static void
draw_impl(struct fd_context *ctx, struct fd_ringbuffer *ring,
struct fd4_emit *emit)
{
const struct pipe_draw_info *info = emit->info;
enum pc_di_primtype primtype = ctx->primtypes[info->mode];
fd4_emit_state(ctx, ring, emit);
if (emit->dirty & (FD_DIRTY_VTXBUF | FD_DIRTY_VTXSTATE))
fd4_emit_vertex_bufs(ring, emit);
OUT_PKT0(ring, REG_A4XX_VFD_INDEX_OFFSET, 2);
OUT_RING(ring, info->indexed ? info->index_bias : info->start); /* VFD_INDEX_OFFSET */
OUT_RING(ring, info->start_instance); /* ??? UNKNOWN_2209 */
OUT_PKT0(ring, REG_A4XX_PC_RESTART_INDEX, 1);
OUT_RING(ring, info->primitive_restart ? /* PC_RESTART_INDEX */
info->restart_index : 0xffffffff);
/* points + psize -> spritelist: */
if (ctx->rasterizer->point_size_per_vertex &&
fd4_emit_get_vp(emit)->writes_psize &&
(info->mode == PIPE_PRIM_POINTS))
primtype = DI_PT_POINTLIST_PSIZE;
fd4_draw_emit(ctx->batch, ring, primtype,
emit->key.binning_pass ? IGNORE_VISIBILITY : USE_VISIBILITY,
info);
}
static void
setup_stages(struct fd4_emit *emit, struct stage *s)
{
unsigned i;
s[VS].v = fd4_emit_get_vp(emit);
s[FS].v = fd4_emit_get_fp(emit);
s[HS].v = s[DS].v = s[GS].v = NULL; /* for now */
for (i = 0; i < MAX_STAGES; i++) {
if (s[i].v) {
s[i].i = &s[i].v->info;
/* constlen is in units of 4 * vec4: */
s[i].constlen = align(s[i].v->constlen, 4) / 4;
/* instrlen is already in units of 16 instr.. although
* probably we should ditch that and not make the compiler
* care about instruction group size of a3xx vs a4xx
*/
s[i].instrlen = s[i].v->instrlen;
} else {
s[i].i = NULL;
s[i].constlen = 0;
s[i].instrlen = 0;
}
}
/* NOTE: at least for gles2, blob partitions VS at bottom of const
* space and FS taking entire remaining space. We probably don't
* need to do that the same way, but for now mimic what the blob
* does to make it easier to diff against register values from blob
*
* NOTE: if VS.instrlen + FS.instrlen > 64, then one or both shaders
* is run from external memory.
*/
if ((s[VS].instrlen + s[FS].instrlen) > 64) {
/* prioritize FS for internal memory: */
if (s[FS].instrlen < 64) {
/* if FS can fit, kick VS out to external memory: */
s[VS].instrlen = 0;
} else if (s[VS].instrlen < 64) {
/* otherwise if VS can fit, kick out FS: */
s[FS].instrlen = 0;
} else {
/* neither can fit, run both from external memory: */
s[VS].instrlen = 0;
s[FS].instrlen = 0;
}
}
s[VS].constlen = 66;
s[FS].constlen = 128 - s[VS].constlen;
s[VS].instroff = 0;
s[VS].constoff = 0;
s[FS].instroff = 64 - s[FS].instrlen;
s[FS].constoff = s[VS].constlen;
s[HS].instroff = s[DS].instroff = s[GS].instroff = s[FS].instroff;
s[HS].constoff = s[DS].constoff = s[GS].constoff = s[FS].constoff;
}
void
fd4_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
struct fd4_emit *emit)
{
struct ir3_shader_variant *vp = fd4_emit_get_vp(emit);
struct ir3_shader_variant *fp = fd4_emit_get_fp(emit);
uint32_t dirty = emit->dirty;
emit_marker(ring, 5);
if ((dirty & (FD_DIRTY_ZSA | FD_DIRTY_PROG)) && !emit->key.binning_pass) {
uint32_t val = fd4_zsa_stateobj(ctx->zsa)->rb_render_control;
/* I suppose if we needed to (which I don't *think* we need
* to), we could emit this for binning pass too. But we
* would need to keep a different patch-list for binning
* vs render pass.
*/
OUT_PKT0(ring, REG_A4XX_RB_RENDER_CONTROL, 1);
OUT_RINGP(ring, val, &fd4_context(ctx)->rbrc_patches);
}
if (dirty & FD_DIRTY_ZSA) {
struct fd4_zsa_stateobj *zsa = fd4_zsa_stateobj(ctx->zsa);
OUT_PKT0(ring, REG_A4XX_RB_ALPHA_CONTROL, 1);
OUT_RING(ring, zsa->rb_alpha_control);
OUT_PKT0(ring, REG_A4XX_RB_STENCIL_CONTROL, 2);
OUT_RING(ring, zsa->rb_stencil_control);
OUT_RING(ring, zsa->rb_stencil_control2);
}
if (dirty & (FD_DIRTY_ZSA | FD_DIRTY_STENCIL_REF)) {
struct fd4_zsa_stateobj *zsa = fd4_zsa_stateobj(ctx->zsa);
struct pipe_stencil_ref *sr = &ctx->stencil_ref;
OUT_PKT0(ring, REG_A4XX_RB_STENCILREFMASK, 2);
OUT_RING(ring, zsa->rb_stencilrefmask |
A4XX_RB_STENCILREFMASK_STENCILREF(sr->ref_value[0]));
OUT_RING(ring, zsa->rb_stencilrefmask_bf |
A4XX_RB_STENCILREFMASK_BF_STENCILREF(sr->ref_value[1]));
}
if (dirty & (FD_DIRTY_ZSA | FD_DIRTY_PROG)) {
struct fd4_zsa_stateobj *zsa = fd4_zsa_stateobj(ctx->zsa);
bool fragz = fp->has_kill | fp->writes_pos;
OUT_PKT0(ring, REG_A4XX_RB_DEPTH_CONTROL, 1);
OUT_RING(ring, zsa->rb_depth_control |
COND(fragz, A4XX_RB_DEPTH_CONTROL_EARLY_Z_DISABLE));
/* maybe this register/bitfield needs a better name.. this
* appears to be just disabling early-z
*/
OUT_PKT0(ring, REG_A4XX_GRAS_ALPHA_CONTROL, 1);
OUT_RING(ring, zsa->gras_alpha_control |
COND(fragz, A4XX_GRAS_ALPHA_CONTROL_ALPHA_TEST_ENABLE));
}
if (dirty & FD_DIRTY_RASTERIZER) {
struct fd4_rasterizer_stateobj *rasterizer =
fd4_rasterizer_stateobj(ctx->rasterizer);
OUT_PKT0(ring, REG_A4XX_GRAS_SU_MODE_CONTROL, 1);
OUT_RING(ring, rasterizer->gras_su_mode_control |
A4XX_GRAS_SU_MODE_CONTROL_RENDERING_PASS);
OUT_PKT0(ring, REG_A4XX_GRAS_SU_POINT_MINMAX, 2);
OUT_RING(ring, rasterizer->gras_su_point_minmax);
OUT_RING(ring, rasterizer->gras_su_point_size);
OUT_PKT0(ring, REG_A4XX_GRAS_SU_POLY_OFFSET_SCALE, 2);
OUT_RING(ring, rasterizer->gras_su_poly_offset_scale);
OUT_RING(ring, rasterizer->gras_su_poly_offset_offset);
OUT_PKT0(ring, REG_A4XX_GRAS_CL_CLIP_CNTL, 1);
OUT_RING(ring, rasterizer->gras_cl_clip_cntl);
}
/* NOTE: since primitive_restart is not actually part of any
* state object, we need to make sure that we always emit
* PRIM_VTX_CNTL.. either that or be more clever and detect
* when it changes.
*/
if (emit->info) {
const struct pipe_draw_info *info = emit->info;
uint32_t val = fd4_rasterizer_stateobj(ctx->rasterizer)
->pc_prim_vtx_cntl;
if (info->indexed && info->primitive_restart)
val |= A4XX_PC_PRIM_VTX_CNTL_PRIMITIVE_RESTART;
val |= COND(vp->writes_psize, A4XX_PC_PRIM_VTX_CNTL_PSIZE);
if (fp->total_in > 0) {
uint32_t varout = align(fp->total_in, 16) / 16;
if (varout > 1)
varout = align(varout, 2);
val |= A4XX_PC_PRIM_VTX_CNTL_VAROUT(varout);
}
OUT_PKT0(ring, REG_A4XX_PC_PRIM_VTX_CNTL, 2);
OUT_RING(ring, val);
OUT_RING(ring, 0x12); /* XXX UNKNOWN_21C5 */
}
if (dirty & FD_DIRTY_SCISSOR) {
struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
OUT_PKT0(ring, REG_A4XX_GRAS_SC_WINDOW_SCISSOR_BR, 2);
OUT_RING(ring, A4XX_GRAS_SC_WINDOW_SCISSOR_BR_X(scissor->maxx - 1) |
A4XX_GRAS_SC_WINDOW_SCISSOR_BR_Y(scissor->maxy - 1));
OUT_RING(ring, A4XX_GRAS_SC_WINDOW_SCISSOR_TL_X(scissor->minx) |
A4XX_GRAS_SC_WINDOW_SCISSOR_TL_Y(scissor->miny));
ctx->max_scissor.minx = MIN2(ctx->max_scissor.minx, scissor->minx);
ctx->max_scissor.miny = MIN2(ctx->max_scissor.miny, scissor->miny);
ctx->max_scissor.maxx = MAX2(ctx->max_scissor.maxx, scissor->maxx);
ctx->max_scissor.maxy = MAX2(ctx->max_scissor.maxy, scissor->maxy);
}
if (dirty & FD_DIRTY_VIEWPORT) {
fd_wfi(ctx, ring);
OUT_PKT0(ring, REG_A4XX_GRAS_CL_VPORT_XOFFSET_0, 6);
OUT_RING(ring, A4XX_GRAS_CL_VPORT_XOFFSET_0(ctx->viewport.translate[0]));
OUT_RING(ring, A4XX_GRAS_CL_VPORT_XSCALE_0(ctx->viewport.scale[0]));
OUT_RING(ring, A4XX_GRAS_CL_VPORT_YOFFSET_0(ctx->viewport.translate[1]));
OUT_RING(ring, A4XX_GRAS_CL_VPORT_YSCALE_0(ctx->viewport.scale[1]));
OUT_RING(ring, A4XX_GRAS_CL_VPORT_ZOFFSET_0(ctx->viewport.translate[2]));
OUT_RING(ring, A4XX_GRAS_CL_VPORT_ZSCALE_0(ctx->viewport.scale[2]));
}
if (dirty & FD_DIRTY_PROG)
fd4_program_emit(ring, emit);
if ((dirty & (FD_DIRTY_PROG | FD_DIRTY_CONSTBUF)) &&
/* evil hack to deal sanely with clear path: */
(emit->prog == &ctx->prog)) {
fd_wfi(ctx, ring);
emit_constants(ring, SB_VERT_SHADER,
&ctx->constbuf[PIPE_SHADER_VERTEX],
vp, emit->prog->dirty & FD_SHADER_DIRTY_VP);
if (!emit->key.binning_pass) {
emit_constants(ring, SB_FRAG_SHADER,
&ctx->constbuf[PIPE_SHADER_FRAGMENT],
fp, emit->prog->dirty & FD_SHADER_DIRTY_FP);
}
}
/* emit driver params every time */
if (emit->info && emit->prog == &ctx->prog) {
uint32_t vertex_params[4] = {
emit->info->indexed ? emit->info->index_bias : emit->info->start,
0,
0,
0
};
if (vp->constlen >= vp->first_driver_param + 4) {
fd4_emit_constant(ring, SB_VERT_SHADER,
(vp->first_driver_param + 4) * 4,
0, 4, vertex_params, NULL);
}
}
if ((dirty & FD_DIRTY_BLEND) && ctx->blend) {
struct fd4_blend_stateobj *blend = fd4_blend_stateobj(ctx->blend);
uint32_t i;
for (i = 0; i < 8; i++) {
OUT_PKT0(ring, REG_A4XX_RB_MRT_CONTROL(i), 1);
OUT_RING(ring, blend->rb_mrt[i].control);
OUT_PKT0(ring, REG_A4XX_RB_MRT_BLEND_CONTROL(i), 1);
OUT_RING(ring, blend->rb_mrt[i].blend_control);
}
OUT_PKT0(ring, REG_A4XX_RB_FS_OUTPUT, 1);
OUT_RING(ring, blend->rb_fs_output |
A4XX_RB_FS_OUTPUT_SAMPLE_MASK(0xffff));
}
if (dirty & FD_DIRTY_BLEND_COLOR) {
struct pipe_blend_color *bcolor = &ctx->blend_color;
OUT_PKT0(ring, REG_A4XX_RB_BLEND_RED, 4);
OUT_RING(ring, A4XX_RB_BLEND_RED_UINT(bcolor->color[0] * 255.0) |
A4XX_RB_BLEND_RED_FLOAT(bcolor->color[0]));
OUT_RING(ring, A4XX_RB_BLEND_GREEN_UINT(bcolor->color[1] * 255.0) |
A4XX_RB_BLEND_GREEN_FLOAT(bcolor->color[1]));
OUT_RING(ring, A4XX_RB_BLEND_BLUE_UINT(bcolor->color[2] * 255.0) |
A4XX_RB_BLEND_BLUE_FLOAT(bcolor->color[2]));
OUT_RING(ring, A4XX_RB_BLEND_ALPHA_UINT(bcolor->color[3] * 255.0) |
A4XX_RB_BLEND_ALPHA_FLOAT(bcolor->color[3]));
}
if (dirty & FD_DIRTY_VERTTEX) {
if (vp->has_samp)
emit_textures(ctx, ring, SB_VERT_TEX, &ctx->verttex);
else
dirty &= ~FD_DIRTY_VERTTEX;
}
if (dirty & FD_DIRTY_FRAGTEX) {
if (fp->has_samp)
emit_textures(ctx, ring, SB_FRAG_TEX, &ctx->fragtex);
else
dirty &= ~FD_DIRTY_FRAGTEX;
}
ctx->dirty &= ~dirty;
}
void
fd4_emit_vertex_bufs(struct fd_ringbuffer *ring, struct fd4_emit *emit)
{
int32_t i, j, last = -1;
uint32_t total_in = 0;
const struct fd_vertex_state *vtx = emit->vtx;
struct ir3_shader_variant *vp = fd4_emit_get_vp(emit);
unsigned vertex_regid = regid(63, 0), instance_regid = regid(63, 0);
for (i = 0; i < vp->inputs_count; i++) {
uint8_t semantic = sem2name(vp->inputs[i].semantic);
if (semantic == TGSI_SEMANTIC_VERTEXID_NOBASE)
vertex_regid = vp->inputs[i].regid;
else if (semantic == TGSI_SEMANTIC_INSTANCEID)
instance_regid = vp->inputs[i].regid;
else if ((i < vtx->vtx->num_elements) && vp->inputs[i].compmask)
last = i;
}
/* hw doesn't like to be configured for zero vbo's, it seems: */
if ((vtx->vtx->num_elements == 0) &&
(vertex_regid == regid(63, 0)) &&
(instance_regid == regid(63, 0)))
return;
for (i = 0, j = 0; i <= last; i++) {
assert(sem2name(vp->inputs[i].semantic) == 0);
if (vp->inputs[i].compmask) {
struct pipe_vertex_element *elem = &vtx->vtx->pipe[i];
const struct pipe_vertex_buffer *vb =
&vtx->vertexbuf.vb[elem->vertex_buffer_index];
struct fd_resource *rsc = fd_resource(vb->buffer);
enum pipe_format pfmt = elem->src_format;
enum a4xx_vtx_fmt fmt = fd4_pipe2vtx(pfmt);
bool switchnext = (i != last) ||
(vertex_regid != regid(63, 0)) ||
(instance_regid != regid(63, 0));
bool isint = util_format_is_pure_integer(pfmt);
uint32_t fs = util_format_get_blocksize(pfmt);
uint32_t off = vb->buffer_offset + elem->src_offset;
uint32_t size = fd_bo_size(rsc->bo) - off;
debug_assert(fmt != ~0);
OUT_PKT0(ring, REG_A4XX_VFD_FETCH(j), 4);
OUT_RING(ring, A4XX_VFD_FETCH_INSTR_0_FETCHSIZE(fs - 1) |
A4XX_VFD_FETCH_INSTR_0_BUFSTRIDE(vb->stride) |
COND(elem->instance_divisor, A4XX_VFD_FETCH_INSTR_0_INSTANCED) |
COND(switchnext, A4XX_VFD_FETCH_INSTR_0_SWITCHNEXT));
OUT_RELOC(ring, rsc->bo, off, 0, 0);
OUT_RING(ring, A4XX_VFD_FETCH_INSTR_2_SIZE(size));
OUT_RING(ring, A4XX_VFD_FETCH_INSTR_3_STEPRATE(MAX2(1, elem->instance_divisor)));
OUT_PKT0(ring, REG_A4XX_VFD_DECODE_INSTR(j), 1);
OUT_RING(ring, A4XX_VFD_DECODE_INSTR_CONSTFILL |
A4XX_VFD_DECODE_INSTR_WRITEMASK(vp->inputs[i].compmask) |
A4XX_VFD_DECODE_INSTR_FORMAT(fmt) |
A4XX_VFD_DECODE_INSTR_SWAP(fd4_pipe2swap(pfmt)) |
A4XX_VFD_DECODE_INSTR_REGID(vp->inputs[i].regid) |
A4XX_VFD_DECODE_INSTR_SHIFTCNT(fs) |
A4XX_VFD_DECODE_INSTR_LASTCOMPVALID |
COND(isint, A4XX_VFD_DECODE_INSTR_INT) |
COND(switchnext, A4XX_VFD_DECODE_INSTR_SWITCHNEXT));
total_in += vp->inputs[i].ncomp;
j++;
}
}
OUT_PKT0(ring, REG_A4XX_VFD_CONTROL_0, 5);
OUT_RING(ring, A4XX_VFD_CONTROL_0_TOTALATTRTOVS(total_in) |
0xa0000 | /* XXX */
A4XX_VFD_CONTROL_0_STRMDECINSTRCNT(j) |
A4XX_VFD_CONTROL_0_STRMFETCHINSTRCNT(j));
OUT_RING(ring, A4XX_VFD_CONTROL_1_MAXSTORAGE(129) | // XXX
A4XX_VFD_CONTROL_1_REGID4VTX(vertex_regid) |
A4XX_VFD_CONTROL_1_REGID4INST(instance_regid));
OUT_RING(ring, 0x00000000); /* XXX VFD_CONTROL_2 */
OUT_RING(ring, A4XX_VFD_CONTROL_3_REGID_VTXCNT(regid(63, 0)));
OUT_RING(ring, 0x00000000); /* XXX VFD_CONTROL_4 */
/* cache invalidate, otherwise vertex fetch could see
* stale vbo contents:
*/
OUT_PKT0(ring, REG_A4XX_UCHE_INVALIDATE0, 2);
OUT_RING(ring, 0x00000000);
OUT_RING(ring, 0x00000012);
}
static bool
fd4_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info)
{
struct fd4_context *fd4_ctx = fd4_context(ctx);
struct fd4_emit emit = {
.debug = &ctx->debug,
.vtx = &ctx->vtx,
.prog = &ctx->prog,
.info = info,
.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,
// TODO set .half_precision based on render target format,
// ie. float16 and smaller use half, float32 use full..
.half_precision = !!(fd_mesa_debug & FD_DBG_FRAGHALF),
.ucp_enables = ctx->rasterizer->clip_plane_enable,
.has_per_samp = (fd4_ctx->fsaturate || fd4_ctx->vsaturate ||
fd4_ctx->fastc_srgb || fd4_ctx->vastc_srgb),
.vsaturate_s = fd4_ctx->vsaturate_s,
.vsaturate_t = fd4_ctx->vsaturate_t,
.vsaturate_r = fd4_ctx->vsaturate_r,
.fsaturate_s = fd4_ctx->fsaturate_s,
.fsaturate_t = fd4_ctx->fsaturate_t,
.fsaturate_r = fd4_ctx->fsaturate_r,
.vastc_srgb = fd4_ctx->vastc_srgb,
.fastc_srgb = fd4_ctx->fastc_srgb,
},
.rasterflat = ctx->rasterizer->flatshade,
.sprite_coord_enable = ctx->rasterizer->sprite_coord_enable,
.sprite_coord_mode = ctx->rasterizer->sprite_coord_mode,
};
fixup_shader_state(ctx, &emit.key);
unsigned dirty = ctx->dirty;
/* do regular pass first, since that is more likely to fail compiling: */
if (!(fd4_emit_get_vp(&emit) && fd4_emit_get_fp(&emit)))
return false;
emit.key.binning_pass = false;
emit.dirty = dirty;
struct fd_ringbuffer *ring = ctx->batch->draw;
if (ctx->rasterizer->rasterizer_discard) {
fd_wfi(ctx->batch, ring);
OUT_PKT3(ring, CP_REG_RMW, 3);
OUT_RING(ring, REG_A4XX_RB_RENDER_CONTROL);
OUT_RING(ring, ~A4XX_RB_RENDER_CONTROL_DISABLE_COLOR_PIPE);
OUT_RING(ring, A4XX_RB_RENDER_CONTROL_DISABLE_COLOR_PIPE);
}
draw_impl(ctx, ctx->batch->draw, &emit);
if (ctx->rasterizer->rasterizer_discard) {
fd_wfi(ctx->batch, ring);
OUT_PKT3(ring, CP_REG_RMW, 3);
OUT_RING(ring, REG_A4XX_RB_RENDER_CONTROL);
OUT_RING(ring, ~A4XX_RB_RENDER_CONTROL_DISABLE_COLOR_PIPE);
OUT_RING(ring, 0);
}
/* and now binning pass: */
emit.key.binning_pass = true;
emit.dirty = dirty & ~(FD_DIRTY_BLEND);
emit.vp = NULL; /* we changed key so need to refetch vp */
emit.fp = NULL;
draw_impl(ctx, ctx->batch->binning, &emit);
return true;
}
