/*
* Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
+ * Copyright 2018 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
#include "si_shader_internal.h"
#include "sid.h"
+#include "ac_llvm_util.h"
#include "radeon/radeon_uvd.h"
-#include "util/hash_table.h"
+#include "gallivm/lp_bld_misc.h"
+#include "util/disk_cache.h"
#include "util/u_log.h"
#include "util/u_memory.h"
#include "util/u_suballoc.h"
#include "util/u_tests.h"
+#include "util/u_upload_mgr.h"
#include "util/xmlconfig.h"
#include "vl/vl_decoder.h"
-#include "../ddebug/dd_util.h"
+#include "driver_ddebug/dd_util.h"
static const struct debug_named_value debug_options[] = {
/* Shader logging options: */
/* Shader compiler options the shader cache should be aware of: */
{ "unsafemath", DBG(UNSAFE_MATH), "Enable unsafe math shader optimizations" },
{ "sisched", DBG(SI_SCHED), "Enable LLVM SI Machine Instruction Scheduler." },
+ { "gisel", DBG(GISEL), "Enable LLVM global instruction selector." },
/* Shader compiler options (with no effect on the shader cache): */
{ "checkir", DBG(CHECK_IR), "Enable additional sanity checks on shader IR" },
{ "nowc", DBG(NO_WC), "Disable GTT write combining" },
{ "check_vm", DBG(CHECK_VM), "Check VM faults and dump debug info." },
{ "reserve_vmid", DBG(RESERVE_VMID), "Force VMID reservation per context." },
+ { "zerovram", DBG(ZERO_VRAM), "Clear VRAM allocations." },
/* 3D engine options: */
{ "switch_on_eop", DBG(SWITCH_ON_EOP), "Program WD/IA to switch on end-of-packet." },
{ "nodccclear", DBG(NO_DCC_CLEAR), "Disable DCC fast clear." },
{ "nodccfb", DBG(NO_DCC_FB), "Disable separate DCC on the main framebuffer" },
{ "nodccmsaa", DBG(NO_DCC_MSAA), "Disable DCC for MSAA" },
- { "dccmsaa", DBG(DCC_MSAA), "Enable DCC for MSAA" },
+ { "nofmask", DBG(NO_FMASK), "Disable MSAA compression" },
/* Tests: */
{ "testdma", DBG(TEST_DMA), "Invoke SDMA tests and exit." },
{ "testvmfaultcp", DBG(TEST_VMFAULT_CP), "Invoke a CP VM fault test and exit." },
{ "testvmfaultsdma", DBG(TEST_VMFAULT_SDMA), "Invoke a SDMA VM fault test and exit." },
{ "testvmfaultshader", DBG(TEST_VMFAULT_SHADER), "Invoke a shader VM fault test and exit." },
+ { "testdmaperf", DBG(TEST_DMA_PERF), "Test DMA performance" },
+ { "testgds", DBG(TEST_GDS), "Test GDS." },
+ { "testgdsmm", DBG(TEST_GDS_MM), "Test GDS memory management." },
+ { "testgdsoamm", DBG(TEST_GDS_OA_MM), "Test GDS OA memory management." },
DEBUG_NAMED_VALUE_END /* must be last */
};
+static void si_init_compiler(struct si_screen *sscreen,
+ struct ac_llvm_compiler *compiler)
+{
+ /* Only create the less-optimizing version of the compiler on APUs
+ * predating Ryzen (Raven). */
+ bool create_low_opt_compiler = !sscreen->info.has_dedicated_vram &&
+ sscreen->info.chip_class <= VI;
+
+ enum ac_target_machine_options tm_options =
+ (sscreen->debug_flags & DBG(SI_SCHED) ? AC_TM_SISCHED : 0) |
+ (sscreen->debug_flags & DBG(GISEL) ? AC_TM_ENABLE_GLOBAL_ISEL : 0) |
+ (sscreen->info.chip_class >= GFX9 ? AC_TM_FORCE_ENABLE_XNACK : 0) |
+ (sscreen->info.chip_class < GFX9 ? AC_TM_FORCE_DISABLE_XNACK : 0) |
+ (!sscreen->llvm_has_working_vgpr_indexing ? AC_TM_PROMOTE_ALLOCA_TO_SCRATCH : 0) |
+ (sscreen->debug_flags & DBG(CHECK_IR) ? AC_TM_CHECK_IR : 0) |
+ (create_low_opt_compiler ? AC_TM_CREATE_LOW_OPT : 0);
+
+ ac_init_llvm_once();
+ ac_init_llvm_compiler(compiler, sscreen->info.family, tm_options);
+ compiler->passes = ac_create_llvm_passes(compiler->tm);
+
+ if (compiler->low_opt_tm)
+ compiler->low_opt_passes = ac_create_llvm_passes(compiler->low_opt_tm);
+}
+
+static void si_destroy_compiler(struct ac_llvm_compiler *compiler)
+{
+ ac_destroy_llvm_passes(compiler->passes);
+ ac_destroy_llvm_passes(compiler->low_opt_passes);
+ ac_destroy_llvm_compiler(compiler);
+}
+
/*
* pipe_context
*/
pipe_resource_reference(&sctx->esgs_ring, NULL);
pipe_resource_reference(&sctx->gsvs_ring, NULL);
- pipe_resource_reference(&sctx->tf_ring, NULL);
- pipe_resource_reference(&sctx->tess_offchip_ring, NULL);
+ pipe_resource_reference(&sctx->tess_rings, NULL);
pipe_resource_reference(&sctx->null_const_buf.buffer, NULL);
+ pipe_resource_reference(&sctx->sample_pos_buffer, NULL);
r600_resource_reference(&sctx->border_color_buffer, NULL);
free(sctx->border_color_table);
r600_resource_reference(&sctx->scratch_buffer, NULL);
si_pm4_delete_state(sctx, vgt_shader_config, sctx->vgt_shader_config[i]);
if (sctx->fixed_func_tcs_shader.cso)
- sctx->b.b.delete_tcs_state(&sctx->b.b, sctx->fixed_func_tcs_shader.cso);
+ sctx->b.delete_tcs_state(&sctx->b, sctx->fixed_func_tcs_shader.cso);
if (sctx->custom_dsa_flush)
- sctx->b.b.delete_depth_stencil_alpha_state(&sctx->b.b, sctx->custom_dsa_flush);
+ sctx->b.delete_depth_stencil_alpha_state(&sctx->b, sctx->custom_dsa_flush);
if (sctx->custom_blend_resolve)
- sctx->b.b.delete_blend_state(&sctx->b.b, sctx->custom_blend_resolve);
+ sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_resolve);
if (sctx->custom_blend_fmask_decompress)
- sctx->b.b.delete_blend_state(&sctx->b.b, sctx->custom_blend_fmask_decompress);
+ sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_fmask_decompress);
if (sctx->custom_blend_eliminate_fastclear)
- sctx->b.b.delete_blend_state(&sctx->b.b, sctx->custom_blend_eliminate_fastclear);
+ sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_eliminate_fastclear);
if (sctx->custom_blend_dcc_decompress)
- sctx->b.b.delete_blend_state(&sctx->b.b, sctx->custom_blend_dcc_decompress);
+ sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_dcc_decompress);
if (sctx->vs_blit_pos)
- sctx->b.b.delete_vs_state(&sctx->b.b, sctx->vs_blit_pos);
+ sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_pos);
if (sctx->vs_blit_pos_layered)
- sctx->b.b.delete_vs_state(&sctx->b.b, sctx->vs_blit_pos_layered);
+ sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_pos_layered);
if (sctx->vs_blit_color)
- sctx->b.b.delete_vs_state(&sctx->b.b, sctx->vs_blit_color);
+ sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_color);
if (sctx->vs_blit_color_layered)
- sctx->b.b.delete_vs_state(&sctx->b.b, sctx->vs_blit_color_layered);
+ sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_color_layered);
if (sctx->vs_blit_texcoord)
- sctx->b.b.delete_vs_state(&sctx->b.b, sctx->vs_blit_texcoord);
+ sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_texcoord);
+ if (sctx->cs_clear_buffer)
+ sctx->b.delete_compute_state(&sctx->b, sctx->cs_clear_buffer);
+ if (sctx->cs_copy_buffer)
+ sctx->b.delete_compute_state(&sctx->b, sctx->cs_copy_buffer);
if (sctx->blitter)
util_blitter_destroy(sctx->blitter);
- si_common_context_cleanup(&sctx->b);
+ /* Release DCC stats. */
+ for (int i = 0; i < ARRAY_SIZE(sctx->dcc_stats); i++) {
+ assert(!sctx->dcc_stats[i].query_active);
+
+ for (int j = 0; j < ARRAY_SIZE(sctx->dcc_stats[i].ps_stats); j++)
+ if (sctx->dcc_stats[i].ps_stats[j])
+ sctx->b.destroy_query(&sctx->b,
+ sctx->dcc_stats[i].ps_stats[j]);
+
+ si_texture_reference(&sctx->dcc_stats[i].tex, NULL);
+ }
+
+ if (sctx->query_result_shader)
+ sctx->b.delete_compute_state(&sctx->b, sctx->query_result_shader);
+
+ if (sctx->gfx_cs)
+ sctx->ws->cs_destroy(sctx->gfx_cs);
+ if (sctx->dma_cs)
+ sctx->ws->cs_destroy(sctx->dma_cs);
+ if (sctx->ctx)
+ sctx->ws->ctx_destroy(sctx->ctx);
- LLVMDisposeTargetMachine(sctx->tm);
+ if (sctx->b.stream_uploader)
+ u_upload_destroy(sctx->b.stream_uploader);
+ if (sctx->b.const_uploader)
+ u_upload_destroy(sctx->b.const_uploader);
+ if (sctx->cached_gtt_allocator)
+ u_upload_destroy(sctx->cached_gtt_allocator);
+
+ slab_destroy_child(&sctx->pool_transfers);
+ slab_destroy_child(&sctx->pool_transfers_unsync);
+
+ if (sctx->allocator_zeroed_memory)
+ u_suballocator_destroy(sctx->allocator_zeroed_memory);
+
+ sctx->ws->fence_reference(&sctx->last_gfx_fence, NULL);
+ sctx->ws->fence_reference(&sctx->last_sdma_fence, NULL);
+ r600_resource_reference(&sctx->eop_bug_scratch, NULL);
+
+ si_destroy_compiler(&sctx->compiler);
si_saved_cs_reference(&sctx->current_saved_cs, NULL);
FREE(sctx);
}
-static enum pipe_reset_status
-si_amdgpu_get_reset_status(struct pipe_context *ctx)
+static enum pipe_reset_status si_get_reset_status(struct pipe_context *ctx)
+{
+ struct si_context *sctx = (struct si_context *)ctx;
+
+ if (sctx->screen->info.has_gpu_reset_status_query)
+ return sctx->ws->ctx_query_reset_status(sctx->ctx);
+
+ if (sctx->screen->info.has_gpu_reset_counter_query) {
+ unsigned latest = sctx->ws->query_value(sctx->ws,
+ RADEON_GPU_RESET_COUNTER);
+
+ if (sctx->gpu_reset_counter == latest)
+ return PIPE_NO_RESET;
+
+ sctx->gpu_reset_counter = latest;
+ return PIPE_UNKNOWN_CONTEXT_RESET;
+ }
+
+ return PIPE_NO_RESET;
+}
+
+static void si_set_device_reset_callback(struct pipe_context *ctx,
+ const struct pipe_device_reset_callback *cb)
{
struct si_context *sctx = (struct si_context *)ctx;
- return sctx->b.ws->ctx_query_reset_status(sctx->b.ctx);
+ if (cb)
+ sctx->device_reset_callback = *cb;
+ else
+ memset(&sctx->device_reset_callback, 0,
+ sizeof(sctx->device_reset_callback));
+}
+
+bool si_check_device_reset(struct si_context *sctx)
+{
+ enum pipe_reset_status status;
+
+ if (!sctx->device_reset_callback.reset)
+ return false;
+
+ if (!sctx->b.get_device_reset_status)
+ return false;
+
+ status = sctx->b.get_device_reset_status(&sctx->b);
+ if (status == PIPE_NO_RESET)
+ return false;
+
+ sctx->device_reset_callback.reset(sctx->device_reset_callback.data, status);
+ return true;
}
/* Apitrace profiling:
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
-si_create_llvm_target_machine(struct si_screen *sscreen)
-{
- enum ac_target_machine_options tm_options =
- (sscreen->debug_flags & DBG(SI_SCHED) ? AC_TM_SISCHED : 0) |
- (sscreen->info.chip_class >= GFX9 ? AC_TM_FORCE_ENABLE_XNACK : 0) |
- (sscreen->info.chip_class < GFX9 ? AC_TM_FORCE_DISABLE_XNACK : 0) |
- (!sscreen->llvm_has_working_vgpr_indexing ? AC_TM_PROMOTE_ALLOCA_TO_SCRATCH : 0);
-
- return ac_create_target_machine(sscreen->info.family, tm_options);
+ if (sctx->log)
+ u_log_printf(sctx->log, "\nString marker: %*s\n", len, string);
}
static void si_set_debug_callback(struct pipe_context *ctx,
struct u_log_context *log)
{
struct si_context *sctx = (struct si_context *)ctx;
- sctx->b.log = log;
+ sctx->log = log;
if (log)
u_log_add_auto_logger(log, si_auto_log_cs, sctx);
}
+static void si_set_context_param(struct pipe_context *ctx,
+ enum pipe_context_param param,
+ unsigned value)
+{
+ struct radeon_winsys *ws = ((struct si_context *)ctx)->ws;
+
+ switch (param) {
+ case PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE:
+ ws->pin_threads_to_L3_cache(ws, value);
+ break;
+ default:;
+ }
+}
+
static struct pipe_context *si_create_context(struct pipe_screen *screen,
unsigned flags)
{
struct si_screen* sscreen = (struct si_screen *)screen;
struct radeon_winsys *ws = sscreen->ws;
int shader, i;
+ bool stop_exec_on_failure = (flags & PIPE_CONTEXT_LOSE_CONTEXT_ON_RESET) != 0;
if (!sctx)
return NULL;
if (flags & PIPE_CONTEXT_DEBUG)
sscreen->record_llvm_ir = true; /* racy but not critical */
- sctx->b.b.screen = screen; /* this must be set first */
- 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_debug_callback = si_set_debug_callback;
- sctx->b.b.set_log_context = si_set_log_context;
- sctx->b.set_atom_dirty = (void *)si_set_atom_dirty;
+ sctx->b.screen = screen; /* this must be set first */
+ sctx->b.priv = NULL;
+ sctx->b.destroy = si_destroy_context;
+ sctx->b.emit_string_marker = si_emit_string_marker;
+ sctx->b.set_debug_callback = si_set_debug_callback;
+ sctx->b.set_log_context = si_set_log_context;
+ sctx->b.set_context_param = si_set_context_param;
sctx->screen = sscreen; /* Easy accessing of screen/winsys. */
sctx->is_debug = (flags & PIPE_CONTEXT_DEBUG) != 0;
- if (!si_common_context_init(&sctx->b, sscreen, flags))
+ slab_create_child(&sctx->pool_transfers, &sscreen->pool_transfers);
+ slab_create_child(&sctx->pool_transfers_unsync, &sscreen->pool_transfers);
+
+ sctx->ws = sscreen->ws;
+ sctx->family = sscreen->info.family;
+ sctx->chip_class = sscreen->info.chip_class;
+
+ if (sscreen->info.has_gpu_reset_counter_query) {
+ sctx->gpu_reset_counter =
+ sctx->ws->query_value(sctx->ws, RADEON_GPU_RESET_COUNTER);
+ }
+
+ sctx->b.get_device_reset_status = si_get_reset_status;
+ sctx->b.set_device_reset_callback = si_set_device_reset_callback;
+
+ si_init_context_texture_functions(sctx);
+ si_init_query_functions(sctx);
+
+ if (sctx->chip_class == CIK ||
+ sctx->chip_class == VI ||
+ sctx->chip_class == GFX9) {
+ sctx->eop_bug_scratch = r600_resource(
+ pipe_buffer_create(&sscreen->b, 0, PIPE_USAGE_DEFAULT,
+ 16 * sscreen->info.num_render_backends));
+ if (!sctx->eop_bug_scratch)
+ goto fail;
+ }
+
+ sctx->allocator_zeroed_memory =
+ u_suballocator_create(&sctx->b, sscreen->info.gart_page_size,
+ 0, PIPE_USAGE_DEFAULT,
+ SI_RESOURCE_FLAG_SO_FILLED_SIZE, true);
+ if (!sctx->allocator_zeroed_memory)
+ goto fail;
+
+ sctx->b.stream_uploader = u_upload_create(&sctx->b, 1024 * 1024,
+ 0, PIPE_USAGE_STREAM,
+ SI_RESOURCE_FLAG_READ_ONLY);
+ if (!sctx->b.stream_uploader)
+ goto fail;
+
+ sctx->b.const_uploader = u_upload_create(&sctx->b, 128 * 1024,
+ 0, PIPE_USAGE_DEFAULT,
+ SI_RESOURCE_FLAG_32BIT |
+ (sscreen->cpdma_prefetch_writes_memory ?
+ 0 : SI_RESOURCE_FLAG_READ_ONLY));
+ if (!sctx->b.const_uploader)
goto fail;
- if (sscreen->info.drm_major == 3)
- sctx->b.b.get_device_reset_status = si_amdgpu_get_reset_status;
+ sctx->cached_gtt_allocator = u_upload_create(&sctx->b, 16 * 1024,
+ 0, PIPE_USAGE_STAGING, 0);
+ if (!sctx->cached_gtt_allocator)
+ goto fail;
+
+ sctx->ctx = sctx->ws->ctx_create(sctx->ws);
+ if (!sctx->ctx)
+ goto fail;
+
+ if (sscreen->info.num_sdma_rings && !(sscreen->debug_flags & DBG(NO_ASYNC_DMA))) {
+ sctx->dma_cs = sctx->ws->cs_create(sctx->ctx, RING_DMA,
+ (void*)si_flush_dma_cs,
+ sctx, stop_exec_on_failure);
+ }
si_init_buffer_functions(sctx);
si_init_clear_functions(sctx);
si_init_blit_functions(sctx);
si_init_compute_functions(sctx);
- si_init_cp_dma_functions(sctx);
+ si_init_compute_blit_functions(sctx);
si_init_debug_functions(sctx);
si_init_msaa_functions(sctx);
si_init_streamout_functions(sctx);
if (sscreen->info.has_hw_decode) {
- sctx->b.b.create_video_codec = si_uvd_create_decoder;
- sctx->b.b.create_video_buffer = si_video_buffer_create;
+ sctx->b.create_video_codec = si_uvd_create_decoder;
+ sctx->b.create_video_buffer = si_video_buffer_create;
} else {
- sctx->b.b.create_video_codec = vl_create_decoder;
- sctx->b.b.create_video_buffer = vl_video_buffer_create;
+ sctx->b.create_video_codec = vl_create_decoder;
+ sctx->b.create_video_buffer = vl_video_buffer_create;
}
- sctx->b.gfx.cs = ws->cs_create(sctx->b.ctx, RING_GFX,
- si_context_gfx_flush, sctx);
- sctx->b.gfx.flush = si_context_gfx_flush;
+ sctx->gfx_cs = ws->cs_create(sctx->ctx, RING_GFX,
+ (void*)si_flush_gfx_cs, sctx, stop_exec_on_failure);
/* Border colors. */
sctx->border_color_table = malloc(SI_MAX_BORDER_COLORS *
if (!sctx->border_color_table)
goto fail;
- sctx->border_color_buffer = (struct r600_resource*)
+ sctx->border_color_buffer = r600_resource(
pipe_buffer_create(screen, 0, PIPE_USAGE_DEFAULT,
SI_MAX_BORDER_COLORS *
- sizeof(*sctx->border_color_table));
+ sizeof(*sctx->border_color_table)));
if (!sctx->border_color_buffer)
goto fail;
si_init_state_functions(sctx);
si_init_shader_functions(sctx);
si_init_viewport_functions(sctx);
- si_init_ia_multi_vgt_param_table(sctx);
- if (sctx->b.chip_class >= CIK)
+ if (sctx->chip_class >= CIK)
cik_init_sdma_functions(sctx);
else
si_init_dma_functions(sctx);
if (sscreen->debug_flags & DBG(FORCE_DMA))
- sctx->b.b.resource_copy_region = sctx->b.dma_copy;
+ sctx->b.resource_copy_region = sctx->dma_copy;
- sctx->blitter = util_blitter_create(&sctx->b.b);
+ sctx->blitter = util_blitter_create(&sctx->b);
if (sctx->blitter == NULL)
goto fail;
- sctx->blitter->draw_rectangle = si_draw_rectangle;
sctx->blitter->skip_viewport_restore = true;
- sctx->sample_mask.sample_mask = 0xffff;
+ si_init_draw_functions(sctx);
- /* these must be last */
- si_begin_new_cs(sctx);
+ sctx->sample_mask = 0xffff;
- if (sctx->b.chip_class >= GFX9) {
- sctx->wait_mem_scratch = (struct r600_resource*)
- pipe_buffer_create(screen, 0, PIPE_USAGE_DEFAULT, 4);
+ if (sctx->chip_class >= GFX9) {
+ sctx->wait_mem_scratch = 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);
+ si_cp_write_data(sctx, sctx->wait_mem_scratch, 0, 4,
+ V_370_MEM, V_370_ME, &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) {
+ if (sctx->chip_class == CIK) {
sctx->null_const_buf.buffer =
- si_aligned_buffer_create(screen,
- R600_RESOURCE_FLAG_32BIT,
+ pipe_aligned_buffer_create(screen,
+ SI_RESOURCE_FLAG_32BIT,
PIPE_USAGE_DEFAULT, 16,
sctx->screen->info.tcc_cache_line_size);
if (!sctx->null_const_buf.buffer)
for (shader = 0; shader < SI_NUM_SHADERS; shader++) {
for (i = 0; i < SI_NUM_CONST_BUFFERS; i++) {
- sctx->b.b.set_constant_buffer(&sctx->b.b, shader, i,
+ sctx->b.set_constant_buffer(&sctx->b, shader, i,
&sctx->null_const_buf);
}
}
&sctx->null_const_buf);
si_set_rw_buffer(sctx, SI_PS_CONST_SAMPLE_POSITIONS,
&sctx->null_const_buf);
-
- /* Clear the NULL constant buffer, because loads should return zeros. */
- si_clear_buffer(&sctx->b.b, sctx->null_const_buf.buffer, 0,
- sctx->null_const_buf.buffer->width0, 0,
- R600_COHERENCY_SHADER);
}
uint64_t max_threads_per_block;
sctx->scratch_waves = MAX2(32 * sscreen->info.num_good_compute_units,
max_threads_per_block / 64);
- sctx->tm = si_create_llvm_target_machine(sscreen);
+ si_init_compiler(sscreen, &sctx->compiler);
/* Bindless handles. */
sctx->tex_handles = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
util_dynarray_init(&sctx->resident_img_needs_color_decompress, NULL);
util_dynarray_init(&sctx->resident_tex_needs_depth_decompress, NULL);
- return &sctx->b.b;
+ sctx->sample_pos_buffer =
+ pipe_buffer_create(sctx->b.screen, 0, PIPE_USAGE_DEFAULT,
+ sizeof(sctx->sample_positions));
+ pipe_buffer_write(&sctx->b, sctx->sample_pos_buffer, 0,
+ sizeof(sctx->sample_positions), &sctx->sample_positions);
+
+ /* this must be last */
+ si_begin_new_gfx_cs(sctx);
+
+ if (sctx->chip_class == CIK) {
+ /* Clear the NULL constant buffer, because loads should return zeros. */
+ uint32_t clear_value = 0;
+ si_clear_buffer(sctx, sctx->null_const_buf.buffer, 0,
+ sctx->null_const_buf.buffer->width0,
+ &clear_value, 4, SI_COHERENCY_SHADER);
+ }
+ return &sctx->b;
fail:
fprintf(stderr, "radeonsi: Failed to create a context.\n");
- si_destroy_context(&sctx->b.b);
+ si_destroy_context(&sctx->b);
return NULL;
}
return threaded_context_create(ctx, &sscreen->pool_transfers,
si_replace_buffer_storage,
sscreen->info.drm_major >= 3 ? si_create_fence : NULL,
- &((struct si_context*)ctx)->b.tc);
+ &((struct si_context*)ctx)->tc);
}
/*
util_queue_destroy(&sscreen->shader_compiler_queue);
util_queue_destroy(&sscreen->shader_compiler_queue_low_priority);
- for (i = 0; i < ARRAY_SIZE(sscreen->tm); i++)
- if (sscreen->tm[i])
- LLVMDisposeTargetMachine(sscreen->tm[i]);
+ for (i = 0; i < ARRAY_SIZE(sscreen->compiler); i++)
+ si_destroy_compiler(&sscreen->compiler[i]);
- for (i = 0; i < ARRAY_SIZE(sscreen->tm_low_priority); i++)
- if (sscreen->tm_low_priority[i])
- LLVMDisposeTargetMachine(sscreen->tm_low_priority[i]);
+ for (i = 0; i < ARRAY_SIZE(sscreen->compiler_lowp); i++)
+ si_destroy_compiler(&sscreen->compiler_lowp[i]);
/* Free shader parts. */
for (i = 0; i < ARRAY_SIZE(parts); i++) {
mtx_destroy(&sscreen->shader_parts_mutex);
si_destroy_shader_cache(sscreen);
- si_perfcounters_destroy(sscreen);
+ si_destroy_perfcounters(sscreen);
si_gpu_load_kill_thread(sscreen);
mtx_destroy(&sscreen->gpu_load_mutex);
FREE(sscreen);
}
-static bool si_init_gs_info(struct si_screen *sscreen)
+static void si_init_gs_info(struct si_screen *sscreen)
{
- /* gs_table_depth is not used by GFX9 */
- if (sscreen->info.chip_class >= GFX9)
- return true;
-
- switch (sscreen->info.family) {
- case CHIP_OLAND:
- case CHIP_HAINAN:
- case CHIP_KAVERI:
- case CHIP_KABINI:
- case CHIP_MULLINS:
- case CHIP_ICELAND:
- case CHIP_CARRIZO:
- case CHIP_STONEY:
- sscreen->gs_table_depth = 16;
- return true;
- case CHIP_TAHITI:
- case CHIP_PITCAIRN:
- case CHIP_VERDE:
- case CHIP_BONAIRE:
- case CHIP_HAWAII:
- case CHIP_TONGA:
- case CHIP_FIJI:
- case CHIP_POLARIS10:
- case CHIP_POLARIS11:
- case CHIP_POLARIS12:
- sscreen->gs_table_depth = 32;
- return true;
- default:
- return false;
- }
-}
-
-static void si_handle_env_var_force_family(struct si_screen *sscreen)
-{
- const char *family = debug_get_option("SI_FORCE_FAMILY", NULL);
- unsigned i;
-
- if (!family)
- return;
-
- for (i = CHIP_TAHITI; i < CHIP_LAST; i++) {
- if (!strcmp(family, ac_get_llvm_processor_name(i))) {
- /* Override family and chip_class. */
- sscreen->info.family = i;
-
- if (i >= CHIP_VEGA10)
- sscreen->info.chip_class = GFX9;
- else if (i >= CHIP_TONGA)
- sscreen->info.chip_class = VI;
- else if (i >= CHIP_BONAIRE)
- sscreen->info.chip_class = CIK;
- else
- sscreen->info.chip_class = SI;
-
- /* Don't submit any IBs. */
- setenv("RADEON_NOOP", "1", 1);
- return;
- }
- }
-
- fprintf(stderr, "radeonsi: Unknown family: %s\n", family);
- exit(1);
+ sscreen->gs_table_depth = ac_get_gs_table_depth(sscreen->info.chip_class,
+ sscreen->info.family);
}
static void si_test_vmfault(struct si_screen *sscreen)
r600_resource(buf)->gpu_address = 0; /* cause a VM fault */
if (sscreen->debug_flags & DBG(TEST_VMFAULT_CP)) {
- si_copy_buffer(sctx, buf, buf, 0, 4, 4, 0);
+ si_cp_dma_copy_buffer(sctx, buf, buf, 0, 4, 4, 0,
+ SI_COHERENCY_NONE, L2_BYPASS);
ctx->flush(ctx, NULL, 0);
puts("VM fault test: CP - done.");
}
if (sscreen->debug_flags & DBG(TEST_VMFAULT_SDMA)) {
- sctx->b.dma_clear_buffer(ctx, buf, 0, 4, 0);
+ si_sdma_clear_buffer(sctx, buf, 0, 4, 0);
ctx->flush(ctx, NULL, 0);
puts("VM fault test: SDMA - done.");
}
exit(0);
}
+static void si_test_gds_memory_management(struct si_context *sctx,
+ unsigned alloc_size, unsigned alignment,
+ enum radeon_bo_domain domain)
+{
+ struct radeon_winsys *ws = sctx->ws;
+ struct radeon_cmdbuf *cs[8];
+ struct pb_buffer *gds_bo[ARRAY_SIZE(cs)];
+
+ for (unsigned i = 0; i < ARRAY_SIZE(cs); i++) {
+ cs[i] = ws->cs_create(sctx->ctx, RING_COMPUTE,
+ NULL, NULL, false);
+ gds_bo[i] = ws->buffer_create(ws, alloc_size, alignment, domain, 0);
+ assert(gds_bo[i]);
+ }
+
+ for (unsigned iterations = 0; iterations < 20000; iterations++) {
+ for (unsigned i = 0; i < ARRAY_SIZE(cs); i++) {
+ /* This clears GDS with CP DMA.
+ *
+ * We don't care if GDS is present. Just add some packet
+ * to make the GPU busy for a moment.
+ */
+ si_cp_dma_clear_buffer(sctx, cs[i], NULL, 0, alloc_size, 0,
+ SI_CPDMA_SKIP_BO_LIST_UPDATE |
+ SI_CPDMA_SKIP_CHECK_CS_SPACE |
+ SI_CPDMA_SKIP_GFX_SYNC, 0, 0);
+
+ ws->cs_add_buffer(cs[i], gds_bo[i], domain,
+ RADEON_USAGE_READWRITE, 0);
+ ws->cs_flush(cs[i], PIPE_FLUSH_ASYNC, NULL);
+ }
+ }
+ exit(0);
+}
+
static void si_disk_cache_create(struct si_screen *sscreen)
{
/* Don't use the cache if shader dumping is enabled. */
if (sscreen->debug_flags & DBG_ALL_SHADERS)
return;
- uint32_t mesa_timestamp;
- if (disk_cache_get_function_timestamp(si_disk_cache_create,
- &mesa_timestamp)) {
- char *timestamp_str;
- int res = -1;
- uint32_t llvm_timestamp;
-
- if (disk_cache_get_function_timestamp(LLVMInitializeAMDGPUTargetInfo,
- &llvm_timestamp)) {
- res = asprintf(×tamp_str, "%u_%u",
- mesa_timestamp, llvm_timestamp);
- }
+ struct mesa_sha1 ctx;
+ unsigned char sha1[20];
+ char cache_id[20 * 2 + 1];
- if (res != -1) {
- /* These flags affect shader compilation. */
- #define ALL_FLAGS (DBG(FS_CORRECT_DERIVS_AFTER_KILL) | \
- DBG(SI_SCHED) | \
- DBG(UNSAFE_MATH) | \
- DBG(NIR))
- uint64_t shader_debug_flags = sscreen->debug_flags &
- ALL_FLAGS;
-
- /* Add the high bits of 32-bit addresses, which affects
- * how 32-bit addresses are expanded to 64 bits.
- */
- STATIC_ASSERT(ALL_FLAGS <= UINT_MAX);
- shader_debug_flags |= (uint64_t)sscreen->info.address32_hi << 32;
-
- sscreen->disk_shader_cache =
- disk_cache_create(si_get_family_name(sscreen),
- timestamp_str,
- shader_debug_flags);
- free(timestamp_str);
- }
- }
+ _mesa_sha1_init(&ctx);
+
+ if (!disk_cache_get_function_identifier(si_disk_cache_create, &ctx) ||
+ !disk_cache_get_function_identifier(LLVMInitializeAMDGPUTargetInfo,
+ &ctx))
+ return;
+
+ _mesa_sha1_final(&ctx, sha1);
+ disk_cache_format_hex_id(cache_id, sha1, 20 * 2);
+
+ /* These flags affect shader compilation. */
+ #define ALL_FLAGS (DBG(FS_CORRECT_DERIVS_AFTER_KILL) | \
+ DBG(SI_SCHED) | \
+ DBG(GISEL) | \
+ DBG(UNSAFE_MATH) | \
+ DBG(NIR))
+ uint64_t shader_debug_flags = sscreen->debug_flags &
+ ALL_FLAGS;
+
+ /* Add the high bits of 32-bit addresses, which affects
+ * how 32-bit addresses are expanded to 64 bits.
+ */
+ STATIC_ASSERT(ALL_FLAGS <= UINT_MAX);
+ shader_debug_flags |= (uint64_t)sscreen->info.address32_hi << 32;
+
+ sscreen->disk_shader_cache =
+ disk_cache_create(sscreen->info.name,
+ cache_id,
+ shader_debug_flags);
}
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;
+ unsigned hw_threads, num_comp_hi_threads, num_comp_lo_threads, i;
if (!sscreen) {
return NULL;
sscreen->ws = ws;
ws->query_info(ws, &sscreen->info);
+ if (sscreen->info.chip_class >= GFX9) {
+ sscreen->se_tile_repeat = 32 * sscreen->info.max_se;
+ } else {
+ ac_get_raster_config(&sscreen->info,
+ &sscreen->pa_sc_raster_config,
+ &sscreen->pa_sc_raster_config_1,
+ &sscreen->se_tile_repeat);
+ }
+
sscreen->debug_flags = debug_get_flags_option("R600_DEBUG",
debug_options, 0);
ac_print_gpu_info(&sscreen->info);
slab_create_parent(&sscreen->pool_transfers,
- sizeof(struct r600_transfer), 64);
+ sizeof(struct si_transfer), 64);
sscreen->force_aniso = MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
if (sscreen->force_aniso >= 0) {
(void) mtx_init(&sscreen->aux_context_lock, mtx_plain);
(void) mtx_init(&sscreen->gpu_load_mutex, mtx_plain);
- if (!si_init_gs_info(sscreen) ||
- !si_init_shader_cache(sscreen)) {
+ si_init_gs_info(sscreen);
+ if (!si_init_shader_cache(sscreen)) {
FREE(sscreen);
return NULL;
}
si_disk_cache_create(sscreen);
- /* Only enable as many threads as we have target machines, but at most
- * the number of CPUs - 1 if there is more than one.
- */
- num_threads = sysconf(_SC_NPROCESSORS_ONLN);
- num_threads = MAX2(1, num_threads - 1);
- num_compiler_threads = MIN2(num_threads, ARRAY_SIZE(sscreen->tm));
- num_compiler_threads_lowprio =
- MIN2(num_threads, ARRAY_SIZE(sscreen->tm_low_priority));
-
- if (!util_queue_init(&sscreen->shader_compiler_queue, "si_shader",
- 32, num_compiler_threads,
- UTIL_QUEUE_INIT_RESIZE_IF_FULL)) {
+ /* Determine the number of shader compiler threads. */
+ hw_threads = sysconf(_SC_NPROCESSORS_ONLN);
+
+ if (hw_threads >= 12) {
+ num_comp_hi_threads = hw_threads * 3 / 4;
+ num_comp_lo_threads = hw_threads / 3;
+ } else if (hw_threads >= 6) {
+ num_comp_hi_threads = hw_threads - 2;
+ num_comp_lo_threads = hw_threads / 2;
+ } else if (hw_threads >= 2) {
+ num_comp_hi_threads = hw_threads - 1;
+ num_comp_lo_threads = hw_threads / 2;
+ } else {
+ num_comp_hi_threads = 1;
+ num_comp_lo_threads = 1;
+ }
+
+ num_comp_hi_threads = MIN2(num_comp_hi_threads,
+ ARRAY_SIZE(sscreen->compiler));
+ num_comp_lo_threads = MIN2(num_comp_lo_threads,
+ ARRAY_SIZE(sscreen->compiler_lowp));
+
+ if (!util_queue_init(&sscreen->shader_compiler_queue, "sh",
+ 64, num_comp_hi_threads,
+ UTIL_QUEUE_INIT_RESIZE_IF_FULL |
+ UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY)) {
si_destroy_shader_cache(sscreen);
FREE(sscreen);
return NULL;
}
if (!util_queue_init(&sscreen->shader_compiler_queue_low_priority,
- "si_shader_low",
- 32, num_compiler_threads_lowprio,
+ "shlo",
+ 64, num_comp_lo_threads,
UTIL_QUEUE_INIT_RESIZE_IF_FULL |
+ UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY |
UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY)) {
si_destroy_shader_cache(sscreen);
FREE(sscreen);
return NULL;
}
- si_handle_env_var_force_family(sscreen);
-
if (!debug_get_bool_option("RADEON_DISABLE_PERFCOUNTERS", false))
si_init_perfcounters(sscreen);
/* This must be one less than the maximum number due to a hw limitation.
* Various hardware bugs in SI, CIK, and GFX9 need this.
*/
- unsigned max_offchip_buffers_per_se = double_offchip_buffers ? 127 : 63;
+ unsigned max_offchip_buffers_per_se;
+
+ /* Only certain chips can use the maximum value. */
+ if (sscreen->info.family == CHIP_VEGA12 ||
+ sscreen->info.family == CHIP_VEGA20)
+ max_offchip_buffers_per_se = double_offchip_buffers ? 128 : 64;
+ else
+ max_offchip_buffers_per_se = double_offchip_buffers ? 127 : 63;
+
unsigned max_offchip_buffers = max_offchip_buffers_per_se *
sscreen->info.max_se;
unsigned offchip_granularity;
}
/* The mere presense of CLEAR_STATE in the IB causes random GPU hangs
- * on SI. */
- sscreen->has_clear_state = sscreen->info.chip_class >= CIK;
+ * on SI. Some CLEAR_STATE cause asic hang on radeon kernel, etc.
+ * SPI_VS_OUT_CONFIG. So only enable CI CLEAR_STATE on amdgpu kernel.*/
+ sscreen->has_clear_state = sscreen->info.chip_class >= CIK &&
+ sscreen->info.drm_major == 3;
sscreen->has_distributed_tess =
sscreen->info.chip_class >= VI &&
sscreen->info.family == CHIP_RAVEN;
sscreen->has_ls_vgpr_init_bug = sscreen->info.family == CHIP_VEGA10 ||
sscreen->info.family == CHIP_RAVEN;
+ sscreen->has_dcc_constant_encode = sscreen->info.family == CHIP_RAVEN2;
+
+ /* Only enable primitive binning on APUs by default. */
+ sscreen->dpbb_allowed = sscreen->info.family == CHIP_RAVEN ||
+ sscreen->info.family == CHIP_RAVEN2;
+ sscreen->dfsm_allowed = sscreen->info.family == CHIP_RAVEN ||
+ sscreen->info.family == CHIP_RAVEN2;
+
+ /* Process DPBB enable flags. */
if (sscreen->debug_flags & DBG(DPBB)) {
sscreen->dpbb_allowed = true;
- } else {
- /* Only enable primitive binning on Raven by default. */
- sscreen->dpbb_allowed = sscreen->info.family == CHIP_RAVEN &&
- !(sscreen->debug_flags & DBG(NO_DPBB));
+ if (sscreen->debug_flags & DBG(DFSM))
+ sscreen->dfsm_allowed = true;
}
- if (sscreen->debug_flags & DBG(DFSM)) {
- sscreen->dfsm_allowed = sscreen->dpbb_allowed;
- } else {
- sscreen->dfsm_allowed = sscreen->dpbb_allowed &&
- !(sscreen->debug_flags & DBG(NO_DFSM));
+ /* Process DPBB disable flags. */
+ if (sscreen->debug_flags & DBG(NO_DPBB)) {
+ sscreen->dpbb_allowed = false;
+ sscreen->dfsm_allowed = false;
+ } else if (sscreen->debug_flags & DBG(NO_DFSM)) {
+ sscreen->dfsm_allowed = false;
}
/* While it would be nice not to have this flag, we are constrained
sscreen->rbplus_allowed =
!(sscreen->debug_flags & DBG(NO_RB_PLUS)) &&
(sscreen->info.family == CHIP_STONEY ||
- sscreen->info.family == CHIP_RAVEN);
+ sscreen->info.family == CHIP_VEGA12 ||
+ sscreen->info.family == CHIP_RAVEN ||
+ sscreen->info.family == CHIP_RAVEN2);
}
sscreen->dcc_msaa_allowed =
- !(sscreen->debug_flags & DBG(NO_DCC_MSAA)) &&
- (sscreen->debug_flags & DBG(DCC_MSAA) ||
- sscreen->info.chip_class == VI);
+ !(sscreen->debug_flags & DBG(NO_DCC_MSAA));
sscreen->cpdma_prefetch_writes_memory = sscreen->info.chip_class <= VI;
if (debug_get_bool_option("RADEON_DUMP_SHADERS", false))
sscreen->debug_flags |= DBG_ALL_SHADERS;
- for (i = 0; i < num_compiler_threads; i++)
- sscreen->tm[i] = si_create_llvm_target_machine(sscreen);
- for (i = 0; i < num_compiler_threads_lowprio; i++)
- sscreen->tm_low_priority[i] = si_create_llvm_target_machine(sscreen);
+ /* Syntax:
+ * EQAA=s,z,c
+ * Example:
+ * EQAA=8,4,2
+
+ * That means 8 coverage samples, 4 Z/S samples, and 2 color samples.
+ * Constraints:
+ * s >= z >= c (ignoring this only wastes memory)
+ * s = [2..16]
+ * z = [2..8]
+ * c = [2..8]
+ *
+ * Only MSAA color and depth buffers are overriden.
+ */
+ if (sscreen->info.has_eqaa_surface_allocator) {
+ const char *eqaa = debug_get_option("EQAA", NULL);
+ unsigned s,z,f;
+
+ if (eqaa && sscanf(eqaa, "%u,%u,%u", &s, &z, &f) == 3 && s && z && f) {
+ sscreen->eqaa_force_coverage_samples = s;
+ sscreen->eqaa_force_z_samples = z;
+ sscreen->eqaa_force_color_samples = f;
+ }
+ }
+
+ for (i = 0; i < num_comp_hi_threads; i++)
+ si_init_compiler(sscreen, &sscreen->compiler[i]);
+ for (i = 0; i < num_comp_lo_threads; i++)
+ si_init_compiler(sscreen, &sscreen->compiler_lowp[i]);
/* Create the auxiliary context. This must be done last. */
sscreen->aux_context = si_create_context(&sscreen->b, 0);
if (sscreen->debug_flags & DBG(TEST_DMA))
si_test_dma(sscreen);
+ if (sscreen->debug_flags & DBG(TEST_DMA_PERF)) {
+ si_test_dma_perf(sscreen);
+ }
+
if (sscreen->debug_flags & (DBG(TEST_VMFAULT_CP) |
DBG(TEST_VMFAULT_SDMA) |
DBG(TEST_VMFAULT_SHADER)))
si_test_vmfault(sscreen);
+ if (sscreen->debug_flags & DBG(TEST_GDS))
+ si_test_gds((struct si_context*)sscreen->aux_context);
+
+ if (sscreen->debug_flags & DBG(TEST_GDS_MM)) {
+ si_test_gds_memory_management((struct si_context*)sscreen->aux_context,
+ 32 * 1024, 4, RADEON_DOMAIN_GDS);
+ }
+ if (sscreen->debug_flags & DBG(TEST_GDS_OA_MM)) {
+ si_test_gds_memory_management((struct si_context*)sscreen->aux_context,
+ 4, 1, RADEON_DOMAIN_OA);
+ }
+
return &sscreen->b;
}
projects / mesa.git / blobdiff