#include "sid.h"
#include "radeon/radeon_uvd.h"
+#include "util/hash_table.h"
+#include "util/u_log.h"
#include "util/u_memory.h"
#include "util/u_suballoc.h"
#include "util/u_tests.h"
+#include "util/xmlconfig.h"
#include "vl/vl_decoder.h"
#include "../ddebug/dd_util.h"
+#include "compiler/nir/nir.h"
+
/*
* pipe_context
*/
si_release_all_descriptors(sctx);
- if (sctx->ce_suballocator)
- u_suballocator_destroy(sctx->ce_suballocator);
-
- r600_resource_reference(&sctx->ce_ram_saved_buffer, NULL);
pipe_resource_reference(&sctx->esgs_ring, NULL);
pipe_resource_reference(&sctx->gsvs_ring, NULL);
pipe_resource_reference(&sctx->tf_ring, NULL);
free(sctx->border_color_table);
r600_resource_reference(&sctx->scratch_buffer, NULL);
r600_resource_reference(&sctx->compute_scratch_buffer, NULL);
+ r600_resource_reference(&sctx->wait_mem_scratch, NULL);
si_pm4_free_state(sctx, sctx->init_config, ~0);
if (sctx->init_config_gs_rings)
LLVMDisposeTargetMachine(sctx->tm);
- r600_resource_reference(&sctx->trace_buf, NULL);
- r600_resource_reference(&sctx->last_trace_buf, NULL);
- radeon_clear_saved_cs(&sctx->last_gfx);
+ si_saved_cs_reference(&sctx->current_saved_cs, NULL);
+ _mesa_hash_table_destroy(sctx->tex_handles, NULL);
+ _mesa_hash_table_destroy(sctx->img_handles, NULL);
+
+ util_dynarray_fini(&sctx->resident_tex_handles);
+ util_dynarray_fini(&sctx->resident_img_handles);
+ util_dynarray_fini(&sctx->resident_tex_needs_color_decompress);
+ util_dynarray_fini(&sctx->resident_img_needs_color_decompress);
+ util_dynarray_fini(&sctx->resident_tex_needs_depth_decompress);
FREE(sctx);
}
struct si_context *sctx = (struct si_context *)ctx;
dd_parse_apitrace_marker(string, len, &sctx->apitrace_call_number);
+
+ if (sctx->b.log)
+ u_log_printf(sctx->b.log, "\nString marker: %*s\n", len, string);
}
static LLVMTargetMachineRef
char features[256];
snprintf(features, sizeof(features),
- "+DumpCode,+vgpr-spilling,-fp32-denormals,+fp64-denormals%s%s",
+ "+DumpCode,+vgpr-spilling,-fp32-denormals,+fp64-denormals%s%s%s",
sscreen->b.chip_class >= GFX9 ? ",+xnack" : ",-xnack",
+ sscreen->llvm_has_working_vgpr_indexing ? "" : ",-promote-alloca",
sscreen->b.debug_flags & DBG_SI_SCHED ? ",+si-scheduler" : "");
- return LLVMCreateTargetMachine(si_llvm_get_amdgpu_target(triple), triple,
+ return LLVMCreateTargetMachine(ac_get_llvm_target(triple), triple,
r600_get_llvm_processor_name(sscreen->b.family),
features,
LLVMCodeGenLevelDefault,
LLVMCodeModelDefault);
}
+static void si_set_log_context(struct pipe_context *ctx,
+ struct u_log_context *log)
+{
+ struct si_context *sctx = (struct si_context *)ctx;
+ sctx->b.log = log;
+
+ if (log)
+ u_log_add_auto_logger(log, si_auto_log_cs, sctx);
+}
+
static struct pipe_context *si_create_context(struct pipe_screen *screen,
unsigned flags)
{
if (!sctx)
return NULL;
- if (sscreen->b.debug_flags & DBG_CHECK_VM)
- flags |= PIPE_CONTEXT_DEBUG;
-
if (flags & PIPE_CONTEXT_DEBUG)
sscreen->record_llvm_ir = true; /* racy but not critical */
sctx->b.b.priv = NULL;
sctx->b.b.destroy = si_destroy_context;
sctx->b.b.emit_string_marker = si_emit_string_marker;
+ sctx->b.b.set_log_context = si_set_log_context;
sctx->b.set_atom_dirty = (void *)si_set_atom_dirty;
sctx->screen = sscreen; /* Easy accessing of screen/winsys. */
sctx->is_debug = (flags & PIPE_CONTEXT_DEBUG) != 0;
sctx->b.gfx.cs = ws->cs_create(sctx->b.ctx, RING_GFX,
si_context_gfx_flush, sctx);
-
- /* SI + AMDGPU + CE = GPU hang */
- if (!(sscreen->b.debug_flags & DBG_NO_CE) && ws->cs_add_const_ib &&
- sscreen->b.chip_class != SI &&
- /* These can't use CE due to a power gating bug in the kernel. */
- sscreen->b.family != CHIP_CARRIZO &&
- sscreen->b.family != CHIP_STONEY &&
- /* Some CE bug is causing green screen corruption w/ MPV video
- * playback and occasional corruption w/ 3D. */
- sscreen->b.chip_class != GFX9) {
- sctx->ce_ib = ws->cs_add_const_ib(sctx->b.gfx.cs);
- if (!sctx->ce_ib)
- goto fail;
-
- if (ws->cs_add_const_preamble_ib) {
- sctx->ce_preamble_ib =
- ws->cs_add_const_preamble_ib(sctx->b.gfx.cs);
-
- if (!sctx->ce_preamble_ib)
- goto fail;
- }
-
- sctx->ce_suballocator =
- u_suballocator_create(&sctx->b.b, 1024 * 1024, 0,
- PIPE_USAGE_DEFAULT,
- R600_RESOURCE_FLAG_UNMAPPABLE, false);
- if (!sctx->ce_suballocator)
- goto fail;
- }
-
sctx->b.gfx.flush = si_context_gfx_flush;
/* Border colors. */
/* these must be last */
si_begin_new_cs(sctx);
+ if (sctx->b.chip_class >= GFX9) {
+ sctx->wait_mem_scratch = (struct r600_resource*)
+ pipe_buffer_create(screen, 0, PIPE_USAGE_DEFAULT, 4);
+ if (!sctx->wait_mem_scratch)
+ goto fail;
+
+ /* Initialize the memory. */
+ struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
+ radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
+ radeon_emit(cs, S_370_DST_SEL(V_370_MEMORY_SYNC) |
+ S_370_WR_CONFIRM(1) |
+ S_370_ENGINE_SEL(V_370_ME));
+ radeon_emit(cs, sctx->wait_mem_scratch->gpu_address);
+ radeon_emit(cs, sctx->wait_mem_scratch->gpu_address >> 32);
+ radeon_emit(cs, sctx->wait_mem_number);
+ }
+
/* CIK cannot unbind a constant buffer (S_BUFFER_LOAD doesn't skip loads
* if NUM_RECORDS == 0). We need to use a dummy buffer instead. */
if (sctx->b.chip_class == CIK) {
si_set_rw_buffer(sctx, SI_HS_CONST_DEFAULT_TESS_LEVELS,
&sctx->null_const_buf);
+ si_set_rw_buffer(sctx, SI_VS_CONST_INSTANCE_DIVISORS,
+ &sctx->null_const_buf);
si_set_rw_buffer(sctx, SI_VS_CONST_CLIP_PLANES,
&sctx->null_const_buf);
si_set_rw_buffer(sctx, SI_PS_CONST_POLY_STIPPLE,
sctx->tm = si_create_llvm_target_machine(sscreen);
+ /* Bindless handles. */
+ sctx->tex_handles = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+ sctx->img_handles = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+
+ util_dynarray_init(&sctx->resident_tex_handles, NULL);
+ util_dynarray_init(&sctx->resident_img_handles, NULL);
+ util_dynarray_init(&sctx->resident_tex_needs_color_decompress, NULL);
+ util_dynarray_init(&sctx->resident_img_needs_color_decompress, NULL);
+ util_dynarray_init(&sctx->resident_tex_needs_depth_decompress, NULL);
+
return &sctx->b.b;
fail:
fprintf(stderr, "radeonsi: Failed to create a context.\n");
void *priv, unsigned flags)
{
struct si_screen *sscreen = (struct si_screen *)screen;
- struct pipe_context *ctx = si_create_context(screen, flags);
+ struct pipe_context *ctx;
+
+ if (sscreen->b.debug_flags & DBG_CHECK_VM)
+ flags |= PIPE_CONTEXT_DEBUG;
+
+ ctx = si_create_context(screen, flags);
if (!(flags & PIPE_CONTEXT_PREFER_THREADED))
return ctx;
case PIPE_CAP_DOUBLES:
case PIPE_CAP_TGSI_TEX_TXF_LZ:
case PIPE_CAP_TGSI_TES_LAYER_VIEWPORT:
+ case PIPE_CAP_BINDLESS_TEXTURE:
+ case PIPE_CAP_QUERY_TIMESTAMP:
+ case PIPE_CAP_QUERY_TIME_ELAPSED:
+ case PIPE_CAP_NIR_SAMPLERS_AS_DEREF:
+ case PIPE_CAP_QUERY_SO_OVERFLOW:
+ case PIPE_CAP_MEMOBJ:
+ return 1;
+
case PIPE_CAP_INT64:
case PIPE_CAP_INT64_DIVMOD:
case PIPE_CAP_TGSI_CLOCK:
return 4;
case PIPE_CAP_GLSL_FEATURE_LEVEL:
+ if (sscreen->b.debug_flags & DBG_NIR)
+ return 140; /* no geometry and tessellation shaders yet */
if (si_have_tgsi_compute(sscreen))
return 450;
return 420;
sscreen->b.info.drm_minor < 50);
case PIPE_CAP_SPARSE_BUFFER_PAGE_SIZE:
+ /* TODO: GFX9 hangs. */
+ if (sscreen->b.chip_class >= GFX9)
+ return 0;
/* Disable on SI due to VM faults in CP DMA. Enable once these
* faults are mitigated in software.
*/
case PIPE_CAP_MAX_RENDER_TARGETS:
return 8;
- /* Timer queries, present when the clock frequency is non zero. */
- case PIPE_CAP_QUERY_TIMESTAMP:
- case PIPE_CAP_QUERY_TIME_ELAPSED:
- return sscreen->b.info.clock_crystal_freq != 0;
-
case PIPE_CAP_MIN_TEXTURE_GATHER_OFFSET:
case PIPE_CAP_MIN_TEXEL_OFFSET:
return -32;
case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT:
return 32;
case PIPE_SHADER_CAP_PREFERRED_IR:
+ if (sscreen->b.debug_flags & DBG_NIR &&
+ (shader == PIPE_SHADER_VERTEX ||
+ shader == PIPE_SHADER_FRAGMENT))
+ return PIPE_SHADER_IR_NIR;
return PIPE_SHADER_IR_TGSI;
case PIPE_SHADER_CAP_LOWER_IF_THRESHOLD:
return 3;
/* Supported boolean features. */
case PIPE_SHADER_CAP_TGSI_CONT_SUPPORTED:
case PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED:
- case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR:
case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR:
case PIPE_SHADER_CAP_INDIRECT_CONST_ADDR:
case PIPE_SHADER_CAP_INTEGERS:
return 1;
case PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR:
- /* TODO: Indirection of geometry shader input dimension is not
- * handled yet
- */
- return shader != PIPE_SHADER_GEOMETRY;
+ /* TODO: Indirect indexing of GS inputs is unimplemented. */
+ return shader != PIPE_SHADER_GEOMETRY &&
+ (sscreen->llvm_has_working_vgpr_indexing ||
+ /* TCS and TES load inputs directly from LDS or
+ * offchip memory, so indirect indexing is trivial. */
+ shader == PIPE_SHADER_TESS_CTRL ||
+ shader == PIPE_SHADER_TESS_EVAL);
+
+ case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR:
+ return sscreen->llvm_has_working_vgpr_indexing ||
+ /* TCS stores outputs directly to memory. */
+ shader == PIPE_SHADER_TESS_CTRL;
/* Unsupported boolean features. */
case PIPE_SHADER_CAP_SUBROUTINES:
return 0;
}
+static const struct nir_shader_compiler_options nir_options = {
+ .vertex_id_zero_based = true,
+ .lower_scmp = true,
+ .lower_flrp32 = true,
+ .lower_fsat = true,
+ .lower_fdiv = true,
+ .lower_sub = true,
+ .lower_pack_snorm_2x16 = true,
+ .lower_pack_snorm_4x8 = true,
+ .lower_pack_unorm_2x16 = true,
+ .lower_pack_unorm_4x8 = true,
+ .lower_unpack_snorm_2x16 = true,
+ .lower_unpack_snorm_4x8 = true,
+ .lower_unpack_unorm_2x16 = true,
+ .lower_unpack_unorm_4x8 = true,
+ .lower_extract_byte = true,
+ .lower_extract_word = true,
+ .max_unroll_iterations = 32,
+ .native_integers = true,
+};
+
+static const void *
+si_get_compiler_options(struct pipe_screen *screen,
+ enum pipe_shader_ir ir,
+ enum pipe_shader_type shader)
+{
+ assert(ir == PIPE_SHADER_IR_NIR);
+ return &nir_options;
+}
+
static void si_destroy_screen(struct pipe_screen* pscreen)
{
struct si_screen *sscreen = (struct si_screen *)pscreen;
exit(0);
}
-struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws)
+static void radeonsi_get_driver_uuid(struct pipe_screen *pscreen, char *uuid)
+{
+ ac_compute_driver_uuid(uuid, PIPE_UUID_SIZE);
+}
+
+static void radeonsi_get_device_uuid(struct pipe_screen *pscreen, char *uuid)
+{
+ struct r600_common_screen *rscreen = (struct r600_common_screen *)pscreen;
+
+ ac_compute_device_uuid(&rscreen->info, uuid, PIPE_UUID_SIZE);
+}
+
+struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
+ const struct pipe_screen_config *config)
{
struct si_screen *sscreen = CALLOC_STRUCT(si_screen);
unsigned num_threads, num_compiler_threads, num_compiler_threads_lowprio, i;
sscreen->b.b.destroy = si_destroy_screen;
sscreen->b.b.get_param = si_get_param;
sscreen->b.b.get_shader_param = si_get_shader_param;
+ sscreen->b.b.get_compiler_options = si_get_compiler_options;
+ sscreen->b.b.get_device_uuid = radeonsi_get_device_uuid;
+ sscreen->b.b.get_driver_uuid = radeonsi_get_driver_uuid;
sscreen->b.b.resource_create = r600_resource_create_common;
si_init_screen_state_functions(sscreen);
+ /* Set these flags in debug_flags early, so that the shader cache takes
+ * them into account.
+ */
+ if (driQueryOptionb(config->options,
+ "glsl_correct_derivatives_after_discard"))
+ sscreen->b.debug_flags |= DBG_FS_CORRECT_DERIVS_AFTER_KILL;
+ if (driQueryOptionb(config->options, "radeonsi_enable_sisched"))
+ sscreen->b.debug_flags |= DBG_SI_SCHED;
+
if (!r600_common_screen_init(&sscreen->b, ws) ||
!si_init_gs_info(sscreen) ||
!si_init_shader_cache(sscreen)) {
MIN2(num_threads, ARRAY_SIZE(sscreen->tm_low_priority));
if (!util_queue_init(&sscreen->shader_compiler_queue, "si_shader",
- 32, num_compiler_threads, 0)) {
+ 32, num_compiler_threads,
+ UTIL_QUEUE_INIT_RESIZE_IF_FULL)) {
si_destroy_shader_cache(sscreen);
FREE(sscreen);
return NULL;
}
- /* The queue must be large enough so that adding optimized shaders
- * doesn't stall draw calls when the queue is full. Especially varying
- * packing generates a very high volume of optimized shader compilation
- * jobs.
- */
if (!util_queue_init(&sscreen->shader_compiler_queue_low_priority,
"si_shader_low",
- 1024, num_compiler_threads,
+ 32, num_compiler_threads_lowprio,
+ UTIL_QUEUE_INIT_RESIZE_IF_FULL |
UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY)) {
si_destroy_shader_cache(sscreen);
FREE(sscreen);
sscreen->tess_offchip_block_dw_size =
sscreen->b.family == CHIP_HAWAII ? 4096 : 8192;
+ /* The mere presense of CLEAR_STATE in the IB causes random GPU hangs
+ * on SI. */
+ sscreen->has_clear_state = sscreen->b.chip_class >= CIK;
+
sscreen->has_distributed_tess =
sscreen->b.chip_class >= VI &&
sscreen->b.info.max_se >= 2;
sscreen->b.info.pfp_fw_version >= 211 &&
sscreen->b.info.me_fw_version >= 173) ||
(sscreen->b.chip_class == SI &&
- sscreen->b.info.pfp_fw_version >= 121 &&
- sscreen->b.info.me_fw_version >= 87);
+ sscreen->b.info.pfp_fw_version >= 79 &&
+ sscreen->b.info.me_fw_version >= 142);
sscreen->has_ds_bpermute = sscreen->b.chip_class >= VI;
sscreen->has_msaa_sample_loc_bug = (sscreen->b.family >= CHIP_POLARIS10 &&
sscreen->b.family <= CHIP_POLARIS12) ||
sscreen->b.family == CHIP_VEGA10 ||
sscreen->b.family == CHIP_RAVEN;
+ /* While it would be nice not to have this flag, we are constrained
+ * by the reality that LLVM 5.0 doesn't have working VGPR indexing
+ * on GFX9.
+ */
+ sscreen->llvm_has_working_vgpr_indexing = sscreen->b.chip_class <= VI;
sscreen->b.has_cp_dma = true;
sscreen->b.has_streamout = true;
(sscreen->b.debug_flags & DBG_MONOLITHIC_SHADERS) != 0;
sscreen->b.barrier_flags.cp_to_L2 = SI_CONTEXT_INV_SMEM_L1 |
- SI_CONTEXT_INV_VMEM_L1 |
- SI_CONTEXT_INV_GLOBAL_L2;
+ SI_CONTEXT_INV_VMEM_L1;
+ if (sscreen->b.chip_class <= VI)
+ sscreen->b.barrier_flags.cp_to_L2 |= SI_CONTEXT_INV_GLOBAL_L2;
+
sscreen->b.barrier_flags.compute_to_L2 = SI_CONTEXT_CS_PARTIAL_FLUSH;
if (debug_get_bool_option("RADEON_DUMP_SHADERS", false))