#include <stdio.h>
#include <xf86drm.h>
-#include <amdgpu_drm.h>
+#include "drm-uapi/amdgpu_drm.h"
#include <amdgpu.h>
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)
/* 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;
#define identify_chip2(asic, chipname) \
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 clock from KHz to MHz */
+ /* 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->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->has_graphics = gfx.available_rings > 0;
- info->num_sdma_rings = util_bitcount(dma.available_rings);
- info->num_compute_rings = util_bitcount(compute.available_rings);
+ 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);
/* 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.
*/
info->has_clear_state = info->chip_class >= GFX7;
- info->has_distributed_tess = info->chip_class >= GFX8 &&
- info->max_se >= 2;
+ 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->family == CHIP_RENOIR);
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_double_rate_fp16 = 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 &&
/* 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++) {
+ info->cu_mask[i][j] = amdinfo->cu_bitmap[i][j];
info->num_good_compute_units +=
- util_bitcount(amdinfo->cu_bitmap[i][j]);
+ util_bitcount(info->cu_mask[i][j]);
+ }
+ }
info->num_good_cu_per_sh = info->num_good_compute_units /
(info->max_se * info->max_sh_per_se);
assert(ib_align);
info->ib_start_alignment = ib_align;
- if (info->drm_minor >= 31 &&
- (info->family == CHIP_RAVEN ||
- info->family == CHIP_RAVEN2 ||
- info->family == CHIP_RENOIR)) {
+ 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))) {
if (info->num_render_backends == 1)
info->use_display_dcc_unaligned = true;
else
assert(0);
}
+ info->pc_lines = pc_lines;
+
if (info->chip_class >= GFX10) {
info->pbb_max_alloc_count = pc_lines / 3;
} else {
/* The number is per SIMD. There is enough SGPRs for the maximum number
* of Wave32, which is double the number for Wave64.
*/
- if (info->chip_class >= GFX10)
+ if (info->chip_class >= GFX10) {
info->num_physical_sgprs_per_simd = 128 * info->max_wave64_per_simd * 2;
- else if (info->chip_class >= GFX8)
+ 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;
- else
+ 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(" 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(" family_id = %i\n", info->family_id);
printf("Features:\n");
printf(" has_graphics = %i\n", info->has_graphics);
- printf(" num_compute_rings = %u\n", info->num_compute_rings);
- printf(" num_sdma_rings = %i\n", info->num_sdma_rings);
+ 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(" 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(" 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(" 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);