/* command buffer handling for SI */
#include "radv_private.h"
+#include "radv_shader.h"
#include "radv_cs.h"
#include "sid.h"
#include "gfx9d.h"
S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE));
if (physical_device->rad_info.chip_class >= GFX9) {
- radeon_set_context_reg(cs, R_030920_VGT_MAX_VTX_INDX, ~0);
- radeon_set_context_reg(cs, R_030924_VGT_MIN_VTX_INDX, 0);
- radeon_set_context_reg(cs, R_030928_VGT_INDX_OFFSET, 0);
+ radeon_set_uconfig_reg(cs, R_030920_VGT_MAX_VTX_INDX, ~0);
+ radeon_set_uconfig_reg(cs, R_030924_VGT_MIN_VTX_INDX, 0);
+ radeon_set_uconfig_reg(cs, R_030928_VGT_INDX_OFFSET, 0);
} else {
radeon_set_context_reg(cs, R_028400_VGT_MAX_VTX_INDX, ~0);
radeon_set_context_reg(cs, R_028404_VGT_MIN_VTX_INDX, 0);
radeon_set_context_reg(cs, R_028C5C_VGT_OUT_DEALLOC_CNTL, 16);
}
- if (physical_device->rad_info.family == CHIP_STONEY)
+ if (physical_device->has_rbplus)
radeon_set_context_reg(cs, R_028C40_PA_SC_SHADER_CONTROL, 0);
if (physical_device->rad_info.chip_class >= GFX9) {
S_028C48_MAX_PRIM_PER_BATCH(1023));
radeon_set_context_reg(cs, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
- radeon_set_context_reg(cs, R_030968_VGT_INSTANCE_BASE_ID, 0);
+ radeon_set_uconfig_reg(cs, R_030968_VGT_INSTANCE_BASE_ID, 0);
}
si_emit_compute(physical_device, cs);
}
enum radeon_family family = cmd_buffer->device->physical_device->rad_info.family;
struct radeon_info *info = &cmd_buffer->device->physical_device->rad_info;
unsigned prim = cmd_buffer->state.pipeline->graphics.prim;
- unsigned primgroup_size = 128; /* recommended without a GS */
unsigned max_primgroup_in_wave = 2;
/* SWITCH_ON_EOP(0) is always preferable. */
bool wd_switch_on_eop = false;
bool ia_switch_on_eoi = false;
bool partial_vs_wave = false;
bool partial_es_wave = false;
- uint32_t num_prims = radv_prims_for_vertices(&cmd_buffer->state.pipeline->graphics.prim_vertex_count, draw_vertex_count);
bool multi_instances_smaller_than_primgroup;
- if (radv_pipeline_has_tess(cmd_buffer->state.pipeline))
- primgroup_size = cmd_buffer->state.pipeline->graphics.tess.num_patches;
- else if (radv_pipeline_has_gs(cmd_buffer->state.pipeline))
- primgroup_size = 64; /* recommended with a GS */
+ multi_instances_smaller_than_primgroup = indirect_draw;
+ if (!multi_instances_smaller_than_primgroup && instanced_draw) {
+ uint32_t num_prims = radv_prims_for_vertices(&cmd_buffer->state.pipeline->graphics.prim_vertex_count, draw_vertex_count);
+ if (num_prims < cmd_buffer->state.pipeline->graphics.primgroup_size)
+ multi_instances_smaller_than_primgroup = true;
+ }
+
+ if (cmd_buffer->state.pipeline->shaders[MESA_SHADER_FRAGMENT]->info.fs.prim_id_input)
+ ia_switch_on_eoi = true;
- multi_instances_smaller_than_primgroup = indirect_draw || (instanced_draw &&
- num_prims < primgroup_size);
if (radv_pipeline_has_tess(cmd_buffer->state.pipeline)) {
/* SWITCH_ON_EOI must be set if PrimID is used. */
if (cmd_buffer->state.pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.uses_prim_id ||
/* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
if (cmd_buffer->device->has_distributed_tess) {
if (radv_pipeline_has_gs(cmd_buffer->state.pipeline)) {
- partial_es_wave = true;
+ if (chip_class <= VI)
+ partial_es_wave = true;
if (family == CHIP_TONGA ||
family == CHIP_FIJI ||
assert(wd_switch_on_eop || !ia_switch_on_eop);
}
/* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
- if (ia_switch_on_eoi)
+ if (chip_class <= VI && ia_switch_on_eoi)
partial_es_wave = true;
if (radv_pipeline_has_gs(cmd_buffer->state.pipeline)) {
+
+ if (radv_pipeline_has_gs(cmd_buffer->state.pipeline) &&
+ cmd_buffer->state.pipeline->shaders[MESA_SHADER_GEOMETRY]->info.gs.uses_prim_id)
+ ia_switch_on_eoi = true;
+
/* GS requirement. */
- if (SI_GS_PER_ES / primgroup_size >= cmd_buffer->device->gs_table_depth - 3)
+ if (SI_GS_PER_ES / cmd_buffer->state.pipeline->graphics.primgroup_size >= cmd_buffer->device->gs_table_depth - 3)
partial_es_wave = true;
- /* Hw bug with single-primitive instances and SWITCH_ON_EOI
- * on multi-SE chips. */
- if (info->max_se >= 2 && ia_switch_on_eoi &&
- ((instanced_draw || indirect_draw) &&
- num_prims <= 1))
- cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_VGT_FLUSH;
+ /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
+ * The hw doc says all multi-SE chips are affected, but amdgpu-pro Vulkan
+ * only applies it to Hawaii. Do what amdgpu-pro Vulkan does.
+ */
+ if (family == CHIP_HAWAII && ia_switch_on_eoi) {
+ bool set_vgt_flush = indirect_draw;
+ if (!set_vgt_flush && instanced_draw) {
+ uint32_t num_prims = radv_prims_for_vertices(&cmd_buffer->state.pipeline->graphics.prim_vertex_count, draw_vertex_count);
+ if (num_prims <= 1)
+ set_vgt_flush = true;
+ }
+ if (set_vgt_flush)
+ cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_VGT_FLUSH;
+ }
}
return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop) |
S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi) |
S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave) |
S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave) |
- S_028AA8_PRIMGROUP_SIZE(primgroup_size - 1) |
+ S_028AA8_PRIMGROUP_SIZE(cmd_buffer->state.pipeline->graphics.primgroup_size - 1) |
S_028AA8_WD_SWITCH_ON_EOP(chip_class >= CIK ? wd_switch_on_eop : 0) |
- S_028AA8_MAX_PRIMGRP_IN_WAVE(chip_class >= VI ?
- max_primgroup_in_wave : 0);
+ /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
+ S_028AA8_MAX_PRIMGRP_IN_WAVE(chip_class == VI ?
+ max_primgroup_in_wave : 0) |
+ S_030960_EN_INST_OPT_BASIC(chip_class >= GFX9) |
+ S_030960_EN_INST_OPT_ADV(chip_class >= GFX9);
}
void si_cs_emit_write_event_eop(struct radeon_winsys_cs *cs,
+ bool predicated,
enum chip_class chip_class,
bool is_mec,
unsigned event, unsigned event_flags,
unsigned is_gfx8_mec = is_mec && chip_class < GFX9;
if (chip_class >= GFX9 || is_gfx8_mec) {
- radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, is_gfx8_mec ? 5 : 6, 0));
+ radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, is_gfx8_mec ? 5 : 6, predicated));
radeon_emit(cs, op);
radeon_emit(cs, EOP_DATA_SEL(data_sel));
radeon_emit(cs, va); /* address lo */
* (and optional cache flushes executed) before the timestamp
* is written.
*/
- radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, predicated));
radeon_emit(cs, op);
radeon_emit(cs, va);
radeon_emit(cs, ((va >> 32) & 0xffff) | EOP_DATA_SEL(data_sel));
radeon_emit(cs, 0); /* unused */
}
- radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, predicated));
radeon_emit(cs, op);
radeon_emit(cs, va);
radeon_emit(cs, ((va >> 32) & 0xffff) | EOP_DATA_SEL(data_sel));
void
si_emit_wait_fence(struct radeon_winsys_cs *cs,
+ bool predicated,
uint64_t va, uint32_t ref,
uint32_t mask)
{
- radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, 0));
+ radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, predicated));
radeon_emit(cs, WAIT_REG_MEM_EQUAL | WAIT_REG_MEM_MEM_SPACE(1));
radeon_emit(cs, va);
radeon_emit(cs, va >> 32);
static void
si_emit_acquire_mem(struct radeon_winsys_cs *cs,
- bool is_mec, bool is_gfx9,
+ bool is_mec,
+ bool predicated,
+ bool is_gfx9,
unsigned cp_coher_cntl)
{
if (is_mec || is_gfx9) {
uint32_t hi_val = is_gfx9 ? 0xffffff : 0xff;
- radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, 0) |
+ radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, predicated) |
PKT3_SHADER_TYPE_S(is_mec));
radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
radeon_emit(cs, 0x0000000A); /* POLL_INTERVAL */
} else {
/* ACQUIRE_MEM is only required on a compute ring. */
- radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, 0));
+ radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, predicated));
radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
radeon_emit(cs, 0); /* CP_COHER_BASE */
void
si_cs_emit_cache_flush(struct radeon_winsys_cs *cs,
+ bool predicated,
enum chip_class chip_class,
uint32_t *flush_cnt,
uint64_t flush_va,
/* Necessary for DCC */
if (chip_class >= VI) {
si_cs_emit_write_event_eop(cs,
+ predicated,
chip_class,
is_mec,
V_028A90_FLUSH_AND_INV_CB_DATA_TS,
}
if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB_META) {
- radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, predicated));
radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
}
if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB_META) {
- radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, predicated));
radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
}
if (!flush_cb_db) {
if (flush_bits & RADV_CMD_FLAG_PS_PARTIAL_FLUSH) {
- radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, predicated));
radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
} else if (flush_bits & RADV_CMD_FLAG_VS_PARTIAL_FLUSH) {
- radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, predicated));
radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
}
}
if (flush_bits & RADV_CMD_FLAG_CS_PARTIAL_FLUSH) {
- radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, predicated));
radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
}
assert(flush_cnt);
uint32_t old_fence = (*flush_cnt)++;
- si_cs_emit_write_event_eop(cs, chip_class, false, cb_db_event, tc_flags, 1,
+ si_cs_emit_write_event_eop(cs, predicated, chip_class, false, cb_db_event, tc_flags, 1,
flush_va, old_fence, *flush_cnt);
- si_emit_wait_fence(cs, flush_va, *flush_cnt, 0xffffffff);
+ si_emit_wait_fence(cs, predicated, flush_va, *flush_cnt, 0xffffffff);
}
/* VGT state sync */
if (flush_bits & RADV_CMD_FLAG_VGT_FLUSH) {
- radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, predicated));
radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
}
RADV_CMD_FLAG_INV_GLOBAL_L2 |
RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2))) &&
!is_mec) {
- radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
+ radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, predicated));
radeon_emit(cs, 0);
}
if ((flush_bits & RADV_CMD_FLAG_INV_GLOBAL_L2) ||
(chip_class <= CIK && (flush_bits & RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2))) {
- si_emit_acquire_mem(cs, is_mec, chip_class >= GFX9,
+ si_emit_acquire_mem(cs, is_mec, predicated, chip_class >= GFX9,
cp_coher_cntl |
S_0085F0_TC_ACTION_ENA(1) |
S_0085F0_TCL1_ACTION_ENA(1) |
*
* WB doesn't work without NC.
*/
- si_emit_acquire_mem(cs, is_mec, chip_class >= GFX9,
+ si_emit_acquire_mem(cs, is_mec, predicated,
+ chip_class >= GFX9,
cp_coher_cntl |
S_0301F0_TC_WB_ACTION_ENA(1) |
S_0301F0_TC_NC_ACTION_ENA(1));
cp_coher_cntl = 0;
}
if (flush_bits & RADV_CMD_FLAG_INV_VMEM_L1) {
- si_emit_acquire_mem(cs, is_mec, chip_class >= GFX9,
+ si_emit_acquire_mem(cs, is_mec,
+ predicated, chip_class >= GFX9,
cp_coher_cntl |
S_0085F0_TCL1_ACTION_ENA(1));
cp_coher_cntl = 0;
* Therefore, it should be last. Done in PFP.
*/
if (cp_coher_cntl)
- si_emit_acquire_mem(cs, is_mec, chip_class >= GFX9, cp_coher_cntl);
+ si_emit_acquire_mem(cs, is_mec, predicated, chip_class >= GFX9, cp_coher_cntl);
}
void
si_emit_cache_flush(struct radv_cmd_buffer *cmd_buffer)
{
bool is_compute = cmd_buffer->queue_family_index == RADV_QUEUE_COMPUTE;
- enum chip_class chip_class = cmd_buffer->device->physical_device->rad_info.chip_class;
+
if (is_compute)
cmd_buffer->state.flush_bits &= ~(RADV_CMD_FLAG_FLUSH_AND_INV_CB |
RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
RADV_CMD_FLAG_VS_PARTIAL_FLUSH |
RADV_CMD_FLAG_VGT_FLUSH);
+ if (!cmd_buffer->state.flush_bits)
+ return;
+
+ enum chip_class chip_class = cmd_buffer->device->physical_device->rad_info.chip_class;
radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 128);
uint32_t *ptr = NULL;
ptr = &cmd_buffer->gfx9_fence_idx;
}
si_cs_emit_cache_flush(cmd_buffer->cs,
+ cmd_buffer->state.predicating,
cmd_buffer->device->physical_device->rad_info.chip_class,
ptr, va,
radv_cmd_buffer_uses_mec(cmd_buffer),
cmd_buffer->state.flush_bits);
- if (cmd_buffer->state.flush_bits)
- radv_cmd_buffer_trace_emit(cmd_buffer);
+ radv_cmd_buffer_trace_emit(cmd_buffer);
cmd_buffer->state.flush_bits = 0;
}
+/* sets the CP predication state using a boolean stored at va */
+void
+si_emit_set_predication_state(struct radv_cmd_buffer *cmd_buffer, uint64_t va)
+{
+ uint32_t op = 0;
+
+ if (va)
+ op = PRED_OP(PREDICATION_OP_BOOL64) | PREDICATION_DRAW_VISIBLE;
+ if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9) {
+ radeon_emit(cmd_buffer->cs, PKT3(PKT3_SET_PREDICATION, 2, 0));
+ radeon_emit(cmd_buffer->cs, op);
+ radeon_emit(cmd_buffer->cs, va);
+ radeon_emit(cmd_buffer->cs, va >> 32);
+ } else {
+ radeon_emit(cmd_buffer->cs, PKT3(PKT3_SET_PREDICATION, 1, 0));
+ radeon_emit(cmd_buffer->cs, va);
+ radeon_emit(cmd_buffer->cs, op | ((va >> 32) & 0xFF));
+ }
+}
/* Set this if you want the 3D engine to wait until CP DMA is done.
* It should be set on the last CP DMA packet. */
header |= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2);
if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK) {
- radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, 0));
+ radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, cmd_buffer->state.predicating));
radeon_emit(cs, header);
radeon_emit(cs, src_va); /* SRC_ADDR_LO [31:0] */
radeon_emit(cs, src_va >> 32); /* SRC_ADDR_HI [31:0] */
radeon_emit(cs, dst_va >> 32); /* DST_ADDR_HI [31:0] */
radeon_emit(cs, command);
} else {
+ assert(!(flags & CP_DMA_USE_L2));
header |= S_411_SRC_ADDR_HI(src_va >> 32);
- radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, 0));
+ radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, cmd_buffer->state.predicating));
radeon_emit(cs, src_va); /* SRC_ADDR_LO [31:0] */
radeon_emit(cs, header); /* SRC_ADDR_HI [15:0] + flags. */
radeon_emit(cs, dst_va); /* DST_ADDR_LO [31:0] */
* should precede it.
*/
if ((flags & CP_DMA_SYNC) && cmd_buffer->queue_family_index == RADV_QUEUE_GENERAL) {
- radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
+ radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, cmd_buffer->state.predicating));
radeon_emit(cs, 0);
}
projects / mesa.git / blobdiff