g->enable_mask & ~ENABLE_ALL) == 0);
if (n == 0) {
OUT_RING(ring, CP_SET_DRAW_STATE__0_COUNT(0) |
CP_SET_DRAW_STATE__0_DISABLE |
g->enable_mask |
CP_SET_DRAW_STATE__0_GROUP_ID(g->group_id));
OUT_RING(ring, 0x00000000);
OUT_RING(ring, 0x00000000);
} else {
OUT_RING(ring, CP_SET_DRAW_STATE__0_COUNT(n) |
g->enable_mask |
CP_SET_DRAW_STATE__0_GROUP_ID(g->group_id));
OUT_RB(ring, g->stateobj);
}
if (g->stateobj)
fd_ringbuffer_del(g->stateobj);
}
emit->num_groups = 0;
}
}
kernel code¶
2 UMD、KMD、显卡数据交互流程¶
@startuml
actor platform
participant adreno_device
participant dispatcher
participant ringbuffer
participant trace
participant profile
== UMD 向 KMD 发送命令 ==
platform -> dispatcher++ : adreno_dispatcher_queue_cmds(drawobjs[])
loop foreach drawobj
dispatcher -> dispatcher++ : _queue_xxxobj
dispatcher -> dispatcher++ : _queue_drawobj(drawobj)
dispatcher -> dispatcher : ctxt->drawqueue[tail] = drawobj
dispatcher -> trace : trace_adreno_cmdbatch_queued()
dispatcher--
dispatcher--
dispatcher--
end loop
== [kthread_worker] KMD 与显卡通信 ==
loop kthread_worker
platform -> dispatcher++ : adreno_dispatcher_work()
== retire ==
loop foreach rb : device.ringbuffers
dispatcher -> dispatcher++ : adreno_dispatch_process_drawqueue(rb->dispatch_q)
dispatcher -> dispatcher++ : adreno_dispatch_retire_drawqueue(rb->dispatch_q)
loop foreach drawobj : dispatch_q->cmd_q
dispatcher -> dispatcher++ : retire_obj(drawobj)
dispatcher -> profile : cmdobj_profile_ticks()
dispatcher -> trace : trace_adreno_cmdbatch_retired()
dispatcher--
end loop
dispatcher--
== submit ==
dispatcher -> dispatcher++ : _adreno_dispatcher_issue_cmds(dev)
loop foreach context
dispatcher -> dispatcher++ : dispatcher_context_sendcmds(drawctx)
loop foreach drawobj : ctxt->drawqueue[]
dispatcher -> dispatcher: _retire_markerobj
dispatcher -> dispatcher: _retire_syncobj
dispatcher -> dispatcher++: sendcmd(drawobj as cmdobj)
dispatcher -> ringbuffer++: adreno_ringbuffer_submitcmd
ringbuffer -> profile: kernel started
ringbuffer -> profile: user submitted
loop foreach ib in cmdobj
ringbuffer -> ringbuffer: copy to cmd
end loop
ringbuffer -> ringbuffer++: adreno_drawctxt_switch(rb, drawctxt);
ringbuffer -> profile: kernel retired
ringbuffer -> profile: user retired
ringbuffer -> ringbuffer++: adreno_ringbuffer_addcmds(cmd)
ringbuffer -> profile : adreno_profile_pre_ib_processing //prepend记录perfcounter指令
ringbuffer -> profile : adreno_profile_post_ib_processing //记录perfcounter
ringbuffer -> ringbuffer : adreno_ringbufferr_submit,修改write ptr
ringbuffer--
ringbuffer -> trace: trace_kgsl_issueibcmds(timestamps, numibs)
ringbuffer--
dispatcher -> trace: trace_adreno_cmdbatch_submitted
dispatcher -> dispatcher : dispatch_q->cmd_q[tail] = cmdobj
end loop
dispatcher--
end loop
dispatcher--
end loop
@enduml
3 数据交互中重要的类图¶
UMD提交给KMD context的drawqueue的是kgsl_drawobj对象。
dispatcher从drawqueue中获取drawobj，转换并提交cmd给显卡，然后将drawobj放入了dispatch_q的cmd_q队列里的是kgsl_drawobj_cmd。实际这两个对象是继承关系。
@startuml
class kgsl_drawobj {
struct kgsl_device *device;
struct kgsl_context *context;
uint32_t type;
uint32_t timestamp;
unsigned long flags;
struct kref refcount;
}
class kgsl_drawobj_cmd {
struct kgsl_drawobj base;
unsigned long priv;
unsigned int global_ts;
unsigned long fault_policy;
unsigned long fault_recovery;
struct list_head cmdlist;
struct list_head memlist;
unsigned int marker_timestamp;
struct kgsl_mem_entry *profiling_buf_entry;
uint64_t profiling_buffer_gpuaddr;
unsigned int profile_index;
uint64_t submit_ticks;
}
kgsl_drawobj <|-down- kgsl_drawobj_cmd
class adreno_dispatcher_drawqueue {
struct kgsl_drawobj_cmd *cmd_q[ADRENO_DISPATCH_DRAWQUEUE_SIZE];
unsigned int inflight;
unsigned int head;
unsigned int tail;
int active_context_count;
unsigned long expires;
}
adreno_dispatcher_drawqueue "has many" -> kgsl_drawobj_cmd
class adreno_ringbuffer {
struct adreno_dispatcher_drawqueue dispatch_q
}
adreno_ringbuffer *-down-> adreno_dispatcher_drawqueue : dispatch_q
class adreno_device {
struct kgsl_device dev;
unsigned long priv;
unsigned int *gpmu_cmds;
struct adreno_ringbuffer ringbuffers[KGSL_PRIORITY_MAX_RB_LEVELS];
}
adreno_device *-down-> adreno_ringbuffer : has 4
class adreno_context {
struct kgsl_context base;
struct kgsl_drawobj *drawqueue[ADRENO_CONTEXT_DRAWQUEUE_SIZE];
unsigned int drawqueue_head;
unsigned int drawqueue_tail;
}
adreno_context -left-> kgsl_drawobj : drawqueue
@enduml
4 drawobj是如何retire的¶
device为每个context存了一个devmemstore结构，gpu会把endofpipe timestamp写进去。 如果endofpipe timestamp大于drawobj->timestamp，说明已经处理完了。
struct kgsl_devmemstore {
volatile unsigned int soptimestamp;
unsigned int sbz;
volatile unsigned int eoptimestamp;
unsigned int sbz2;
volatile unsigned int preempted;
unsigned int sbz3;
volatile unsigned int ref_wait_ts;
unsigned int sbz4;
unsigned int current_context;
unsigned int sbz5;
};
向endofpipe timestamp写入值的微码是在adreno_ringbuffer_addcmds函数中追加的：
*ringcmds++ = cp_mem_packet(adreno_dev, CP_EVENT_WRITE, 3, 1);
if (drawctxt || is_internal_cmds(flags))
*ringcmds++ = CACHE_FLUSH_TS | (1 << 31);
else
*ringcmds++ = CACHE_FLUSH_TS;
if (drawctxt && !is_internal_cmds(flags)) {
ringcmds += cp_gpuaddr(adreno_dev, ringcmds,
MEMSTORE_ID_GPU_ADDR(device, context_id, eoptimestamp));
*ringcmds++ = timestamp;
/* Write the end of pipeline timestamp to the ringbuffer too */
*ringcmds++ = cp_mem_packet(adreno_dev, CP_EVENT_WRITE, 3, 1);
*ringcmds++ = CACHE_FLUSH_TS;
ringcmds += cp_gpuaddr(adreno_dev, ringcmds,
MEMSTORE_RB_GPU_ADDR(device, rb, eoptimestamp));
*ringcmds++ = rb->timestamp;
} else {