#include "sid.h"
#include "compiler/nir/nir_serialize.h"
+#include "nir/tgsi_to_nir.h"
#include "tgsi/tgsi_parse.h"
#include "util/hash_table.h"
#include "util/crc32.h"
radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
shader->ctx_reg.ngg.vgt_esgs_ring_itemsize);
- radeon_opt_set_context_reg(sctx, R_028AB4_VGT_REUSE_OFF,
- SI_TRACKED_VGT_REUSE_OFF,
- shader->ctx_reg.ngg.vgt_reuse_off);
radeon_opt_set_context_reg(sctx, R_0286C4_SPI_VS_OUT_CONFIG,
SI_TRACKED_SPI_VS_OUT_CONFIG,
shader->ctx_reg.ngg.spi_vs_out_config);
gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
}
-static void si_set_ge_pc_alloc(struct si_screen *sscreen,
- struct si_pm4_state *pm4, bool culling)
-{
- si_pm4_set_reg(pm4, R_030980_GE_PC_ALLOC,
- S_030980_OVERSUB_EN(1) |
- S_030980_NUM_PC_LINES((culling ? 256 : 128) * sscreen->info.max_se - 1));
-}
-
-unsigned si_get_input_prim(const struct si_shader_selector *gs,
- unsigned default_worst_case)
+unsigned si_get_input_prim(const struct si_shader_selector *gs)
{
if (gs->type == PIPE_SHADER_GEOMETRY)
return gs->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
}
/* TODO: Set this correctly if the primitive type is set in the shader key. */
- return default_worst_case;
+ return PIPE_PRIM_TRIANGLES; /* worst case for all callers */
}
/**
gs_info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
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);
- /* Anything above TRIANGLES has the same effect as TRIANGLES here. */
- unsigned input_prim = si_get_input_prim(gs_sel, PIPE_PRIM_TRIANGLES);
+ 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 (!pm4)
S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5) |
S_00B22C_OC_LDS_EN(es_type == PIPE_SHADER_TESS_EVAL) |
S_00B22C_LDS_SIZE(shader->config.lds_size));
- si_set_ge_pc_alloc(sscreen, pm4, false);
nparams = MAX2(shader->info.nr_param_exports, 1);
shader->ctx_reg.ngg.spi_vs_out_config =
S_03096C_VERT_GRP_SIZE(shader->ngg.hw_max_esverts) |
S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi);
+ /* Bug workaround for a possible hang with non-tessellation cases.
+ * Tessellation always sets GE_CNTL.VERT_GRP_SIZE = 0
+ *
+ * Requirement: GE_CNTL.VERT_GRP_SIZE = VGT_GS_ONCHIP_CNTL.ES_VERTS_PER_SUBGRP - 5
+ */
+ if ((sscreen->info.family == CHIP_NAVI10 ||
+ sscreen->info.family == CHIP_NAVI12 ||
+ sscreen->info.family == CHIP_NAVI14) &&
+ (es_type == PIPE_SHADER_VERTEX || gs_type == PIPE_SHADER_VERTEX) && /* = no tess */
+ shader->ngg.hw_max_esverts != 256) {
+ shader->ge_cntl &= C_03096C_VERT_GRP_SIZE;
+
+ if (shader->ngg.hw_max_esverts > 5) {
+ shader->ge_cntl |=
+ S_03096C_VERT_GRP_SIZE(shader->ngg.hw_max_esverts - 5);
+ }
+ }
+
if (window_space) {
shader->ctx_reg.ngg.pa_cl_vte_cntl =
S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
}
-
- shader->ctx_reg.ngg.vgt_reuse_off =
- S_028AB4_REUSE_OFF(sscreen->info.family == CHIP_NAVI10 &&
- sscreen->info.chip_external_rev == 0x1 &&
- es_type == PIPE_SHADER_TESS_EVAL);
}
static void si_emit_shader_vs(struct si_context *sctx)
vgpr_comp_cnt = 0; /* only VertexID is needed for GS-COPY. */
num_user_sgprs = SI_GSCOPY_NUM_USER_SGPR;
} else if (shader->selector->type == PIPE_SHADER_VERTEX) {
- /* VGPR0-3: (VertexID, InstanceID / StepRate0, PrimID, InstanceID)
- * If PrimID is disabled. InstanceID / StepRate1 is loaded instead.
- * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
- */
- vgpr_comp_cnt = enable_prim_id ? 2 : (shader->info.uses_instanceid ? 1 : 0);
+ if (sscreen->info.chip_class >= GFX10) {
+ vgpr_comp_cnt = shader->info.uses_instanceid ? 3 : (enable_prim_id ? 2 : 0);
+ } else {
+ /* VGPR0-3: (VertexID, InstanceID / StepRate0, PrimID, InstanceID)
+ * If PrimID is disabled. InstanceID / StepRate1 is loaded instead.
+ * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
+ */
+ vgpr_comp_cnt = enable_prim_id ? 2 : (shader->info.uses_instanceid ? 1 : 0);
+ }
if (info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
num_user_sgprs = SI_SGPR_VS_BLIT_DATA +
si_pm4_set_reg(pm4, R_00B120_SPI_SHADER_PGM_LO_VS, va >> 8);
si_pm4_set_reg(pm4, R_00B124_SPI_SHADER_PGM_HI_VS, S_00B124_MEM_BASE(va >> 40));
- if (sscreen->info.chip_class >= GFX10)
- si_set_ge_pc_alloc(sscreen, pm4, false);
uint32_t rsrc1 = S_00B128_VGPRS((shader->config.num_vgprs - 1) /
(sscreen->ge_wave_size == 32 ? 8 : 4)) |
static unsigned si_get_alpha_test_func(struct si_context *sctx)
{
/* Alpha-test should be disabled if colorbuffer 0 is integer. */
- if (sctx->queued.named.dsa)
- return sctx->queued.named.dsa->alpha_func;
-
- return PIPE_FUNC_ALWAYS;
+ return sctx->queued.named.dsa->alpha_func;
}
void si_shader_selector_key_vs(struct si_context *sctx,
/* Find out if PS is disabled. */
bool ps_disabled = true;
if (ps) {
- const struct si_state_blend *blend = sctx->queued.named.blend;
- bool alpha_to_coverage = blend && blend->alpha_to_coverage;
bool ps_modifies_zs = ps->info.uses_kill ||
ps->info.writes_z ||
ps->info.writes_stencil ||
ps->info.writes_samplemask ||
- alpha_to_coverage ||
+ sctx->queued.named.blend->alpha_to_coverage ||
si_get_alpha_test_func(sctx) != PIPE_FUNC_ALWAYS;
unsigned ps_colormask = si_get_total_colormask(sctx);
key->mono.u.ff_tcs_inputs_to_copy = sctx->vs_shader.cso->outputs_written;
break;
case PIPE_SHADER_TESS_EVAL:
+ key->as_ngg = stages_key.u.ngg;
+
if (sctx->gs_shader.cso)
key->as_es = 1;
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)
sel->info.colors_written == 0x1)
key->part.ps.epilog.last_cbuf = MAX2(sctx->framebuffer.state.nr_cbufs, 1) - 1;
- if (blend) {
- /* Select the shader color format based on whether
- * blending or alpha are needed.
- */
- key->part.ps.epilog.spi_shader_col_format =
- (blend->blend_enable_4bit & blend->need_src_alpha_4bit &
- sctx->framebuffer.spi_shader_col_format_blend_alpha) |
- (blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
- sctx->framebuffer.spi_shader_col_format_blend) |
- (~blend->blend_enable_4bit & blend->need_src_alpha_4bit &
- sctx->framebuffer.spi_shader_col_format_alpha) |
- (~blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
- sctx->framebuffer.spi_shader_col_format);
- key->part.ps.epilog.spi_shader_col_format &= blend->cb_target_enabled_4bit;
-
- /* The output for dual source blending should have
- * the same format as the first output.
- */
- if (blend->dual_src_blend)
- key->part.ps.epilog.spi_shader_col_format |=
- (key->part.ps.epilog.spi_shader_col_format & 0xf) << 4;
- } else
- key->part.ps.epilog.spi_shader_col_format = sctx->framebuffer.spi_shader_col_format;
+ /* Select the shader color format based on whether
+ * blending or alpha are needed.
+ */
+ key->part.ps.epilog.spi_shader_col_format =
+ (blend->blend_enable_4bit & blend->need_src_alpha_4bit &
+ sctx->framebuffer.spi_shader_col_format_blend_alpha) |
+ (blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
+ sctx->framebuffer.spi_shader_col_format_blend) |
+ (~blend->blend_enable_4bit & blend->need_src_alpha_4bit &
+ sctx->framebuffer.spi_shader_col_format_alpha) |
+ (~blend->blend_enable_4bit & ~blend->need_src_alpha_4bit &
+ sctx->framebuffer.spi_shader_col_format);
+ key->part.ps.epilog.spi_shader_col_format &= blend->cb_target_enabled_4bit;
+
+ /* The output for dual source blending should have
+ * the same format as the first output.
+ */
+ if (blend->dual_src_blend) {
+ key->part.ps.epilog.spi_shader_col_format |=
+ (key->part.ps.epilog.spi_shader_col_format & 0xf) << 4;
+ }
/* If alpha-to-coverage is enabled, we have to export alpha
* even if there is no color buffer.
*/
if (!(key->part.ps.epilog.spi_shader_col_format & 0xf) &&
- blend && blend->alpha_to_coverage)
+ blend->alpha_to_coverage)
key->part.ps.epilog.spi_shader_col_format |= V_028710_SPI_SHADER_32_AR;
/* On GFX6 and GFX7 except Hawaii, the CB doesn't clamp outputs
key->part.ps.prolog.color_two_side = rs->two_side && sel->info.colors_read;
key->part.ps.prolog.flatshade_colors = rs->flatshade && sel->info.colors_read;
- if (sctx->queued.named.blend) {
- key->part.ps.epilog.alpha_to_one = sctx->queued.named.blend->alpha_to_one &&
- rs->multisample_enable;
- }
+ key->part.ps.epilog.alpha_to_one = blend->alpha_to_one &&
+ rs->multisample_enable;
key->part.ps.prolog.poly_stipple = rs->poly_stipple_enable && is_poly;
key->part.ps.epilog.poly_line_smoothing = ((is_poly && rs->poly_smooth) ||
sel->info.uses_linear_centroid +
sel->info.uses_linear_sample > 1;
- if (sel->info.opcode_count[TGSI_OPCODE_INTERP_SAMPLE])
+ if (sel->info.uses_persp_opcode_interp_sample ||
+ sel->info.uses_linear_opcode_interp_sample)
key->mono.u.ps.interpolate_at_sample_force_center = 1;
}
else
assert(0);
+ if (sel->type == PIPE_SHADER_GEOMETRY &&
+ previous_stage_sel->type == PIPE_SHADER_TESS_EVAL)
+ shader1_key.as_ngg = key->as_ngg;
+
mtx_lock(&previous_stage_sel->mutex);
ok = si_check_missing_main_part(sscreen,
previous_stage_sel,
if (sel->nir) {
/* TODO: GS always sets wave size = default. Legacy GS will have
* incorrect subgroup_size and ballot_bit_size. */
- si_lower_nir(sel, si_get_wave_size(sscreen, sel->type, true));
+ si_lower_nir(sel, si_get_wave_size(sscreen, sel->type, true, false));
}
/* Compile the main shader part for use with a prolog and/or epilog.
sel->so.num_outputs != 0,
&shader->key);
if (sscreen->info.chip_class >= GFX10 &&
- (((sel->type == PIPE_SHADER_VERTEX ||
- sel->type == PIPE_SHADER_TESS_EVAL) &&
+ ((sel->type == PIPE_SHADER_VERTEX &&
!shader->key.as_ls && !shader->key.as_es) ||
+ sel->type == PIPE_SHADER_TESS_EVAL ||
sel->type == PIPE_SHADER_GEOMETRY))
shader->key.as_ngg = 1;
sel->so = state->stream_output;
- if (state->type == PIPE_SHADER_IR_TGSI) {
+ if (state->type == PIPE_SHADER_IR_TGSI &&
+ !sscreen->options.always_nir) {
sel->tokens = tgsi_dup_tokens(state->tokens);
if (!sel->tokens) {
FREE(sel);
tgsi_scan_shader(state->tokens, &sel->info);
tgsi_scan_tess_ctrl(state->tokens, &sel->info, &sel->tcs_info);
- } else {
- assert(state->type == PIPE_SHADER_IR_NIR);
- sel->nir = state->ir.nir;
+ /* Fixup for TGSI: Set which opcode uses which (i,j) pair. */
+ if (sel->info.uses_persp_opcode_interp_centroid)
+ sel->info.uses_persp_centroid = true;
+
+ if (sel->info.uses_linear_opcode_interp_centroid)
+ sel->info.uses_linear_centroid = true;
+
+ if (sel->info.uses_persp_opcode_interp_offset ||
+ sel->info.uses_persp_opcode_interp_sample)
+ sel->info.uses_persp_center = true;
+
+ if (sel->info.uses_linear_opcode_interp_offset ||
+ sel->info.uses_linear_opcode_interp_sample)
+ sel->info.uses_linear_center = true;
+ } else {
+ if (state->type == PIPE_SHADER_IR_TGSI) {
+ sel->nir = tgsi_to_nir(state->tokens, ctx->screen);
+ } else {
+ assert(state->type == PIPE_SHADER_IR_NIR);
+ sel->nir = state->ir.nir;
+ }
+ si_nir_lower_ps_inputs(sel->nir);
si_nir_opts(sel->nir);
si_nir_scan_shader(sel->nir, &sel->info);
si_nir_scan_tess_ctrl(sel->nir, &sel->tcs_info);
sel->pos_writes_edgeflag = true;
}
- /* Set which opcode uses which (i,j) pair. */
- if (sel->info.uses_persp_opcode_interp_centroid)
- sel->info.uses_persp_centroid = true;
-
- if (sel->info.uses_linear_opcode_interp_centroid)
- sel->info.uses_linear_centroid = true;
-
- if (sel->info.uses_persp_opcode_interp_offset ||
- sel->info.uses_persp_opcode_interp_sample)
- sel->info.uses_persp_center = true;
-
- if (sel->info.uses_linear_opcode_interp_offset ||
- sel->info.uses_linear_opcode_interp_sample)
- sel->info.uses_linear_center = true;
-
switch (sel->type) {
case PIPE_SHADER_GEOMETRY:
sel->gs_output_prim =
new_ngg = false;
if (new_ngg != sctx->ngg) {
+ /* Transitioning from NGG to legacy GS requires VGT_FLUSH on Navi10-14.
+ * VGT_FLUSH is also emitted at the beginning of IBs when legacy GS ring
+ * pointers are set.
+ */
+ if ((sctx->family == CHIP_NAVI10 ||
+ sctx->family == CHIP_NAVI12 ||
+ sctx->family == CHIP_NAVI14) &&
+ !new_ngg)
+ sctx->flags |= SI_CONTEXT_VGT_FLUSH;
+
sctx->ngg = new_ngg;
sctx->last_rast_prim = -1; /* reset this so that it gets updated */
return true;
if (sctx->gs_shader.cso)
key.u.gs = 1;
- if (sctx->chip_class >= GFX10) {
- key.u.ngg = sctx->ngg;
-
- if (sctx->gs_shader.cso)
- key.u.streamout = !!sctx->gs_shader.cso->so.num_outputs;
- else if (sctx->tes_shader.cso)
- key.u.streamout = !!sctx->tes_shader.cso->so.num_outputs;
- else
- key.u.streamout = !!sctx->vs_shader.cso->so.num_outputs;
+ if (sctx->ngg) {
+ key.u.ngg = 1;
+ key.u.streamout = !!si_get_vs(sctx)->cso->so.num_outputs;
}
/* Update TCS and TES. */