shader_variant_flags |= 1 << 1;
if (si_get_wave_size(sel->screen, sel->info.stage, ngg, es, false, false) == 32)
shader_variant_flags |= 1 << 2;
- if (sel->info.stage == MESA_SHADER_FRAGMENT && sel->info.uses_derivatives && sel->info.uses_kill &&
+ if (sel->info.stage == MESA_SHADER_FRAGMENT &&
+ /* Derivatives imply helper invocations so check for needs_helper_invocations. */
+ sel->info.base.fs.needs_helper_invocations &&
+ sel->info.base.fs.uses_discard &&
sel->screen->debug_flags & DBG(FS_CORRECT_DERIVS_AFTER_KILL))
shader_variant_flags |= 1 << 3;
void gfx9_get_gs_info(struct si_shader_selector *es, struct si_shader_selector *gs,
struct gfx9_gs_info *out)
{
- unsigned gs_num_invocations = MAX2(gs->gs_num_invocations, 1);
+ unsigned gs_num_invocations = MAX2(gs->info.base.gs.invocations, 1);
unsigned input_prim = gs->info.base.gs.input_primitive;
bool uses_adjacency =
input_prim >= PIPE_PRIM_LINES_ADJACENCY && input_prim <= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY;
/* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
* Make sure we don't go over the maximum value.
*/
- if (gs->gs_max_out_vertices > 0) {
+ if (gs->info.base.gs.vertices_out > 0) {
max_gs_prims =
- MIN2(max_gs_prims, max_out_prims / (gs->gs_max_out_vertices * gs_num_invocations));
+ MIN2(max_gs_prims, max_out_prims / (gs->info.base.gs.vertices_out * gs_num_invocations));
}
assert(max_gs_prims > 0);
out->es_verts_per_subgroup = es_verts;
out->gs_prims_per_subgroup = gs_prims;
out->gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations;
- out->max_prims_per_subgroup = out->gs_inst_prims_in_subgroup * gs->gs_max_out_vertices;
+ out->max_prims_per_subgroup = out->gs_inst_prims_in_subgroup * gs->info.base.gs.vertices_out;
out->esgs_ring_size = esgs_lds_size;
assert(out->max_prims_per_subgroup <= max_out_prims);
{
struct si_shader_selector *sel = shader->selector;
const ubyte *num_components = sel->info.num_stream_output_components;
- unsigned gs_num_invocations = sel->gs_num_invocations;
+ unsigned gs_num_invocations = sel->info.base.gs.invocations;
struct si_pm4_state *pm4;
uint64_t va;
- unsigned max_stream = sel->max_gs_stream;
+ unsigned max_stream = util_last_bit(sel->info.base.gs.active_stream_mask);
unsigned offset;
pm4 = si_get_shader_pm4_state(shader);
pm4->atom.emit = si_emit_shader_gs;
- offset = num_components[0] * sel->gs_max_out_vertices;
+ offset = num_components[0] * sel->info.base.gs.vertices_out;
shader->ctx_reg.gs.vgt_gsvs_ring_offset_1 = offset;
- if (max_stream >= 1)
- offset += num_components[1] * sel->gs_max_out_vertices;
+ if (max_stream >= 2)
+ offset += num_components[1] * sel->info.base.gs.vertices_out;
shader->ctx_reg.gs.vgt_gsvs_ring_offset_2 = offset;
- if (max_stream >= 2)
- offset += num_components[2] * sel->gs_max_out_vertices;
+ if (max_stream >= 3)
+ offset += num_components[2] * sel->info.base.gs.vertices_out;
shader->ctx_reg.gs.vgt_gsvs_ring_offset_3 = offset;
- if (max_stream >= 3)
- offset += num_components[3] * sel->gs_max_out_vertices;
+ if (max_stream >= 4)
+ offset += num_components[3] * sel->info.base.gs.vertices_out;
shader->ctx_reg.gs.vgt_gsvs_ring_itemsize = offset;
/* The GSVS_RING_ITEMSIZE register takes 15 bits */
assert(offset < (1 << 15));
- shader->ctx_reg.gs.vgt_gs_max_vert_out = sel->gs_max_out_vertices;
+ shader->ctx_reg.gs.vgt_gs_max_vert_out = sel->info.base.gs.vertices_out;
shader->ctx_reg.gs.vgt_gs_vert_itemsize = num_components[0];
- shader->ctx_reg.gs.vgt_gs_vert_itemsize_1 = (max_stream >= 1) ? num_components[1] : 0;
- shader->ctx_reg.gs.vgt_gs_vert_itemsize_2 = (max_stream >= 2) ? num_components[2] : 0;
- shader->ctx_reg.gs.vgt_gs_vert_itemsize_3 = (max_stream >= 3) ? num_components[3] : 0;
+ shader->ctx_reg.gs.vgt_gs_vert_itemsize_1 = (max_stream >= 2) ? num_components[1] : 0;
+ shader->ctx_reg.gs.vgt_gs_vert_itemsize_2 = (max_stream >= 3) ? num_components[2] : 0;
+ shader->ctx_reg.gs.vgt_gs_vert_itemsize_3 = (max_stream >= 4) ? num_components[3] : 0;
shader->ctx_reg.gs.vgt_gs_instance_cnt =
S_028B90_CNT(MIN2(gs_num_invocations, 127)) | S_028B90_ENABLE(gs_num_invocations > 0);
return PIPE_PRIM_TRIANGLES; /* worst case for all callers */
}
-static unsigned si_get_vs_out_cntl(const struct si_shader_selector *sel, bool ngg)
+static unsigned si_get_vs_out_cntl(const struct si_shader_selector *sel,
+ const struct si_shader *shader, bool ngg)
{
- bool misc_vec_ena = sel->info.writes_psize || (sel->info.writes_edgeflag && !ngg) ||
+ bool writes_psize = sel->info.writes_psize;
+
+ if (shader)
+ writes_psize &= !shader->key.opt.kill_pointsize;
+
+ bool misc_vec_ena = writes_psize || (sel->info.writes_edgeflag && !ngg) ||
sel->info.writes_layer || sel->info.writes_viewport_index;
- return S_02881C_USE_VTX_POINT_SIZE(sel->info.writes_psize) |
+ return S_02881C_USE_VTX_POINT_SIZE(writes_psize) |
S_02881C_USE_VTX_EDGE_FLAG(sel->info.writes_edgeflag && !ngg) |
S_02881C_USE_VTX_RENDER_TARGET_INDX(sel->info.writes_layer) |
S_02881C_USE_VTX_VIEWPORT_INDX(sel->info.writes_viewport_index) |
bool window_space = gs_info->stage == MESA_SHADER_VERTEX ?
gs_info->base.vs.window_space_position : 0;
bool es_enable_prim_id = shader->key.mono.u.vs_export_prim_id || es_info->uses_primid;
- unsigned gs_num_invocations = MAX2(gs_sel->gs_num_invocations, 1);
+ unsigned gs_num_invocations = MAX2(gs_sel->info.base.gs.invocations, 1);
unsigned input_prim = si_get_input_prim(gs_sel);
bool break_wave_at_eoi = false;
struct si_pm4_state *pm4 = si_get_shader_pm4_state(shader);
if (gs_stage == MESA_SHADER_GEOMETRY) {
shader->ctx_reg.ngg.vgt_esgs_ring_itemsize = es_sel->esgs_itemsize / 4;
- shader->ctx_reg.ngg.vgt_gs_max_vert_out = gs_sel->gs_max_out_vertices;
+ shader->ctx_reg.ngg.vgt_gs_max_vert_out = gs_sel->info.base.gs.vertices_out;
} else {
shader->ctx_reg.ngg.vgt_esgs_ring_itemsize = 1;
}
S_028838_INDEX_BUF_EDGE_FLAG_ENA(gs_stage == MESA_SHADER_VERTEX) |
/* Reuse for NGG. */
S_028838_VERTEX_REUSE_DEPTH(sscreen->info.chip_class >= GFX10_3 ? 30 : 0);
- shader->pa_cl_vs_out_cntl = si_get_vs_out_cntl(gs_sel, true);
+ shader->pa_cl_vs_out_cntl = si_get_vs_out_cntl(shader->selector, shader, true);
/* Oversubscribe PC. This improves performance when there are too many varyings. */
float oversub_pc_factor = 0.25;
shader->ctx_reg.vs.vgt_primitiveid_en = enable_prim_id;
} else {
shader->ctx_reg.vs.vgt_gs_mode =
- ac_vgt_gs_mode(gs->gs_max_out_vertices, sscreen->info.chip_class);
+ ac_vgt_gs_mode(gs->info.base.gs.vertices_out, sscreen->info.chip_class);
shader->ctx_reg.vs.vgt_primitiveid_en = 0;
}
: V_02870C_SPI_SHADER_NONE);
shader->ctx_reg.vs.ge_pc_alloc = S_030980_OVERSUB_EN(sscreen->info.use_late_alloc) |
S_030980_NUM_PC_LINES(sscreen->info.pc_lines / 4 - 1);
- shader->pa_cl_vs_out_cntl = si_get_vs_out_cntl(shader->selector, false);
+ shader->pa_cl_vs_out_cntl = si_get_vs_out_cntl(shader->selector, shader, false);
oc_lds_en = shader->selector->info.stage == MESA_SHADER_TESS_EVAL ? 1 : 0;
*/
spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
- if (info->properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] == TGSI_FS_COORD_PIXEL_CENTER_INTEGER)
+ if (info->base.fs.pixel_center_integer)
spi_baryc_cntl |= S_0286E0_POS_FLOAT_ULC(1);
spi_shader_col_format = si_get_spi_shader_col_format(shader);
* the color and Z formats to SPI_SHADER_ZERO. The hw will skip export
* instructions if any are present.
*/
- if ((sscreen->info.chip_class <= GFX9 || info->uses_kill ||
+ if ((sscreen->info.chip_class <= GFX9 || info->base.fs.uses_discard ||
shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS) &&
!spi_shader_col_format && !info->writes_z && !info->writes_stencil &&
!info->writes_samplemask)
struct si_shader_selector *ps = sctx->ps_shader.cso;
key->opt.clip_disable = sctx->queued.named.rasterizer->clip_plane_enable == 0 &&
- (vs->info.clipdist_writemask || vs->info.writes_clipvertex) &&
- !vs->info.culldist_writemask;
+ (vs->info.base.clip_distance_array_size || vs->info.writes_clipvertex) &&
+ !vs->info.base.cull_distance_array_size;
/* Find out if PS is disabled. */
bool ps_disabled = true;
if (ps) {
- bool ps_modifies_zs = ps->info.uses_kill || ps->info.writes_z || ps->info.writes_stencil ||
+ bool ps_modifies_zs = ps->info.base.fs.uses_discard || ps->info.writes_z || ps->info.writes_stencil ||
ps->info.writes_samplemask ||
sctx->queued.named.blend->alpha_to_coverage ||
si_get_alpha_test_func(sctx) != PIPE_FUNC_ALWAYS;
unsigned ps_colormask = si_get_total_colormask(sctx);
ps_disabled = sctx->queued.named.rasterizer->rasterizer_discard ||
- (!ps_colormask && !ps_modifies_zs && !ps->info.writes_memory);
+ (!ps_colormask && !ps_modifies_zs && !ps->info.base.writes_memory);
}
/* Find out which VS outputs aren't used by the PS. */
key->opt.kill_outputs = ~linked & outputs_written;
key->opt.ngg_culling = sctx->ngg_culling;
+
+ if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
+ key->mono.u.vs_export_prim_id = 1;
+
+ /* We need PKT3_CONTEXT_REG_RMW, which we currently only use on GFX10+. */
+ if (sctx->chip_class >= GFX10 &&
+ vs->info.writes_psize &&
+ sctx->current_rast_prim != PIPE_PRIM_POINTS &&
+ !sctx->queued.named.rasterizer->polygon_mode_is_points)
+ key->opt.kill_pointsize = 1;
}
/* Compute the key for the hw shader variant */
} else {
key->as_ngg = stages_key.u.ngg;
si_shader_selector_key_hw_vs(sctx, sel, key);
-
- if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
- key->mono.u.vs_export_prim_id = 1;
}
break;
case MESA_SHADER_TESS_CTRL:
key->as_es = 1;
else {
si_shader_selector_key_hw_vs(sctx, sel, key);
-
- if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
- key->mono.u.vs_export_prim_id = 1;
}
break;
case MESA_SHADER_GEOMETRY:
struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
struct si_state_blend *blend = sctx->queued.named.blend;
- if (sel->info.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
+ if (sel->info.color0_writes_all_cbufs &&
sel->info.colors_written == 0x1)
key->part.ps.epilog.last_cbuf = MAX2(sctx->framebuffer.state.nr_cbufs, 1) - 1;
sel->info.uses_linear_sample >
1;
- if (sel->info.uses_persp_opcode_interp_sample ||
- sel->info.uses_linear_opcode_interp_sample)
+ if (sel->info.uses_interp_at_sample)
key->mono.u.ps.interpolate_at_sample_force_center = 1;
}
{
unsigned start, num_shaderbufs, num_constbufs, num_images, num_msaa_images, num_samplers;
- num_shaderbufs = util_last_bit(info->shader_buffers_declared);
- num_constbufs = util_last_bit(info->const_buffers_declared);
+ num_shaderbufs = info->base.num_ssbos;
+ num_constbufs = info->base.num_ubos;
/* two 8-byte images share one 16-byte slot */
- num_images = align(util_last_bit(info->images_declared), 2);
- num_msaa_images = align(util_last_bit(info->msaa_images_declared), 2);
- num_samplers = util_last_bit(info->samplers_declared);
+ num_images = align(info->base.num_images, 2);
+ num_msaa_images = align(util_last_bit(info->base.msaa_images), 2);
+ num_samplers = util_last_bit(info->base.textures_used);
/* The layout is: sb[last] ... sb[0], cb[0] ... cb[last] */
start = si_get_shaderbuf_slot(num_shaderbufs - 1);
sel->prim_discard_cs_allowed =
sel->info.stage == MESA_SHADER_VERTEX && !sel->info.uses_bindless_images &&
- !sel->info.uses_bindless_samplers && !sel->info.writes_memory &&
+ !sel->info.uses_bindless_samplers && !sel->info.base.writes_memory &&
!sel->info.writes_viewport_index &&
!sel->info.base.vs.window_space_position && !sel->so.num_outputs;
switch (sel->info.stage) {
case MESA_SHADER_GEOMETRY:
- sel->gs_output_prim = sel->info.base.gs.output_primitive;
-
/* Only possibilities: POINTS, LINE_STRIP, TRIANGLES */
- sel->rast_prim = sel->gs_output_prim;
+ sel->rast_prim = sel->info.base.gs.output_primitive;
if (util_rast_prim_is_triangles(sel->rast_prim))
sel->rast_prim = PIPE_PRIM_TRIANGLES;
- sel->gs_max_out_vertices = sel->info.base.gs.vertices_out;
- sel->gs_num_invocations = sel->info.base.gs.invocations;
sel->gsvs_vertex_size = sel->info.num_outputs * 16;
- sel->max_gsvs_emit_size = sel->gsvs_vertex_size * sel->gs_max_out_vertices;
-
- sel->max_gs_stream = 0;
- for (i = 0; i < sel->so.num_outputs; i++)
- sel->max_gs_stream = MAX2(sel->max_gs_stream, sel->so.output[i].stream);
-
+ sel->max_gsvs_emit_size = sel->gsvs_vertex_size * sel->info.base.gs.vertices_out;
sel->gs_input_verts_per_prim =
u_vertices_per_prim(sel->info.base.gs.input_primitive);
/* EN_MAX_VERT_OUT_PER_GS_INSTANCE does not work with tesselation so
* we can't split workgroups. Disable ngg if any of the following conditions is true:
- * - num_invocations * gs_max_out_vertices > 256
+ * - num_invocations * gs.vertices_out > 256
* - LDS usage is too high
*/
sel->tess_turns_off_ngg = sscreen->info.chip_class >= GFX10 &&
- (sel->gs_num_invocations * sel->gs_max_out_vertices > 256 ||
- sel->gs_num_invocations * sel->gs_max_out_vertices *
+ (sel->info.base.gs.invocations * sel->info.base.gs.vertices_out > 256 ||
+ sel->info.base.gs.invocations * sel->info.base.gs.vertices_out *
(sel->info.num_outputs * 4 + 1) > 6500 /* max dw per GS primitive */);
break;
sscreen->always_use_ngg_culling_tess))) &&
sel->info.writes_position &&
!sel->info.writes_viewport_index && /* cull only against viewport 0 */
- !sel->info.writes_memory && !sel->so.num_outputs &&
+ !sel->info.base.writes_memory && !sel->so.num_outputs &&
(sel->info.stage != MESA_SHADER_VERTEX ||
(!sel->info.base.vs.blit_sgprs_amd &&
!sel->info.base.vs.window_space_position));
/* PA_CL_VS_OUT_CNTL */
if (sctx->chip_class <= GFX9)
- sel->pa_cl_vs_out_cntl = si_get_vs_out_cntl(sel, false);
+ sel->pa_cl_vs_out_cntl = si_get_vs_out_cntl(sel, NULL, false);
- sel->clipdist_mask = sel->info.writes_clipvertex ? SIX_BITS : sel->info.clipdist_writemask;
- sel->culldist_mask = sel->info.culldist_writemask << sel->info.num_written_clipdistance;
+ sel->clipdist_mask = sel->info.writes_clipvertex ? SIX_BITS :
+ u_bit_consecutive(0, sel->info.base.clip_distance_array_size);
+ sel->culldist_mask = u_bit_consecutive(0, sel->info.base.cull_distance_array_size) <<
+ sel->info.base.clip_distance_array_size;
/* DB_SHADER_CONTROL */
sel->db_shader_control = S_02880C_Z_EXPORT_ENABLE(sel->info.writes_z) |
S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(sel->info.writes_stencil) |
S_02880C_MASK_EXPORT_ENABLE(sel->info.writes_samplemask) |
- S_02880C_KILL_ENABLE(sel->info.uses_kill);
-
- switch (sel->info.properties[TGSI_PROPERTY_FS_DEPTH_LAYOUT]) {
- case TGSI_FS_DEPTH_LAYOUT_GREATER:
- sel->db_shader_control |= S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z);
- break;
- case TGSI_FS_DEPTH_LAYOUT_LESS:
- sel->db_shader_control |= S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z);
- break;
- }
+ S_02880C_KILL_ENABLE(sel->info.base.fs.uses_discard);
if (sel->info.stage == MESA_SHADER_FRAGMENT) {
+ switch (sel->info.base.fs.depth_layout) {
+ case FRAG_DEPTH_LAYOUT_GREATER:
+ sel->db_shader_control |= S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z);
+ break;
+ case FRAG_DEPTH_LAYOUT_LESS:
+ sel->db_shader_control |= S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z);
+ break;
+ default:;
+ }
+
/* Z_ORDER, EXEC_ON_HIER_FAIL and EXEC_ON_NOOP should be set as following:
*
* | early Z/S | writes_mem | allow_ReZ? | Z_ORDER | EXEC_ON_HIER_FAIL | EXEC_ON_NOOP
/* Cases 3, 4. */
sel->db_shader_control |= S_02880C_DEPTH_BEFORE_SHADER(1) |
S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z) |
- S_02880C_EXEC_ON_NOOP(sel->info.writes_memory);
- } else if (sel->info.writes_memory) {
+ S_02880C_EXEC_ON_NOOP(sel->info.base.writes_memory);
+ } else if (sel->info.base.writes_memory) {
/* Case 2. */
sel->db_shader_control |= S_02880C_Z_ORDER(V_02880C_LATE_Z) | S_02880C_EXEC_ON_HIER_FAIL(1);
} else {
sel->db_shader_control |= S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z);
}
- if (sel->info.properties[TGSI_PROPERTY_FS_POST_DEPTH_COVERAGE])
+ if (sel->info.base.fs.post_depth_coverage)
sel->db_shader_control |= S_02880C_PRE_SHADER_DEPTH_COVERAGE_ENABLE(1);
}
si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
if (sctx->screen->has_out_of_order_rast &&
- (!old_sel || old_sel->info.writes_memory != sel->info.writes_memory ||
+ (!old_sel || old_sel->info.base.writes_memory != sel->info.base.writes_memory ||
old_sel->info.base.fs.early_fragment_tests !=
sel->info.base.fs.early_fragment_tests))
si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
projects / mesa.git / blobdiff