#include "radeon_drm_cs.h"
#include "radeon_drm_public.h"
+#include "util/u_cpu_detect.h"
#include "util/u_memory.h"
#include "util/u_hash_table.h"
ws->info.drm_major = version->version_major;
ws->info.drm_minor = version->version_minor;
ws->info.drm_patchlevel = version->version_patchlevel;
+ ws->info.is_amdgpu = false;
drmFreeVersion(version);
/* Get PCI ID. */
#include "pci_ids/r600_pci_ids.h"
#undef CHIPSET
-#define CHIPSET(pci_id, cfamily) case pci_id: ws->info.family = CHIP_##cfamily; ws->gen = DRV_SI; break;
+#define CHIPSET(pci_id, cfamily) \
+ case pci_id: \
+ ws->info.family = CHIP_##cfamily; \
+ ws->info.name = #cfamily; \
+ ws->gen = DRV_SI; \
+ break;
#include "pci_ids/radeonsi_pci_ids.h"
#undef CHIPSET
case CHIP_VERDE:
case CHIP_OLAND:
case CHIP_HAINAN:
- ws->info.chip_class = SI;
+ ws->info.chip_class = GFX6;
break;
case CHIP_BONAIRE:
case CHIP_KAVERI:
case CHIP_KABINI:
case CHIP_HAWAII:
- case CHIP_MULLINS:
- ws->info.chip_class = CIK;
+ ws->info.chip_class = GFX7;
break;
}
case CHIP_ARUBA:
case CHIP_KAVERI:
case CHIP_KABINI:
- case CHIP_MULLINS:
ws->info.has_dedicated_vram = false;
break;
if (ws->info.drm_minor < 49)
ws->info.vram_vis_size = MIN2(ws->info.vram_vis_size, 256*1024*1024);
- /* Radeon allocates all buffers as contigous, which makes large allocations
+ /* Radeon allocates all buffers contiguously, which makes large allocations
* unlikely to succeed. */
- ws->info.max_alloc_size = MAX2(ws->info.vram_size, ws->info.gart_size) * 0.7;
if (ws->info.has_dedicated_vram)
- ws->info.max_alloc_size = MIN2(ws->info.vram_size * 0.7, ws->info.max_alloc_size);
+ ws->info.max_alloc_size = ws->info.vram_size * 0.7;
+ else
+ ws->info.max_alloc_size = ws->info.gart_size * 0.7;
+
if (ws->info.drm_minor < 40)
ws->info.max_alloc_size = MIN2(ws->info.max_alloc_size, 256*1024*1024);
+ /* Both 32-bit and 64-bit address spaces only have 4GB. */
+ ws->info.max_alloc_size = MIN2(ws->info.max_alloc_size, 3ull*1024*1024*1024);
/* Get max clock frequency info and convert it to MHz */
radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_SCLK, NULL,
radeon_get_drm_value(ws->fd, RADEON_INFO_SI_BACKEND_ENABLED_MASK, NULL,
&ws->info.enabled_rb_mask);
- ws->info.has_virtual_memory = false;
+ ws->info.r600_has_virtual_memory = false;
if (ws->info.drm_minor >= 13) {
uint32_t ib_vm_max_size;
- ws->info.has_virtual_memory = true;
+ ws->info.r600_has_virtual_memory = true;
if (!radeon_get_drm_value(ws->fd, RADEON_INFO_VA_START, NULL,
&ws->va_start))
- ws->info.has_virtual_memory = false;
+ ws->info.r600_has_virtual_memory = false;
if (!radeon_get_drm_value(ws->fd, RADEON_INFO_IB_VM_MAX_SIZE, NULL,
&ib_vm_max_size))
- ws->info.has_virtual_memory = false;
+ ws->info.r600_has_virtual_memory = false;
radeon_get_drm_value(ws->fd, RADEON_INFO_VA_UNMAP_WORKING, NULL,
&ws->va_unmap_working);
}
if (ws->gen == DRV_R600 && !debug_get_bool_option("RADEON_VA", false))
- ws->info.has_virtual_memory = false;
+ ws->info.r600_has_virtual_memory = false;
}
/* Get max pipes, this is only needed for compute shaders. All evergreen+
radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_SE, NULL,
&ws->info.max_se);
+ switch (ws->info.family) {
+ case CHIP_HAINAN:
+ case CHIP_KABINI:
+ ws->info.num_tcc_blocks = 2;
+ break;
+ case CHIP_VERDE:
+ case CHIP_OLAND:
+ case CHIP_BONAIRE:
+ case CHIP_KAVERI:
+ ws->info.num_tcc_blocks = 4;
+ break;
+ case CHIP_PITCAIRN:
+ ws->info.num_tcc_blocks = 8;
+ break;
+ case CHIP_TAHITI:
+ ws->info.num_tcc_blocks = 12;
+ break;
+ case CHIP_HAWAII:
+ ws->info.num_tcc_blocks = 16;
+ break;
+ default:
+ ws->info.num_tcc_blocks = 0;
+ break;
+ }
+
if (!ws->info.max_se) {
switch (ws->info.family) {
default:
radeon_get_drm_value(ws->fd, RADEON_INFO_MAX_SH_PER_SE, NULL,
&ws->info.max_sh_per_se);
+ if (ws->gen == DRV_SI) {
+ ws->info.num_good_cu_per_sh = ws->info.num_good_compute_units /
+ (ws->info.max_se * ws->info.max_sh_per_se);
+ }
radeon_get_drm_value(ws->fd, RADEON_INFO_ACCEL_WORKING2, NULL,
&ws->accel_working2);
return false;
}
- if (ws->info.chip_class == CIK) {
+ if (ws->info.chip_class == GFX7) {
if (!radeon_get_drm_value(ws->fd, RADEON_INFO_CIK_MACROTILE_MODE_ARRAY, NULL,
ws->info.cik_macrotile_mode_array)) {
- fprintf(stderr, "radeon: Kernel 3.13 is required for CIK support.\n");
+ fprintf(stderr, "radeon: Kernel 3.13 is required for Sea Islands support.\n");
return false;
}
}
- if (ws->info.chip_class >= SI) {
+ if (ws->info.chip_class >= GFX6) {
if (!radeon_get_drm_value(ws->fd, RADEON_INFO_SI_TILE_MODE_ARRAY, NULL,
ws->info.si_tile_mode_array)) {
- fprintf(stderr, "radeon: Kernel 3.10 is required for SI support.\n");
+ fprintf(stderr, "radeon: Kernel 3.10 is required for Southern Islands support.\n");
return false;
}
}
/* Hawaii with old firmware needs type2 nop packet.
* accel_working2 with value 3 indicates the new firmware.
*/
- ws->info.gfx_ib_pad_with_type2 = ws->info.chip_class <= SI ||
+ ws->info.gfx_ib_pad_with_type2 = ws->info.chip_class <= GFX6 ||
(ws->info.family == CHIP_HAWAII &&
ws->accel_working2 < 3);
ws->info.tcc_cache_line_size = 64; /* TC L2 line size on GCN */
-
- ws->check_vm = strstr(debug_get_option("R600_DEBUG", ""), "check_vm") != NULL;
+ ws->info.ib_start_alignment = 4096;
+ ws->info.kernel_flushes_hdp_before_ib = ws->info.drm_minor >= 40;
+ /* HTILE is broken with 1D tiling on old kernels and GFX7. */
+ ws->info.htile_cmask_support_1d_tiling = ws->info.chip_class != GFX7 ||
+ ws->info.drm_minor >= 38;
+ ws->info.si_TA_CS_BC_BASE_ADDR_allowed = ws->info.drm_minor >= 48;
+ ws->info.has_bo_metadata = false;
+ ws->info.has_gpu_reset_status_query = ws->info.drm_minor >= 43;
+ ws->info.has_eqaa_surface_allocator = false;
+ ws->info.has_format_bc1_through_bc7 = ws->info.drm_minor >= 31;
+ ws->info.kernel_flushes_tc_l2_after_ib = true;
+ /* Old kernels disallowed register writes via COPY_DATA
+ * that are used for indirect compute dispatches. */
+ ws->info.has_indirect_compute_dispatch = ws->info.chip_class == GFX7 ||
+ (ws->info.chip_class == GFX6 &&
+ ws->info.drm_minor >= 45);
+ /* GFX6 doesn't support unaligned loads. */
+ ws->info.has_unaligned_shader_loads = ws->info.chip_class == GFX7 &&
+ ws->info.drm_minor >= 50;
+ ws->info.has_sparse_vm_mappings = false;
+ /* 2D tiling on GFX7 is supported since DRM 2.35.0 */
+ ws->info.has_2d_tiling = ws->info.chip_class <= GFX6 || ws->info.drm_minor >= 35;
+ ws->info.has_read_registers_query = ws->info.drm_minor >= 42;
+ ws->info.max_alignment = 1024*1024;
+ ws->info.has_graphics = true;
+ ws->info.cpdma_prefetch_writes_memory = true;
+ ws->info.max_wave64_per_simd = 10;
+ ws->info.num_physical_sgprs_per_simd = 512;
+ ws->info.num_physical_wave64_vgprs_per_simd = 256;
+
+ ws->check_vm = strstr(debug_get_option("R600_DEBUG", ""), "check_vm") != NULL ||
+ strstr(debug_get_option("AMD_DEBUG", ""), "check_vm") != NULL;
return true;
}
mtx_destroy(&ws->hyperz_owner_mutex);
mtx_destroy(&ws->cmask_owner_mutex);
- if (ws->info.has_virtual_memory)
+ if (ws->info.r600_has_virtual_memory)
pb_slabs_deinit(&ws->bo_slabs);
pb_cache_deinit(&ws->bo_cache);
util_hash_table_destroy(ws->bo_handles);
util_hash_table_destroy(ws->bo_vas);
mtx_destroy(&ws->bo_handles_mutex);
+ mtx_destroy(&ws->vm32.mutex);
mtx_destroy(&ws->vm64.mutex);
mtx_destroy(&ws->bo_fence_lock);
*info = ((struct radeon_drm_winsys *)rws)->info;
}
-static bool radeon_cs_request_feature(struct radeon_winsys_cs *rcs,
+static bool radeon_cs_request_feature(struct radeon_cmdbuf *rcs,
enum radeon_feature_id fid,
bool enable)
{
return false;
}
+uint32_t radeon_drm_get_gpu_reset_counter(struct radeon_drm_winsys *ws)
+{
+ uint64_t retval = 0;
+
+ if (!ws->info.has_gpu_reset_status_query)
+ return 0;
+
+ radeon_get_drm_value(ws->fd, RADEON_INFO_GPU_RESET_COUNTER,
+ "gpu-reset-counter", (uint32_t*)&retval);
+ return retval;
+}
+
static uint64_t radeon_query_value(struct radeon_winsys *rws,
enum radeon_value_id value)
{
radeon_get_drm_value(ws->fd, RADEON_INFO_CURRENT_GPU_MCLK,
"current-gpu-mclk", (uint32_t*)&retval);
return retval;
- case RADEON_GPU_RESET_COUNTER:
- radeon_get_drm_value(ws->fd, RADEON_INFO_GPU_RESET_COUNTER,
- "gpu-reset-counter", (uint32_t*)&retval);
- return retval;
case RADEON_CS_THREAD_TIME:
return util_queue_get_thread_time_nano(&ws->cs_queue, 0);
}
mtx_lock(&fd_tab_mutex);
destroy = pipe_reference(&rws->reference, NULL);
- if (destroy && fd_tab)
+ if (destroy && fd_tab) {
util_hash_table_remove(fd_tab, intptr_to_pointer(rws->fd));
+ if (util_hash_table_count(fd_tab) == 0) {
+ util_hash_table_destroy(fd_tab);
+ fd_tab = NULL;
+ }
+ }
mtx_unlock(&fd_tab_mutex);
return destroy;
return PTR_TO_UINT(key1) != PTR_TO_UINT(key2);
}
+static void radeon_pin_threads_to_L3_cache(struct radeon_winsys *ws,
+ unsigned cache)
+{
+ struct radeon_drm_winsys *rws = (struct radeon_drm_winsys*)ws;
+
+ if (util_queue_is_initialized(&rws->cs_queue)) {
+ util_pin_thread_to_L3(rws->cs_queue.threads[0], cache,
+ util_cpu_caps.cores_per_L3);
+ }
+}
+
PUBLIC struct radeon_winsys *
radeon_drm_winsys_create(int fd, const struct pipe_screen_config *config,
radeon_screen_create_t screen_create)
radeon_bo_destroy,
radeon_bo_can_reclaim);
- if (ws->info.has_virtual_memory) {
+ if (ws->info.r600_has_virtual_memory) {
/* There is no fundamental obstacle to using slab buffer allocation
* without GPUVM, but enabling it requires making sure that the drivers
* honor the address offset.
ws->base.unref = radeon_winsys_unref;
ws->base.destroy = radeon_winsys_destroy;
ws->base.query_info = radeon_query_info;
+ ws->base.pin_threads_to_L3_cache = radeon_pin_threads_to_L3_cache;
ws->base.cs_request_feature = radeon_cs_request_feature;
ws->base.query_value = radeon_query_value;
ws->base.read_registers = radeon_read_registers;
ws->bo_handles = util_hash_table_create(handle_hash, handle_compare);
ws->bo_vas = util_hash_table_create(handle_hash, handle_compare);
(void) mtx_init(&ws->bo_handles_mutex, mtx_plain);
+ (void) mtx_init(&ws->vm32.mutex, mtx_plain);
(void) mtx_init(&ws->vm64.mutex, mtx_plain);
(void) mtx_init(&ws->bo_fence_lock, mtx_plain);
+ list_inithead(&ws->vm32.holes);
list_inithead(&ws->vm64.holes);
- ws->vm64.start = ws->va_start;
+ /* The kernel currently returns 8MB. Make sure this doesn't change. */
+ if (ws->va_start > 8 * 1024 * 1024) {
+ /* Not enough 32-bit address space. */
+ radeon_winsys_destroy(&ws->base);
+ mtx_unlock(&fd_tab_mutex);
+ return NULL;
+ }
+
+ ws->vm32.start = ws->va_start;
+ ws->vm32.end = 1ull << 32;
+
+ /* The maximum is 8GB of virtual address space limited by the kernel.
+ * It's obviously not enough for bigger cards, like Hawaiis with 4GB
+ * and 8GB of physical memory and 4GB of GART.
+ *
+ * Older kernels set the limit to 4GB, which is even worse, so they only
+ * have 32-bit address space.
+ */
+ if (ws->info.drm_minor >= 41) {
+ ws->vm64.start = 1ull << 32;
+ ws->vm64.end = 1ull << 33;
+ }
/* TTM aligns the BO size to the CPU page size */
ws->info.gart_page_size = sysconf(_SC_PAGESIZE);
if (ws->num_cpus > 1 && debug_get_option_thread())
- util_queue_init(&ws->cs_queue, "radeon_cs", 8, 1, 0);
+ util_queue_init(&ws->cs_queue, "rcs", 8, 1, 0);
/* Create the screen at the end. The winsys must be initialized
* completely.
return &ws->base;
fail_slab:
- if (ws->info.has_virtual_memory)
+ if (ws->info.r600_has_virtual_memory)
pb_slabs_deinit(&ws->bo_slabs);
fail_cache:
pb_cache_deinit(&ws->bo_cache);