#include "radv_cs.h"
#include "sid.h"
#include "radv_util.h"
-#include "main/macros.h"
static void
si_write_harvested_raster_configs(struct radv_physical_device *physical_device,
}
void
-si_emit_graphics(struct radv_physical_device *physical_device,
+si_emit_graphics(struct radv_device *device,
struct radeon_cmdbuf *cs)
{
+ struct radv_physical_device *physical_device = device->physical_device;
+
bool has_clear_state = physical_device->rad_info.has_clear_state;
int i;
- /* Since amdgpu version 3.6.0, CONTEXT_CONTROL is emitted by the kernel */
- if (physical_device->rad_info.drm_minor < 6) {
- 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, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
+ radeon_emit(cs, CC0_UPDATE_LOAD_ENABLES(1));
+ radeon_emit(cs, CC1_UPDATE_SHADOW_ENABLES(1));
if (has_clear_state) {
radeon_emit(cs, PKT3(PKT3_CLEAR_STATE, 0, 0));
}
/* Compute LATE_ALLOC_VS.LIMIT. */
- unsigned num_cu_per_sh = physical_device->rad_info.num_good_cu_per_sh;
- unsigned late_alloc_limit; /* The limit is per SH. */
-
- if (physical_device->rad_info.family == CHIP_KABINI) {
- late_alloc_limit = 0; /* Potential hang on Kabini. */
- } else if (num_cu_per_sh <= 4) {
- /* Too few available compute units per SH. 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_limit = 2;
- } else {
- /* This is a good initial value, allowing 1 late_alloc
- * wave per SIMD on num_cu - 2.
- */
- late_alloc_limit = (num_cu_per_sh - 2) * 4;
- }
-
- unsigned late_alloc_limit_gs = late_alloc_limit;
+ unsigned num_cu_per_sh = physical_device->rad_info.min_good_cu_per_sa;
+ unsigned late_alloc_wave64 = 0; /* The limit is per SH. */
+ unsigned late_alloc_wave64_gs = 0;
unsigned cu_mask_vs = 0xffff;
unsigned cu_mask_gs = 0xffff;
- if (late_alloc_limit > 2) {
- if (physical_device->rad_info.chip_class >= GFX10) {
+ 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.
+ */
+ 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.
+ */
+ if (!physical_device->use_ngg ||
+ physical_device->rad_info.family == CHIP_NAVI14) {
+ late_alloc_wave64_gs = 0;
+ cu_mask_gs = 0xffff;
+ }
+ } 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 SH.
+ * 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 {
- cu_mask_vs = 0xfffe; /* 1 CU disabled */
+ /* 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;
}
- }
- /* Don't use late alloc for NGG on Navi14 due to a hw bug. */
- if (physical_device->rad_info.family == CHIP_NAVI14) {
- late_alloc_limit_gs = 0;
- cu_mask_gs = 0xffff;
+ 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_limit));
+ 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_limit_gs));
+ 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,
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);
+
+ /* Enable CMASK/FMASK/HTILE/DCC caching in L2 for small chips. */
+ unsigned meta_write_policy, meta_read_policy;
+
+ /* TODO: investigate whether LRU improves performance on other chips too */
+ if (physical_device->rad_info.num_render_backends <= 4) {
+ meta_write_policy = V_02807C_CACHE_LRU_WR; /* cache writes */
+ meta_read_policy = V_02807C_CACHE_LRU_RD; /* cache reads */
+ } else {
+ meta_write_policy = V_02807C_CACHE_STREAM_WR; /* write combine */
+ meta_read_policy = V_02807C_CACHE_NOA_RD; /* don't cache reads */
+ }
+
radeon_set_context_reg(cs, R_02807C_DB_RMI_L2_CACHE_CONTROL,
S_02807C_Z_WR_POLICY(V_02807C_CACHE_STREAM_WR) |
S_02807C_S_WR_POLICY(V_02807C_CACHE_STREAM_WR) |
- S_02807C_HTILE_WR_POLICY(V_02807C_CACHE_STREAM_WR) |
+ S_02807C_HTILE_WR_POLICY(meta_write_policy) |
S_02807C_ZPCPSD_WR_POLICY(V_02807C_CACHE_STREAM_WR) |
S_02807C_Z_RD_POLICY(V_02807C_CACHE_NOA_RD) |
S_02807C_S_RD_POLICY(V_02807C_CACHE_NOA_RD) |
- S_02807C_HTILE_RD_POLICY(V_02807C_CACHE_NOA_RD));
+ S_02807C_HTILE_RD_POLICY(meta_read_policy));
radeon_set_context_reg(cs, R_028410_CB_RMI_GL2_CACHE_CONTROL,
- S_028410_CMASK_WR_POLICY(V_028410_CACHE_STREAM_WR) |
- S_028410_FMASK_WR_POLICY(V_028410_CACHE_STREAM_WR) |
- S_028410_DCC_WR_POLICY(V_028410_CACHE_STREAM_WR) |
+ S_028410_CMASK_WR_POLICY(meta_write_policy) |
+ S_028410_FMASK_WR_POLICY(meta_write_policy) |
+ S_028410_DCC_WR_POLICY(meta_write_policy) |
S_028410_COLOR_WR_POLICY(V_028410_CACHE_STREAM_WR) |
- S_028410_CMASK_RD_POLICY(V_028410_CACHE_NOA_RD) |
- S_028410_FMASK_RD_POLICY(V_028410_CACHE_NOA_RD) |
- S_028410_DCC_RD_POLICY(V_028410_CACHE_NOA_RD) |
+ S_028410_CMASK_RD_POLICY(meta_read_policy) |
+ S_028410_FMASK_RD_POLICY(meta_read_policy) |
+ S_028410_DCC_RD_POLICY(meta_read_policy) |
S_028410_COLOR_RD_POLICY(V_028410_CACHE_NOA_RD));
radeon_set_context_reg(cs, R_028428_CB_COVERAGE_OUT_CONTROL, 0);
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.family == CHIP_NAVI10 ||
- physical_device->rad_info.family == CHIP_NAVI12 ||
- physical_device->rad_info.family == CHIP_NAVI14) {
+ 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));
/* TODO: For culling, replace 128 with 256. */
radeon_set_uconfig_reg(cs, R_030980_GE_PC_ALLOC,
- S_030980_OVERSUB_EN(1) |
+ 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 >= GFX8) {
+ 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;
vgt_tess_distribution = S_028B50_ACCUM_ISOLINE(32) |
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;
- unsigned max_alloc_count = 0;
-
- switch (physical_device->rad_info.family) {
- case CHIP_VEGA10:
- case CHIP_VEGA12:
- case CHIP_VEGA20:
- pc_lines = 4096;
- break;
- case CHIP_RAVEN:
- case CHIP_RAVEN2:
- case CHIP_RENOIR:
- case CHIP_NAVI10:
- case CHIP_NAVI12:
- pc_lines = 1024;
- break;
- case CHIP_NAVI14:
- pc_lines = 512;
- break;
- default:
- assert(0);
- }
+ if (device->border_color_data.bo) {
+ uint64_t border_color_va = radv_buffer_get_va(device->border_color_data.bo);
- if (physical_device->rad_info.chip_class >= GFX10) {
- max_alloc_count = pc_lines / 3;
- } else {
- max_alloc_count = MIN2(128, pc_lines / (4 * num_se));
+ 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(max_alloc_count - 1) |
+ 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));
if (!cs)
return;
- si_emit_graphics(device->physical_device, cs);
+ si_emit_graphics(device, cs);
while (cs->cdw & 7) {
if (device->physical_device->rad_info.gfx_ib_pad_with_type2)
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;
}
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_GLM_INV(1);
+ 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. */
- gcr_cntl |= S_586_GL2_WB(1);
+ /* 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+.
- if (flush_bits & RADV_CMD_FLAG_INV_L2_METADATA)
- gcr_cntl |= S_586_GLM_INV(1);
+ 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)) {
projects / mesa.git / blobdiff