*/
#include "si_pipe.h"
-#include "si_shader.h"
#include "si_public.h"
+#include "si_shader_internal.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 "../ddebug/dd_util.h"
#define SI_LLVM_DEFAULT_FEATURES \
- "+DumpCode,+vgpr-spilling,-fp32-denormals,+fp64-denormals"
+ "+DumpCode,+vgpr-spilling,-fp32-denormals,-xnack"
/*
* pipe_context
{
const char *triple = "amdgcn--";
- return LLVMCreateTargetMachine(radeon_llvm_get_r600_target(triple), triple,
+ return LLVMCreateTargetMachine(si_llvm_get_amdgpu_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,
/* SI + AMDGPU + CE = GPU hang */
if (!(sscreen->b.debug_flags & DBG_NO_CE) && ws->cs_add_const_ib &&
- sscreen->b.chip_class != SI) {
+ 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) {
sctx->ce_ib = ws->cs_add_const_ib(sctx->b.gfx.cs);
if (!sctx->ce_ib)
goto fail;
}
sctx->ce_suballocator =
- u_suballocator_create(&sctx->b.b, 1024 * 1024,
- PIPE_BIND_CUSTOM,
- PIPE_USAGE_DEFAULT, false);
+ u_suballocator_create(&sctx->b.b, 1024 * 1024, 0,
+ PIPE_USAGE_DEFAULT,
+ R600_RESOURCE_FLAG_UNMAPPABLE, false);
if (!sctx->ce_suballocator)
goto fail;
}
goto fail;
sctx->border_color_buffer = (struct r600_resource*)
- pipe_buffer_create(screen, PIPE_BIND_CUSTOM, PIPE_USAGE_DEFAULT,
+ pipe_buffer_create(screen, 0, PIPE_USAGE_DEFAULT,
SI_MAX_BORDER_COLORS *
sizeof(*sctx->border_color_table));
if (!sctx->border_color_buffer)
si_init_all_descriptors(sctx);
si_init_state_functions(sctx);
si_init_shader_functions(sctx);
+ si_init_ia_multi_vgt_param_table(sctx);
if (sctx->b.chip_class >= CIK)
cik_init_sdma_functions(sctx);
/* these must be last */
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 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);
+ sctx->null_const_buf.buffer =
+ r600_aligned_buffer_create(screen,
+ R600_RESOURCE_FLAG_UNMAPPABLE,
+ PIPE_USAGE_DEFAULT, 16,
+ sctx->screen->b.info.tcc_cache_line_size);
if (!sctx->null_const_buf.buffer)
goto fail;
sctx->null_const_buf.buffer_size = sctx->null_const_buf.buffer->width0;
/*
* pipe_screen
*/
+static bool si_have_tgsi_compute(struct si_screen *sscreen)
+{
+ /* Old kernels disallowed some register writes for SI
+ * that are used for indirect dispatches. */
+ return 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));
+}
static int si_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
{
switch (param) {
/* Supported features (boolean caps). */
+ case PIPE_CAP_ACCELERATED:
case PIPE_CAP_TWO_SIDED_STENCIL:
case PIPE_CAP_MAX_DUAL_SOURCE_RENDER_TARGETS:
case PIPE_CAP_ANISOTROPIC_FILTER:
case PIPE_CAP_INDEP_BLEND_FUNC:
case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
case PIPE_CAP_VERTEX_COLOR_UNCLAMPED:
- case PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY:
- case PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY:
- case PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY:
- case PIPE_CAP_USER_INDEX_BUFFERS:
case PIPE_CAP_USER_CONSTANT_BUFFERS:
case PIPE_CAP_START_INSTANCE:
case PIPE_CAP_NPOT_TEXTURES:
case PIPE_CAP_POLYGON_OFFSET_UNITS_UNSCALED:
case PIPE_CAP_STRING_MARKER:
case PIPE_CAP_CLEAR_TEXTURE:
+ case PIPE_CAP_CULL_DISTANCE:
+ case PIPE_CAP_TGSI_ARRAY_COMPONENTS:
+ case PIPE_CAP_TGSI_CAN_READ_OUTPUTS:
+ case PIPE_CAP_GLSL_OPTIMIZE_CONSERVATIVELY:
+ case PIPE_CAP_STREAM_OUTPUT_PAUSE_RESUME:
+ case PIPE_CAP_STREAM_OUTPUT_INTERLEAVE_BUFFERS:
+ case PIPE_CAP_DOUBLES:
+ case PIPE_CAP_TGSI_TEX_TXF_LZ:
return 1;
+ case PIPE_CAP_INT64:
+ case PIPE_CAP_INT64_DIVMOD:
+ case PIPE_CAP_TGSI_CLOCK:
+ return HAVE_LLVM >= 0x0309;
+
+ case PIPE_CAP_TGSI_VOTE:
+ return HAVE_LLVM >= 0x0400;
+
case PIPE_CAP_RESOURCE_FROM_USER_MEMORY:
return !SI_BIG_ENDIAN && sscreen->b.info.has_userptr;
case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
case PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT:
case PIPE_CAP_MAX_TEXTURE_GATHER_COMPONENTS:
+ case PIPE_CAP_MAX_STREAM_OUTPUT_BUFFERS:
+ case PIPE_CAP_MAX_VERTEX_STREAMS:
return 4;
+
case PIPE_CAP_SHADER_BUFFER_OFFSET_ALIGNMENT:
return HAVE_LLVM >= 0x0309 ? 4 : 0;
case PIPE_CAP_GLSL_FEATURE_LEVEL:
- 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;
+ if (sscreen->b.chip_class >= GFX9)
+ return 140;
+ if (si_have_tgsi_compute(sscreen))
+ return 450;
+ return HAVE_LLVM >= 0x0309 ? 420 : 410;
case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE:
return MIN2(sscreen->b.info.max_alloc_size, INT_MAX);
- case PIPE_CAP_BUFFER_SAMPLER_VIEW_RGBA_ONLY:
- return 0;
+ case PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY:
+ case PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY:
+ case PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY:
+ /* SI doesn't support unaligned loads.
+ * CIK needs DRM 2.50.0 on radeon. */
+ return sscreen->b.chip_class == SI ||
+ (sscreen->b.info.drm_major == 2 &&
+ sscreen->b.info.drm_minor < 50);
/* Unsupported features. */
+ case PIPE_CAP_BUFFER_SAMPLER_VIEW_RGBA_ONLY:
case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT:
case PIPE_CAP_TGSI_CAN_COMPACT_CONSTANTS:
case PIPE_CAP_USER_VERTEX_BUFFERS:
case PIPE_CAP_FAKE_SW_MSAA:
case PIPE_CAP_TEXTURE_GATHER_OFFSETS:
case PIPE_CAP_VERTEXID_NOBASE:
- 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:
+ case PIPE_CAP_NATIVE_FENCE_FD:
+ case PIPE_CAP_TGSI_FS_FBFETCH:
+ case PIPE_CAP_TGSI_MUL_ZERO_WINS:
+ case PIPE_CAP_UMA:
return 0;
+ case PIPE_CAP_QUERY_BUFFER_OBJECT:
+ return si_have_tgsi_compute(sscreen);
+
case PIPE_CAP_DRAW_PARAMETERS:
case PIPE_CAP_MULTI_DRAW_INDIRECT:
case PIPE_CAP_MULTI_DRAW_INDIRECT_PARAMS:
return 30;
case PIPE_CAP_TEXTURE_BORDER_COLOR_QUIRK:
- return PIPE_QUIRK_TEXTURE_BORDER_COLOR_SWIZZLE_R600;
+ return sscreen->b.chip_class <= VI ?
+ PIPE_QUIRK_TEXTURE_BORDER_COLOR_SWIZZLE_R600 : 0;
/* Stream output. */
- case PIPE_CAP_MAX_STREAM_OUTPUT_BUFFERS:
- return sscreen->b.has_streamout ? 4 : 0;
- case PIPE_CAP_STREAM_OUTPUT_PAUSE_RESUME:
- return sscreen->b.has_streamout ? 1 : 0;
case PIPE_CAP_MAX_STREAM_OUTPUT_SEPARATE_COMPONENTS:
case PIPE_CAP_MAX_STREAM_OUTPUT_INTERLEAVED_COMPONENTS:
- return sscreen->b.has_streamout ? 32*4 : 0;
+ return 32*4;
/* Geometry shader output. */
case PIPE_CAP_MAX_GEOMETRY_OUTPUT_VERTICES:
return 1024;
case PIPE_CAP_MAX_GEOMETRY_TOTAL_OUTPUT_COMPONENTS:
return 4095;
- case PIPE_CAP_MAX_VERTEX_STREAMS:
- return 4;
case PIPE_CAP_MAX_VERTEX_ATTRIB_STRIDE:
return 2048;
/* textures support 8192, but layered rendering supports 2048 */
return 2048;
- /* Render targets. */
- case PIPE_CAP_MAX_RENDER_TARGETS:
- return 8;
-
+ /* Viewports and render targets. */
case PIPE_CAP_MAX_VIEWPORTS:
return R600_MAX_VIEWPORTS;
case PIPE_CAP_VIEWPORT_SUBPIXEL_BITS:
+ case PIPE_CAP_MAX_RENDER_TARGETS:
return 8;
/* Timer queries, present when the clock frequency is non zero. */
return ATI_VENDOR_ID;
case PIPE_CAP_DEVICE_ID:
return sscreen->b.info.pci_id;
- case PIPE_CAP_ACCELERATED:
- return 1;
case PIPE_CAP_VIDEO_MEMORY:
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 0;
}
-static int si_get_shader_param(struct pipe_screen* pscreen, unsigned shader, enum pipe_shader_cap param)
+static int si_get_shader_param(struct pipe_screen* pscreen,
+ enum pipe_shader_type shader,
+ enum pipe_shader_cap param)
{
struct si_screen *sscreen = (struct si_screen *)pscreen;
{
case PIPE_SHADER_FRAGMENT:
case PIPE_SHADER_VERTEX:
- case PIPE_SHADER_GEOMETRY:
break;
+ case PIPE_SHADER_GEOMETRY:
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 && MESA_LLVM_VERSION_PATCH < 2)
+ if (sscreen->b.chip_class >= GFX9)
return 0;
break;
case PIPE_SHADER_COMPUTE:
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)))
+ if (si_have_tgsi_compute(sscreen))
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;
}
switch (param) {
+ /* Shader limits. */
case PIPE_SHADER_CAP_MAX_INSTRUCTIONS:
case PIPE_SHADER_CAP_MAX_ALU_INSTRUCTIONS:
case PIPE_SHADER_CAP_MAX_TEX_INSTRUCTIONS:
case PIPE_SHADER_CAP_MAX_TEX_INDIRECTIONS:
- return 16384;
case PIPE_SHADER_CAP_MAX_CONTROL_FLOW_DEPTH:
- return 32;
+ return 16384;
case PIPE_SHADER_CAP_MAX_INPUTS:
- return shader == PIPE_SHADER_VERTEX ? SI_NUM_VERTEX_BUFFERS : 32;
+ return shader == PIPE_SHADER_VERTEX ? SI_MAX_ATTRIBS : 32;
case PIPE_SHADER_CAP_MAX_OUTPUTS:
return shader == PIPE_SHADER_FRAGMENT ? 8 : 32;
case PIPE_SHADER_CAP_MAX_TEMPS:
return 4096 * sizeof(float[4]); /* actually only memory limits this */
case PIPE_SHADER_CAP_MAX_CONST_BUFFERS:
return SI_NUM_CONST_BUFFERS;
- case PIPE_SHADER_CAP_MAX_PREDS:
- return 0; /* FIXME */
+ case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
+ case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
+ return SI_NUM_SAMPLERS;
+ case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
+ return HAVE_LLVM >= 0x0309 ? SI_NUM_SHADER_BUFFERS : 0;
+ case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
+ return HAVE_LLVM >= 0x0309 ? SI_NUM_IMAGES : 0;
+ case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT:
+ return 32;
+ case PIPE_SHADER_CAP_PREFERRED_IR:
+ return PIPE_SHADER_IR_TGSI;
+ case PIPE_SHADER_CAP_LOWER_IF_THRESHOLD:
+ return 3;
+
+ /* Supported boolean features. */
case PIPE_SHADER_CAP_TGSI_CONT_SUPPORTED:
- return 1;
case PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED:
- return 1;
- case PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR:
- /* Indirection of geometry shader input dimension is not
- * handled yet
- */
- return shader != PIPE_SHADER_GEOMETRY;
case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR:
case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR:
case PIPE_SHADER_CAP_INDIRECT_CONST_ADDR:
- return 1;
case PIPE_SHADER_CAP_INTEGERS:
+ case PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED:
+ case PIPE_SHADER_CAP_TGSI_ANY_INOUT_DECL_RANGE:
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;
+
+ /* Unsupported boolean features. */
+ case PIPE_SHADER_CAP_MAX_PREDS:
case PIPE_SHADER_CAP_SUBROUTINES:
- return 0;
- case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
- case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
- 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_TGSI_DFRACEXP_DLDEXP_SUPPORTED:
return 0;
- case PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED:
- case PIPE_SHADER_CAP_TGSI_ANY_INOUT_DECL_RANGE:
- return 1;
- case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT:
- return 32;
- case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
- 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;
}
sscreen->vs_prologs,
sscreen->vs_epilogs,
sscreen->tcs_epilogs,
+ sscreen->gs_prologs,
sscreen->ps_prologs,
sscreen->ps_epilogs
};
unsigned i;
- if (!sscreen)
- return;
-
if (!sscreen->b.ws->unref(sscreen->b.ws))
return;
- if (util_queue_is_initialized(&sscreen->shader_compiler_queue))
- util_queue_destroy(&sscreen->shader_compiler_queue);
+ util_queue_destroy(&sscreen->shader_compiler_queue);
for (i = 0; i < ARRAY_SIZE(sscreen->tm); i++)
if (sscreen->tm[i])
FREE(part);
}
}
- pipe_mutex_destroy(sscreen->shader_parts_mutex);
+ mtx_destroy(&sscreen->shader_parts_mutex);
si_destroy_shader_cache(sscreen);
r600_destroy_common_screen(&sscreen->b);
}
case CHIP_FIJI:
case CHIP_POLARIS10:
case CHIP_POLARIS11:
+ case CHIP_POLARIS12:
+ case CHIP_VEGA10:
sscreen->gs_table_depth = 32;
return true;
default:
/* Override family and chip_class. */
sscreen->b.family = sscreen->b.info.family = i;
- if (i >= CHIP_TONGA)
+ if (i >= CHIP_VEGA10)
+ sscreen->b.chip_class = sscreen->b.info.chip_class = GFX9;
+ else 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;
return NULL;
}
+ /* 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));
+
+ if (!util_queue_init(&sscreen->shader_compiler_queue, "si_shader",
+ 32, num_compiler_threads)) {
+ 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))
sscreen->b.info.pfp_fw_version >= 121 &&
sscreen->b.info.me_fw_version >= 87);
+ sscreen->has_ds_bpermute = HAVE_LLVM >= 0x0309 &&
+ sscreen->b.chip_class >= VI;
+
sscreen->b.has_cp_dma = true;
sscreen->b.has_streamout = true;
- pipe_mutex_init(sscreen->shader_parts_mutex);
+
+ /* Some chips have RB+ registers, but don't support RB+. Those must
+ * always disable it.
+ */
+ if (sscreen->b.family == CHIP_STONEY ||
+ sscreen->b.chip_class >= GFX9) {
+ sscreen->b.has_rbplus = true;
+
+ sscreen->b.rbplus_allowed =
+ !(sscreen->b.debug_flags & DBG_NO_RB_PLUS) &&
+ sscreen->b.family == CHIP_STONEY;
+ }
+
+ (void) mtx_init(&sscreen->shader_parts_mutex, mtx_plain);
sscreen->use_monolithic_shaders =
- HAVE_LLVM < 0x0308 ||
(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;
+ sscreen->b.barrier_flags.compute_to_L2 = SI_CONTEXT_CS_PARTIAL_FLUSH;
+
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);
projects / mesa.git / blobdiff