/* command buffer handling for SI */
#include "radv_private.h"
+#include "radv_shader.h"
#include "radv_cs.h"
#include "sid.h"
+#include "gfx9d.h"
#include "radv_util.h"
#include "main/macros.h"
-#define SI_GS_PER_ES 128
-
static void
si_write_harvested_raster_configs(struct radv_physical_device *physical_device,
struct radeon_winsys_cs *cs,
/* GRBM_GFX_INDEX has a different offset on SI and CI+ */
if (physical_device->rad_info.chip_class < CIK)
- radeon_set_config_reg(cs, GRBM_GFX_INDEX,
- SE_INDEX(se) | SH_BROADCAST_WRITES |
- INSTANCE_BROADCAST_WRITES);
+ radeon_set_config_reg(cs, R_00802C_GRBM_GFX_INDEX,
+ S_00802C_SE_INDEX(se) |
+ S_00802C_SH_BROADCAST_WRITES(1) |
+ S_00802C_INSTANCE_BROADCAST_WRITES(1));
else
radeon_set_uconfig_reg(cs, R_030800_GRBM_GFX_INDEX,
S_030800_SE_INDEX(se) | S_030800_SH_BROADCAST_WRITES(1) |
/* GRBM_GFX_INDEX has a different offset on SI and CI+ */
if (physical_device->rad_info.chip_class < CIK)
- radeon_set_config_reg(cs, GRBM_GFX_INDEX,
- SE_BROADCAST_WRITES | SH_BROADCAST_WRITES |
- INSTANCE_BROADCAST_WRITES);
+ radeon_set_config_reg(cs, R_00802C_GRBM_GFX_INDEX,
+ S_00802C_SE_BROADCAST_WRITES(1) |
+ S_00802C_SH_BROADCAST_WRITES(1) |
+ S_00802C_INSTANCE_BROADCAST_WRITES(1));
else
radeon_set_uconfig_reg(cs, R_030800_GRBM_GFX_INDEX,
S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
radeon_emit(cs, 0);
radeon_emit(cs, 0);
- radeon_set_sh_reg_seq(cs, R_00B854_COMPUTE_RESOURCE_LIMITS, 3);
+ radeon_set_sh_reg_seq(cs, R_00B854_COMPUTE_RESOURCE_LIMITS,
+ S_00B854_WAVES_PER_SH(0x3));
radeon_emit(cs, 0);
/* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1 */
radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
si_emit_compute(physical_device, cmd_buffer->cs);
}
+/* 12.4 fixed-point */
+static unsigned radv_pack_float_12p4(float x)
+{
+ return x <= 0 ? 0 :
+ x >= 4096 ? 0xffff : x * 16;
+}
+
static void
-si_emit_config(struct radv_physical_device *physical_device,
- struct radeon_winsys_cs *cs)
+si_set_raster_config(struct radv_physical_device *physical_device,
+ struct radeon_winsys_cs *cs)
{
unsigned num_rb = MIN2(physical_device->rad_info.num_render_backends, 16);
unsigned rb_mask = physical_device->rad_info.enabled_rb_mask;
unsigned raster_config, raster_config_1;
- int i;
-
- radeon_emit(cs, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
- radeon_emit(cs, CONTEXT_CONTROL_LOAD_ENABLE(1));
- radeon_emit(cs, CONTEXT_CONTROL_SHADOW_ENABLE(1));
-
- radeon_set_context_reg(cs, R_028A18_VGT_HOS_MAX_TESS_LEVEL, fui(64));
- radeon_set_context_reg(cs, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
-
- /* FIXME calculate these values somehow ??? */
- radeon_set_context_reg(cs, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
- radeon_set_context_reg(cs, R_028A58_VGT_ES_PER_GS, 0x40);
- radeon_set_context_reg(cs, R_028A5C_VGT_GS_PER_VS, 0x2);
-
- radeon_set_context_reg(cs, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
- radeon_set_context_reg(cs, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
-
- radeon_set_context_reg(cs, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
- radeon_set_context_reg(cs, R_028AB8_VGT_VTX_CNT_EN, 0x0);
- if (physical_device->rad_info.chip_class < CIK)
- radeon_set_config_reg(cs, R_008A14_PA_CL_ENHANCE, S_008A14_NUM_CLIP_SEQ(3) |
- S_008A14_CLIP_VTX_REORDER_ENA(1));
-
- radeon_set_context_reg(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 0x76543210);
- radeon_set_context_reg(cs, R_028BD8_PA_SC_CENTROID_PRIORITY_1, 0xfedcba98);
-
- radeon_set_context_reg(cs, R_02882C_PA_SU_PRIM_FILTER_CNTL, 0);
-
- for (i = 0; i < 16; i++) {
- radeon_set_context_reg(cs, R_0282D0_PA_SC_VPORT_ZMIN_0 + i*8, 0);
- radeon_set_context_reg(cs, R_0282D4_PA_SC_VPORT_ZMAX_0 + i*8, fui(1.0));
- }
switch (physical_device->rad_info.family) {
case CHIP_TAHITI:
break;
default:
fprintf(stderr,
- "radeonsi: Unknown GPU, using 0 for raster_config\n");
+ "radv: Unknown GPU, using 0 for raster_config\n");
raster_config = 0x00000000;
raster_config_1 = 0x00000000;
break;
radeon_set_context_reg(cs, R_028354_PA_SC_RASTER_CONFIG_1,
raster_config_1);
} else {
- si_write_harvested_raster_configs(physical_device, cs, raster_config, raster_config_1);
+ si_write_harvested_raster_configs(physical_device, cs,
+ raster_config,
+ raster_config_1);
+ }
+}
+
+static void
+si_emit_config(struct radv_physical_device *physical_device,
+ struct radeon_winsys_cs *cs)
+{
+ int i;
+
+ /* Only SI can disable CLEAR_STATE for now. */
+ assert(physical_device->has_clear_state ||
+ physical_device->rad_info.chip_class == SI);
+
+ radeon_emit(cs, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
+ radeon_emit(cs, CONTEXT_CONTROL_LOAD_ENABLE(1));
+ radeon_emit(cs, CONTEXT_CONTROL_SHADOW_ENABLE(1));
+
+ if (physical_device->has_clear_state) {
+ radeon_emit(cs, PKT3(PKT3_CLEAR_STATE, 0, 0));
+ radeon_emit(cs, 0);
+ }
+
+ if (physical_device->rad_info.chip_class <= VI)
+ si_set_raster_config(physical_device, cs);
+
+ radeon_set_context_reg(cs, R_028A18_VGT_HOS_MAX_TESS_LEVEL, fui(64));
+ if (!physical_device->has_clear_state)
+ radeon_set_context_reg(cs, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
+
+ /* FIXME calculate these values somehow ??? */
+ if (physical_device->rad_info.chip_class <= VI) {
+ radeon_set_context_reg(cs, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
+ radeon_set_context_reg(cs, R_028A58_VGT_ES_PER_GS, 0x40);
+ }
+
+ if (!physical_device->has_clear_state) {
+ radeon_set_context_reg(cs, R_028A5C_VGT_GS_PER_VS, 0x2);
+ radeon_set_context_reg(cs, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
+ radeon_set_context_reg(cs, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
+ }
+
+ radeon_set_context_reg(cs, R_028AA0_VGT_INSTANCE_STEP_RATE_0, 1);
+ if (!physical_device->has_clear_state)
+ radeon_set_context_reg(cs, R_028AB8_VGT_VTX_CNT_EN, 0x0);
+ if (physical_device->rad_info.chip_class < CIK)
+ radeon_set_config_reg(cs, R_008A14_PA_CL_ENHANCE, S_008A14_NUM_CLIP_SEQ(3) |
+ S_008A14_CLIP_VTX_REORDER_ENA(1));
+
+ radeon_set_context_reg(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 0x76543210);
+ radeon_set_context_reg(cs, R_028BD8_PA_SC_CENTROID_PRIORITY_1, 0xfedcba98);
+
+ if (!physical_device->has_clear_state)
+ radeon_set_context_reg(cs, R_02882C_PA_SU_PRIM_FILTER_CNTL, 0);
+
+ /* CLEAR_STATE doesn't clear these correctly on certain generations.
+ * I don't know why. Deduced by trial and error.
+ */
+ if (physical_device->rad_info.chip_class <= CIK) {
+ radeon_set_context_reg(cs, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
+ radeon_set_context_reg(cs, R_028204_PA_SC_WINDOW_SCISSOR_TL,
+ S_028204_WINDOW_OFFSET_DISABLE(1));
+ radeon_set_context_reg(cs, R_028240_PA_SC_GENERIC_SCISSOR_TL,
+ S_028240_WINDOW_OFFSET_DISABLE(1));
+ radeon_set_context_reg(cs, R_028244_PA_SC_GENERIC_SCISSOR_BR,
+ S_028244_BR_X(16384) | S_028244_BR_Y(16384));
+ radeon_set_context_reg(cs, R_028030_PA_SC_SCREEN_SCISSOR_TL, 0);
+ radeon_set_context_reg(cs, R_028034_PA_SC_SCREEN_SCISSOR_BR,
+ S_028034_BR_X(16384) | S_028034_BR_Y(16384));
+ }
+
+ if (!physical_device->has_clear_state) {
+ for (i = 0; i < 16; i++) {
+ radeon_set_context_reg(cs, R_0282D0_PA_SC_VPORT_ZMIN_0 + i*8, 0);
+ radeon_set_context_reg(cs, R_0282D4_PA_SC_VPORT_ZMAX_0 + i*8, fui(1.0));
+ }
+ }
+
+ if (!physical_device->has_clear_state) {
+ radeon_set_context_reg(cs, R_02820C_PA_SC_CLIPRECT_RULE, 0xFFFF);
+ radeon_set_context_reg(cs, R_028230_PA_SC_EDGERULE, 0xAAAAAAAA);
+ /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
+ radeon_set_context_reg(cs, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET, 0);
+ radeon_set_context_reg(cs, R_028820_PA_CL_NANINF_CNTL, 0);
+ radeon_set_context_reg(cs, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 0x0);
+ radeon_set_context_reg(cs, R_028AC4_DB_SRESULTS_COMPARE_STATE1, 0x0);
+ radeon_set_context_reg(cs, R_028AC8_DB_PRELOAD_CONTROL, 0x0);
}
- radeon_set_context_reg(cs, R_028204_PA_SC_WINDOW_SCISSOR_TL, S_028204_WINDOW_OFFSET_DISABLE(1));
- radeon_set_context_reg(cs, R_028240_PA_SC_GENERIC_SCISSOR_TL, S_028240_WINDOW_OFFSET_DISABLE(1));
- radeon_set_context_reg(cs, R_028244_PA_SC_GENERIC_SCISSOR_BR,
- S_028244_BR_X(16384) | S_028244_BR_Y(16384));
- radeon_set_context_reg(cs, R_028030_PA_SC_SCREEN_SCISSOR_TL, 0);
- radeon_set_context_reg(cs, R_028034_PA_SC_SCREEN_SCISSOR_BR,
- S_028034_BR_X(16384) | S_028034_BR_Y(16384));
-
- radeon_set_context_reg(cs, R_02820C_PA_SC_CLIPRECT_RULE, 0xFFFF);
- radeon_set_context_reg(cs, R_028230_PA_SC_EDGERULE, 0xAAAAAAAA);
- /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
- radeon_set_context_reg(cs, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET, 0);
- radeon_set_context_reg(cs, R_028820_PA_CL_NANINF_CNTL, 0);
-
- radeon_set_context_reg(cs, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 0x0);
- radeon_set_context_reg(cs, R_028AC4_DB_SRESULTS_COMPARE_STATE1, 0x0);
- radeon_set_context_reg(cs, R_028AC8_DB_PRELOAD_CONTROL, 0x0);
radeon_set_context_reg(cs, R_02800C_DB_RENDER_OVERRIDE,
S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE));
- 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_028408_VGT_INDX_OFFSET, 0);
-
- if (physical_device->rad_info.chip_class >= CIK) {
- /* If this is 0, Bonaire can hang even if GS isn't being used.
- * Other chips are unaffected. These are suboptimal values,
- * but we don't use on-chip GS.
+ if (physical_device->rad_info.chip_class >= GFX9) {
+ 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 {
+ /* These registers, when written, also overwrite the
+ * CLEAR_STATE context, so we can't rely on CLEAR_STATE setting
+ * them. It would be an issue if there was another UMD
+ * changing them.
*/
- radeon_set_context_reg(cs, R_028A44_VGT_GS_ONCHIP_CNTL,
- S_028A44_ES_VERTS_PER_SUBGRP(64) |
- S_028A44_GS_PRIMS_PER_SUBGRP(4));
+ 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_028408_VGT_INDX_OFFSET, 0);
+ }
- radeon_set_sh_reg(cs, R_00B51C_SPI_SHADER_PGM_RSRC3_LS, S_00B51C_CU_EN(0xffff));
- radeon_set_sh_reg(cs, R_00B41C_SPI_SHADER_PGM_RSRC3_HS, 0);
- radeon_set_sh_reg(cs, R_00B31C_SPI_SHADER_PGM_RSRC3_ES, S_00B31C_CU_EN(0xffff));
- radeon_set_sh_reg(cs, R_00B21C_SPI_SHADER_PGM_RSRC3_GS, S_00B21C_CU_EN(0xffff));
+ if (physical_device->rad_info.chip_class >= CIK) {
+ if (physical_device->rad_info.chip_class >= GFX9) {
+ radeon_set_sh_reg(cs, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
+ S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
+ } else {
+ radeon_set_sh_reg(cs, R_00B51C_SPI_SHADER_PGM_RSRC3_LS,
+ S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
+ radeon_set_sh_reg(cs, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
+ S_00B41C_WAVE_LIMIT(0x3F));
+ radeon_set_sh_reg(cs, R_00B31C_SPI_SHADER_PGM_RSRC3_ES,
+ S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
+ /* If this is 0, Bonaire can hang even if GS isn't being used.
+ * Other chips are unaffected. These are suboptimal values,
+ * but we don't use on-chip GS.
+ */
+ radeon_set_context_reg(cs, R_028A44_VGT_GS_ONCHIP_CNTL,
+ S_028A44_ES_VERTS_PER_SUBGRP(64) |
+ S_028A44_GS_PRIMS_PER_SUBGRP(4));
+ }
+ radeon_set_sh_reg(cs, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
+ S_00B21C_CU_EN(0xffff) | S_00B21C_WAVE_LIMIT(0x3F));
if (physical_device->rad_info.num_good_compute_units /
(physical_device->rad_info.max_se * physical_device->rad_info.max_sh_per_se) <= 4) {
*
* LATE_ALLOC_VS = 2 is the highest safe number.
*/
- radeon_set_sh_reg(cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS, S_00B118_CU_EN(0xffff));
+ radeon_set_sh_reg(cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
+ S_00B118_CU_EN(0xffff) | S_00B118_WAVE_LIMIT(0x3F) );
radeon_set_sh_reg(cs, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(2));
} else {
/* Set LATE_ALLOC_VS == 31. It should be less than
* - VS can't execute on CU0.
* - If HS writes outputs to LDS, LS can't execute on CU0.
*/
- radeon_set_sh_reg(cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS, S_00B118_CU_EN(0xfffe));
+ radeon_set_sh_reg(cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
+ S_00B118_CU_EN(0xfffe) | S_00B118_WAVE_LIMIT(0x3F));
radeon_set_sh_reg(cs, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(31));
}
- radeon_set_sh_reg(cs, R_00B01C_SPI_SHADER_PGM_RSRC3_PS, S_00B01C_CU_EN(0xffff));
+ radeon_set_sh_reg(cs, R_00B01C_SPI_SHADER_PGM_RSRC3_PS,
+ S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
}
if (physical_device->rad_info.chip_class >= VI) {
radeon_set_context_reg(cs, R_028424_CB_DCC_CONTROL,
S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
S_028424_OVERWRITE_COMBINER_WATERMARK(4));
- if (physical_device->rad_info.family < CHIP_POLARIS10)
- radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 30);
- radeon_set_context_reg(cs, R_028C5C_VGT_OUT_DEALLOC_CNTL, 32);
vgt_tess_distribution = S_028B50_ACCUM_ISOLINE(32) |
S_028B50_ACCUM_TRI(11) |
radeon_set_context_reg(cs, R_028B50_VGT_TESS_DISTRIBUTION,
vgt_tess_distribution);
- } else {
+ } else if (!physical_device->has_clear_state) {
radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
radeon_set_context_reg(cs, R_028C5C_VGT_OUT_DEALLOC_CNTL, 16);
}
- if (physical_device->rad_info.family == CHIP_STONEY)
- radeon_set_context_reg(cs, R_028C40_PA_SC_SHADER_CONTROL, 0);
+ if (physical_device->rad_info.chip_class >= GFX9) {
+ unsigned num_se = physical_device->rad_info.max_se;
+ unsigned pc_lines = 0;
+
+ switch (physical_device->rad_info.family) {
+ case CHIP_VEGA10:
+ pc_lines = 4096;
+ break;
+ case CHIP_RAVEN:
+ pc_lines = 1024;
+ break;
+ default:
+ assert(0);
+ }
+
+ radeon_set_context_reg(cs, R_028C48_PA_SC_BINNER_CNTL_1,
+ S_028C48_MAX_ALLOC_COUNT(MIN2(128, pc_lines / (4 * num_se))) |
+ 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_uconfig_reg(cs, R_030968_VGT_INSTANCE_BASE_ID, 0);
+ }
+
+ unsigned tmp = (unsigned)(1.0 * 8.0);
+ radeon_set_context_reg_seq(cs, R_028A00_PA_SU_POINT_SIZE, 1);
+ radeon_emit(cs, S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp));
+ radeon_set_context_reg_seq(cs, R_028A04_PA_SU_POINT_MINMAX, 1);
+ radeon_emit(cs, S_028A04_MIN_SIZE(radv_pack_float_12p4(0)) |
+ S_028A04_MAX_SIZE(radv_pack_float_12p4(8192/2)));
+
+ if (!physical_device->has_clear_state) {
+ radeon_set_context_reg(cs, R_028004_DB_COUNT_CONTROL,
+ S_028004_ZPASS_INCREMENT_DISABLE(1));
+ }
si_emit_compute(physical_device, cs);
}
device->gfx_init = device->ws->buffer_create(device->ws,
cs->cdw * 4, 4096,
RADEON_DOMAIN_GTT,
- RADEON_FLAG_CPU_ACCESS);
+ RADEON_FLAG_CPU_ACCESS|
+ RADEON_FLAG_NO_INTERPROCESS_SHARING |
+ RADEON_FLAG_READ_ONLY);
if (!device->gfx_init)
goto fail;
int i;
float scale[3], translate[3], guardband_x = INFINITY, guardband_y = INFINITY;
const float max_range = 32767.0f;
- assert(count);
+ if (!count)
+ return;
radeon_set_context_reg_seq(cs, R_028250_PA_SC_VPORT_SCISSOR_0_TL + first * 4 * 2, count * 2);
for (i = 0; i < count; i++) {
enum chip_class chip_class = cmd_buffer->device->physical_device->rad_info.chip_class;
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;
+ const unsigned max_primgroup_in_wave = 2;
/* SWITCH_ON_EOP(0) is always preferable. */
bool wd_switch_on_eop = false;
bool ia_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 partial_es_wave = cmd_buffer->state.pipeline->graphics.partial_es_wave;
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 || (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 ||
- cmd_buffer->state.pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.uses_prim_id)
- ia_switch_on_eoi = true;
-
- /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
- if ((family == CHIP_TAHITI ||
- family == CHIP_PITCAIRN ||
- family == CHIP_BONAIRE) &&
- radv_pipeline_has_gs(cmd_buffer->state.pipeline))
- partial_vs_wave = true;
-
- /* 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 (family == CHIP_TONGA ||
- family == CHIP_FIJI ||
- family == CHIP_POLARIS10 ||
- family == CHIP_POLARIS11 ||
- family == CHIP_POLARIS12)
- partial_vs_wave = true;
- } else {
- partial_vs_wave = true;
- }
- }
+ 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;
}
- /* TODO linestipple */
+
+ ia_switch_on_eoi = cmd_buffer->state.pipeline->graphics.ia_switch_on_eoi;
+ partial_vs_wave = cmd_buffer->state.pipeline->graphics.partial_vs_wave;
if (chip_class >= CIK) {
- /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
- * 4 shader engines. Set 1 to pass the assertion below.
- * The other cases are hardware requirements. */
- if (info->max_se < 4 ||
- prim == V_008958_DI_PT_POLYGON ||
- prim == V_008958_DI_PT_LINELOOP ||
- prim == V_008958_DI_PT_TRIFAN ||
- prim == V_008958_DI_PT_TRISTRIP_ADJ ||
- (cmd_buffer->state.pipeline->graphics.prim_restart_enable &&
- (family < CHIP_POLARIS10 ||
- (prim != V_008958_DI_PT_POINTLIST &&
- prim != V_008958_DI_PT_LINESTRIP &&
- prim != V_008958_DI_PT_TRISTRIP))))
- wd_switch_on_eop = true;
+ wd_switch_on_eop = cmd_buffer->state.pipeline->graphics.wd_switch_on_eop;
/* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
* We don't know that for indirect drawing, so treat it as
if (ia_switch_on_eoi &&
(family == CHIP_HAWAII ||
(chip_class == VI &&
+ /* max primgroup in wave is always 2 - leave this for documentation */
(radv_pipeline_has_gs(cmd_buffer->state.pipeline) || max_primgroup_in_wave != 2))))
partial_vs_wave = true;
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)) {
- /* GS requirement. */
- if (SI_GS_PER_ES / 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) |
+ return cmd_buffer->state.pipeline->graphics.base_ia_multi_vgt_param |
+ 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_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);
+ S_028AA8_WD_SWITCH_ON_EOP(chip_class >= CIK ? wd_switch_on_eop : 0);
}
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 op = EVENT_TYPE(event) |
EVENT_INDEX(5) |
event_flags;
+ unsigned is_gfx8_mec = is_mec && chip_class < GFX9;
- if (is_mec) {
- radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, 5, 0));
+ if (chip_class >= GFX9 || is_gfx8_mec) {
+ 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 */
radeon_emit(cs, va >> 32); /* address hi */
radeon_emit(cs, new_fence); /* immediate data lo */
radeon_emit(cs, 0); /* immediate data hi */
+ if (!is_gfx8_mec)
+ radeon_emit(cs, 0); /* unused */
} else {
if (chip_class == CIK ||
chip_class == VI) {
/* Two EOP events are required to make all engines go idle
- * (and optional cache flushes executed) before the timstamp
+ * (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 predicated,
+ bool is_gfx9,
unsigned cp_coher_cntl)
{
- if (is_mec) {
- radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, 0) |
- PKT3_SHADER_TYPE_S(1));
+ if (is_mec || is_gfx9) {
+ uint32_t hi_val = is_gfx9 ? 0xffffff : 0xff;
+ 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, 0xff); /* CP_COHER_SIZE_HI */
+ radeon_emit(cs, hi_val); /* CP_COHER_SIZE_HI */
radeon_emit(cs, 0); /* CP_COHER_BASE */
radeon_emit(cs, 0); /* CP_COHER_BASE_HI */
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,
bool is_mec,
enum radv_cmd_flush_bits flush_bits)
{
unsigned cp_coher_cntl = 0;
-
+ uint32_t flush_cb_db = flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
+ RADV_CMD_FLAG_FLUSH_AND_INV_DB);
+
if (flush_bits & RADV_CMD_FLAG_INV_ICACHE)
cp_coher_cntl |= S_0085F0_SH_ICACHE_ACTION_ENA(1);
if (flush_bits & RADV_CMD_FLAG_INV_SMEM_L1)
cp_coher_cntl |= S_0085F0_SH_KCACHE_ACTION_ENA(1);
- if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB) {
- cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
- S_0085F0_CB0_DEST_BASE_ENA(1) |
- S_0085F0_CB1_DEST_BASE_ENA(1) |
- S_0085F0_CB2_DEST_BASE_ENA(1) |
- S_0085F0_CB3_DEST_BASE_ENA(1) |
- S_0085F0_CB4_DEST_BASE_ENA(1) |
- S_0085F0_CB5_DEST_BASE_ENA(1) |
- S_0085F0_CB6_DEST_BASE_ENA(1) |
- S_0085F0_CB7_DEST_BASE_ENA(1);
-
- /* Necessary for DCC */
- if (chip_class >= VI) {
- si_cs_emit_write_event_eop(cs,
- chip_class,
- is_mec,
- V_028A90_FLUSH_AND_INV_CB_DATA_TS,
- 0, 0, 0, 0, 0);
+ if (chip_class <= VI) {
+ if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB) {
+ cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
+ S_0085F0_CB0_DEST_BASE_ENA(1) |
+ S_0085F0_CB1_DEST_BASE_ENA(1) |
+ S_0085F0_CB2_DEST_BASE_ENA(1) |
+ S_0085F0_CB3_DEST_BASE_ENA(1) |
+ S_0085F0_CB4_DEST_BASE_ENA(1) |
+ S_0085F0_CB5_DEST_BASE_ENA(1) |
+ S_0085F0_CB6_DEST_BASE_ENA(1) |
+ S_0085F0_CB7_DEST_BASE_ENA(1);
+
+ /* 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,
+ 0, 0, 0, 0, 0);
+ }
+ }
+ if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB) {
+ cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
+ S_0085F0_DB_DEST_BASE_ENA(1);
}
- }
-
- if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB) {
- cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
- S_0085F0_DB_DEST_BASE_ENA(1);
}
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_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
- RADV_CMD_FLAG_FLUSH_AND_INV_DB))) {
- if (flush_bits & RADV_CMD_FLAG_PS_PARTIAL_FLUSH) {
- radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
- 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, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
- }
+ if (flush_bits & RADV_CMD_FLAG_PS_PARTIAL_FLUSH) {
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ 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, 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));
}
+ if (chip_class >= GFX9 && flush_cb_db) {
+ unsigned cb_db_event, tc_flags;
+
+#if 0
+ /* This breaks a bunch of:
+ dEQP-VK.renderpass.dedicated_allocation.formats.d32_sfloat_s8_uint.input*.
+ use the big hammer always.
+ */
+ /* Set the CB/DB flush event. */
+ switch (flush_cb_db) {
+ case RADV_CMD_FLAG_FLUSH_AND_INV_CB:
+ cb_db_event = V_028A90_FLUSH_AND_INV_CB_DATA_TS;
+ break;
+ case RADV_CMD_FLAG_FLUSH_AND_INV_DB:
+ cb_db_event = V_028A90_FLUSH_AND_INV_DB_DATA_TS;
+ break;
+ default:
+ /* both CB & DB */
+ cb_db_event = V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT;
+ }
+#else
+ cb_db_event = V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT;
+#endif
+ /* These are the only allowed combinations. If you need to
+ * do multiple operations at once, do them separately.
+ * All operations that invalidate L2 also seem to invalidate
+ * metadata. Volatile (VOL) and WC flushes are not listed here.
+ *
+ * TC | TC_WB = writeback & invalidate L2 & L1
+ * TC | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
+ * TC_WB | TC_NC = writeback L2 for MTYPE == NC
+ * TC | TC_NC = invalidate L2 for MTYPE == NC
+ * TC | TC_MD = writeback & invalidate L2 metadata (DCC, etc.)
+ * TCL1 = invalidate L1
+ */
+ tc_flags = EVENT_TC_ACTION_ENA |
+ EVENT_TC_MD_ACTION_ENA;
+
+ /* Ideally flush TC together with CB/DB. */
+ if (flush_bits & RADV_CMD_FLAG_INV_GLOBAL_L2) {
+ /* Writeback and invalidate everything in L2 & L1. */
+ tc_flags = EVENT_TC_ACTION_ENA |
+ EVENT_TC_WB_ACTION_ENA;
+
+
+ /* Clear the flags. */
+ flush_bits &= ~(RADV_CMD_FLAG_INV_GLOBAL_L2 |
+ RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2 |
+ RADV_CMD_FLAG_INV_VMEM_L1);
+ }
+ assert(flush_cnt);
+ uint32_t old_fence = (*flush_cnt)++;
+
+ 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, 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));
}
/* Make sure ME is idle (it executes most packets) before continuing.
* This prevents read-after-write hazards between PFP and ME.
*/
- if ((cp_coher_cntl || (flush_bits & RADV_CMD_FLAG_CS_PARTIAL_FLUSH)) &&
+ if ((cp_coher_cntl ||
+ (flush_bits & (RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
+ RADV_CMD_FLAG_INV_VMEM_L1 |
+ 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))) {
- cp_coher_cntl |= S_0085F0_TC_ACTION_ENA(1);
- if (chip_class >= VI)
- cp_coher_cntl |= S_0301F0_TC_WB_ACTION_ENA(1);
- } else if(flush_bits & RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2) {
- cp_coher_cntl |= S_0301F0_TC_WB_ACTION_ENA(1) |
- S_0301F0_TC_NC_ACTION_ENA(1);
-
- /* L2 writeback doesn't combine with L1 invalidate */
- si_emit_acquire_mem(cs, is_mec, cp_coher_cntl);
-
+ 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) |
+ S_0301F0_TC_WB_ACTION_ENA(chip_class >= VI));
cp_coher_cntl = 0;
+ } else {
+ if(flush_bits & RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2) {
+ /* WB = write-back
+ * NC = apply to non-coherent MTYPEs
+ * (i.e. MTYPE <= 1, which is what we use everywhere)
+ *
+ * WB doesn't work without NC.
+ */
+ 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,
+ predicated, chip_class >= GFX9,
+ cp_coher_cntl |
+ S_0085F0_TCL1_ACTION_ENA(1));
+ cp_coher_cntl = 0;
+ }
}
- if (flush_bits & RADV_CMD_FLAG_INV_VMEM_L1)
- cp_coher_cntl |= S_0085F0_TCL1_ACTION_ENA(1);
-
/* When one of the DEST_BASE flags is set, SURFACE_SYNC waits for idle.
* Therefore, it should be last. Done in PFP.
*/
if (cp_coher_cntl)
- si_emit_acquire_mem(cs, is_mec, cp_coher_cntl);
+ si_emit_acquire_mem(cs, is_mec, predicated, chip_class >= GFX9, cp_coher_cntl);
}
void
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;
+ uint64_t va = 0;
+ if (chip_class == GFX9) {
+ va = radv_buffer_get_va(cmd_buffer->gfx9_fence_bo) + cmd_buffer->gfx9_fence_offset;
+ 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)
+ if (unlikely(cmd_buffer->device->trace_bo))
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. */
/* The max number of bytes that can be copied per packet. */
static inline unsigned cp_dma_max_byte_count(struct radv_cmd_buffer *cmd_buffer)
{
- unsigned max = S_414_BYTE_COUNT_GFX6(~0u);
+ unsigned max = cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9 ?
+ S_414_BYTE_COUNT_GFX9(~0u) :
+ S_414_BYTE_COUNT_GFX6(~0u);
/* make it aligned for optimal performance */
return max & ~(SI_CPDMA_ALIGNMENT - 1);
assert(size <= cp_dma_max_byte_count(cmd_buffer));
radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 9);
-
- command |= S_414_BYTE_COUNT_GFX6(size);
+ if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9)
+ command |= S_414_BYTE_COUNT_GFX9(size);
+ else
+ command |= S_414_BYTE_COUNT_GFX6(size);
/* Sync flags. */
if (flags & CP_DMA_SYNC)
header |= S_411_CP_SYNC(1);
else {
- command |= S_414_DISABLE_WR_CONFIRM_GFX6(1);
+ if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9)
+ command |= S_414_DISABLE_WR_CONFIRM_GFX9(1);
+ else
+ command |= S_414_DISABLE_WR_CONFIRM_GFX6(1);
}
if (flags & CP_DMA_RAW_WAIT)
command |= S_414_RAW_WAIT(1);
/* Src and dst flags. */
- if (flags & CP_DMA_USE_L2)
+ if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9 &&
+ !(flags & CP_DMA_CLEAR) &&
+ src_va == dst_va)
+ header |= S_411_DSL_SEL(V_411_NOWHERE); /* prefetch only */
+ else if (flags & CP_DMA_USE_L2)
header |= S_411_DSL_SEL(V_411_DST_ADDR_TC_L2);
if (flags & CP_DMA_CLEAR)
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);
}
- radv_cmd_buffer_trace_emit(cmd_buffer);
+ if (unlikely(cmd_buffer->device->trace_bo))
+ radv_cmd_buffer_trace_emit(cmd_buffer);
}
void si_cp_dma_prefetch(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
radv_cmd_buffer_upload_alloc(cmd_buffer, buf_size, SI_CPDMA_ALIGNMENT, &offset, &ptr);
- va = cmd_buffer->device->ws->buffer_get_va(cmd_buffer->upload.upload_bo);
+ va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
va += offset;
si_cp_dma_prepare(cmd_buffer, size, size, &dma_flags);
switch (nr_samples) {
default:
case 1:
- radeon_set_context_reg(cs, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 0);
- radeon_set_context_reg(cs, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, 0);
- radeon_set_context_reg(cs, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, 0);
- radeon_set_context_reg(cs, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, 0);
+ radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 0);
+ radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, 0);
+ radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, 0);
+ radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, 0);
break;
case 2:
- radeon_set_context_reg(cs, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, eg_sample_locs_2x[0]);
- radeon_set_context_reg(cs, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, eg_sample_locs_2x[1]);
- radeon_set_context_reg(cs, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, eg_sample_locs_2x[2]);
- radeon_set_context_reg(cs, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, eg_sample_locs_2x[3]);
+ radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, eg_sample_locs_2x[0]);
+ radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, eg_sample_locs_2x[1]);
+ radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, eg_sample_locs_2x[2]);
+ radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, eg_sample_locs_2x[3]);
break;
case 4:
- radeon_set_context_reg(cs, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, eg_sample_locs_4x[0]);
- radeon_set_context_reg(cs, CM_R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, eg_sample_locs_4x[1]);
- radeon_set_context_reg(cs, CM_R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, eg_sample_locs_4x[2]);
- radeon_set_context_reg(cs, CM_R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, eg_sample_locs_4x[3]);
+ radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, eg_sample_locs_4x[0]);
+ radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, eg_sample_locs_4x[1]);
+ radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, eg_sample_locs_4x[2]);
+ radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, eg_sample_locs_4x[3]);
break;
case 8:
- radeon_set_context_reg_seq(cs, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 14);
+ radeon_set_context_reg_seq(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 14);
radeon_emit(cs, cm_sample_locs_8x[0]);
radeon_emit(cs, cm_sample_locs_8x[4]);
radeon_emit(cs, 0);
radeon_emit(cs, cm_sample_locs_8x[7]);
break;
case 16:
- radeon_set_context_reg_seq(cs, CM_R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 16);
+ radeon_set_context_reg_seq(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, 16);
radeon_emit(cs, cm_sample_locs_16x[0]);
radeon_emit(cs, cm_sample_locs_16x[4]);
radeon_emit(cs, cm_sample_locs_16x[8]);