/* SHADER STATES */
static void si_set_tesseval_regs(struct si_screen *sscreen,
- struct si_shader *shader,
+ struct si_shader_selector *tes,
struct si_pm4_state *pm4)
{
- struct tgsi_shader_info *info = &shader->selector->info;
+ struct tgsi_shader_info *info = &tes->info;
unsigned tes_prim_mode = info->properties[TGSI_PROPERTY_TES_PRIM_MODE];
unsigned tes_spacing = info->properties[TGSI_PROPERTY_TES_SPACING];
bool tes_vertex_order_cw = info->properties[TGSI_PROPERTY_TES_VERTEX_ORDER_CW];
* VS as ES | ES -> GS -> VS | 30
* TES as VS | LS -> HS -> VS | 14 or 30
* TES as ES | LS -> HS -> ES -> GS -> VS | 14 or 30
+ *
+ * If "shader" is NULL, it's assumed it's not LS or GS copy shader.
*/
static void polaris_set_vgt_vertex_reuse(struct si_screen *sscreen,
+ struct si_shader_selector *sel,
struct si_shader *shader,
struct si_pm4_state *pm4)
{
- unsigned type = shader->selector->type;
+ unsigned type = sel->type;
if (sscreen->b.family < CHIP_POLARIS10)
return;
/* VS as VS, or VS as ES: */
if ((type == PIPE_SHADER_VERTEX &&
- !shader->key.as_ls &&
- !shader->is_gs_copy_shader) ||
+ (!shader ||
+ (!shader->key.as_ls && !shader->is_gs_copy_shader))) ||
/* TES as VS, or TES as ES: */
type == PIPE_SHADER_TESS_EVAL) {
unsigned vtx_reuse_depth = 30;
if (type == PIPE_SHADER_TESS_EVAL &&
- shader->selector->info.properties[TGSI_PROPERTY_TES_SPACING] ==
+ sel->info.properties[TGSI_PROPERTY_TES_SPACING] ==
PIPE_TESS_SPACING_FRACTIONAL_ODD)
vtx_reuse_depth = 14;
si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
/* We need at least 2 components for LS.
- * VGPR0-3: (VertexID, RelAutoindex, ???, InstanceID). */
- vgpr_comp_cnt = shader->info.uses_instanceid ? 3 : 1;
+ * VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
+ * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
+ */
+ vgpr_comp_cnt = shader->info.uses_instanceid ? 2 : 1;
si_pm4_set_reg(pm4, R_00B520_SPI_SHADER_PGM_LO_LS, va >> 8);
si_pm4_set_reg(pm4, R_00B524_SPI_SHADER_PGM_HI_LS, va >> 40);
{
struct si_pm4_state *pm4;
uint64_t va;
+ unsigned ls_vgpr_comp_cnt = 0;
pm4 = si_get_shader_pm4_state(shader);
if (!pm4)
va = shader->bo->gpu_address;
si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
- si_pm4_set_reg(pm4, R_00B420_SPI_SHADER_PGM_LO_HS, va >> 8);
- si_pm4_set_reg(pm4, R_00B424_SPI_SHADER_PGM_HI_HS, va >> 40);
+ if (sscreen->b.chip_class >= GFX9) {
+ si_pm4_set_reg(pm4, R_00B410_SPI_SHADER_PGM_LO_LS, va >> 8);
+ si_pm4_set_reg(pm4, R_00B414_SPI_SHADER_PGM_HI_LS, va >> 40);
+
+ /* We need at least 2 components for LS.
+ * VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
+ * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
+ */
+ ls_vgpr_comp_cnt = shader->info.uses_instanceid ? 2 : 1;
+
+ if (shader->config.scratch_bytes_per_wave) {
+ fprintf(stderr, "HS: scratch buffer unsupported");
+ abort();
+ }
+
+ shader->config.rsrc2 =
+ S_00B42C_USER_SGPR(GFX9_TCS_NUM_USER_SGPR) |
+ S_00B42C_USER_SGPR_MSB(GFX9_TCS_NUM_USER_SGPR >> 5) |
+ S_00B42C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
+ } else {
+ si_pm4_set_reg(pm4, R_00B420_SPI_SHADER_PGM_LO_HS, va >> 8);
+ si_pm4_set_reg(pm4, R_00B424_SPI_SHADER_PGM_HI_HS, va >> 40);
+
+ shader->config.rsrc2 =
+ S_00B42C_USER_SGPR(GFX6_TCS_NUM_USER_SGPR) |
+ S_00B42C_OC_LDS_EN(1) |
+ S_00B42C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0);
+ }
+
si_pm4_set_reg(pm4, R_00B428_SPI_SHADER_PGM_RSRC1_HS,
S_00B428_VGPRS((shader->config.num_vgprs - 1) / 4) |
S_00B428_SGPRS((shader->config.num_sgprs - 1) / 8) |
S_00B428_DX10_CLAMP(1) |
- S_00B428_FLOAT_MODE(shader->config.float_mode));
- si_pm4_set_reg(pm4, R_00B42C_SPI_SHADER_PGM_RSRC2_HS,
- S_00B42C_USER_SGPR(SI_TCS_NUM_USER_SGPR) |
- S_00B42C_OC_LDS_EN(sscreen->b.chip_class <= VI) |
- S_00B42C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
+ S_00B428_FLOAT_MODE(shader->config.float_mode) |
+ S_00B428_LS_VGPR_COMP_CNT(ls_vgpr_comp_cnt));
+
+ if (sscreen->b.chip_class <= VI) {
+ si_pm4_set_reg(pm4, R_00B42C_SPI_SHADER_PGM_RSRC2_HS,
+ shader->config.rsrc2);
+ }
}
static void si_shader_es(struct si_screen *sscreen, struct si_shader *shader)
si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
if (shader->selector->type == PIPE_SHADER_VERTEX) {
- vgpr_comp_cnt = shader->info.uses_instanceid ? 3 : 0;
+ /* VGPR0-3: (VertexID, InstanceID / StepRate0, ...) */
+ vgpr_comp_cnt = shader->info.uses_instanceid ? 1 : 0;
num_user_sgprs = SI_VS_NUM_USER_SGPR;
} else if (shader->selector->type == PIPE_SHADER_TESS_EVAL) {
- vgpr_comp_cnt = 3; /* all components are needed for TES */
+ vgpr_comp_cnt = shader->selector->info.uses_primid ? 3 : 2;
num_user_sgprs = SI_TES_NUM_USER_SGPR;
} else
unreachable("invalid shader selector type");
S_00B32C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
- si_set_tesseval_regs(sscreen, shader, pm4);
+ si_set_tesseval_regs(sscreen, shader->selector, pm4);
- polaris_set_vgt_vertex_reuse(sscreen, shader, pm4);
+ polaris_set_vgt_vertex_reuse(sscreen, shader->selector, shader, pm4);
}
/**
*/
static uint32_t si_vgt_gs_mode(struct si_shader_selector *sel)
{
+ enum chip_class chip_class = sel->screen->b.chip_class;
unsigned gs_max_vert_out = sel->gs_max_out_vertices;
unsigned cut_mode;
return S_028A40_MODE(V_028A40_GS_SCENARIO_G) |
S_028A40_CUT_MODE(cut_mode)|
- S_028A40_ES_WRITE_OPTIMIZE(1) |
- S_028A40_GS_WRITE_OPTIMIZE(1);
+ S_028A40_ES_WRITE_OPTIMIZE(chip_class <= VI) |
+ S_028A40_GS_WRITE_OPTIMIZE(1) |
+ S_028A40_ONCHIP(chip_class >= GFX9 ? 1 : 0);
+}
+
+struct gfx9_gs_info {
+ unsigned es_verts_per_subgroup;
+ unsigned gs_prims_per_subgroup;
+ unsigned gs_inst_prims_in_subgroup;
+ unsigned max_prims_per_subgroup;
+ unsigned lds_size;
+};
+
+static 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 input_prim = gs->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
+ bool uses_adjacency = input_prim >= PIPE_PRIM_LINES_ADJACENCY &&
+ input_prim <= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY;
+
+ /* All these are in dwords: */
+ /* We can't allow using the whole LDS, because GS waves compete with
+ * other shader stages for LDS space. */
+ const unsigned max_lds_size = 8 * 1024;
+ const unsigned esgs_itemsize = es->esgs_itemsize / 4;
+ unsigned esgs_lds_size;
+
+ /* All these are per subgroup: */
+ const unsigned max_out_prims = 32 * 1024;
+ const unsigned max_es_verts = 255;
+ const unsigned ideal_gs_prims = 64;
+ unsigned max_gs_prims, gs_prims;
+ unsigned min_es_verts, es_verts, worst_case_es_verts;
+
+ assert(gs_num_invocations <= 32); /* GL maximum */
+
+ if (uses_adjacency || gs_num_invocations > 1)
+ max_gs_prims = 127 / gs_num_invocations;
+ else
+ max_gs_prims = 255;
+
+ /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
+ * Make sure we don't go over the maximum value.
+ */
+ max_gs_prims = MIN2(max_gs_prims,
+ max_out_prims /
+ (gs->gs_max_out_vertices * gs_num_invocations));
+ assert(max_gs_prims > 0);
+
+ /* If the primitive has adjacency, halve the number of vertices
+ * that will be reused in multiple primitives.
+ */
+ min_es_verts = gs->gs_input_verts_per_prim / (uses_adjacency ? 2 : 1);
+
+ gs_prims = MIN2(ideal_gs_prims, max_gs_prims);
+ worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts);
+
+ /* Compute ESGS LDS size based on the worst case number of ES vertices
+ * needed to create the target number of GS prims per subgroup.
+ */
+ esgs_lds_size = esgs_itemsize * worst_case_es_verts;
+
+ /* If total LDS usage is too big, refactor partitions based on ratio
+ * of ESGS item sizes.
+ */
+ if (esgs_lds_size > max_lds_size) {
+ /* Our target GS Prims Per Subgroup was too large. Calculate
+ * the maximum number of GS Prims Per Subgroup that will fit
+ * into LDS, capped by the maximum that the hardware can support.
+ */
+ gs_prims = MIN2((max_lds_size / (esgs_itemsize * min_es_verts)),
+ max_gs_prims);
+ assert(gs_prims > 0);
+ worst_case_es_verts = MIN2(min_es_verts * gs_prims,
+ max_es_verts);
+
+ esgs_lds_size = esgs_itemsize * worst_case_es_verts;
+ assert(esgs_lds_size <= max_lds_size);
+ }
+
+ /* Now calculate remaining ESGS information. */
+ if (esgs_lds_size)
+ es_verts = MIN2(esgs_lds_size / esgs_itemsize, max_es_verts);
+ else
+ es_verts = max_es_verts;
+
+ /* Vertices for adjacency primitives are not always reused, so restore
+ * it for ES_VERTS_PER_SUBGRP.
+ */
+ min_es_verts = gs->gs_input_verts_per_prim;
+
+ /* For normal primitives, the VGT only checks if they are past the ES
+ * verts per subgroup after allocating a full GS primitive and if they
+ * are, kick off a new subgroup. But if those additional ES verts are
+ * unique (e.g. not reused) we need to make sure there is enough LDS
+ * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
+ */
+ es_verts -= min_es_verts - 1;
+
+ 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->lds_size = align(esgs_lds_size, 128) / 128;
+
+ assert(out->max_prims_per_subgroup <= max_out_prims);
}
-static void si_shader_gs(struct si_shader *shader)
+static void si_shader_gs(struct si_screen *sscreen, struct si_shader *shader)
{
struct si_shader_selector *sel = shader->selector;
const ubyte *num_components = sel->info.num_stream_output_components;
/* The GSVS_RING_ITEMSIZE register takes 15 bits */
assert(offset < (1 << 15));
- si_pm4_set_reg(pm4, R_028B38_VGT_GS_MAX_VERT_OUT, shader->selector->gs_max_out_vertices);
+ si_pm4_set_reg(pm4, R_028B38_VGT_GS_MAX_VERT_OUT, sel->gs_max_out_vertices);
si_pm4_set_reg(pm4, R_028B5C_VGT_GS_VERT_ITEMSIZE, num_components[0]);
si_pm4_set_reg(pm4, R_028B60_VGT_GS_VERT_ITEMSIZE_1, (max_stream >= 1) ? num_components[1] : 0);
va = shader->bo->gpu_address;
si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
- si_pm4_set_reg(pm4, R_00B220_SPI_SHADER_PGM_LO_GS, va >> 8);
- si_pm4_set_reg(pm4, R_00B224_SPI_SHADER_PGM_HI_GS, va >> 40);
- si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
- S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
- S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
- S_00B228_DX10_CLAMP(1) |
- S_00B228_FLOAT_MODE(shader->config.float_mode));
- si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
- S_00B22C_USER_SGPR(SI_GS_NUM_USER_SGPR) |
- S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
+ if (sscreen->b.chip_class >= GFX9) {
+ unsigned input_prim = sel->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
+ unsigned es_type = shader->key.part.gs.es->type;
+ unsigned es_vgpr_comp_cnt, gs_vgpr_comp_cnt;
+ struct gfx9_gs_info gs_info;
+
+ if (es_type == PIPE_SHADER_VERTEX)
+ /* VGPR0-3: (VertexID, InstanceID / StepRate0, ...) */
+ es_vgpr_comp_cnt = shader->info.uses_instanceid ? 1 : 0;
+ else if (es_type == PIPE_SHADER_TESS_EVAL)
+ es_vgpr_comp_cnt = shader->key.part.gs.es->info.uses_primid ? 3 : 2;
+ else
+ unreachable("invalid shader selector type");
+
+ /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
+ * VGPR[0:4] are always loaded.
+ */
+ if (sel->info.uses_invocationid)
+ gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID. */
+ else if (sel->info.uses_primid)
+ gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
+ else if (input_prim >= PIPE_PRIM_TRIANGLES)
+ gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
+ else
+ gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
+
+ gfx9_get_gs_info(shader->key.part.gs.es, sel, &gs_info);
+
+ si_pm4_set_reg(pm4, R_00B210_SPI_SHADER_PGM_LO_ES, va >> 8);
+ si_pm4_set_reg(pm4, R_00B214_SPI_SHADER_PGM_HI_ES, va >> 40);
+
+ si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
+ S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
+ S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
+ S_00B228_DX10_CLAMP(1) |
+ S_00B228_FLOAT_MODE(shader->config.float_mode) |
+ S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt));
+ si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
+ S_00B22C_USER_SGPR(GFX9_GS_NUM_USER_SGPR) |
+ S_00B22C_USER_SGPR_MSB(GFX9_GS_NUM_USER_SGPR >> 5) |
+ S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
+ S_00B22C_OC_LDS_EN(es_type == PIPE_SHADER_TESS_EVAL) |
+ S_00B22C_LDS_SIZE(gs_info.lds_size) |
+ S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
+
+ si_pm4_set_reg(pm4, R_028A44_VGT_GS_ONCHIP_CNTL,
+ S_028A44_ES_VERTS_PER_SUBGRP(gs_info.es_verts_per_subgroup) |
+ S_028A44_GS_PRIMS_PER_SUBGRP(gs_info.gs_prims_per_subgroup) |
+ S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_info.gs_inst_prims_in_subgroup));
+ si_pm4_set_reg(pm4, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP,
+ S_028A94_MAX_PRIMS_PER_SUBGROUP(gs_info.max_prims_per_subgroup));
+ si_pm4_set_reg(pm4, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
+ shader->key.part.gs.es->esgs_itemsize / 4);
+
+ if (es_type == PIPE_SHADER_TESS_EVAL)
+ si_set_tesseval_regs(sscreen, shader->key.part.gs.es, pm4);
+
+ polaris_set_vgt_vertex_reuse(sscreen, shader->key.part.gs.es,
+ NULL, pm4);
+
+ if (shader->config.scratch_bytes_per_wave) {
+ fprintf(stderr, "GS: scratch buffer unsupported");
+ abort();
+ }
+ } else {
+ si_pm4_set_reg(pm4, R_00B220_SPI_SHADER_PGM_LO_GS, va >> 8);
+ si_pm4_set_reg(pm4, R_00B224_SPI_SHADER_PGM_HI_GS, va >> 40);
+
+ si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
+ S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
+ S_00B228_SGPRS((shader->config.num_sgprs - 1) / 8) |
+ S_00B228_DX10_CLAMP(1) |
+ S_00B228_FLOAT_MODE(shader->config.float_mode));
+ si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
+ S_00B22C_USER_SGPR(GFX6_GS_NUM_USER_SGPR) |
+ S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0));
+ }
}
/**
unsigned oc_lds_en;
unsigned window_space =
shader->selector->info.properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
- bool enable_prim_id = si_vs_exports_prim_id(shader);
+ bool enable_prim_id = shader->key.mono.vs_export_prim_id;
pm4 = si_get_shader_pm4_state(shader);
if (!pm4)
* not sent again.
*/
if (!gs) {
- si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE,
- S_028A40_MODE(enable_prim_id ? V_028A40_GS_SCENARIO_A : 0));
+ unsigned mode = 0;
+
+ /* PrimID needs GS scenario A.
+ * GFX9 also needs it when ViewportIndex is enabled.
+ */
+ if (enable_prim_id ||
+ (sscreen->b.chip_class >= GFX9 &&
+ shader->selector->info.writes_viewport_index))
+ mode = V_028A40_GS_SCENARIO_A;
+
+ si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE, S_028A40_MODE(mode));
si_pm4_set_reg(pm4, R_028A84_VGT_PRIMITIVEID_EN, enable_prim_id);
} else {
si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE, si_vgt_gs_mode(gs));
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) {
- vgpr_comp_cnt = shader->info.uses_instanceid ? 3 : (enable_prim_id ? 2 : 0);
+ /* 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);
num_user_sgprs = SI_VS_NUM_USER_SGPR;
} else if (shader->selector->type == PIPE_SHADER_TESS_EVAL) {
- vgpr_comp_cnt = 3; /* all components are needed for TES */
+ vgpr_comp_cnt = shader->selector->info.uses_primid ? 3 : 2;
num_user_sgprs = SI_TES_NUM_USER_SGPR;
} else
unreachable("invalid shader selector type");
S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
- si_set_tesseval_regs(sscreen, shader, pm4);
+ si_set_tesseval_regs(sscreen, shader->selector, pm4);
- polaris_set_vgt_vertex_reuse(sscreen, shader, pm4);
+ polaris_set_vgt_vertex_reuse(sscreen, shader->selector, shader, pm4);
}
static unsigned si_get_ps_num_interp(struct si_shader *ps)
si_shader_vs(sscreen, shader, NULL);
break;
case PIPE_SHADER_GEOMETRY:
- si_shader_gs(shader);
+ si_shader_gs(sscreen, shader);
break;
case PIPE_SHADER_FRAGMENT:
si_shader_ps(shader);
si_shader_selector_key_hw_vs(sctx, sel, key);
if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
- key->part.vs.epilog.export_prim_id = 1;
+ key->mono.vs_export_prim_id = 1;
}
break;
case PIPE_SHADER_TESS_CTRL:
si_shader_selector_key_hw_vs(sctx, sel, key);
if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
- key->part.tes.epilog.export_prim_id = 1;
+ key->mono.vs_export_prim_id = 1;
}
break;
case PIPE_SHADER_GEOMETRY:
+ if (sctx->b.chip_class >= GFX9) {
+ if (sctx->tes_shader.cso) {
+ key->part.gs.es = sctx->tes_shader.cso;
+ } else {
+ si_shader_selector_key_vs(sctx, sctx->vs_shader.cso,
+ key, &key->part.gs.vs_prolog);
+ key->part.gs.es = sctx->vs_shader.cso;
+ }
+
+ /* Merged ES-GS can have unbalanced wave usage.
+ *
+ * ES threads are per-vertex, while GS threads are
+ * per-primitive. So without any amplification, there
+ * are fewer GS threads than ES threads, which can result
+ * in empty (no-op) GS waves. With too much amplification,
+ * there are more GS threads than ES threads, which
+ * can result in empty (no-op) ES waves.
+ *
+ * Non-monolithic shaders are implemented by setting EXEC
+ * at the beginning of shader parts, and don't jump to
+ * the end if EXEC is 0.
+ *
+ * Monolithic shaders use conditional blocks, so they can
+ * jump and skip empty waves of ES or GS. So set this to
+ * always use optimized variants, which are monolithic.
+ */
+ key->opt.prefer_mono = 1;
+ }
key->part.gs.prolog.tri_strip_adj_fix = sctx->gs_tri_strip_adj_fix;
break;
case PIPE_SHADER_FRAGMENT: {
si_shader_init_pm4_state(sscreen, shader);
}
+static const struct si_shader_key zeroed;
+
+static bool si_check_missing_main_part(struct si_screen *sscreen,
+ struct si_shader_selector *sel,
+ struct si_compiler_ctx_state *compiler_state,
+ struct si_shader_key *key)
+{
+ struct si_shader **mainp = si_get_main_shader_part(sel, key);
+
+ if (!*mainp) {
+ struct si_shader *main_part = CALLOC_STRUCT(si_shader);
+
+ if (!main_part)
+ return false;
+
+ main_part->selector = sel;
+ main_part->key.as_es = key->as_es;
+ main_part->key.as_ls = key->as_ls;
+
+ if (si_compile_tgsi_shader(sscreen, compiler_state->tm,
+ main_part, false,
+ &compiler_state->debug) != 0) {
+ FREE(main_part);
+ return false;
+ }
+ *mainp = main_part;
+ }
+ return true;
+}
+
+static void si_destroy_shader_selector(struct si_context *sctx,
+ struct si_shader_selector *sel);
+
+static void si_shader_selector_reference(struct si_context *sctx,
+ struct si_shader_selector **dst,
+ struct si_shader_selector *src)
+{
+ if (pipe_reference(&(*dst)->reference, &src->reference))
+ si_destroy_shader_selector(sctx, *dst);
+
+ *dst = src;
+}
+
/* Select the hw shader variant depending on the current state. */
static int si_shader_select_with_key(struct si_screen *sscreen,
struct si_shader_ctx_state *state,
struct si_shader_key *key,
int thread_index)
{
- static const struct si_shader_key zeroed;
struct si_shader_selector *sel = state->cso;
+ struct si_shader_selector *previous_stage_sel = NULL;
struct si_shader *current = state->current;
struct si_shader *iter, *shader = NULL;
shader->key = *key;
shader->compiler_ctx_state = *compiler_state;
+ /* If this is a merged shader, get the first shader's selector. */
+ if (sscreen->b.chip_class >= GFX9) {
+ if (sel->type == PIPE_SHADER_TESS_CTRL)
+ previous_stage_sel = key->part.tcs.ls;
+ else if (sel->type == PIPE_SHADER_GEOMETRY)
+ previous_stage_sel = key->part.gs.es;
+ }
+
/* Compile the main shader part if it doesn't exist. This can happen
* if the initial guess was wrong. */
- struct si_shader **mainp = si_get_main_shader_part(sel, key);
bool is_pure_monolithic =
sscreen->use_monolithic_shaders ||
memcmp(&key->mono, &zeroed.mono, sizeof(key->mono)) != 0;
- if (!*mainp && !is_pure_monolithic) {
- struct si_shader *main_part = CALLOC_STRUCT(si_shader);
+ if (!is_pure_monolithic) {
+ bool ok;
- if (!main_part) {
- FREE(shader);
- mtx_unlock(&sel->mutex);
- return -ENOMEM; /* skip the draw call */
- }
+ /* Make sure the main shader part is present. This is needed
+ * for shaders that can be compiled as VS, LS, or ES, and only
+ * one of them is compiled at creation.
+ *
+ * For merged shaders, check that the starting shader's main
+ * part is present.
+ */
+ if (previous_stage_sel) {
+ struct si_shader_key shader1_key = zeroed;
- main_part->selector = sel;
- main_part->key.as_es = key->as_es;
- main_part->key.as_ls = key->as_ls;
+ if (sel->type == PIPE_SHADER_TESS_CTRL)
+ shader1_key.as_ls = 1;
+ else if (sel->type == PIPE_SHADER_GEOMETRY)
+ shader1_key.as_es = 1;
+ else
+ assert(0);
- if (si_compile_tgsi_shader(sscreen, compiler_state->tm,
- main_part, false,
- &compiler_state->debug) != 0) {
- FREE(main_part);
+ ok = si_check_missing_main_part(sscreen,
+ previous_stage_sel,
+ compiler_state, &shader1_key);
+ } else {
+ ok = si_check_missing_main_part(sscreen, sel,
+ compiler_state, key);
+ }
+ if (!ok) {
FREE(shader);
mtx_unlock(&sel->mutex);
return -ENOMEM; /* skip the draw call */
}
- *mainp = main_part;
}
+ /* Keep the reference to the 1st shader of merged shaders, so that
+ * Gallium can't destroy it before we destroy the 2nd shader.
+ *
+ * Set sctx = NULL, because it's unused if we're not releasing
+ * the shader, and we don't have any sctx here.
+ */
+ si_shader_selector_reference(NULL, &shader->previous_stage_sel,
+ previous_stage_sel);
+
/* Monolithic-only shaders don't make a distinction between optimized
* and unoptimized. */
shader->is_monolithic =
if (!sel)
return NULL;
+ pipe_reference_init(&sel->reference, 1);
sel->screen = sscreen;
sel->compiler_ctx_state.tm = sctx->tm;
sel->compiler_ctx_state.debug = sctx->b.debug;
}
}
sel->esgs_itemsize = util_last_bit64(sel->outputs_written) * 16;
+
+ /* For the ESGS ring in LDS, add 1 dword to reduce LDS bank
+ * conflicts, i.e. each vertex will start at a different bank.
+ */
+ if (sctx->b.chip_class >= GFX9)
+ sel->esgs_itemsize += 4;
break;
case PIPE_SHADER_FRAGMENT:
}
}
+ si_shader_selector_reference(sctx, &shader->previous_stage_sel, NULL);
si_shader_destroy(shader);
free(shader);
}
-static void si_delete_shader_selector(struct pipe_context *ctx, void *state)
+static void si_destroy_shader_selector(struct si_context *sctx,
+ struct si_shader_selector *sel)
{
- struct si_context *sctx = (struct si_context *)ctx;
- struct si_shader_selector *sel = (struct si_shader_selector *)state;
struct si_shader *p = sel->first_variant, *c;
struct si_shader_ctx_state *current_shader[SI_NUM_SHADERS] = {
[PIPE_SHADER_VERTEX] = &sctx->vs_shader,
free(sel);
}
+static void si_delete_shader_selector(struct pipe_context *ctx, void *state)
+{
+ struct si_context *sctx = (struct si_context *)ctx;
+ struct si_shader_selector *sel = (struct si_shader_selector *)state;
+
+ si_shader_selector_reference(sctx, &sel, NULL);
+}
+
static unsigned si_get_ps_input_cntl(struct si_context *sctx,
struct si_shader *vs, unsigned name,
unsigned index, unsigned interpolate)
assert(sctx->scratch_buffer);
- si_shader_apply_scratch_relocs(sctx, shader, &shader->config, scratch_va);
+ si_shader_apply_scratch_relocs(shader, scratch_va);
/* Replace the shader bo with a new bo that has the relocs applied. */
r = si_shader_binary_upload(sctx->screen, shader);
}
assert(!sctx->tf_ring);
+ /* Use 64K alignment for both rings, so that we can pass the address
+ * to shaders as one SGPR containing bits [16:47].
+ */
sctx->tf_ring = r600_aligned_buffer_create(sctx->b.b.screen,
R600_RESOURCE_FLAG_UNMAPPABLE,
PIPE_USAGE_DEFAULT,
32768 * sctx->screen->b.info.max_se,
- 256);
+ 64 * 1024);
if (!sctx->tf_ring)
return;
PIPE_USAGE_DEFAULT,
max_offchip_buffers *
sctx->screen->tess_offchip_block_dw_size * 4,
- 256);
+ 64 * 1024);
if (!sctx->tess_offchip_ring)
return;
si_init_config_add_vgt_flush(sctx);
+ uint64_t offchip_va = r600_resource(sctx->tess_offchip_ring)->gpu_address;
+ uint64_t factor_va = r600_resource(sctx->tf_ring)->gpu_address;
+ assert((offchip_va & 0xffff) == 0);
+ assert((factor_va & 0xffff) == 0);
+
+ si_pm4_add_bo(sctx->init_config, r600_resource(sctx->tess_offchip_ring),
+ RADEON_USAGE_READWRITE, RADEON_PRIO_SHADER_RINGS);
+ si_pm4_add_bo(sctx->init_config, r600_resource(sctx->tf_ring),
+ RADEON_USAGE_READWRITE, RADEON_PRIO_SHADER_RINGS);
+
/* Append these registers to the init config state. */
if (sctx->b.chip_class >= CIK) {
if (sctx->b.chip_class >= VI)
si_pm4_set_reg(sctx->init_config, R_030938_VGT_TF_RING_SIZE,
S_030938_SIZE(sctx->tf_ring->width0 / 4));
si_pm4_set_reg(sctx->init_config, R_030940_VGT_TF_MEMORY_BASE,
- r600_resource(sctx->tf_ring)->gpu_address >> 8);
+ factor_va >> 8);
if (sctx->b.chip_class >= GFX9)
si_pm4_set_reg(sctx->init_config, R_030944_VGT_TF_MEMORY_BASE_HI,
- r600_resource(sctx->tf_ring)->gpu_address >> 40);
+ factor_va >> 40);
si_pm4_set_reg(sctx->init_config, R_03093C_VGT_HS_OFFCHIP_PARAM,
S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers) |
S_03093C_OFFCHIP_GRANULARITY(offchip_granularity));
si_pm4_set_reg(sctx->init_config, R_008988_VGT_TF_RING_SIZE,
S_008988_SIZE(sctx->tf_ring->width0 / 4));
si_pm4_set_reg(sctx->init_config, R_0089B8_VGT_TF_MEMORY_BASE,
- r600_resource(sctx->tf_ring)->gpu_address >> 8);
+ factor_va >> 8);
si_pm4_set_reg(sctx->init_config, R_0089B0_VGT_HS_OFFCHIP_PARAM,
S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers));
}
+ if (sctx->b.chip_class >= GFX9) {
+ si_pm4_set_reg(sctx->init_config,
+ R_00B430_SPI_SHADER_USER_DATA_LS_0 +
+ GFX9_SGPR_TCS_OFFCHIP_ADDR_BASE64K * 4,
+ offchip_va >> 16);
+ si_pm4_set_reg(sctx->init_config,
+ R_00B430_SPI_SHADER_USER_DATA_LS_0 +
+ GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K * 4,
+ factor_va >> 16);
+ } else {
+ si_pm4_set_reg(sctx->init_config,
+ R_00B430_SPI_SHADER_USER_DATA_HS_0 +
+ GFX6_SGPR_TCS_OFFCHIP_ADDR_BASE64K * 4,
+ offchip_va >> 16);
+ si_pm4_set_reg(sctx->init_config,
+ R_00B430_SPI_SHADER_USER_DATA_HS_0 +
+ GFX6_SGPR_TCS_FACTOR_ADDR_BASE64K * 4,
+ factor_va >> 16);
+ }
+
/* Flush the context to re-emit the init_config state.
* This is done only once in a lifetime of a context.
*/
si_pm4_upload_indirect_buffer(sctx, sctx->init_config);
sctx->b.initial_gfx_cs_size = 0; /* force flush */
si_context_gfx_flush(sctx, RADEON_FLUSH_ASYNC, NULL);
-
- si_set_ring_buffer(&sctx->b.b, SI_HS_RING_TESS_FACTOR, sctx->tf_ring,
- 0, sctx->tf_ring->width0, false, false, 0, 0, 0);
-
- si_set_ring_buffer(&sctx->b.b, SI_HS_RING_TESS_OFFCHIP,
- sctx->tess_offchip_ring, 0,
- sctx->tess_offchip_ring->width0, false, false, 0, 0, 0);
}
/**