*/
#include "ac_gpu_info.h"
+#include "addrlib/src/amdgpu_asic_addr.h"
#include "sid.h"
-#include "gfx9d.h"
+#include "util/macros.h"
#include "util/u_math.h"
#include <stdio.h>
#include <xf86drm.h>
-#include <amdgpu_drm.h>
+#include "drm-uapi/amdgpu_drm.h"
#include <amdgpu.h>
case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16:
return 16;
default:
- fprintf(stderr, "Invalid CIK pipe configuration, assuming P2\n");
+ fprintf(stderr, "Invalid GFX7 pipe configuration, assuming P2\n");
assert(!"this should never occur");
return 2;
}
return value ? true : false;
}
-bool ac_query_gpu_info(int fd, amdgpu_device_handle dev,
+static bool has_timeline_syncobj(int fd)
+{
+ uint64_t value;
+ if (drmGetCap(fd, DRM_CAP_SYNCOBJ_TIMELINE, &value))
+ return false;
+ return value ? true : false;
+}
+
+static uint64_t fix_vram_size(uint64_t size)
+{
+ /* The VRAM size is underreported, so we need to fix it, because
+ * it's used to compute the number of memory modules for harvesting.
+ */
+ return align64(size, 256*1024*1024);
+}
+
+static uint32_t
+get_l2_cache_size(enum radeon_family family)
+{
+ switch (family) {
+ case CHIP_KABINI:
+ case CHIP_STONEY:
+ return 128 * 1024;
+ case CHIP_OLAND:
+ case CHIP_HAINAN:
+ case CHIP_ICELAND:
+ return 256 * 1024;
+ case CHIP_PITCAIRN:
+ case CHIP_VERDE:
+ case CHIP_BONAIRE:
+ case CHIP_KAVERI:
+ case CHIP_POLARIS12:
+ case CHIP_CARRIZO:
+ return 512 * 1024;
+ case CHIP_TAHITI:
+ case CHIP_TONGA:
+ return 768 * 1024;
+ break;
+ case CHIP_HAWAII:
+ case CHIP_POLARIS11:
+ return 1024 * 1024;
+ case CHIP_FIJI:
+ case CHIP_POLARIS10:
+ return 2048 * 1024;
+ break;
+ default:
+ return 4096 * 1024;
+ }
+}
+
+bool ac_query_gpu_info(int fd, void *dev_p,
struct radeon_info *info,
struct amdgpu_gpu_info *amdinfo)
{
+ struct drm_amdgpu_info_device device_info = {};
struct amdgpu_buffer_size_alignments alignment_info = {};
- struct amdgpu_heap_info vram, vram_vis, gtt;
struct drm_amdgpu_info_hw_ip dma = {}, compute = {}, uvd = {};
- struct drm_amdgpu_info_hw_ip uvd_enc = {}, vce = {}, vcn_dec = {};
+ struct drm_amdgpu_info_hw_ip uvd_enc = {}, vce = {}, vcn_dec = {}, vcn_jpeg = {};
struct drm_amdgpu_info_hw_ip vcn_enc = {}, gfx = {};
struct amdgpu_gds_resource_info gds = {};
uint32_t vce_version = 0, vce_feature = 0, uvd_version = 0, uvd_feature = 0;
int r, i, j;
+ amdgpu_device_handle dev = dev_p;
drmDevicePtr devinfo;
/* Get PCI info. */
info->pci_func = devinfo->businfo.pci->func;
drmFreeDevice(&devinfo);
+ assert(info->drm_major == 3);
+ info->is_amdgpu = true;
+
/* Query hardware and driver information. */
r = amdgpu_query_gpu_info(dev, amdinfo);
if (r) {
return false;
}
- r = amdgpu_query_buffer_size_alignment(dev, &alignment_info);
- if (r) {
- fprintf(stderr, "amdgpu: amdgpu_query_buffer_size_alignment failed.\n");
- return false;
- }
-
- r = amdgpu_query_heap_info(dev, AMDGPU_GEM_DOMAIN_VRAM, 0, &vram);
- if (r) {
- fprintf(stderr, "amdgpu: amdgpu_query_heap_info(vram) failed.\n");
- return false;
- }
-
- r = amdgpu_query_heap_info(dev, AMDGPU_GEM_DOMAIN_VRAM,
- AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
- &vram_vis);
+ r = amdgpu_query_info(dev, AMDGPU_INFO_DEV_INFO, sizeof(device_info),
+ &device_info);
if (r) {
- fprintf(stderr, "amdgpu: amdgpu_query_heap_info(vram_vis) failed.\n");
+ fprintf(stderr, "amdgpu: amdgpu_query_info(dev_info) failed.\n");
return false;
}
- r = amdgpu_query_heap_info(dev, AMDGPU_GEM_DOMAIN_GTT, 0, >t);
+ r = amdgpu_query_buffer_size_alignment(dev, &alignment_info);
if (r) {
- fprintf(stderr, "amdgpu: amdgpu_query_heap_info(gtt) failed.\n");
+ fprintf(stderr, "amdgpu: amdgpu_query_buffer_size_alignment failed.\n");
return false;
}
return false;
}
- if (info->drm_major == 3 && info->drm_minor >= 17) {
+ if (info->drm_minor >= 17) {
r = amdgpu_query_hw_ip_info(dev, AMDGPU_HW_IP_UVD_ENC, 0, &uvd_enc);
if (r) {
fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(uvd_enc) failed.\n");
}
}
- if (info->drm_major == 3 && info->drm_minor >= 17) {
+ if (info->drm_minor >= 17) {
r = amdgpu_query_hw_ip_info(dev, AMDGPU_HW_IP_VCN_DEC, 0, &vcn_dec);
if (r) {
fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(vcn_dec) failed.\n");
}
}
- if (info->drm_major == 3 && info->drm_minor >= 17) {
+ if (info->drm_minor >= 17) {
r = amdgpu_query_hw_ip_info(dev, AMDGPU_HW_IP_VCN_ENC, 0, &vcn_enc);
if (r) {
fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(vcn_enc) failed.\n");
}
}
+ if (info->drm_minor >= 27) {
+ r = amdgpu_query_hw_ip_info(dev, AMDGPU_HW_IP_VCN_JPEG, 0, &vcn_jpeg);
+ if (r) {
+ fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(vcn_jpeg) failed.\n");
+ return false;
+ }
+ }
+
r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_GFX_ME, 0, 0,
&info->me_fw_version,
&info->me_fw_feature);
return false;
}
+ if (info->drm_minor >= 9) {
+ struct drm_amdgpu_memory_info meminfo = {};
+
+ r = amdgpu_query_info(dev, AMDGPU_INFO_MEMORY, sizeof(meminfo), &meminfo);
+ if (r) {
+ fprintf(stderr, "amdgpu: amdgpu_query_info(memory) failed.\n");
+ return false;
+ }
+
+ /* Note: usable_heap_size values can be random and can't be relied on. */
+ info->gart_size = meminfo.gtt.total_heap_size;
+ info->vram_size = fix_vram_size(meminfo.vram.total_heap_size);
+ info->vram_vis_size = meminfo.cpu_accessible_vram.total_heap_size;
+ } else {
+ /* This is a deprecated interface, which reports usable sizes
+ * (total minus pinned), but the pinned size computation is
+ * buggy, so the values returned from these functions can be
+ * random.
+ */
+ struct amdgpu_heap_info vram, vram_vis, gtt;
+
+ r = amdgpu_query_heap_info(dev, AMDGPU_GEM_DOMAIN_VRAM, 0, &vram);
+ if (r) {
+ fprintf(stderr, "amdgpu: amdgpu_query_heap_info(vram) failed.\n");
+ return false;
+ }
+
+ r = amdgpu_query_heap_info(dev, AMDGPU_GEM_DOMAIN_VRAM,
+ AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
+ &vram_vis);
+ if (r) {
+ fprintf(stderr, "amdgpu: amdgpu_query_heap_info(vram_vis) failed.\n");
+ return false;
+ }
+
+ r = amdgpu_query_heap_info(dev, AMDGPU_GEM_DOMAIN_GTT, 0, >t);
+ if (r) {
+ fprintf(stderr, "amdgpu: amdgpu_query_heap_info(gtt) failed.\n");
+ return false;
+ }
+
+ info->gart_size = gtt.heap_size;
+ info->vram_size = fix_vram_size(vram.heap_size);
+ info->vram_vis_size = vram_vis.heap_size;
+ }
+
/* Set chip identification. */
info->pci_id = amdinfo->asic_id; /* TODO: is this correct? */
+ info->pci_rev_id = amdinfo->pci_rev_id;
info->vce_harvest_config = amdinfo->vce_harvest_config;
- switch (info->pci_id) {
-#define CHIPSET(pci_id, cfamily) case pci_id: info->family = CHIP_##cfamily; break;
-#include "pci_ids/radeonsi_pci_ids.h"
-#undef CHIPSET
+#define identify_chip2(asic, chipname) \
+ if (ASICREV_IS(amdinfo->chip_external_rev, asic)) { \
+ info->family = CHIP_##chipname; \
+ info->name = #chipname; \
+ }
+#define identify_chip(chipname) identify_chip2(chipname, chipname)
+
+ switch (amdinfo->family_id) {
+ case FAMILY_SI:
+ identify_chip(TAHITI);
+ identify_chip(PITCAIRN);
+ identify_chip2(CAPEVERDE, VERDE);
+ identify_chip(OLAND);
+ identify_chip(HAINAN);
+ break;
+ case FAMILY_CI:
+ identify_chip(BONAIRE);
+ identify_chip(HAWAII);
+ break;
+ case FAMILY_KV:
+ identify_chip2(SPECTRE, KAVERI);
+ identify_chip2(SPOOKY, KAVERI);
+ identify_chip2(KALINDI, KABINI);
+ identify_chip2(GODAVARI, KABINI);
+ break;
+ case FAMILY_VI:
+ identify_chip(ICELAND);
+ identify_chip(TONGA);
+ identify_chip(FIJI);
+ identify_chip(POLARIS10);
+ identify_chip(POLARIS11);
+ identify_chip(POLARIS12);
+ identify_chip(VEGAM);
+ break;
+ case FAMILY_CZ:
+ identify_chip(CARRIZO);
+ identify_chip(STONEY);
+ break;
+ case FAMILY_AI:
+ identify_chip(VEGA10);
+ identify_chip(VEGA12);
+ identify_chip(VEGA20);
+ identify_chip(ARCTURUS);
+ break;
+ case FAMILY_RV:
+ identify_chip(RAVEN);
+ identify_chip(RAVEN2);
+ identify_chip(RENOIR);
+ break;
+ case FAMILY_NV:
+ identify_chip(NAVI10);
+ identify_chip(NAVI12);
+ identify_chip(NAVI14);
+ identify_chip(SIENNA_CICHLID);
+ identify_chip(NAVY_FLOUNDER);
+ break;
+ }
- default:
- fprintf(stderr, "amdgpu: Invalid PCI ID.\n");
+ if (!info->name) {
+ fprintf(stderr, "amdgpu: unknown (family_id, chip_external_rev): (%u, %u)\n",
+ amdinfo->family_id, amdinfo->chip_external_rev);
return false;
}
- if (info->family >= CHIP_VEGA10)
+ if (info->family >= CHIP_SIENNA_CICHLID)
+ info->chip_class = GFX10_3;
+ else if (info->family >= CHIP_NAVI10)
+ info->chip_class = GFX10;
+ else if (info->family >= CHIP_VEGA10)
info->chip_class = GFX9;
else if (info->family >= CHIP_TONGA)
- info->chip_class = VI;
+ info->chip_class = GFX8;
else if (info->family >= CHIP_BONAIRE)
- info->chip_class = CIK;
+ info->chip_class = GFX7;
else if (info->family >= CHIP_TAHITI)
- info->chip_class = SI;
+ info->chip_class = GFX6;
else {
fprintf(stderr, "amdgpu: Unknown family.\n");
return false;
}
+ info->family_id = amdinfo->family_id;
+ info->chip_external_rev = amdinfo->chip_external_rev;
+ info->marketing_name = amdgpu_get_marketing_name(dev);
+ info->is_pro_graphics = info->marketing_name &&
+ (!strcmp(info->marketing_name, "Pro") ||
+ !strcmp(info->marketing_name, "PRO") ||
+ !strcmp(info->marketing_name, "Frontier"));
+
/* Set which chips have dedicated VRAM. */
info->has_dedicated_vram =
!(amdinfo->ids_flags & AMDGPU_IDS_FLAGS_FUSION);
- /* Set hardware information. */
- info->gart_size = gtt.heap_size;
- info->vram_size = vram.heap_size;
- info->vram_vis_size = vram_vis.heap_size;
- info->gds_size = gds.gds_total_size;
- info->gds_gfx_partition_size = gds.gds_gfx_partition_size;
/* The kernel can split large buffers in VRAM but not in GTT, so large
* allocations can fail or cause buffer movement failures in the kernel.
*/
- info->max_alloc_size = MIN2(info->vram_size * 0.9, info->gart_size * 0.7);
- /* convert the shader clock from KHz to MHz */
+ if (info->has_dedicated_vram)
+ info->max_alloc_size = info->vram_size * 0.8;
+ else
+ info->max_alloc_size = info->gart_size * 0.7;
+
+ info->vram_type = amdinfo->vram_type;
+ info->vram_bit_width = amdinfo->vram_bit_width;
+ info->ce_ram_size = amdinfo->ce_ram_size;
+
+ info->l2_cache_size = get_l2_cache_size(info->family);
+ info->l1_cache_size = 16384;
+
+ /* Set which chips have uncached device memory. */
+ info->has_l2_uncached = info->chip_class >= GFX9;
+
+ /* Set hardware information. */
+ info->gds_size = gds.gds_total_size;
+ info->gds_gfx_partition_size = gds.gds_gfx_partition_size;
+ /* convert the shader/memory clocks from KHz to MHz */
info->max_shader_clock = amdinfo->max_engine_clk / 1000;
+ info->max_memory_clock = amdinfo->max_memory_clk / 1000;
+ info->num_tcc_blocks = device_info.num_tcc_blocks;
info->max_se = amdinfo->num_shader_engines;
info->max_sh_per_se = amdinfo->num_shader_arrays_per_engine;
info->has_hw_decode =
- (uvd.available_rings != 0) || (vcn_dec.available_rings != 0);
+ (uvd.available_rings != 0) || (vcn_dec.available_rings != 0) ||
+ (vcn_jpeg.available_rings != 0);
info->uvd_fw_version =
uvd.available_rings ? uvd_version : 0;
info->vce_fw_version =
uvd_enc.available_rings ? true : false;
info->has_userptr = true;
info->has_syncobj = has_syncobj(fd);
+ info->has_timeline_syncobj = has_timeline_syncobj(fd);
info->has_syncobj_wait_for_submit = info->has_syncobj && info->drm_minor >= 20;
info->has_fence_to_handle = info->has_syncobj && info->drm_minor >= 21;
info->has_ctx_priority = info->drm_minor >= 22;
- /* TODO: Enable this once the kernel handles it efficiently. */
- info->has_local_buffers = info->drm_minor >= 20 &&
- !info->has_dedicated_vram;
+ info->has_local_buffers = info->drm_minor >= 20;
info->kernel_flushes_hdp_before_ib = true;
info->htile_cmask_support_1d_tiling = true;
info->si_TA_CS_BC_BASE_ADDR_allowed = true;
info->has_bo_metadata = true;
info->has_gpu_reset_status_query = true;
- info->has_gpu_reset_counter_query = false;
info->has_eqaa_surface_allocator = true;
info->has_format_bc1_through_bc7 = true;
-
+ /* DRM 3.1.0 doesn't flush TC for GFX8 correctly. */
+ info->kernel_flushes_tc_l2_after_ib = info->chip_class != GFX8 ||
+ info->drm_minor >= 2;
+ info->has_indirect_compute_dispatch = true;
+ /* GFX6 doesn't support unaligned loads. */
+ info->has_unaligned_shader_loads = info->chip_class != GFX6;
+ /* Disable sparse mappings on GFX6 due to VM faults in CP DMA. Enable them once
+ * these faults are mitigated in software.
+ */
+ info->has_sparse_vm_mappings = info->chip_class >= GFX7 && info->drm_minor >= 13;
+ info->has_2d_tiling = true;
+ info->has_read_registers_query = true;
+ info->has_scheduled_fence_dependency = info->drm_minor >= 28;
+ info->mid_command_buffer_preemption_enabled =
+ amdinfo->ids_flags & AMDGPU_IDS_FLAGS_PREEMPTION;
+
+ info->pa_sc_tile_steering_override = device_info.pa_sc_tile_steering_override;
info->num_render_backends = amdinfo->rb_pipes;
/* The value returned by the kernel driver was wrong. */
if (info->family == CHIP_KAVERI)
fprintf(stderr, "amdgpu: clock crystal frequency is 0, timestamps will be wrong\n");
info->clock_crystal_freq = 1;
}
- info->tcc_cache_line_size = 64; /* TC L2 line size on GCN */
+ if (info->chip_class >= GFX10) {
+ info->tcc_cache_line_size = 128;
+
+ if (info->drm_minor >= 35) {
+ info->tcc_harvested = device_info.tcc_disabled_mask != 0;
+ } else {
+ /* This is a hack, but it's all we can do without a kernel upgrade. */
+ info->tcc_harvested =
+ (info->vram_size / info->num_tcc_blocks) != 512*1024*1024;
+ }
+ } else {
+ info->tcc_cache_line_size = 64;
+ }
info->gb_addr_config = amdinfo->gb_addr_cfg;
- if (info->chip_class == GFX9) {
+ if (info->chip_class >= GFX9) {
info->num_tile_pipes = 1 << G_0098F8_NUM_PIPES(amdinfo->gb_addr_cfg);
info->pipe_interleave_bytes =
256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(amdinfo->gb_addr_cfg);
}
info->r600_has_virtual_memory = true;
+ /* LDS is 64KB per CU (4 SIMDs), which is 16KB per SIMD (usage above
+ * 16KB makes some SIMDs unoccupied).
+ *
+ * LDS is 128KB in WGP mode and 64KB in CU mode. Assume the WGP mode is used.
+ */
+ info->lds_size_per_workgroup = info->chip_class >= GFX10 ? 128 * 1024 : 64 * 1024;
+ info->lds_granularity = info->chip_class >= GFX7 ? 128 * 4 : 64 * 4;
+
assert(util_is_power_of_two_or_zero(dma.available_rings + 1));
assert(util_is_power_of_two_or_zero(compute.available_rings + 1));
- info->num_sdma_rings = util_bitcount(dma.available_rings);
- info->num_compute_rings = util_bitcount(compute.available_rings);
+ info->has_graphics = gfx.available_rings > 0;
+ info->num_rings[RING_GFX] = util_bitcount(gfx.available_rings);
+ info->num_rings[RING_COMPUTE] = util_bitcount(compute.available_rings);
+ info->num_rings[RING_DMA] = util_bitcount(dma.available_rings);
+ info->num_rings[RING_UVD] = util_bitcount(uvd.available_rings);
+ info->num_rings[RING_VCE] = util_bitcount(vce.available_rings);
+ info->num_rings[RING_UVD_ENC] = util_bitcount(uvd_enc.available_rings);
+ info->num_rings[RING_VCN_DEC] = util_bitcount(vcn_dec.available_rings);
+ info->num_rings[RING_VCN_ENC] = util_bitcount(vcn_enc.available_rings);
+ info->num_rings[RING_VCN_JPEG] = util_bitcount(vcn_jpeg.available_rings);
+
+ /* This is "align_mask" copied from the kernel, maximums of all IP versions. */
+ info->ib_pad_dw_mask[RING_GFX] = 0xff;
+ info->ib_pad_dw_mask[RING_COMPUTE] = 0xff;
+ info->ib_pad_dw_mask[RING_DMA] = 0xf;
+ info->ib_pad_dw_mask[RING_UVD] = 0xf;
+ info->ib_pad_dw_mask[RING_VCE] = 0x3f;
+ info->ib_pad_dw_mask[RING_UVD_ENC] = 0x3f;
+ info->ib_pad_dw_mask[RING_VCN_DEC] = 0xf;
+ info->ib_pad_dw_mask[RING_VCN_ENC] = 0x3f;
+ info->ib_pad_dw_mask[RING_VCN_JPEG] = 0xf;
+
+ /* The mere presence of CLEAR_STATE in the IB causes random GPU hangs
+ * on GFX6. Some CLEAR_STATE cause asic hang on radeon kernel, etc.
+ * SPI_VS_OUT_CONFIG. So only enable GFX7 CLEAR_STATE on amdgpu kernel.
+ */
+ info->has_clear_state = info->chip_class >= GFX7;
+
+ info->has_distributed_tess = info->chip_class >= GFX10 ||
+ (info->chip_class >= GFX8 && info->max_se >= 2);
+
+ info->has_dcc_constant_encode = info->family == CHIP_RAVEN2 ||
+ info->family == CHIP_RENOIR ||
+ info->chip_class >= GFX10;
+
+ info->has_rbplus = info->family == CHIP_STONEY ||
+ info->chip_class >= GFX9;
+
+ /* Some chips have RB+ registers, but don't support RB+. Those must
+ * always disable it.
+ */
+ info->rbplus_allowed = info->has_rbplus &&
+ (info->family == CHIP_STONEY ||
+ info->family == CHIP_VEGA12 ||
+ info->family == CHIP_RAVEN ||
+ info->family == CHIP_RAVEN2 ||
+ info->family == CHIP_RENOIR ||
+ info->chip_class >= GFX10_3);
+
+ info->has_out_of_order_rast = info->chip_class >= GFX8 &&
+ info->chip_class <= GFX9 &&
+ info->max_se >= 2;
+
+ /* Whether chips support double rate packed math instructions. */
+ info->has_packed_math_16bit = info->chip_class >= GFX9;
+
+ /* TODO: Figure out how to use LOAD_CONTEXT_REG on GFX6-GFX7. */
+ info->has_load_ctx_reg_pkt = info->chip_class >= GFX9 ||
+ (info->chip_class >= GFX8 &&
+ info->me_fw_feature >= 41);
+
+ info->cpdma_prefetch_writes_memory = info->chip_class <= GFX8;
+
+ info->has_gfx9_scissor_bug = info->family == CHIP_VEGA10 ||
+ info->family == CHIP_RAVEN;
+
+ info->has_tc_compat_zrange_bug = info->chip_class >= GFX8 &&
+ info->chip_class <= GFX9;
+
+ info->has_msaa_sample_loc_bug = (info->family >= CHIP_POLARIS10 &&
+ info->family <= CHIP_POLARIS12) ||
+ info->family == CHIP_VEGA10 ||
+ info->family == CHIP_RAVEN;
+
+ info->has_ls_vgpr_init_bug = info->family == CHIP_VEGA10 ||
+ info->family == CHIP_RAVEN;
/* Get the number of good compute units. */
info->num_good_compute_units = 0;
- for (i = 0; i < info->max_se; i++)
- for (j = 0; j < info->max_sh_per_se; j++)
+ for (i = 0; i < info->max_se; i++) {
+ for (j = 0; j < info->max_sh_per_se; j++) {
+ /*
+ * The cu bitmap in amd gpu info structure is
+ * 4x4 size array, and it's usually suitable for Vega
+ * ASICs which has 4*2 SE/SH layout.
+ * But for Arcturus, SE/SH layout is changed to 8*1.
+ * To mostly reduce the impact, we make it compatible
+ * with current bitmap array as below:
+ * SE4,SH0 --> cu_bitmap[0][1]
+ * SE5,SH0 --> cu_bitmap[1][1]
+ * SE6,SH0 --> cu_bitmap[2][1]
+ * SE7,SH0 --> cu_bitmap[3][1]
+ */
+ info->cu_mask[i%4][j+i/4] = amdinfo->cu_bitmap[i%4][j+i/4];
info->num_good_compute_units +=
- util_bitcount(amdinfo->cu_bitmap[i][j]);
+ util_bitcount(info->cu_mask[i][j]);
+ }
+ }
+
+ /* On GFX10, only whole WGPs (in units of 2 CUs) can be disabled,
+ * and max - min <= 2.
+ */
+ unsigned cu_group = info->chip_class >= GFX10 ? 2 : 1;
+ info->max_good_cu_per_sa = DIV_ROUND_UP(info->num_good_compute_units,
+ (info->max_se * info->max_sh_per_se * cu_group)) * cu_group;
+ info->min_good_cu_per_sa = (info->num_good_compute_units /
+ (info->max_se * info->max_sh_per_se * cu_group)) * cu_group;
memcpy(info->si_tile_mode_array, amdinfo->gb_tile_mode,
sizeof(amdinfo->gb_tile_mode));
info->pte_fragment_size = alignment_info.size_local;
info->gart_page_size = alignment_info.size_remote;
- if (info->chip_class == SI)
- info->gfx_ib_pad_with_type2 = TRUE;
+ if (info->chip_class == GFX6)
+ info->gfx_ib_pad_with_type2 = true;
unsigned ib_align = 0;
ib_align = MAX2(ib_align, gfx.ib_start_alignment);
+ ib_align = MAX2(ib_align, gfx.ib_size_alignment);
ib_align = MAX2(ib_align, compute.ib_start_alignment);
+ ib_align = MAX2(ib_align, compute.ib_size_alignment);
ib_align = MAX2(ib_align, dma.ib_start_alignment);
+ ib_align = MAX2(ib_align, dma.ib_size_alignment);
ib_align = MAX2(ib_align, uvd.ib_start_alignment);
+ ib_align = MAX2(ib_align, uvd.ib_size_alignment);
ib_align = MAX2(ib_align, uvd_enc.ib_start_alignment);
+ ib_align = MAX2(ib_align, uvd_enc.ib_size_alignment);
ib_align = MAX2(ib_align, vce.ib_start_alignment);
+ ib_align = MAX2(ib_align, vce.ib_size_alignment);
ib_align = MAX2(ib_align, vcn_dec.ib_start_alignment);
+ ib_align = MAX2(ib_align, vcn_dec.ib_size_alignment);
ib_align = MAX2(ib_align, vcn_enc.ib_start_alignment);
- assert(ib_align);
- info->ib_start_alignment = ib_align;
+ ib_align = MAX2(ib_align, vcn_enc.ib_size_alignment);
+ ib_align = MAX2(ib_align, vcn_jpeg.ib_start_alignment);
+ ib_align = MAX2(ib_align, vcn_jpeg.ib_size_alignment);
+ /* GFX10 and maybe GFX9 need this alignment for cache coherency. */
+ if (info->chip_class >= GFX9)
+ ib_align = MAX2(ib_align, info->tcc_cache_line_size);
+ /* The kernel pads gfx and compute IBs to 256 dwords since:
+ * 66f3b2d527154bd258a57c8815004b5964aa1cf5
+ * Do the same.
+ */
+ ib_align = MAX2(ib_align, 1024);
+ info->ib_alignment = ib_align;
+
+ if ((info->drm_minor >= 31 &&
+ (info->family == CHIP_RAVEN ||
+ info->family == CHIP_RAVEN2 ||
+ info->family == CHIP_RENOIR)) ||
+ (info->drm_minor >= 34 &&
+ (info->family == CHIP_NAVI12 ||
+ info->family == CHIP_NAVI14)) ||
+ info->chip_class >= GFX10_3) {
+ if (info->num_render_backends == 1)
+ info->use_display_dcc_unaligned = true;
+ else
+ info->use_display_dcc_with_retile_blit = true;
+ }
+
+ info->has_gds_ordered_append = info->chip_class >= GFX7 &&
+ info->drm_minor >= 29;
+
+ if (info->chip_class >= GFX9) {
+ unsigned pc_lines = 0;
+
+ switch (info->family) {
+ case CHIP_VEGA10:
+ case CHIP_VEGA12:
+ case CHIP_VEGA20:
+ pc_lines = 2048;
+ break;
+ case CHIP_RAVEN:
+ case CHIP_RAVEN2:
+ case CHIP_RENOIR:
+ case CHIP_NAVI10:
+ case CHIP_NAVI12:
+ case CHIP_SIENNA_CICHLID:
+ case CHIP_NAVY_FLOUNDER:
+ pc_lines = 1024;
+ break;
+ case CHIP_NAVI14:
+ pc_lines = 512;
+ break;
+ case CHIP_ARCTURUS:
+ break;
+ default:
+ assert(0);
+ }
+
+ info->pc_lines = pc_lines;
+
+ if (info->chip_class >= GFX10) {
+ info->pbb_max_alloc_count = pc_lines / 3;
+ } else {
+ info->pbb_max_alloc_count =
+ MIN2(128, pc_lines / (4 * info->max_se));
+ }
+ }
+
+ /* The number of SDPs is the same as the number of TCCs for now. */
+ if (info->chip_class >= GFX10)
+ info->num_sdp_interfaces = device_info.num_tcc_blocks;
+
+ if (info->chip_class >= GFX10_3)
+ info->max_wave64_per_simd = 16;
+ else if (info->chip_class == GFX10)
+ info->max_wave64_per_simd = 20;
+ else if (info->family >= CHIP_POLARIS10 && info->family <= CHIP_VEGAM)
+ info->max_wave64_per_simd = 8;
+ else
+ info->max_wave64_per_simd = 10;
+
+ if (info->chip_class >= GFX10) {
+ info->num_physical_sgprs_per_simd = 128 * info->max_wave64_per_simd;
+ info->min_sgpr_alloc = 128;
+ info->sgpr_alloc_granularity = 128;
+ /* Don't use late alloc on small chips. */
+ info->use_late_alloc = info->num_render_backends > 4;
+ } else if (info->chip_class >= GFX8) {
+ info->num_physical_sgprs_per_simd = 800;
+ info->min_sgpr_alloc = 16;
+ info->sgpr_alloc_granularity = 16;
+ info->use_late_alloc = true;
+ } else {
+ info->num_physical_sgprs_per_simd = 512;
+ info->min_sgpr_alloc = 8;
+ info->sgpr_alloc_granularity = 8;
+ /* Potential hang on Kabini: */
+ info->use_late_alloc = info->family != CHIP_KABINI;
+ }
+
+ info->max_sgpr_alloc = info->family == CHIP_TONGA ||
+ info->family == CHIP_ICELAND ? 96 : 104;
+
+ info->min_wave64_vgpr_alloc = 4;
+ info->max_vgpr_alloc = 256;
+ info->wave64_vgpr_alloc_granularity = 4;
+
+ info->num_physical_wave64_vgprs_per_simd = info->chip_class >= GFX10 ? 512 : 256;
+ info->num_simd_per_compute_unit = info->chip_class >= GFX10 ? 2 : 4;
return true;
}
printf(" pci (domain:bus:dev.func): %04x:%02x:%02x.%x\n",
info->pci_domain, info->pci_bus,
info->pci_dev, info->pci_func);
+
+ printf(" name = %s\n", info->name);
+ printf(" marketing_name = %s\n", info->marketing_name);
+ printf(" is_pro_graphics = %u\n", info->is_pro_graphics);
printf(" pci_id = 0x%x\n", info->pci_id);
+ printf(" pci_rev_id = 0x%x\n", info->pci_rev_id);
printf(" family = %i\n", info->family);
printf(" chip_class = %i\n", info->chip_class);
- printf(" num_compute_rings = %u\n", info->num_compute_rings);
- printf(" num_sdma_rings = %i\n", info->num_sdma_rings);
+ printf(" family_id = %i\n", info->family_id);
+ printf(" chip_external_rev = %i\n", info->chip_external_rev);
printf(" clock_crystal_freq = %i\n", info->clock_crystal_freq);
- printf(" tcc_cache_line_size = %u\n", info->tcc_cache_line_size);
+
+ printf("Features:\n");
+ printf(" has_graphics = %i\n", info->has_graphics);
+ printf(" num_rings[RING_GFX] = %i\n", info->num_rings[RING_GFX]);
+ printf(" num_rings[RING_DMA] = %i\n", info->num_rings[RING_DMA]);
+ printf(" num_rings[RING_COMPUTE] = %u\n", info->num_rings[RING_COMPUTE]);
+ printf(" num_rings[RING_UVD] = %i\n", info->num_rings[RING_UVD]);
+ printf(" num_rings[RING_VCE] = %i\n", info->num_rings[RING_VCE]);
+ printf(" num_rings[RING_UVD_ENC] = %i\n", info->num_rings[RING_UVD_ENC]);
+ printf(" num_rings[RING_VCN_DEC] = %i\n", info->num_rings[RING_VCN_DEC]);
+ printf(" num_rings[RING_VCN_ENC] = %i\n", info->num_rings[RING_VCN_ENC]);
+ printf(" num_rings[RING_VCN_JPEG] = %i\n", info->num_rings[RING_VCN_JPEG]);
+ printf(" has_clear_state = %u\n", info->has_clear_state);
+ printf(" has_distributed_tess = %u\n", info->has_distributed_tess);
+ printf(" has_dcc_constant_encode = %u\n", info->has_dcc_constant_encode);
+ printf(" has_rbplus = %u\n", info->has_rbplus);
+ printf(" rbplus_allowed = %u\n", info->rbplus_allowed);
+ printf(" has_load_ctx_reg_pkt = %u\n", info->has_load_ctx_reg_pkt);
+ printf(" has_out_of_order_rast = %u\n", info->has_out_of_order_rast);
+ printf(" cpdma_prefetch_writes_memory = %u\n", info->cpdma_prefetch_writes_memory);
+ printf(" has_gfx9_scissor_bug = %i\n", info->has_gfx9_scissor_bug);
+ printf(" has_tc_compat_zrange_bug = %i\n", info->has_tc_compat_zrange_bug);
+ printf(" has_msaa_sample_loc_bug = %i\n", info->has_msaa_sample_loc_bug);
+ printf(" has_ls_vgpr_init_bug = %i\n", info->has_ls_vgpr_init_bug);
+
+ printf("Display features:\n");
+ printf(" use_display_dcc_unaligned = %u\n", info->use_display_dcc_unaligned);
+ printf(" use_display_dcc_with_retile_blit = %u\n", info->use_display_dcc_with_retile_blit);
printf("Memory info:\n");
printf(" pte_fragment_size = %u\n", info->pte_fragment_size);
printf(" gart_size = %i MB\n", (int)DIV_ROUND_UP(info->gart_size, 1024*1024));
printf(" vram_size = %i MB\n", (int)DIV_ROUND_UP(info->vram_size, 1024*1024));
printf(" vram_vis_size = %i MB\n", (int)DIV_ROUND_UP(info->vram_vis_size, 1024*1024));
+ printf(" vram_type = %i\n", info->vram_type);
+ printf(" vram_bit_width = %i\n", info->vram_bit_width);
printf(" gds_size = %u kB\n", info->gds_size / 1024);
printf(" gds_gfx_partition_size = %u kB\n", info->gds_gfx_partition_size / 1024);
printf(" max_alloc_size = %i MB\n",
printf(" min_alloc_size = %u\n", info->min_alloc_size);
printf(" address32_hi = %u\n", info->address32_hi);
printf(" has_dedicated_vram = %u\n", info->has_dedicated_vram);
+ printf(" num_sdp_interfaces = %u\n", info->num_sdp_interfaces);
+ printf(" num_tcc_blocks = %i\n", info->num_tcc_blocks);
+ printf(" tcc_cache_line_size = %u\n", info->tcc_cache_line_size);
+ printf(" tcc_harvested = %u\n", info->tcc_harvested);
+ printf(" pc_lines = %u\n", info->pc_lines);
+ printf(" lds_size_per_workgroup = %u\n", info->lds_size_per_workgroup);
+ printf(" lds_granularity = %i\n", info->lds_granularity);
+ printf(" max_memory_clock = %i\n", info->max_memory_clock);
+ printf(" ce_ram_size = %i\n", info->ce_ram_size);
+ printf(" l1_cache_size = %i\n", info->l1_cache_size);
+ printf(" l2_cache_size = %i\n", info->l2_cache_size);
printf("CP info:\n");
printf(" gfx_ib_pad_with_type2 = %i\n", info->gfx_ib_pad_with_type2);
- printf(" ib_start_alignment = %u\n", info->ib_start_alignment);
+ printf(" ib_alignment = %u\n", info->ib_alignment);
printf(" me_fw_version = %i\n", info->me_fw_version);
printf(" me_fw_feature = %i\n", info->me_fw_feature);
printf(" pfp_fw_version = %i\n", info->pfp_fw_version);
printf(" has_userptr = %i\n", info->has_userptr);
printf(" has_syncobj = %u\n", info->has_syncobj);
printf(" has_syncobj_wait_for_submit = %u\n", info->has_syncobj_wait_for_submit);
+ printf(" has_timeline_syncobj = %u\n", info->has_timeline_syncobj);
printf(" has_fence_to_handle = %u\n", info->has_fence_to_handle);
printf(" has_ctx_priority = %u\n", info->has_ctx_priority);
printf(" has_local_buffers = %u\n", info->has_local_buffers);
printf(" si_TA_CS_BC_BASE_ADDR_allowed = %u\n", info->si_TA_CS_BC_BASE_ADDR_allowed);
printf(" has_bo_metadata = %u\n", info->has_bo_metadata);
printf(" has_gpu_reset_status_query = %u\n", info->has_gpu_reset_status_query);
- printf(" has_gpu_reset_counter_query = %u\n", info->has_gpu_reset_counter_query);
printf(" has_eqaa_surface_allocator = %u\n", info->has_eqaa_surface_allocator);
printf(" has_format_bc1_through_bc7 = %u\n", info->has_format_bc1_through_bc7);
+ printf(" kernel_flushes_tc_l2_after_ib = %u\n", info->kernel_flushes_tc_l2_after_ib);
+ printf(" has_indirect_compute_dispatch = %u\n", info->has_indirect_compute_dispatch);
+ printf(" has_unaligned_shader_loads = %u\n", info->has_unaligned_shader_loads);
+ printf(" has_sparse_vm_mappings = %u\n", info->has_sparse_vm_mappings);
+ printf(" has_2d_tiling = %u\n", info->has_2d_tiling);
+ printf(" has_read_registers_query = %u\n", info->has_read_registers_query);
+ printf(" has_gds_ordered_append = %u\n", info->has_gds_ordered_append);
+ printf(" has_scheduled_fence_dependency = %u\n", info->has_scheduled_fence_dependency);
+ printf(" mid_command_buffer_preemption_enabled = %u\n", info->mid_command_buffer_preemption_enabled);
printf("Shader core info:\n");
printf(" max_shader_clock = %i\n", info->max_shader_clock);
printf(" num_good_compute_units = %i\n", info->num_good_compute_units);
+ printf(" max_good_cu_per_sa = %i\n", info->max_good_cu_per_sa);
+ printf(" min_good_cu_per_sa = %i\n", info->min_good_cu_per_sa);
printf(" max_se = %i\n", info->max_se);
printf(" max_sh_per_se = %i\n", info->max_sh_per_se);
+ printf(" max_wave64_per_simd = %i\n", info->max_wave64_per_simd);
+ printf(" num_physical_sgprs_per_simd = %i\n", info->num_physical_sgprs_per_simd);
+ printf(" num_physical_wave64_vgprs_per_simd = %i\n", info->num_physical_wave64_vgprs_per_simd);
+ printf(" num_simd_per_compute_unit = %i\n", info->num_simd_per_compute_unit);
+ printf(" min_sgpr_alloc = %i\n", info->min_sgpr_alloc);
+ printf(" max_sgpr_alloc = %i\n", info->max_sgpr_alloc);
+ printf(" sgpr_alloc_granularity = %i\n", info->sgpr_alloc_granularity);
+ printf(" min_wave64_vgpr_alloc = %i\n", info->min_wave64_vgpr_alloc);
+ printf(" max_vgpr_alloc = %i\n", info->max_vgpr_alloc);
+ printf(" wave64_vgpr_alloc_granularity = %i\n", info->wave64_vgpr_alloc_granularity);
printf("Render backend info:\n");
+ printf(" pa_sc_tile_steering_override = 0x%x\n", info->pa_sc_tile_steering_override);
printf(" num_render_backends = %i\n", info->num_render_backends);
printf(" num_tile_pipes = %i\n", info->num_tile_pipes);
printf(" pipe_interleave_bytes = %i\n", info->pipe_interleave_bytes);
printf(" enabled_rb_mask = 0x%x\n", info->enabled_rb_mask);
printf(" max_alignment = %u\n", (unsigned)info->max_alignment);
+ printf(" pbb_max_alloc_count = %u\n", info->pbb_max_alloc_count);
- printf("GB_ADDR_CONFIG:\n");
- if (info->chip_class >= GFX9) {
+ printf("GB_ADDR_CONFIG: 0x%08x\n", info->gb_addr_config);
+ if (info->chip_class >= GFX10) {
+ printf(" num_pipes = %u\n",
+ 1 << G_0098F8_NUM_PIPES(info->gb_addr_config));
+ printf(" pipe_interleave_size = %u\n",
+ 256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(info->gb_addr_config));
+ printf(" max_compressed_frags = %u\n",
+ 1 << G_0098F8_MAX_COMPRESSED_FRAGS(info->gb_addr_config));
+ } else if (info->chip_class == GFX9) {
printf(" num_pipes = %u\n",
1 << G_0098F8_NUM_PIPES(info->gb_addr_config));
printf(" pipe_interleave_size = %u\n",
case CHIP_HAINAN:
case CHIP_KAVERI:
case CHIP_KABINI:
- case CHIP_MULLINS:
case CHIP_ICELAND:
case CHIP_CARRIZO:
case CHIP_STONEY:
void
ac_get_raster_config(struct radeon_info *info,
uint32_t *raster_config_p,
- uint32_t *raster_config_1_p)
+ uint32_t *raster_config_1_p,
+ uint32_t *se_tile_repeat_p)
{
- unsigned raster_config, raster_config_1;
+ unsigned raster_config, raster_config_1, se_tile_repeat;
switch (info->family) {
/* 1 SE / 1 RB */
case CHIP_HAINAN:
case CHIP_KABINI:
- case CHIP_MULLINS:
case CHIP_STONEY:
raster_config = 0x00000000;
raster_config_1 = 0x00000000;
/* drm/radeon on Kaveri is buggy, so disable 1 RB to work around it.
* This decreases performance by up to 50% when the RB is the bottleneck.
*/
- if (info->family == CHIP_KAVERI && info->drm_major == 2)
+ if (info->family == CHIP_KAVERI && !info->is_amdgpu)
raster_config = 0x00000000;
/* Fiji: Old kernels have incorrect tiling config. This decreases
raster_config_1 = 0x0000002a;
}
+ unsigned se_width = 8 << G_028350_SE_XSEL_GFX6(raster_config);
+ unsigned se_height = 8 << G_028350_SE_YSEL_GFX6(raster_config);
+
+ /* I don't know how to calculate this, though this is probably a good guess. */
+ se_tile_repeat = MAX2(se_width, se_height) * info->max_se;
+
*raster_config_p = raster_config;
*raster_config_1_p = raster_config_1;
+ if (se_tile_repeat_p)
+ *se_tile_repeat_p = se_tile_repeat;
}
void
assert(rb_per_pkr == 1 || rb_per_pkr == 2);
- if (info->chip_class >= CIK) {
+ if (info->chip_class >= GFX7) {
unsigned raster_config_1 = *cik_raster_config_1_p;
if ((num_se > 2) && ((!se_mask[0] && !se_mask[1]) ||
(!se_mask[2] && !se_mask[3]))) {
}
}
}
+
+unsigned ac_get_compute_resource_limits(struct radeon_info *info,
+ unsigned waves_per_threadgroup,
+ unsigned max_waves_per_sh,
+ unsigned threadgroups_per_cu)
+{
+ unsigned compute_resource_limits =
+ S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup % 4 == 0);
+
+ if (info->chip_class >= GFX7) {
+ unsigned num_cu_per_se = info->num_good_compute_units /
+ info->max_se;
+
+ /* Force even distribution on all SIMDs in CU if the workgroup
+ * size is 64. This has shown some good improvements if # of CUs
+ * per SE is not a multiple of 4.
+ */
+ if (num_cu_per_se % 4 && waves_per_threadgroup == 1)
+ compute_resource_limits |= S_00B854_FORCE_SIMD_DIST(1);
+
+ assert(threadgroups_per_cu >= 1 && threadgroups_per_cu <= 8);
+ compute_resource_limits |= S_00B854_WAVES_PER_SH(max_waves_per_sh) |
+ S_00B854_CU_GROUP_COUNT(threadgroups_per_cu - 1);
+ } else {
+ /* GFX6 */
+ if (max_waves_per_sh) {
+ unsigned limit_div16 = DIV_ROUND_UP(max_waves_per_sh, 16);
+ compute_resource_limits |= S_00B854_WAVES_PER_SH_GFX6(limit_div16);
+ }
+ }
+ return compute_resource_limits;
+}