*/
#include "si_pipe.h"
+#include "si_shader.h"
#include "si_public.h"
#include "sid.h"
#include "radeon/radeon_llvm_emit.h"
#include "radeon/radeon_uvd.h"
#include "util/u_memory.h"
+#include "util/u_suballoc.h"
#include "vl/vl_decoder.h"
+#include "../ddebug/dd_util.h"
+
+#define SI_LLVM_DEFAULT_FEATURES \
+ "+DumpCode,+vgpr-spilling,-fp32-denormals,+fp64-denormals"
/*
* pipe_context
struct si_context *sctx = (struct si_context *)context;
int i;
+ /* Unreference the framebuffer normally to disable related logic
+ * properly.
+ */
+ struct pipe_framebuffer_state fb = {};
+ context->set_framebuffer_state(context, &fb);
+
si_release_all_descriptors(sctx);
+ if (sctx->ce_suballocator)
+ u_suballocator_destroy(sctx->ce_suballocator);
+
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->null_const_buf.buffer, NULL);
r600_resource_reference(&sctx->border_color_buffer, NULL);
free(sctx->border_color_table);
r600_resource_reference(&sctx->scratch_buffer, NULL);
- sctx->b.ws->fence_reference(&sctx->last_gfx_fence, NULL);
+ r600_resource_reference(&sctx->compute_scratch_buffer, NULL);
si_pm4_free_state(sctx, sctx->init_config, ~0);
if (sctx->init_config_gs_rings)
si_pm4_free_state(sctx, sctx->init_config_gs_rings, ~0);
- for (i = 0; i < Elements(sctx->vgt_shader_config); i++)
+ for (i = 0; i < ARRAY_SIZE(sctx->vgt_shader_config); i++)
si_pm4_delete_state(sctx, vgt_shader_config, sctx->vgt_shader_config[i]);
- if (sctx->pstipple_sampler_state)
- sctx->b.b.delete_sampler_state(&sctx->b.b, sctx->pstipple_sampler_state);
if (sctx->fixed_func_tcs_shader.cso)
sctx->b.b.delete_tcs_state(&sctx->b.b, sctx->fixed_func_tcs_shader.cso);
if (sctx->custom_dsa_flush)
sctx->b.b.delete_blend_state(&sctx->b.b, sctx->custom_blend_decompress);
if (sctx->custom_blend_fastclear)
sctx->b.b.delete_blend_state(&sctx->b.b, sctx->custom_blend_fastclear);
- util_unreference_framebuffer_state(&sctx->framebuffer.state);
+ if (sctx->custom_blend_dcc_decompress)
+ sctx->b.b.delete_blend_state(&sctx->b.b, sctx->custom_blend_dcc_decompress);
if (sctx->blitter)
util_blitter_destroy(sctx->blitter);
r600_common_context_cleanup(&sctx->b);
-#if HAVE_LLVM >= 0x0306
LLVMDisposeTargetMachine(sctx->tm);
-#endif
r600_resource_reference(&sctx->trace_buf, NULL);
r600_resource_reference(&sctx->last_trace_buf, NULL);
- free(sctx->last_ib);
- if (sctx->last_bo_list) {
- for (i = 0; i < sctx->last_bo_count; i++)
- pb_reference(&sctx->last_bo_list[i].buf, NULL);
- free(sctx->last_bo_list);
- }
+ radeon_clear_saved_cs(&sctx->last_gfx);
+
FREE(sctx);
}
return sctx->b.ws->ctx_query_reset_status(sctx->b.ctx);
}
+/* Apitrace profiling:
+ * 1) qapitrace : Tools -> Profile: Measure CPU & GPU times
+ * 2) In the middle panel, zoom in (mouse wheel) on some bad draw call
+ * and remember its number.
+ * 3) In Mesa, enable queries and performance counters around that draw
+ * call and print the results.
+ * 4) glretrace --benchmark --markers ..
+ */
+static void si_emit_string_marker(struct pipe_context *ctx,
+ const char *string, int len)
+{
+ struct si_context *sctx = (struct si_context *)ctx;
+
+ dd_parse_apitrace_marker(string, len, &sctx->apitrace_call_number);
+}
+
+static LLVMTargetMachineRef
+si_create_llvm_target_machine(struct si_screen *sscreen)
+{
+ const char *triple = "amdgcn--";
+
+ return LLVMCreateTargetMachine(radeon_llvm_get_r600_target(triple), triple,
+ r600_get_llvm_processor_name(sscreen->b.family),
+#if HAVE_LLVM >= 0x0308
+ sscreen->b.debug_flags & DBG_SI_SCHED ?
+ SI_LLVM_DEFAULT_FEATURES ",+si-scheduler" :
+#endif
+ SI_LLVM_DEFAULT_FEATURES,
+ LLVMCodeGenLevelDefault,
+ LLVMRelocDefault,
+ LLVMCodeModelDefault);
+}
+
static struct pipe_context *si_create_context(struct pipe_screen *screen,
void *priv, unsigned flags)
{
struct si_context *sctx = CALLOC_STRUCT(si_context);
struct si_screen* sscreen = (struct si_screen *)screen;
struct radeon_winsys *ws = sscreen->b.ws;
- LLVMTargetRef r600_target;
-#if HAVE_LLVM >= 0x0306
- const char *triple = "amdgcn--";
-#endif
int shader, i;
if (!sctx)
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.screen = screen; /* this must be set first */
sctx->b.b.priv = priv;
sctx->b.b.destroy = si_destroy_context;
+ sctx->b.b.emit_string_marker = si_emit_string_marker;
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;
- if (!r600_common_context_init(&sctx->b, &sscreen->b))
+ if (!r600_common_context_init(&sctx->b, &sscreen->b, flags))
goto fail;
if (sscreen->b.info.drm_major == 3)
sctx->b.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, sscreen->b.trace_bo ?
- sscreen->b.trace_bo->buf : NULL);
+ 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) {
+ 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,
+ PIPE_BIND_CUSTOM,
+ PIPE_USAGE_DEFAULT, false);
+ if (!sctx->ce_suballocator)
+ goto fail;
+ }
+
sctx->b.gfx.flush = si_context_gfx_flush;
/* Border colors. */
si_init_state_functions(sctx);
si_init_shader_functions(sctx);
+ if (sctx->b.chip_class >= CIK)
+ cik_init_sdma_functions(sctx);
+ else
+ si_init_dma_functions(sctx);
+
if (sscreen->b.debug_flags & DBG_FORCE_DMA)
sctx->b.b.resource_copy_region = sctx->b.dma_copy;
si_begin_new_cs(sctx);
r600_query_init_backend_mask(&sctx->b); /* this emits commands and must be last */
- /* CIK cannot unbind a constant buffer (S_BUFFER_LOAD is buggy
- * with a NULL buffer). We need to use a dummy buffer instead. */
+ /* 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) {
sctx->null_const_buf.buffer = pipe_buffer_create(screen, PIPE_BIND_CONSTANT_BUFFER,
PIPE_USAGE_DEFAULT, 16);
}
}
+ si_set_rw_buffer(sctx, SI_HS_CONST_DEFAULT_TESS_LEVELS,
+ &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->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. */
sctx->b.clear_buffer(&sctx->b.b, sctx->null_const_buf.buffer, 0,
- sctx->null_const_buf.buffer->width0, 0, false);
+ sctx->null_const_buf.buffer->width0, 0,
+ R600_COHERENCY_SHADER);
}
- /* XXX: This is the maximum value allowed. I'm not sure how to compute
- * this for non-cs shaders. Using the wrong value here can result in
- * GPU lockups, but the maximum value seems to always work.
+ uint64_t max_threads_per_block;
+ screen->get_compute_param(screen, PIPE_SHADER_IR_TGSI,
+ PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK,
+ &max_threads_per_block);
+
+ /* The maximum number of scratch waves. Scratch space isn't divided
+ * evenly between CUs. The number is only a function of the number of CUs.
+ * We can decrease the constant to decrease the scratch buffer size.
+ *
+ * sctx->scratch_waves must be >= the maximum posible size of
+ * 1 threadgroup, so that the hw doesn't hang from being unable
+ * to start any.
+ *
+ * The recommended value is 4 per CU at most. Higher numbers don't
+ * bring much benefit, but they still occupy chip resources (think
+ * async compute). I've seen ~2% performance difference between 4 and 32.
*/
- sctx->scratch_waves = 32 * sscreen->b.info.num_good_compute_units;
-
-#if HAVE_LLVM >= 0x0306
- /* Initialize LLVM TargetMachine */
- r600_target = radeon_llvm_get_r600_target(triple);
- sctx->tm = LLVMCreateTargetMachine(r600_target, triple,
- r600_get_llvm_processor_name(sscreen->b.family),
- "+DumpCode,+vgpr-spilling",
- LLVMCodeGenLevelDefault,
- LLVMRelocDefault,
- LLVMCodeModelDefault);
-#endif
+ sctx->scratch_waves = MAX2(32 * sscreen->b.info.num_good_compute_units,
+ max_threads_per_block / 64);
+
+ sctx->tm = si_create_llvm_target_machine(sscreen);
return &sctx->b.b;
fail:
case PIPE_CAP_START_INSTANCE:
case PIPE_CAP_NPOT_TEXTURES:
case PIPE_CAP_MIXED_FRAMEBUFFER_SIZES:
+ case PIPE_CAP_MIXED_COLOR_DEPTH_BITS:
case PIPE_CAP_VERTEX_COLOR_CLAMPED:
case PIPE_CAP_FRAGMENT_COLOR_CLAMPED:
case PIPE_CAP_PREFER_BLIT_BASED_TEXTURE_TRANSFER:
case PIPE_CAP_TGSI_FS_POSITION_IS_SYSVAL:
case PIPE_CAP_TGSI_FS_FACE_IS_INTEGER_SYSVAL:
case PIPE_CAP_INVALIDATE_BUFFER:
+ case PIPE_CAP_SURFACE_REINTERPRET_BLOCKS:
+ case PIPE_CAP_QUERY_MEMORY_INFO:
+ case PIPE_CAP_TGSI_PACK_HALF_FLOAT:
+ case PIPE_CAP_FRAMEBUFFER_NO_ATTACHMENT:
+ case PIPE_CAP_ROBUST_BUFFER_ACCESS_BEHAVIOR:
+ case PIPE_CAP_GENERATE_MIPMAP:
+ case PIPE_CAP_POLYGON_OFFSET_UNITS_UNSCALED:
+ case PIPE_CAP_STRING_MARKER:
+ case PIPE_CAP_CLEAR_TEXTURE:
return 1;
case PIPE_CAP_RESOURCE_FROM_USER_MEMORY:
case PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT:
case PIPE_CAP_MAX_TEXTURE_GATHER_COMPONENTS:
return 4;
+ case PIPE_CAP_SHADER_BUFFER_OFFSET_ALIGNMENT:
+ return HAVE_LLVM >= 0x0309 ? 4 : 0;
case PIPE_CAP_GLSL_FEATURE_LEVEL:
- return HAVE_LLVM >= 0x0307 ? 410 : 330;
+ if (pscreen->get_shader_param(pscreen, PIPE_SHADER_COMPUTE,
+ PIPE_SHADER_CAP_SUPPORTED_IRS) &
+ (1 << PIPE_SHADER_IR_TGSI))
+ return 430;
+ return HAVE_LLVM >= 0x0309 ? 420 :
+ HAVE_LLVM >= 0x0307 ? 410 : 330;
case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE:
- return MIN2(sscreen->b.info.vram_size, 0xFFFFFFFF);
+ return MIN2(sscreen->b.info.max_alloc_size, INT_MAX);
+
+ case PIPE_CAP_BUFFER_SAMPLER_VIEW_RGBA_ONLY:
+ return 0;
/* Unsupported features. */
case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT:
case PIPE_CAP_FAKE_SW_MSAA:
case PIPE_CAP_TEXTURE_GATHER_OFFSETS:
case PIPE_CAP_VERTEXID_NOBASE:
- case PIPE_CAP_CLEAR_TEXTURE:
+ case PIPE_CAP_QUERY_BUFFER_OBJECT:
+ case PIPE_CAP_CULL_DISTANCE:
+ case PIPE_CAP_PRIMITIVE_RESTART_FOR_PATCHES:
+ case PIPE_CAP_TGSI_VOTE:
+ case PIPE_CAP_MAX_WINDOW_RECTANGLES:
+ return 0;
+
case PIPE_CAP_DRAW_PARAMETERS:
- case PIPE_CAP_TGSI_PACK_HALF_FLOAT:
case PIPE_CAP_MULTI_DRAW_INDIRECT:
case PIPE_CAP_MULTI_DRAW_INDIRECT_PARAMS:
- case PIPE_CAP_SHADER_BUFFER_OFFSET_ALIGNMENT:
- case PIPE_CAP_GENERATE_MIPMAP:
- case PIPE_CAP_STRING_MARKER:
- return 0;
+ return sscreen->has_draw_indirect_multi;
case PIPE_CAP_MAX_SHADER_PATCH_VARYINGS:
return 30;
return 8;
case PIPE_CAP_MAX_VIEWPORTS:
- return SI_MAX_VIEWPORTS;
+ return R600_MAX_VIEWPORTS;
+ case PIPE_CAP_VIEWPORT_SUBPIXEL_BITS:
+ 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.r600_clock_crystal_freq != 0;
+ return sscreen->b.info.clock_crystal_freq != 0;
case PIPE_CAP_MIN_TEXTURE_GATHER_OFFSET:
case PIPE_CAP_MIN_TEXEL_OFFSET:
return PIPE_ENDIAN_LITTLE;
case PIPE_CAP_VENDOR_ID:
- return 0x1002;
+ return ATI_VENDOR_ID;
case PIPE_CAP_DEVICE_ID:
return sscreen->b.info.pci_id;
case PIPE_CAP_ACCELERATED:
return sscreen->b.info.vram_size >> 20;
case PIPE_CAP_UMA:
return 0;
+ case PIPE_CAP_PCI_GROUP:
+ return sscreen->b.info.pci_domain;
+ case PIPE_CAP_PCI_BUS:
+ return sscreen->b.info.pci_bus;
+ case PIPE_CAP_PCI_DEVICE:
+ return sscreen->b.info.pci_dev;
+ case PIPE_CAP_PCI_FUNCTION:
+ return sscreen->b.info.pci_func;
}
return 0;
}
static int si_get_shader_param(struct pipe_screen* pscreen, unsigned shader, enum pipe_shader_cap param)
{
+ struct si_screen *sscreen = (struct si_screen *)pscreen;
+
switch(shader)
{
case PIPE_SHADER_FRAGMENT:
case PIPE_SHADER_TESS_CTRL:
case PIPE_SHADER_TESS_EVAL:
/* LLVM 3.6.2 is required for tessellation because of bug fixes there */
- if (HAVE_LLVM < 0x0306 ||
- (HAVE_LLVM == 0x0306 && MESA_LLVM_VERSION_PATCH < 2))
+ if (HAVE_LLVM == 0x0306 && MESA_LLVM_VERSION_PATCH < 2)
return 0;
break;
case PIPE_SHADER_COMPUTE:
switch (param) {
case PIPE_SHADER_CAP_PREFERRED_IR:
-#if HAVE_LLVM < 0x0306
- return PIPE_SHADER_IR_LLVM;
-#else
return PIPE_SHADER_IR_NATIVE;
-#endif
+
+ case PIPE_SHADER_CAP_SUPPORTED_IRS: {
+ int ir = 1 << PIPE_SHADER_IR_NATIVE;
+
+ /* Old kernels disallowed some register writes for SI
+ * that are used for indirect dispatches. */
+ if (HAVE_LLVM >= 0x309 && (sscreen->b.chip_class >= CIK ||
+ sscreen->b.info.drm_major == 3 ||
+ (sscreen->b.info.drm_major == 2 &&
+ sscreen->b.info.drm_minor >= 45)))
+ ir |= 1 << PIPE_SHADER_IR_TGSI;
+
+ return ir;
+ }
case PIPE_SHADER_CAP_DOUBLES:
return HAVE_LLVM >= 0x0307;
case PIPE_SHADER_CAP_MAX_CONST_BUFFER_SIZE: {
uint64_t max_const_buffer_size;
- pscreen->get_compute_param(pscreen,
+ pscreen->get_compute_param(pscreen, PIPE_SHADER_IR_TGSI,
PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE,
&max_const_buffer_size);
- return max_const_buffer_size;
+ return MIN2(max_const_buffer_size, INT_MAX);
}
default:
/* If compute shaders don't require a special value
case PIPE_SHADER_CAP_MAX_CONST_BUFFER_SIZE:
return 4096 * sizeof(float[4]); /* actually only memory limits this */
case PIPE_SHADER_CAP_MAX_CONST_BUFFERS:
- return SI_NUM_USER_CONST_BUFFERS;
+ return SI_NUM_CONST_BUFFERS;
case PIPE_SHADER_CAP_MAX_PREDS:
return 0; /* FIXME */
case PIPE_SHADER_CAP_TGSI_CONT_SUPPORTED:
return 0;
case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
- return 16;
+ return SI_NUM_SAMPLERS;
case PIPE_SHADER_CAP_PREFERRED_IR:
return PIPE_SHADER_IR_TGSI;
+ case PIPE_SHADER_CAP_SUPPORTED_IRS:
+ return 0;
case PIPE_SHADER_CAP_DOUBLES:
return HAVE_LLVM >= 0x0307;
case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED:
case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT:
return 32;
case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
- return 0;
+ return HAVE_LLVM >= 0x0309 ? SI_NUM_SHADER_BUFFERS : 0;
+ case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
+ return HAVE_LLVM >= 0x0309 ? SI_NUM_IMAGES : 0;
}
return 0;
}
static void si_destroy_screen(struct pipe_screen* pscreen)
{
struct si_screen *sscreen = (struct si_screen *)pscreen;
+ struct si_shader_part *parts[] = {
+ sscreen->vs_prologs,
+ sscreen->vs_epilogs,
+ sscreen->tcs_epilogs,
+ sscreen->ps_prologs,
+ sscreen->ps_epilogs
+ };
+ unsigned i;
if (!sscreen)
return;
if (!sscreen->b.ws->unref(sscreen->b.ws))
return;
- r600_destroy_common_screen(&sscreen->b);
-}
+ if (util_queue_is_initialized(&sscreen->shader_compiler_queue))
+ util_queue_destroy(&sscreen->shader_compiler_queue);
-#define SI_TILE_MODE_COLOR_2D_8BPP 14
+ for (i = 0; i < ARRAY_SIZE(sscreen->tm); i++)
+ if (sscreen->tm[i])
+ LLVMDisposeTargetMachine(sscreen->tm[i]);
-/* Initialize pipe config. This is especially important for GPUs
- * with 16 pipes and more where it's initialized incorrectly by
- * the TILING_CONFIG ioctl. */
-static bool si_initialize_pipe_config(struct si_screen *sscreen)
-{
- unsigned mode2d;
+ /* Free shader parts. */
+ for (i = 0; i < ARRAY_SIZE(parts); i++) {
+ while (parts[i]) {
+ struct si_shader_part *part = parts[i];
- /* This is okay, because there can be no 2D tiling without
- * the tile mode array, so we won't need the pipe config.
- * Return "success".
- */
- if (!sscreen->b.info.si_tile_mode_array_valid)
- return true;
-
- /* The same index is used for the 2D mode on CIK too. */
- mode2d = sscreen->b.info.si_tile_mode_array[SI_TILE_MODE_COLOR_2D_8BPP];
-
- switch (G_009910_PIPE_CONFIG(mode2d)) {
- case V_02803C_ADDR_SURF_P2:
- sscreen->b.tiling_info.num_channels = 2;
- break;
- case V_02803C_X_ADDR_SURF_P4_8X16:
- case V_02803C_X_ADDR_SURF_P4_16X16:
- case V_02803C_X_ADDR_SURF_P4_16X32:
- case V_02803C_X_ADDR_SURF_P4_32X32:
- sscreen->b.tiling_info.num_channels = 4;
- break;
- case V_02803C_X_ADDR_SURF_P8_16X16_8X16:
- case V_02803C_X_ADDR_SURF_P8_16X32_8X16:
- case V_02803C_X_ADDR_SURF_P8_32X32_8X16:
- case V_02803C_X_ADDR_SURF_P8_16X32_16X16:
- case V_02803C_X_ADDR_SURF_P8_32X32_16X16:
- case V_02803C_X_ADDR_SURF_P8_32X32_16X32:
- case V_02803C_X_ADDR_SURF_P8_32X64_32X32:
- sscreen->b.tiling_info.num_channels = 8;
- break;
- case V_02803C_X_ADDR_SURF_P16_32X32_8X16:
- case V_02803C_X_ADDR_SURF_P16_32X32_16X16:
- sscreen->b.tiling_info.num_channels = 16;
- break;
- default:
- assert(0);
- fprintf(stderr, "radeonsi: Unknown pipe config %i.\n",
- G_009910_PIPE_CONFIG(mode2d));
- return false;
+ parts[i] = part->next;
+ radeon_shader_binary_clean(&part->binary);
+ FREE(part);
+ }
}
- return true;
+ pipe_mutex_destroy(sscreen->shader_parts_mutex);
+ si_destroy_shader_cache(sscreen);
+ r600_destroy_common_screen(&sscreen->b);
}
static bool si_init_gs_info(struct si_screen *sscreen)
case CHIP_HAWAII:
case CHIP_TONGA:
case CHIP_FIJI:
+ case CHIP_POLARIS10:
+ case CHIP_POLARIS11:
sscreen->gs_table_depth = 32;
return true;
default:
}
}
+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, r600_get_llvm_processor_name(i))) {
+ /* Override family and chip_class. */
+ sscreen->b.family = sscreen->b.info.family = i;
+
+ if (i >= CHIP_TONGA)
+ sscreen->b.chip_class = sscreen->b.info.chip_class = VI;
+ else if (i >= CHIP_BONAIRE)
+ sscreen->b.chip_class = sscreen->b.info.chip_class = CIK;
+ else
+ sscreen->b.chip_class = sscreen->b.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);
+}
+
struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws)
{
struct si_screen *sscreen = CALLOC_STRUCT(si_screen);
+ unsigned num_cpus, num_compiler_threads, i;
if (!sscreen) {
return NULL;
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.is_format_supported = si_is_format_supported;
sscreen->b.b.resource_create = r600_resource_create_common;
+ si_init_screen_state_functions(sscreen);
+
if (!r600_common_screen_init(&sscreen->b, ws) ||
- !si_initialize_pipe_config(sscreen) ||
- !si_init_gs_info(sscreen)) {
+ !si_init_gs_info(sscreen) ||
+ !si_init_shader_cache(sscreen)) {
FREE(sscreen);
return NULL;
}
- if (!debug_get_bool_option("RADEON_DISABLE_PERFCOUNTERS", FALSE))
+ si_handle_env_var_force_family(sscreen);
+
+ if (!debug_get_bool_option("RADEON_DISABLE_PERFCOUNTERS", false))
si_init_perfcounters(sscreen);
+ /* Hawaii has a bug with offchip buffers > 256 that can be worked
+ * around by setting 4K granularity.
+ */
+ sscreen->tess_offchip_block_dw_size =
+ sscreen->b.family == CHIP_HAWAII ? 4096 : 8192;
+
+ sscreen->has_distributed_tess =
+ sscreen->b.chip_class >= VI &&
+ sscreen->b.info.max_se >= 2;
+
+ sscreen->has_draw_indirect_multi =
+ (sscreen->b.family >= CHIP_POLARIS10) ||
+ (sscreen->b.chip_class == VI &&
+ sscreen->b.info.pfp_fw_version >= 121 &&
+ sscreen->b.info.me_fw_version >= 87) ||
+ (sscreen->b.chip_class == CIK &&
+ 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.has_cp_dma = true;
sscreen->b.has_streamout = true;
+ pipe_mutex_init(sscreen->shader_parts_mutex);
+ sscreen->use_monolithic_shaders =
+ HAVE_LLVM < 0x0308 ||
+ (sscreen->b.debug_flags & DBG_MONOLITHIC_SHADERS) != 0;
- if (debug_get_bool_option("RADEON_DUMP_SHADERS", FALSE))
+ if (debug_get_bool_option("RADEON_DUMP_SHADERS", false))
sscreen->b.debug_flags |= DBG_FS | DBG_VS | DBG_GS | DBG_PS | DBG_CS;
+ /* Only enable as many threads as we have target machines and CPUs. */
+ num_cpus = sysconf(_SC_NPROCESSORS_ONLN);
+ num_compiler_threads = MIN2(num_cpus, ARRAY_SIZE(sscreen->tm));
+
+ for (i = 0; i < num_compiler_threads; i++)
+ sscreen->tm[i] = si_create_llvm_target_machine(sscreen);
+
+ util_queue_init(&sscreen->shader_compiler_queue, "si_shader",
+ 32, num_compiler_threads);
+
/* Create the auxiliary context. This must be done last. */
sscreen->b.aux_context = sscreen->b.b.context_create(&sscreen->b.b, NULL, 0);
+ if (sscreen->b.debug_flags & DBG_TEST_DMA)
+ r600_test_dma(&sscreen->b);
+
return &sscreen->b.b;
}