* IN THE SOFTWARE.
*/
-/* command buffer handling for SI */
+/* command buffer handling for AMD GCN */
#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"
static void
si_write_harvested_raster_configs(struct radv_physical_device *physical_device,
raster_config_se);
for (se = 0; se < num_se; se++) {
- /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
- if (physical_device->rad_info.chip_class < CIK)
+ /* GRBM_GFX_INDEX has a different offset on GFX6 and GFX7+ */
+ if (physical_device->rad_info.chip_class < GFX7)
radeon_set_config_reg(cs, R_00802C_GRBM_GFX_INDEX,
S_00802C_SE_INDEX(se) |
S_00802C_SH_BROADCAST_WRITES(1) |
radeon_set_context_reg(cs, R_028350_PA_SC_RASTER_CONFIG, raster_config_se[se]);
}
- /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
- if (physical_device->rad_info.chip_class < CIK)
+ /* GRBM_GFX_INDEX has a different offset on GFX6 and GFX7+ */
+ if (physical_device->rad_info.chip_class < GFX7)
radeon_set_config_reg(cs, R_00802C_GRBM_GFX_INDEX,
S_00802C_SE_BROADCAST_WRITES(1) |
S_00802C_SH_BROADCAST_WRITES(1) |
S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
S_030800_INSTANCE_BROADCAST_WRITES(1));
- if (physical_device->rad_info.chip_class >= CIK)
+ if (physical_device->rad_info.chip_class >= GFX7)
radeon_set_context_reg(cs, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
}
-static void
-si_emit_compute(struct radv_physical_device *physical_device,
+void
+si_emit_compute(struct radv_device *device,
struct radeon_cmdbuf *cs)
{
radeon_set_sh_reg_seq(cs, R_00B810_COMPUTE_START_X, 3);
radeon_emit(cs, 0);
radeon_emit(cs, 0);
- 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_set_sh_reg_seq(cs, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0, 2);
+ /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1,
+ * renamed COMPUTE_DESTINATION_EN_SEn on gfx10. */
+ radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
- radeon_emit(cs, S_00B85C_SH0_CU_EN(0xffff) | S_00B85C_SH1_CU_EN(0xffff));
- if (physical_device->rad_info.chip_class >= CIK) {
+ if (device->physical_device->rad_info.chip_class >= GFX7) {
/* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
radeon_set_sh_reg_seq(cs,
R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2, 2);
- radeon_emit(cs, S_00B864_SH0_CU_EN(0xffff) |
- S_00B864_SH1_CU_EN(0xffff));
- radeon_emit(cs, S_00B868_SH0_CU_EN(0xffff) |
- S_00B868_SH1_CU_EN(0xffff));
+ radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) |
+ S_00B858_SH1_CU_EN(0xffff));
+ radeon_emit(cs, S_00B858_SH0_CU_EN(0xffff) |
+ S_00B858_SH1_CU_EN(0xffff));
+
+ if (device->border_color_data.bo) {
+ uint64_t bc_va = radv_buffer_get_va(device->border_color_data.bo);
+
+ radeon_set_uconfig_reg_seq(cs, R_030E00_TA_CS_BC_BASE_ADDR, 2);
+ radeon_emit(cs, bc_va >> 8);
+ radeon_emit(cs, S_030E04_ADDRESS(bc_va >> 40));
+ }
+ }
+
+ if (device->physical_device->rad_info.chip_class >= GFX9) {
+ radeon_set_uconfig_reg(cs, R_0301EC_CP_COHER_START_DELAY,
+ device->physical_device->rad_info.chip_class >= GFX10 ? 0x20 : 0);
+ }
+
+ if (device->physical_device->rad_info.chip_class >= GFX10) {
+ radeon_set_sh_reg(cs, R_00B890_COMPUTE_USER_ACCUM_0, 0);
+ radeon_set_sh_reg(cs, R_00B894_COMPUTE_USER_ACCUM_1, 0);
+ radeon_set_sh_reg(cs, R_00B898_COMPUTE_USER_ACCUM_2, 0);
+ radeon_set_sh_reg(cs, R_00B89C_COMPUTE_USER_ACCUM_3, 0);
+ radeon_set_sh_reg(cs, R_00B8A0_COMPUTE_PGM_RSRC3, 0);
+ radeon_set_sh_reg(cs, R_00B9F4_COMPUTE_DISPATCH_TUNNEL, 0);
}
/* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
* kernel if we want to use something other than the default value,
* which is now 0x22f.
*/
- if (physical_device->rad_info.chip_class <= SI) {
+ if (device->physical_device->rad_info.chip_class <= GFX6) {
/* XXX: This should be:
* (number of compute units) * 4 * (waves per simd) - 1 */
radeon_set_sh_reg(cs, R_00B82C_COMPUTE_MAX_WAVE_ID,
0x190 /* Default value */);
- }
-}
-void
-si_init_compute(struct radv_cmd_buffer *cmd_buffer)
-{
- struct radv_physical_device *physical_device = cmd_buffer->device->physical_device;
- si_emit_compute(physical_device, cmd_buffer->cs);
+ if (device->border_color_data.bo) {
+ uint64_t bc_va = radv_buffer_get_va(device->border_color_data.bo);
+ radeon_set_config_reg(cs, R_00950C_TA_CS_BC_BASE_ADDR, bc_va >> 8);
+ }
+ }
}
/* 12.4 fixed-point */
ac_get_raster_config(&physical_device->rad_info,
&raster_config,
- &raster_config_1);
+ &raster_config_1, NULL);
/* Always use the default config when all backends are enabled
* (or when we failed to determine the enabled backends).
if (!rb_mask || util_bitcount(rb_mask) >= num_rb) {
radeon_set_context_reg(cs, R_028350_PA_SC_RASTER_CONFIG,
raster_config);
- if (physical_device->rad_info.chip_class >= CIK)
+ if (physical_device->rad_info.chip_class >= GFX7)
radeon_set_context_reg(cs, R_028354_PA_SC_RASTER_CONFIG_1,
raster_config_1);
} else {
}
}
-static void
-si_emit_config(struct radv_physical_device *physical_device,
- struct radeon_cmdbuf *cs)
+void
+si_emit_graphics(struct radv_device *device,
+ struct radeon_cmdbuf *cs)
{
- int i;
+ struct radv_physical_device *physical_device = device->physical_device;
- /* Only SI can disable CLEAR_STATE for now. */
- assert(physical_device->has_clear_state ||
- physical_device->rad_info.chip_class == SI);
+ bool has_clear_state = physical_device->rad_info.has_clear_state;
+ 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_emit(cs, CC0_UPDATE_LOAD_ENABLES(1));
+ radeon_emit(cs, CC1_UPDATE_SHADOW_ENABLES(1));
- if (physical_device->has_clear_state) {
+ if (has_clear_state) {
radeon_emit(cs, PKT3(PKT3_CLEAR_STATE, 0, 0));
radeon_emit(cs, 0);
}
- if (physical_device->rad_info.chip_class <= VI)
+ if (physical_device->rad_info.chip_class <= GFX8)
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)
+ if (!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) {
+ if (physical_device->rad_info.chip_class <= GFX8) {
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) {
+ if (!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)
+ if (physical_device->rad_info.chip_class <= GFX9)
+ radeon_set_context_reg(cs, R_028AA0_VGT_INSTANCE_STEP_RATE_0, 1);
+ if (!has_clear_state)
radeon_set_context_reg(cs, R_028AB8_VGT_VTX_CNT_EN, 0x0);
- if (physical_device->rad_info.chip_class < CIK)
+ if (physical_device->rad_info.chip_class < GFX7)
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)
+ if (!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) {
+ if (physical_device->rad_info.chip_class <= GFX7 || !has_clear_state) {
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));
S_028034_BR_X(16384) | S_028034_BR_Y(16384));
}
- if (!physical_device->has_clear_state) {
+ if (!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) {
+ if (!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 */
+ /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on GFX6 */
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);
S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE));
- if (physical_device->rad_info.chip_class >= GFX9) {
+ if (physical_device->rad_info.chip_class >= GFX10) {
+ radeon_set_context_reg(cs, R_028A98_VGT_DRAW_PAYLOAD_CNTL, 0);
+ radeon_set_uconfig_reg(cs, R_030964_GE_MAX_VTX_INDX, ~0);
+ radeon_set_uconfig_reg(cs, R_030924_GE_MIN_VTX_INDX, 0);
+ radeon_set_uconfig_reg(cs, R_030928_GE_INDX_OFFSET, 0);
+ radeon_set_uconfig_reg(cs, R_03097C_GE_STEREO_CNTL, 0);
+ radeon_set_uconfig_reg(cs, R_030988_GE_USER_VGPR_EN, 0);
+ } else 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);
radeon_set_context_reg(cs, R_028408_VGT_INDX_OFFSET, 0);
}
- if (physical_device->rad_info.chip_class >= CIK) {
+ if (physical_device->rad_info.chip_class >= GFX7) {
+ if (physical_device->rad_info.chip_class >= GFX10) {
+ /* Logical CUs 16 - 31 */
+ radeon_set_sh_reg_idx(physical_device, cs, R_00B404_SPI_SHADER_PGM_RSRC4_HS,
+ 3, S_00B404_CU_EN(0xffff));
+ radeon_set_sh_reg_idx(physical_device, cs, R_00B104_SPI_SHADER_PGM_RSRC4_VS,
+ 3, S_00B104_CU_EN(0xffff));
+ radeon_set_sh_reg_idx(physical_device, cs, R_00B004_SPI_SHADER_PGM_RSRC4_PS,
+ 3, S_00B004_CU_EN(0xffff));
+ }
+
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));
+ radeon_set_sh_reg_idx(physical_device, cs, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
+ 3, 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));
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) {
- /* Too few available compute units per SH. Disallowing
- * VS to run on CU0 could hurt us more than late VS
- * allocation would help.
- *
- * LATE_ALLOC_VS = 2 is the highest safe number.
+
+ /* Compute LATE_ALLOC_VS.LIMIT. */
+ unsigned num_cu_per_sh = physical_device->rad_info.min_good_cu_per_sa;
+ unsigned late_alloc_wave64 = 0; /* The limit is per SA. */
+ unsigned late_alloc_wave64_gs = 0;
+ unsigned cu_mask_vs = 0xffff;
+ unsigned cu_mask_gs = 0xffff;
+
+ if (physical_device->rad_info.chip_class >= GFX10) {
+ /* For Wave32, the hw will launch twice the number of late
+ * alloc waves, so 1 == 2x wave32.
*/
- 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
- * the number of scratch waves. Limitations:
- * - VS can't execute on CU0.
- * - If HS writes outputs to LDS, LS can't execute on CU0.
+ if (!physical_device->rad_info.use_late_alloc) {
+ late_alloc_wave64 = 0;
+ } else if (num_cu_per_sh <= 6) {
+ late_alloc_wave64 = num_cu_per_sh - 2;
+ } else {
+ late_alloc_wave64 = (num_cu_per_sh - 2) * 4;
+
+ /* CU2 & CU3 disabled because of the dual CU design */
+ cu_mask_vs = 0xfff3;
+ cu_mask_gs = 0xfff3; /* NGG only */
+ }
+
+ late_alloc_wave64_gs = late_alloc_wave64;
+
+ /* Don't use late alloc for NGG on Navi14 due to a hw
+ * bug. If NGG is never used, enable all CUs.
*/
- 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));
+ if (!physical_device->use_ngg ||
+ physical_device->rad_info.family == CHIP_NAVI14) {
+ late_alloc_wave64_gs = 0;
+ cu_mask_gs = 0xffff;
+ }
+
+ /* Limit LATE_ALLOC_GS for prevent a hang (hw bug). */
+ if (physical_device->rad_info.chip_class == GFX10)
+ late_alloc_wave64_gs = MIN2(late_alloc_wave64_gs, 64);
+ } else {
+ if (!physical_device->rad_info.use_late_alloc) {
+ late_alloc_wave64 = 0;
+ } else if (num_cu_per_sh <= 4) {
+ /* Too few available compute units per SA.
+ * Disallowing VS to run on one CU could hurt
+ * us more than late VS allocation would help.
+ *
+ * 2 is the highest safe number that allows us
+ * to keep all CUs enabled.
+ */
+ late_alloc_wave64 = 2;
+ } else {
+ /* This is a good initial value, allowing 1
+ * late_alloc wave per SIMD on num_cu - 2.
+ */
+ late_alloc_wave64 = (num_cu_per_sh - 2) * 4;
+ }
+
+ if (late_alloc_wave64 > 2)
+ cu_mask_vs = 0xfffe; /* 1 CU disabled */
+ }
+
+ radeon_set_sh_reg_idx(physical_device, cs, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
+ 3, S_00B118_CU_EN(cu_mask_vs) |
+ S_00B118_WAVE_LIMIT(0x3F));
+ radeon_set_sh_reg(cs, R_00B11C_SPI_SHADER_LATE_ALLOC_VS,
+ S_00B11C_LIMIT(late_alloc_wave64));
+
+ radeon_set_sh_reg_idx(physical_device, cs, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
+ 3, S_00B21C_CU_EN(cu_mask_gs) | S_00B21C_WAVE_LIMIT(0x3F));
+
+ if (physical_device->rad_info.chip_class >= GFX10) {
+ radeon_set_sh_reg_idx(physical_device, cs, R_00B204_SPI_SHADER_PGM_RSRC4_GS,
+ 3, S_00B204_CU_EN(0xffff) |
+ S_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(late_alloc_wave64_gs));
+ }
+
+ radeon_set_sh_reg_idx(physical_device, cs, R_00B01C_SPI_SHADER_PGM_RSRC3_PS,
+ 3, S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
+ }
+
+ if (physical_device->rad_info.chip_class >= GFX10) {
+ /* Break up a pixel wave if it contains deallocs for more than
+ * half the parameter cache.
+ *
+ * To avoid a deadlock where pixel waves aren't launched
+ * because they're waiting for more pixels while the frontend
+ * is stuck waiting for PC space, the maximum allowed value is
+ * the size of the PC minus the largest possible allocation for
+ * a single primitive shader subgroup.
+ */
+ radeon_set_context_reg(cs, R_028C50_PA_SC_NGG_MODE_CNTL,
+ S_028C50_MAX_DEALLOCS_IN_WAVE(512));
+ radeon_set_context_reg(cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
+ radeon_set_context_reg(cs, R_028428_CB_COVERAGE_OUT_CONTROL, 0);
+
+ radeon_set_sh_reg(cs, R_00B0C8_SPI_SHADER_USER_ACCUM_PS_0, 0);
+ radeon_set_sh_reg(cs, R_00B0CC_SPI_SHADER_USER_ACCUM_PS_1, 0);
+ radeon_set_sh_reg(cs, R_00B0D0_SPI_SHADER_USER_ACCUM_PS_2, 0);
+ radeon_set_sh_reg(cs, R_00B0D4_SPI_SHADER_USER_ACCUM_PS_3, 0);
+ radeon_set_sh_reg(cs, R_00B1C8_SPI_SHADER_USER_ACCUM_VS_0, 0);
+ radeon_set_sh_reg(cs, R_00B1CC_SPI_SHADER_USER_ACCUM_VS_1, 0);
+ radeon_set_sh_reg(cs, R_00B1D0_SPI_SHADER_USER_ACCUM_VS_2, 0);
+ radeon_set_sh_reg(cs, R_00B1D4_SPI_SHADER_USER_ACCUM_VS_3, 0);
+ radeon_set_sh_reg(cs, R_00B2C8_SPI_SHADER_USER_ACCUM_ESGS_0, 0);
+ radeon_set_sh_reg(cs, R_00B2CC_SPI_SHADER_USER_ACCUM_ESGS_1, 0);
+ radeon_set_sh_reg(cs, R_00B2D0_SPI_SHADER_USER_ACCUM_ESGS_2, 0);
+ radeon_set_sh_reg(cs, R_00B2D4_SPI_SHADER_USER_ACCUM_ESGS_3, 0);
+ radeon_set_sh_reg(cs, R_00B4C8_SPI_SHADER_USER_ACCUM_LSHS_0, 0);
+ radeon_set_sh_reg(cs, R_00B4CC_SPI_SHADER_USER_ACCUM_LSHS_1, 0);
+ radeon_set_sh_reg(cs, R_00B4D0_SPI_SHADER_USER_ACCUM_LSHS_2, 0);
+ radeon_set_sh_reg(cs, R_00B4D4_SPI_SHADER_USER_ACCUM_LSHS_3, 0);
+
+ radeon_set_sh_reg(cs, R_00B0C0_SPI_SHADER_REQ_CTRL_PS,
+ S_00B0C0_SOFT_GROUPING_EN(1) |
+ S_00B0C0_NUMBER_OF_REQUESTS_PER_CU(4 - 1));
+ radeon_set_sh_reg(cs, R_00B1C0_SPI_SHADER_REQ_CTRL_VS, 0);
+
+ if (physical_device->rad_info.chip_class >= GFX10_3) {
+ radeon_set_context_reg(cs, R_028750_SX_PS_DOWNCONVERT_CONTROL_GFX103, 0xff);
+ radeon_set_context_reg(cs, 0x28848, 1 << 9); /* This fixes sample shading. */
+ }
+
+ if (physical_device->rad_info.chip_class == GFX10) {
+ /* SQ_NON_EVENT must be emitted before GE_PC_ALLOC is written. */
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, EVENT_TYPE(V_028A90_SQ_NON_EVENT) | EVENT_INDEX(0));
}
- radeon_set_sh_reg(cs, R_00B01C_SPI_SHADER_PGM_RSRC3_PS,
- S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
+ /* TODO: For culling, replace 128 with 256. */
+ radeon_set_uconfig_reg(cs, R_030980_GE_PC_ALLOC,
+ S_030980_OVERSUB_EN(physical_device->rad_info.use_late_alloc) |
+ S_030980_NUM_PC_LINES(128 * physical_device->rad_info.max_se - 1));
}
- if (physical_device->rad_info.chip_class >= VI) {
+ if (physical_device->rad_info.chip_class >= GFX9) {
+ radeon_set_context_reg(cs, R_028B50_VGT_TESS_DISTRIBUTION,
+ S_028B50_ACCUM_ISOLINE(40) |
+ S_028B50_ACCUM_TRI(30) |
+ S_028B50_ACCUM_QUAD(24) |
+ S_028B50_DONUT_SPLIT(24) |
+ S_028B50_TRAP_SPLIT(6));
+ } else if (physical_device->rad_info.chip_class >= GFX8) {
uint32_t vgt_tess_distribution;
- radeon_set_context_reg(cs, R_028424_CB_DCC_CONTROL,
- S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
- S_028424_OVERWRITE_COMBINER_WATERMARK(4));
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 if (!physical_device->has_clear_state) {
+ } else if (!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.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:
- case CHIP_VEGA12:
- pc_lines = 4096;
- break;
- case CHIP_RAVEN:
- pc_lines = 1024;
- break;
- default:
- assert(0);
+ if (device->border_color_data.bo) {
+ uint64_t border_color_va = radv_buffer_get_va(device->border_color_data.bo);
+
+ radeon_set_context_reg(cs, R_028080_TA_BC_BASE_ADDR, border_color_va >> 8);
+ if (physical_device->rad_info.chip_class >= GFX7) {
+ radeon_set_context_reg(cs, R_028084_TA_BC_BASE_ADDR_HI,
+ S_028084_ADDRESS(border_color_va >> 40));
}
+ }
+ if (physical_device->rad_info.chip_class >= GFX9) {
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_ALLOC_COUNT(physical_device->rad_info.pbb_max_alloc_count - 1) |
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_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)));
+ S_028A04_MAX_SIZE(radv_pack_float_12p4(8191.875/2)));
- if (!physical_device->has_clear_state) {
+ if (!has_clear_state) {
radeon_set_context_reg(cs, R_028004_DB_COUNT_CONTROL,
S_028004_ZPASS_INCREMENT_DISABLE(1));
}
small_prim_filter_cntl);
}
- si_emit_compute(physical_device, cs);
-}
-
-void si_init_config(struct radv_cmd_buffer *cmd_buffer)
-{
- struct radv_physical_device *physical_device = cmd_buffer->device->physical_device;
-
- si_emit_config(physical_device, cmd_buffer->cs);
+ radeon_set_context_reg(cs, R_0286D4_SPI_INTERP_CONTROL_0,
+ S_0286D4_FLAT_SHADE_ENA(1) |
+ S_0286D4_PNT_SPRITE_ENA(1) |
+ S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
+ S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) |
+ S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) |
+ S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
+ S_0286D4_PNT_SPRITE_TOP_1(0)); /* vulkan is top to bottom - 1.0 at bottom */
+
+ radeon_set_context_reg(cs, R_028BE4_PA_SU_VTX_CNTL,
+ S_028BE4_PIX_CENTER(1) |
+ S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN) |
+ S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH));
+
+ radeon_set_context_reg(cs, R_028818_PA_CL_VTE_CNTL,
+ S_028818_VTX_W0_FMT(1) |
+ S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(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));
+
+ si_emit_compute(device, cs);
}
void
if (!cs)
return;
- si_emit_config(device->physical_device, cs);
+ si_emit_graphics(device, cs);
while (cs->cdw & 7) {
if (device->physical_device->rad_info.gfx_ib_pad_with_type2)
- radeon_emit(cs, 0x80000000);
+ radeon_emit(cs, PKT2_NOP_PAD);
else
- radeon_emit(cs, 0xffff1000);
+ radeon_emit(cs, PKT3_NOP_PAD);
}
device->gfx_init = device->ws->buffer_create(device->ws,
RADEON_DOMAIN_GTT,
RADEON_FLAG_CPU_ACCESS|
RADEON_FLAG_NO_INTERPROCESS_SHARING |
- RADEON_FLAG_READ_ONLY);
+ RADEON_FLAG_READ_ONLY |
+ RADEON_FLAG_GTT_WC,
+ RADV_BO_PRIORITY_CS);
if (!device->gfx_init)
goto fail;
get_viewport_xform(viewport, scale, translate);
- rect.offset.x = translate[0] - fabs(scale[0]);
- rect.offset.y = translate[1] - fabs(scale[1]);
- rect.extent.width = ceilf(translate[0] + fabs(scale[0])) - rect.offset.x;
- rect.extent.height = ceilf(translate[1] + fabs(scale[1])) - rect.offset.y;
+ rect.offset.x = translate[0] - fabsf(scale[0]);
+ rect.offset.y = translate[1] - fabsf(scale[1]);
+ rect.extent.width = ceilf(translate[0] + fabsf(scale[0])) - rect.offset.x;
+ rect.extent.height = ceilf(translate[1] + fabsf(scale[1])) - rect.offset.y;
return rect;
}
VkRect2D scissor = si_intersect_scissor(&scissors[i], &viewport_scissor);
get_viewport_xform(viewports + i, scale, translate);
- scale[0] = abs(scale[0]);
- scale[1] = abs(scale[1]);
+ scale[0] = fabsf(scale[0]);
+ scale[1] = fabsf(scale[1]);
if (scale[0] < 0.5)
scale[0] = 0.5;
if (scale[1] < 0.5)
scale[1] = 0.5;
- guardband_x = MIN2(guardband_x, (max_range - abs(translate[0])) / scale[0]);
- guardband_y = MIN2(guardband_y, (max_range - abs(translate[1])) / scale[1]);
+ guardband_x = MIN2(guardband_x, (max_range - fabsf(translate[0])) / scale[0]);
+ guardband_y = MIN2(guardband_y, (max_range - fabsf(translate[1])) / scale[1]);
radeon_emit(cs, S_028250_TL_X(scissor.offset.x) |
S_028250_TL_Y(scissor.offset.y) |
return 1 + ((num - info->min) / info->incr);
}
+static const struct radv_prim_vertex_count prim_size_table[] = {
+ [V_008958_DI_PT_NONE] = {0, 0},
+ [V_008958_DI_PT_POINTLIST] = {1, 1},
+ [V_008958_DI_PT_LINELIST] = {2, 2},
+ [V_008958_DI_PT_LINESTRIP] = {2, 1},
+ [V_008958_DI_PT_TRILIST] = {3, 3},
+ [V_008958_DI_PT_TRIFAN] = {3, 1},
+ [V_008958_DI_PT_TRISTRIP] = {3, 1},
+ [V_008958_DI_PT_LINELIST_ADJ] = {4, 4},
+ [V_008958_DI_PT_LINESTRIP_ADJ] = {4, 1},
+ [V_008958_DI_PT_TRILIST_ADJ] = {6, 6},
+ [V_008958_DI_PT_TRISTRIP_ADJ] = {6, 2},
+ [V_008958_DI_PT_RECTLIST] = {3, 3},
+ [V_008958_DI_PT_LINELOOP] = {2, 1},
+ [V_008958_DI_PT_POLYGON] = {3, 1},
+ [V_008958_DI_PT_2D_TRI_STRIP] = {0, 0},
+};
+
uint32_t
si_get_ia_multi_vgt_param(struct radv_cmd_buffer *cmd_buffer,
bool instanced_draw, bool indirect_draw,
- uint32_t draw_vertex_count)
+ bool count_from_stream_output,
+ uint32_t draw_vertex_count,
+ unsigned topology)
{
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;
bool partial_vs_wave = false;
bool partial_es_wave = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.partial_es_wave;
bool multi_instances_smaller_than_primgroup;
+ struct radv_prim_vertex_count prim_vertex_count = prim_size_table[topology];
+
+ if (radv_pipeline_has_tess(cmd_buffer->state.pipeline)) {
+ if (topology == V_008958_DI_PT_PATCH) {
+ prim_vertex_count.min = cmd_buffer->state.pipeline->graphics.tess_patch_control_points;
+ prim_vertex_count.incr = 1;
+ }
+ }
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);
+ uint32_t num_prims = radv_prims_for_vertices(&prim_vertex_count, draw_vertex_count);
if (num_prims < cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.primgroup_size)
multi_instances_smaller_than_primgroup = true;
}
ia_switch_on_eoi = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.ia_switch_on_eoi;
partial_vs_wave = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.partial_vs_wave;
- if (chip_class >= CIK) {
- wd_switch_on_eop = cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.wd_switch_on_eop;
+ if (chip_class >= GFX7) {
+ /* 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 (cmd_buffer->device->physical_device->rad_info.max_se < 4 ||
+ topology == V_008958_DI_PT_POLYGON ||
+ topology == V_008958_DI_PT_LINELOOP ||
+ topology == V_008958_DI_PT_TRIFAN ||
+ topology == V_008958_DI_PT_TRISTRIP_ADJ ||
+ (cmd_buffer->state.pipeline->graphics.prim_restart_enable &&
+ (cmd_buffer->device->physical_device->rad_info.family < CHIP_POLARIS10 ||
+ (topology != V_008958_DI_PT_POINTLIST &&
+ topology != V_008958_DI_PT_LINESTRIP))))
+ wd_switch_on_eop = true;
/* 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
* Assume indirect draws always use small instances.
* This is needed for good VS wave utilization.
*/
- if (chip_class <= VI &&
+ if (chip_class <= GFX8 &&
info->max_se == 4 &&
multi_instances_smaller_than_primgroup)
wd_switch_on_eop = true;
- /* Required on CIK and later. */
+ /* Required on GFX7 and later. */
if (info->max_se > 2 && !wd_switch_on_eop)
ia_switch_on_eoi = true;
- /* Required by Hawaii and, for some special cases, by VI. */
+ /* Required by Hawaii and, for some special cases, by GFX8. */
if (ia_switch_on_eoi &&
(family == CHIP_HAWAII ||
- (chip_class == VI &&
+ (chip_class == GFX8 &&
/* 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;
(instanced_draw || indirect_draw))
partial_vs_wave = true;
+ /* Hardware requirement when drawing primitives from a stream
+ * output buffer.
+ */
+ if (count_from_stream_output)
+ wd_switch_on_eop = true;
+
/* If the WD switch is false, the IA switch must be false too. */
assert(wd_switch_on_eop || !ia_switch_on_eop);
}
/* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
- if (chip_class <= VI && ia_switch_on_eoi)
+ if (chip_class <= GFX8 && ia_switch_on_eoi)
partial_es_wave = true;
if (radv_pipeline_has_gs(cmd_buffer->state.pipeline)) {
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);
+ uint32_t num_prims = radv_prims_for_vertices(&prim_vertex_count, draw_vertex_count);
if (num_prims <= 1)
set_vgt_flush = true;
}
}
}
+ /* Workaround for a VGT hang when strip primitive types are used with
+ * primitive restart.
+ */
+ if (cmd_buffer->state.pipeline->graphics.prim_restart_enable &&
+ (topology == V_008958_DI_PT_LINESTRIP ||
+ topology == V_008958_DI_PT_TRISTRIP ||
+ topology == V_008958_DI_PT_LINESTRIP_ADJ ||
+ topology == V_008958_DI_PT_TRISTRIP_ADJ)) {
+ partial_vs_wave = true;
+ }
+
return cmd_buffer->state.pipeline->graphics.ia_multi_vgt_param.base |
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_WD_SWITCH_ON_EOP(chip_class >= CIK ? wd_switch_on_eop : 0);
+ S_028AA8_WD_SWITCH_ON_EOP(chip_class >= GFX7 ? wd_switch_on_eop : 0);
}
void si_cs_emit_write_event_eop(struct radeon_cmdbuf *cs,
- bool predicated,
enum chip_class chip_class,
bool is_mec,
unsigned event, unsigned event_flags,
- unsigned data_sel,
+ unsigned dst_sel, unsigned data_sel,
uint64_t va,
- uint32_t old_fence,
- uint32_t new_fence)
+ uint32_t new_fence,
+ uint64_t gfx9_eop_bug_va)
{
unsigned op = EVENT_TYPE(event) |
- EVENT_INDEX(5) |
+ EVENT_INDEX(event == V_028A90_CS_DONE ||
+ event == V_028A90_PS_DONE ? 6 : 5) |
event_flags;
unsigned is_gfx8_mec = is_mec && chip_class < GFX9;
- unsigned sel = EOP_DATA_SEL(data_sel);
+ unsigned sel = EOP_DST_SEL(dst_sel) |
+ EOP_DATA_SEL(data_sel);
/* Wait for write confirmation before writing data, but don't send
* an interrupt. */
sel |= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM);
if (chip_class >= GFX9 || is_gfx8_mec) {
- radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, is_gfx8_mec ? 5 : 6, predicated));
+ /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
+ * counters) must immediately precede every timestamp event to
+ * prevent a GPU hang on GFX9.
+ */
+ if (chip_class == GFX9 && !is_mec) {
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
+ radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1));
+ radeon_emit(cs, gfx9_eop_bug_va);
+ radeon_emit(cs, gfx9_eop_bug_va >> 32);
+ }
+
+ radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, is_gfx8_mec ? 5 : 6, false));
radeon_emit(cs, op);
radeon_emit(cs, sel);
radeon_emit(cs, va); /* address lo */
if (!is_gfx8_mec)
radeon_emit(cs, 0); /* unused */
} else {
- if (chip_class == CIK ||
- chip_class == VI) {
+ if (chip_class == GFX7 ||
+ chip_class == GFX8) {
/* Two EOP events are required to make all engines go idle
* (and optional cache flushes executed) before the timestamp
* is written.
*/
- radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, predicated));
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, false));
radeon_emit(cs, op);
radeon_emit(cs, va);
radeon_emit(cs, ((va >> 32) & 0xffff) | sel);
- radeon_emit(cs, old_fence); /* immediate data */
+ radeon_emit(cs, 0); /* immediate data */
radeon_emit(cs, 0); /* unused */
}
- radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, predicated));
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, false));
radeon_emit(cs, op);
radeon_emit(cs, va);
radeon_emit(cs, ((va >> 32) & 0xffff) | sel);
}
void
-si_emit_wait_fence(struct radeon_cmdbuf *cs,
- bool predicated,
- uint64_t va, uint32_t ref,
- uint32_t mask)
+radv_cp_wait_mem(struct radeon_cmdbuf *cs, uint32_t op, uint64_t va,
+ uint32_t ref, uint32_t mask)
{
- radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, predicated));
- radeon_emit(cs, WAIT_REG_MEM_EQUAL | WAIT_REG_MEM_MEM_SPACE(1));
+ assert(op == WAIT_REG_MEM_EQUAL ||
+ op == WAIT_REG_MEM_NOT_EQUAL ||
+ op == WAIT_REG_MEM_GREATER_OR_EQUAL);
+
+ radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, false));
+ radeon_emit(cs, op | WAIT_REG_MEM_MEM_SPACE(1));
radeon_emit(cs, va);
radeon_emit(cs, va >> 32);
radeon_emit(cs, ref); /* reference value */
static void
si_emit_acquire_mem(struct radeon_cmdbuf *cs,
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, predicated) |
+ radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, false) |
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, predicated));
+ radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, false));
radeon_emit(cs, cp_coher_cntl); /* CP_COHER_CNTL */
radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
radeon_emit(cs, 0); /* CP_COHER_BASE */
}
}
+static void
+gfx10_cs_emit_cache_flush(struct radeon_cmdbuf *cs,
+ enum chip_class chip_class,
+ uint32_t *flush_cnt,
+ uint64_t flush_va,
+ bool is_mec,
+ enum radv_cmd_flush_bits flush_bits,
+ uint64_t gfx9_eop_bug_va)
+{
+ uint32_t gcr_cntl = 0;
+ unsigned cb_db_event = 0;
+
+ /* We don't need these. */
+ assert(!(flush_bits & (RADV_CMD_FLAG_VGT_STREAMOUT_SYNC)));
+
+ if (flush_bits & RADV_CMD_FLAG_INV_ICACHE)
+ gcr_cntl |= S_586_GLI_INV(V_586_GLI_ALL);
+ if (flush_bits & RADV_CMD_FLAG_INV_SCACHE) {
+ /* TODO: When writing to the SMEM L1 cache, we need to set SEQ
+ * to FORWARD when both L1 and L2 are written out (WB or INV).
+ */
+ gcr_cntl |= S_586_GL1_INV(1) | S_586_GLK_INV(1);
+ }
+ if (flush_bits & RADV_CMD_FLAG_INV_VCACHE)
+ gcr_cntl |= S_586_GL1_INV(1) | S_586_GLV_INV(1);
+ if (flush_bits & RADV_CMD_FLAG_INV_L2) {
+ /* Writeback and invalidate everything in L2. */
+ gcr_cntl |= S_586_GL2_INV(1) | S_586_GL2_WB(1) |
+ S_586_GLM_INV(1) | S_586_GLM_WB(1);
+ } else if (flush_bits & RADV_CMD_FLAG_WB_L2) {
+ /* Writeback but do not invalidate.
+ * GLM doesn't support WB alone. If WB is set, INV must be set too.
+ */
+ gcr_cntl |= S_586_GL2_WB(1) |
+ S_586_GLM_WB(1) | S_586_GLM_INV(1);
+ }
+
+ /* TODO: Implement this new flag for GFX9+.
+ else if (flush_bits & RADV_CMD_FLAG_INV_L2_METADATA)
+ gcr_cntl |= S_586_GLM_INV(1) | S_586_GLM_WB(1);
+ */
+
+ if (flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB | RADV_CMD_FLAG_FLUSH_AND_INV_DB)) {
+ /* TODO: trigger on RADV_CMD_FLAG_FLUSH_AND_INV_CB_META */
+ if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB) {
+ /* Flush CMASK/FMASK/DCC. Will wait for idle later. */
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) |
+ EVENT_INDEX(0));
+ }
+
+ /* TODO: trigger on RADV_CMD_FLAG_FLUSH_AND_INV_DB_META ? */
+ if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB) {
+ /* Flush HTILE. Will wait for idle later. */
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) |
+ EVENT_INDEX(0));
+ }
+
+ /* First flush CB/DB, then L1/L2. */
+ gcr_cntl |= S_586_SEQ(V_586_SEQ_FORWARD);
+
+ if ((flush_bits & (RADV_CMD_FLAG_FLUSH_AND_INV_CB | RADV_CMD_FLAG_FLUSH_AND_INV_DB)) ==
+ (RADV_CMD_FLAG_FLUSH_AND_INV_CB | RADV_CMD_FLAG_FLUSH_AND_INV_DB)) {
+ cb_db_event = V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT;
+ } else if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_CB) {
+ cb_db_event = V_028A90_FLUSH_AND_INV_CB_DATA_TS;
+ } else if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB) {
+ cb_db_event = V_028A90_FLUSH_AND_INV_DB_DATA_TS;
+ } else {
+ assert(0);
+ }
+ } else {
+ /* Wait for graphics shaders to go idle if requested. */
+ 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, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH | EVENT_INDEX(4)));
+ }
+
+ if (cb_db_event) {
+ /* CB/DB flush and invalidate (or possibly just a wait for a
+ * meta flush) via RELEASE_MEM.
+ *
+ * Combine this with other cache flushes when possible; this
+ * requires affected shaders to be idle, so do it after the
+ * CS_PARTIAL_FLUSH before (VS/PS partial flushes are always
+ * implied).
+ */
+ /* Get GCR_CNTL fields, because the encoding is different in RELEASE_MEM. */
+ unsigned glm_wb = G_586_GLM_WB(gcr_cntl);
+ unsigned glm_inv = G_586_GLM_INV(gcr_cntl);
+ unsigned glv_inv = G_586_GLV_INV(gcr_cntl);
+ unsigned gl1_inv = G_586_GL1_INV(gcr_cntl);
+ assert(G_586_GL2_US(gcr_cntl) == 0);
+ assert(G_586_GL2_RANGE(gcr_cntl) == 0);
+ assert(G_586_GL2_DISCARD(gcr_cntl) == 0);
+ unsigned gl2_inv = G_586_GL2_INV(gcr_cntl);
+ unsigned gl2_wb = G_586_GL2_WB(gcr_cntl);
+ unsigned gcr_seq = G_586_SEQ(gcr_cntl);
+
+ gcr_cntl &= C_586_GLM_WB &
+ C_586_GLM_INV &
+ C_586_GLV_INV &
+ C_586_GL1_INV &
+ C_586_GL2_INV &
+ C_586_GL2_WB; /* keep SEQ */
+
+ assert(flush_cnt);
+ (*flush_cnt)++;
+
+ si_cs_emit_write_event_eop(cs, chip_class, false, cb_db_event,
+ S_490_GLM_WB(glm_wb) |
+ S_490_GLM_INV(glm_inv) |
+ S_490_GLV_INV(glv_inv) |
+ S_490_GL1_INV(gl1_inv) |
+ S_490_GL2_INV(gl2_inv) |
+ S_490_GL2_WB(gl2_wb) |
+ S_490_SEQ(gcr_seq),
+ EOP_DST_SEL_MEM,
+ EOP_DATA_SEL_VALUE_32BIT,
+ flush_va, *flush_cnt,
+ gfx9_eop_bug_va);
+
+ radv_cp_wait_mem(cs, WAIT_REG_MEM_EQUAL, 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, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
+ }
+
+ /* Ignore fields that only modify the behavior of other fields. */
+ if (gcr_cntl & C_586_GL1_RANGE & C_586_GL2_RANGE & C_586_SEQ) {
+ /* Flush caches and wait for the caches to assert idle.
+ * The cache flush is executed in the ME, but the PFP waits
+ * for completion.
+ */
+ radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 6, 0));
+ radeon_emit(cs, 0); /* CP_COHER_CNTL */
+ radeon_emit(cs, 0xffffffff); /* CP_COHER_SIZE */
+ radeon_emit(cs, 0xffffff); /* 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 */
+ radeon_emit(cs, gcr_cntl); /* GCR_CNTL */
+ } else if ((cb_db_event ||
+ (flush_bits & (RADV_CMD_FLAG_VS_PARTIAL_FLUSH |
+ RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
+ RADV_CMD_FLAG_CS_PARTIAL_FLUSH)))
+ && !is_mec) {
+ /* We need to ensure that PFP waits as well. */
+ radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
+ radeon_emit(cs, 0);
+ }
+
+ if (flush_bits & RADV_CMD_FLAG_START_PIPELINE_STATS) {
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_START) |
+ EVENT_INDEX(0));
+ } else if (flush_bits & RADV_CMD_FLAG_STOP_PIPELINE_STATS) {
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP) |
+ EVENT_INDEX(0));
+ }
+}
+
void
si_cs_emit_cache_flush(struct radeon_cmdbuf *cs,
enum chip_class chip_class,
uint32_t *flush_cnt,
uint64_t flush_va,
bool is_mec,
- enum radv_cmd_flush_bits flush_bits)
+ enum radv_cmd_flush_bits flush_bits,
+ uint64_t gfx9_eop_bug_va)
{
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 (chip_class >= GFX10) {
+ /* GFX10 cache flush handling is quite different. */
+ gfx10_cs_emit_cache_flush(cs, chip_class, flush_cnt, flush_va,
+ is_mec, flush_bits, gfx9_eop_bug_va);
+ return;
+ }
+
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)
+ if (flush_bits & RADV_CMD_FLAG_INV_SCACHE)
cp_coher_cntl |= S_0085F0_SH_KCACHE_ACTION_ENA(1);
- if (chip_class <= VI) {
+ if (chip_class <= GFX8) {
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_CB7_DEST_BASE_ENA(1);
/* Necessary for DCC */
- if (chip_class >= VI) {
+ if (chip_class >= GFX8) {
si_cs_emit_write_event_eop(cs,
- false,
chip_class,
is_mec,
V_028A90_FLUSH_AND_INV_CB_DATA_TS,
0,
+ EOP_DST_SEL_MEM,
EOP_DATA_SEL_DISCARD,
- 0, 0, 0);
+ 0, 0,
+ gfx9_eop_bug_va);
}
}
if (flush_bits & RADV_CMD_FLAG_FLUSH_AND_INV_DB) {
radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
}
- if (chip_class >= GFX9 && flush_cb_db) {
+ if (chip_class == GFX9 && flush_cb_db) {
unsigned cb_db_event, tc_flags;
/* Set the CB/DB flush event. */
EVENT_TC_MD_ACTION_ENA;
/* Ideally flush TC together with CB/DB. */
- if (flush_bits & RADV_CMD_FLAG_INV_GLOBAL_L2) {
+ if (flush_bits & RADV_CMD_FLAG_INV_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);
+ flush_bits &= ~(RADV_CMD_FLAG_INV_L2 |
+ RADV_CMD_FLAG_WB_L2 |
+ RADV_CMD_FLAG_INV_VCACHE);
}
assert(flush_cnt);
- uint32_t old_fence = (*flush_cnt)++;
+ (*flush_cnt)++;
- si_cs_emit_write_event_eop(cs, false, chip_class, false, cb_db_event, tc_flags,
+ si_cs_emit_write_event_eop(cs, chip_class, false, cb_db_event, tc_flags,
+ EOP_DST_SEL_MEM,
EOP_DATA_SEL_VALUE_32BIT,
- flush_va, old_fence, *flush_cnt);
- si_emit_wait_fence(cs, false, flush_va, *flush_cnt, 0xffffffff);
+ flush_va, *flush_cnt,
+ gfx9_eop_bug_va);
+ radv_cp_wait_mem(cs, WAIT_REG_MEM_EQUAL, flush_va,
+ *flush_cnt, 0xffffffff);
}
/* VGT state sync */
radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
}
+ /* VGT streamout state sync */
+ if (flush_bits & RADV_CMD_FLAG_VGT_STREAMOUT_SYNC) {
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC) | 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 |
- RADV_CMD_FLAG_INV_VMEM_L1 |
- RADV_CMD_FLAG_INV_GLOBAL_L2 |
- RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2))) &&
+ RADV_CMD_FLAG_INV_VCACHE |
+ RADV_CMD_FLAG_INV_L2 |
+ RADV_CMD_FLAG_WB_L2))) &&
!is_mec) {
radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
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, false, chip_class >= GFX9,
+ if ((flush_bits & RADV_CMD_FLAG_INV_L2) ||
+ (chip_class <= GFX7 && (flush_bits & RADV_CMD_FLAG_WB_L2))) {
+ si_emit_acquire_mem(cs, is_mec, 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));
+ S_0301F0_TC_WB_ACTION_ENA(chip_class >= GFX8));
cp_coher_cntl = 0;
} else {
- if(flush_bits & RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2) {
+ if(flush_bits & RADV_CMD_FLAG_WB_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, false,
- chip_class >= GFX9,
+ si_emit_acquire_mem(cs, is_mec,
+ 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) {
+ if (flush_bits & RADV_CMD_FLAG_INV_VCACHE) {
si_emit_acquire_mem(cs, is_mec,
- false, chip_class >= GFX9,
+ 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, false, chip_class >= GFX9, cp_coher_cntl);
+ si_emit_acquire_mem(cs, is_mec, chip_class == GFX9, cp_coher_cntl);
+
+ if (flush_bits & RADV_CMD_FLAG_START_PIPELINE_STATS) {
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_START) |
+ EVENT_INDEX(0));
+ } else if (flush_bits & RADV_CMD_FLAG_STOP_PIPELINE_STATS) {
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+ radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP) |
+ EVENT_INDEX(0));
+ }
}
void
RADV_CMD_FLAG_FLUSH_AND_INV_DB_META |
RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
RADV_CMD_FLAG_VS_PARTIAL_FLUSH |
- RADV_CMD_FLAG_VGT_FLUSH);
+ RADV_CMD_FLAG_VGT_FLUSH |
+ RADV_CMD_FLAG_START_PIPELINE_STATS |
+ RADV_CMD_FLAG_STOP_PIPELINE_STATS);
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->device->physical_device->rad_info.chip_class,
- ptr, va,
+ &cmd_buffer->gfx9_fence_idx,
+ cmd_buffer->gfx9_fence_va,
radv_cmd_buffer_uses_mec(cmd_buffer),
- cmd_buffer->state.flush_bits);
+ cmd_buffer->state.flush_bits,
+ cmd_buffer->gfx9_eop_bug_va);
if (unlikely(cmd_buffer->device->trace_bo))
radv_cmd_buffer_trace_emit(cmd_buffer);
+ /* Clear the caches that have been flushed to avoid syncing too much
+ * when there is some pending active queries.
+ */
+ cmd_buffer->active_query_flush_bits &= ~cmd_buffer->state.flush_bits;
+
cmd_buffer->state.flush_bits = 0;
+
+ /* If the driver used a compute shader for resetting a query pool, it
+ * should be finished at this point.
+ */
+ cmd_buffer->pending_reset_query = false;
}
/* 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)
+si_emit_set_predication_state(struct radv_cmd_buffer *cmd_buffer,
+ bool draw_visible, uint64_t va)
{
uint32_t op = 0;
- if (va)
- op = PRED_OP(PREDICATION_OP_BOOL64) | PREDICATION_DRAW_VISIBLE;
+ if (va) {
+ op = PRED_OP(PREDICATION_OP_BOOL64);
+
+ /* PREDICATION_DRAW_VISIBLE means that if the 32-bit value is
+ * zero, all rendering commands are discarded. Otherwise, they
+ * are discarded if the value is non zero.
+ */
+ op |= draw_visible ? PREDICATION_DRAW_VISIBLE :
+ PREDICATION_DRAW_NOT_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);
struct radeon_cmdbuf *cs = cmd_buffer->cs;
uint32_t header = 0, command = 0;
- assert(size);
assert(size <= cp_dma_max_byte_count(cmd_buffer));
radeon_check_space(cmd_buffer->device->ws, cmd_buffer->cs, 9);
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 */
+ header |= S_411_DST_SEL(V_411_NOWHERE); /* prefetch only */
else if (flags & CP_DMA_USE_L2)
- header |= S_411_DSL_SEL(V_411_DST_ADDR_TC_L2);
+ header |= S_411_DST_SEL(V_411_DST_ADDR_TC_L2);
if (flags & CP_DMA_CLEAR)
header |= S_411_SRC_SEL(V_411_DATA);
else if (flags & CP_DMA_USE_L2)
header |= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2);
- if (cmd_buffer->device->physical_device->rad_info.chip_class >= CIK) {
+ if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX7) {
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] */
* indices. If we wanted to execute CP DMA in PFP, this packet
* 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, cmd_buffer->state.predicating));
- radeon_emit(cs, 0);
+ if (flags & CP_DMA_SYNC) {
+ if (cmd_buffer->queue_family_index == RADV_QUEUE_GENERAL) {
+ radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, cmd_buffer->state.predicating));
+ radeon_emit(cs, 0);
+ }
+
+ /* CP will see the sync flag and wait for all DMAs to complete. */
+ cmd_buffer->state.dma_is_busy = false;
}
if (unlikely(cmd_buffer->device->trace_bo))
uint64_t main_src_va, main_dest_va;
uint64_t skipped_size = 0, realign_size = 0;
+ /* Assume that we are not going to sync after the last DMA operation. */
+ cmd_buffer->state.dma_is_busy = true;
if (cmd_buffer->device->physical_device->rad_info.family <= CHIP_CARRIZO ||
cmd_buffer->device->physical_device->rad_info.family == CHIP_STONEY) {
unsigned dma_flags = 0;
unsigned byte_count = MIN2(size, cp_dma_max_byte_count(cmd_buffer));
+ if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX10) {
+ /* DMA operations via L2 are coherent and faster.
+ * TODO: GFX7-GFX9 should also support this but it
+ * requires tests/benchmarks.
+ */
+ dma_flags |= CP_DMA_USE_L2;
+ }
+
si_cp_dma_prepare(cmd_buffer, byte_count,
size + skipped_size + realign_size,
&dma_flags);
+ dma_flags &= ~CP_DMA_SYNC;
+
si_emit_cp_dma(cmd_buffer, main_dest_va, main_src_va,
byte_count, dma_flags);
assert(va % 4 == 0 && size % 4 == 0);
+ /* Assume that we are not going to sync after the last DMA operation. */
+ cmd_buffer->state.dma_is_busy = true;
+
while (size) {
unsigned byte_count = MIN2(size, cp_dma_max_byte_count(cmd_buffer));
unsigned dma_flags = CP_DMA_CLEAR;
+ if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX10) {
+ /* DMA operations via L2 are coherent and faster.
+ * TODO: GFX7-GFX9 should also support this but it
+ * requires tests/benchmarks.
+ */
+ dma_flags |= CP_DMA_USE_L2;
+ }
+
si_cp_dma_prepare(cmd_buffer, byte_count, size, &dma_flags);
/* Emit the clear packet. */
}
}
+void si_cp_dma_wait_for_idle(struct radv_cmd_buffer *cmd_buffer)
+{
+ if (cmd_buffer->device->physical_device->rad_info.chip_class < GFX7)
+ return;
+
+ if (!cmd_buffer->state.dma_is_busy)
+ return;
+
+ /* Issue a dummy DMA that copies zero bytes.
+ *
+ * The DMA engine will see that there's no work to do and skip this
+ * DMA request, however, the CP will see the sync flag and still wait
+ * for all DMAs to complete.
+ */
+ si_emit_cp_dma(cmd_buffer, 0, 0, 0, CP_DMA_SYNC);
+
+ cmd_buffer->state.dma_is_busy = false;
+}
+
/* For MSAA sample positions. */
#define FILL_SREG(s0x, s0y, s1x, s1y, s2x, s2y, s3x, s3y) \
- (((s0x) & 0xf) | (((unsigned)(s0y) & 0xf) << 4) | \
- (((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) | \
- (((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) | \
+ ((((unsigned)(s0x) & 0xf) << 0) | (((unsigned)(s0y) & 0xf) << 4) | \
+ (((unsigned)(s1x) & 0xf) << 8) | (((unsigned)(s1y) & 0xf) << 12) | \
+ (((unsigned)(s2x) & 0xf) << 16) | (((unsigned)(s2y) & 0xf) << 20) | \
(((unsigned)(s3x) & 0xf) << 24) | (((unsigned)(s3y) & 0xf) << 28))
-
-/* 2xMSAA
- * There are two locations (4, 4), (-4, -4). */
-const uint32_t eg_sample_locs_2x[4] = {
- FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
- FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
- FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
- FILL_SREG(4, 4, -4, -4, 4, 4, -4, -4),
-};
-const unsigned eg_max_dist_2x = 4;
-/* 4xMSAA
- * There are 4 locations: (-2, 6), (6, -2), (-6, 2), (2, 6). */
-const uint32_t eg_sample_locs_4x[4] = {
- FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
- FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
- FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
- FILL_SREG(-2, -6, 6, -2, -6, 2, 2, 6),
-};
-const unsigned eg_max_dist_4x = 6;
-
-/* Cayman 8xMSAA */
-static const uint32_t cm_sample_locs_8x[] = {
- FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
- FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
- FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
- FILL_SREG( 1, -3, -1, 3, 5, 1, -3, -5),
- FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
- FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
- FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
- FILL_SREG(-5, 5, -7, -1, 3, 7, 7, -7),
-};
-static const unsigned cm_max_dist_8x = 8;
-/* Cayman 16xMSAA */
-static const uint32_t cm_sample_locs_16x[] = {
- FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
- FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
- FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
- FILL_SREG( 1, 1, -1, -3, -3, 2, 4, -1),
- FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
- FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
- FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
- FILL_SREG(-5, -2, 2, 5, 5, 3, 3, -5),
- FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
- FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
- FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
- FILL_SREG(-2, 6, 0, -7, -4, -6, -6, 4),
- FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
- FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
- FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
- FILL_SREG(-8, 0, 7, -4, 6, 7, -7, -8),
+/* For obtaining location coordinates from registers */
+#define SEXT4(x) ((int)((x) | ((x) & 0x8 ? 0xfffffff0 : 0)))
+#define GET_SFIELD(reg, index) SEXT4(((reg) >> ((index) * 4)) & 0xf)
+#define GET_SX(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2)
+#define GET_SY(reg, index) GET_SFIELD((reg)[(index) / 4], ((index) % 4) * 2 + 1)
+
+/* 1x MSAA */
+static const uint32_t sample_locs_1x =
+ FILL_SREG(0, 0, 0, 0, 0, 0, 0, 0);
+static const unsigned max_dist_1x = 0;
+static const uint64_t centroid_priority_1x = 0x0000000000000000ull;
+
+/* 2xMSAA */
+static const uint32_t sample_locs_2x =
+ FILL_SREG(4,4, -4, -4, 0, 0, 0, 0);
+static const unsigned max_dist_2x = 4;
+static const uint64_t centroid_priority_2x = 0x1010101010101010ull;
+
+/* 4xMSAA */
+static const uint32_t sample_locs_4x =
+ FILL_SREG(-2,-6, 6, -2, -6, 2, 2, 6);
+static const unsigned max_dist_4x = 6;
+static const uint64_t centroid_priority_4x = 0x3210321032103210ull;
+
+/* 8xMSAA */
+static const uint32_t sample_locs_8x[] = {
+ FILL_SREG( 1,-3, -1, 3, 5, 1, -3,-5),
+ FILL_SREG(-5, 5, -7,-1, 3, 7, 7,-7),
+ /* The following are unused by hardware, but we emit them to IBs
+ * instead of multiple SET_CONTEXT_REG packets. */
+ 0,
+ 0,
};
-static const unsigned cm_max_dist_16x = 8;
+static const unsigned max_dist_8x = 7;
+static const uint64_t centroid_priority_8x = 0x7654321076543210ull;
-unsigned radv_cayman_get_maxdist(int log_samples)
+unsigned radv_get_default_max_sample_dist(int log_samples)
{
unsigned max_dist[] = {
- 0,
- eg_max_dist_2x,
- eg_max_dist_4x,
- cm_max_dist_8x,
- cm_max_dist_16x
+ max_dist_1x,
+ max_dist_2x,
+ max_dist_4x,
+ max_dist_8x,
};
return max_dist[log_samples];
}
-void radv_cayman_emit_msaa_sample_locs(struct radeon_cmdbuf *cs, int nr_samples)
+void radv_emit_default_sample_locations(struct radeon_cmdbuf *cs, int nr_samples)
{
switch (nr_samples) {
default:
case 1:
- 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);
+ radeon_set_context_reg_seq(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 2);
+ radeon_emit(cs, (uint32_t)centroid_priority_1x);
+ radeon_emit(cs, centroid_priority_1x >> 32);
+ radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, sample_locs_1x);
+ radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, sample_locs_1x);
+ radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, sample_locs_1x);
+ radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, sample_locs_1x);
break;
case 2:
- 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]);
+ radeon_set_context_reg_seq(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 2);
+ radeon_emit(cs, (uint32_t)centroid_priority_2x);
+ radeon_emit(cs, centroid_priority_2x >> 32);
+ radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, sample_locs_2x);
+ radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, sample_locs_2x);
+ radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, sample_locs_2x);
+ radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, sample_locs_2x);
break;
case 4:
- 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]);
+ radeon_set_context_reg_seq(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 2);
+ radeon_emit(cs, (uint32_t)centroid_priority_4x);
+ radeon_emit(cs, centroid_priority_4x >> 32);
+ radeon_set_context_reg(cs, R_028BF8_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y0_0, sample_locs_4x);
+ radeon_set_context_reg(cs, R_028C08_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y0_0, sample_locs_4x);
+ radeon_set_context_reg(cs, R_028C18_PA_SC_AA_SAMPLE_LOCS_PIXEL_X0Y1_0, sample_locs_4x);
+ radeon_set_context_reg(cs, R_028C28_PA_SC_AA_SAMPLE_LOCS_PIXEL_X1Y1_0, sample_locs_4x);
break;
case 8:
+ radeon_set_context_reg_seq(cs, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 2);
+ radeon_emit(cs, (uint32_t)centroid_priority_8x);
+ radeon_emit(cs, centroid_priority_8x >> 32);
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, 0);
- radeon_emit(cs, cm_sample_locs_8x[1]);
- radeon_emit(cs, cm_sample_locs_8x[5]);
- radeon_emit(cs, 0);
- radeon_emit(cs, 0);
- radeon_emit(cs, cm_sample_locs_8x[2]);
- radeon_emit(cs, cm_sample_locs_8x[6]);
- radeon_emit(cs, 0);
- radeon_emit(cs, 0);
- radeon_emit(cs, cm_sample_locs_8x[3]);
- radeon_emit(cs, cm_sample_locs_8x[7]);
- break;
- case 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]);
- radeon_emit(cs, cm_sample_locs_16x[12]);
- radeon_emit(cs, cm_sample_locs_16x[1]);
- radeon_emit(cs, cm_sample_locs_16x[5]);
- radeon_emit(cs, cm_sample_locs_16x[9]);
- radeon_emit(cs, cm_sample_locs_16x[13]);
- radeon_emit(cs, cm_sample_locs_16x[2]);
- radeon_emit(cs, cm_sample_locs_16x[6]);
- radeon_emit(cs, cm_sample_locs_16x[10]);
- radeon_emit(cs, cm_sample_locs_16x[14]);
- radeon_emit(cs, cm_sample_locs_16x[3]);
- radeon_emit(cs, cm_sample_locs_16x[7]);
- radeon_emit(cs, cm_sample_locs_16x[11]);
- radeon_emit(cs, cm_sample_locs_16x[15]);
+ radeon_emit_array(cs, sample_locs_8x, 4);
+ radeon_emit_array(cs, sample_locs_8x, 4);
+ radeon_emit_array(cs, sample_locs_8x, 4);
+ radeon_emit_array(cs, sample_locs_8x, 2);
break;
}
}
-static void radv_cayman_get_sample_position(struct radv_device *device,
- unsigned sample_count,
- unsigned sample_index, float *out_value)
+static void radv_get_sample_position(struct radv_device *device,
+ unsigned sample_count,
+ unsigned sample_index, float *out_value)
{
- int offset, index;
- struct {
- int idx:4;
- } val;
+ const uint32_t *sample_locs;
+
switch (sample_count) {
case 1:
default:
- out_value[0] = out_value[1] = 0.5;
+ sample_locs = &sample_locs_1x;
break;
case 2:
- offset = 4 * (sample_index * 2);
- val.idx = (eg_sample_locs_2x[0] >> offset) & 0xf;
- out_value[0] = (float)(val.idx + 8) / 16.0f;
- val.idx = (eg_sample_locs_2x[0] >> (offset + 4)) & 0xf;
- out_value[1] = (float)(val.idx + 8) / 16.0f;
+ sample_locs = &sample_locs_2x;
break;
case 4:
- offset = 4 * (sample_index * 2);
- val.idx = (eg_sample_locs_4x[0] >> offset) & 0xf;
- out_value[0] = (float)(val.idx + 8) / 16.0f;
- val.idx = (eg_sample_locs_4x[0] >> (offset + 4)) & 0xf;
- out_value[1] = (float)(val.idx + 8) / 16.0f;
+ sample_locs = &sample_locs_4x;
break;
case 8:
- offset = 4 * (sample_index % 4 * 2);
- index = (sample_index / 4) * 4;
- val.idx = (cm_sample_locs_8x[index] >> offset) & 0xf;
- out_value[0] = (float)(val.idx + 8) / 16.0f;
- val.idx = (cm_sample_locs_8x[index] >> (offset + 4)) & 0xf;
- out_value[1] = (float)(val.idx + 8) / 16.0f;
- break;
- case 16:
- offset = 4 * (sample_index % 4 * 2);
- index = (sample_index / 4) * 4;
- val.idx = (cm_sample_locs_16x[index] >> offset) & 0xf;
- out_value[0] = (float)(val.idx + 8) / 16.0f;
- val.idx = (cm_sample_locs_16x[index] >> (offset + 4)) & 0xf;
- out_value[1] = (float)(val.idx + 8) / 16.0f;
+ sample_locs = sample_locs_8x;
break;
}
+
+ out_value[0] = (GET_SX(sample_locs, sample_index) + 8) / 16.0f;
+ out_value[1] = (GET_SY(sample_locs, sample_index) + 8) / 16.0f;
}
void radv_device_init_msaa(struct radv_device *device)
{
int i;
- radv_cayman_get_sample_position(device, 1, 0, device->sample_locations_1x[0]);
+
+ radv_get_sample_position(device, 1, 0, device->sample_locations_1x[0]);
for (i = 0; i < 2; i++)
- radv_cayman_get_sample_position(device, 2, i, device->sample_locations_2x[i]);
+ radv_get_sample_position(device, 2, i, device->sample_locations_2x[i]);
for (i = 0; i < 4; i++)
- radv_cayman_get_sample_position(device, 4, i, device->sample_locations_4x[i]);
+ radv_get_sample_position(device, 4, i, device->sample_locations_4x[i]);
for (i = 0; i < 8; i++)
- radv_cayman_get_sample_position(device, 8, i, device->sample_locations_8x[i]);
- for (i = 0; i < 16; i++)
- radv_cayman_get_sample_position(device, 16, i, device->sample_locations_16x[i]);
+ radv_get_sample_position(device, 8, i, device->sample_locations_8x[i]);
}
projects / mesa.git / blobdiff