#include "isl/isl.h"
#include "blorp/blorp.h"
-#ifdef __cplusplus
-extern "C" {
- /* Evil hack for using libdrm in a c++ compiler. */
- #define virtual virt
-#endif
-
#include <intel_bufmgr.h>
-#ifdef __cplusplus
- #undef virtual
-}
-#endif
-#ifdef __cplusplus
-extern "C" {
-#endif
#include "intel_debug.h"
#include "intel_screen.h"
#include "intel_tex_obj.h"
#include "intel_resolve_map.h"
+#ifdef __cplusplus
+extern "C" {
+#endif
/* Glossary:
*
* URB - uniform resource buffer. A mid-sized buffer which is
};
enum brw_state_id {
- /* brw_cache_ids must come first - see brw_state_cache.c */
+ /* brw_cache_ids must come first - see brw_program_cache.c */
BRW_STATE_URB_FENCE = BRW_MAX_CACHE,
BRW_STATE_FRAGMENT_PROGRAM,
BRW_STATE_GEOMETRY_PROGRAM,
BRW_STATE_CC_STATE,
BRW_STATE_BLORP,
BRW_STATE_VIEWPORT_COUNT,
+ BRW_STATE_CONSERVATIVE_RASTERIZATION,
BRW_NUM_STATE_BITS
};
#define BRW_NEW_FS_PROG_DATA (1ull << BRW_CACHE_FS_PROG)
/* XXX: The BRW_NEW_BLORP_BLIT_PROG_DATA dirty bit is unused (as BLORP doesn't
* use the normal state upload paths), but the cache is still used. To avoid
- * polluting the brw_state_cache code with special cases, we retain the dirty
- * bit for now. It should eventually be removed.
+ * polluting the brw_program_cache code with special cases, we retain the
+ * dirty bit for now. It should eventually be removed.
*/
#define BRW_NEW_BLORP_BLIT_PROG_DATA (1ull << BRW_CACHE_BLORP_PROG)
#define BRW_NEW_SF_PROG_DATA (1ull << BRW_CACHE_SF_PROG)
#define BRW_NEW_URB_SIZE (1ull << BRW_STATE_URB_SIZE)
#define BRW_NEW_CC_STATE (1ull << BRW_STATE_CC_STATE)
#define BRW_NEW_BLORP (1ull << BRW_STATE_BLORP)
+#define BRW_NEW_CONSERVATIVE_RASTERIZATION (1ull << BRW_STATE_CONSERVATIVE_RASTERIZATION)
struct brw_state_flags {
/** State update flags signalled by mesa internals */
struct brw_program {
struct gl_program program;
GLuint id;
-};
-
-
-/** Subclass of Mesa fragment program */
-struct brw_fragment_program {
- struct gl_program program;
- GLuint id; /**< serial no. to identify frag progs, never re-used */
-};
-
-
-struct gen4_fragment_program {
- struct brw_fragment_program base;
-
- bool contains_flat_varying;
- bool contains_noperspective_varying;
-
- /*
- * Mapping of varying slots to interpolation modes.
- * Used Gen4/5 by the clip|sf|wm stages.
- */
- unsigned char interp_mode[BRW_VARYING_SLOT_COUNT];
-};
-
-
-/** Subclass of Mesa compute program */
-struct brw_compute_program {
- struct gl_program program;
- unsigned id; /**< serial no. to identify compute progs, never re-used */
-};
-
-
-struct brw_shader {
- struct gl_linked_shader base;
bool compiled_once;
};
-/**
- * Bitmask indicating which fragment shader inputs represent varyings (and
- * hence have to be delivered to the fragment shader by the SF/SBE stage).
- */
-#define BRW_FS_VARYING_INPUT_MASK \
- (BITFIELD64_RANGE(0, VARYING_SLOT_MAX) & \
- ~VARYING_BIT_POS & ~VARYING_BIT_FACE)
-
struct brw_sf_prog_data {
GLuint urb_read_length;
};
-/**
- * We always program SF to start reading at an offset of 1 (2 varying slots)
- * from the start of the vertex URB entry. This causes it to skip:
- * - VARYING_SLOT_PSIZ and BRW_VARYING_SLOT_NDC on gen4-5
- * - VARYING_SLOT_PSIZ and VARYING_SLOT_POS on gen6+
- */
-#define BRW_SF_URB_ENTRY_READ_OFFSET 1
-
-
struct brw_clip_prog_data {
GLuint curb_read_length; /* user planes? */
GLuint clip_mode;
/** Number of texture sampler units */
#define BRW_MAX_TEX_UNIT 32
-/** Max number of render targets in a shader */
-#define BRW_MAX_DRAW_BUFFERS 8
-
/** Max number of UBOs in a shader */
#define BRW_MAX_UBO 14
/** Max number of image uniforms in a shader */
#define BRW_MAX_IMAGES 32
-/**
- * Max number of binding table entries used for stream output.
- *
- * From the OpenGL 3.0 spec, table 6.44 (Transform Feedback State), the
- * minimum value of MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS is 64.
- *
- * On Gen6, the size of transform feedback data is limited not by the number
- * of components but by the number of binding table entries we set aside. We
- * use one binding table entry for a float, one entry for a vector, and one
- * entry per matrix column. Since the only way we can communicate our
- * transform feedback capabilities to the client is via
- * MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS, we need to plan for the
- * worst case, in which all the varyings are floats, so we use up one binding
- * table entry per component. Therefore we need to set aside at least 64
- * binding table entries for use by transform feedback.
- *
- * Note: since we don't currently pack varyings, it is currently impossible
- * for the client to actually use up all of these binding table entries--if
- * all of their varyings were floats, they would run out of varying slots and
- * fail to link. But that's a bug, so it seems prudent to go ahead and
- * allocate the number of binding table entries we will need once the bug is
- * fixed.
- */
-#define BRW_MAX_SOL_BINDINGS 64
-
/** Maximum number of actual buffers used for stream output */
#define BRW_MAX_SOL_BUFFERS 4
2 + /* shader time, pull constants */ \
1 /* cs num work groups */)
-#define SURF_INDEX_GEN6_SOL_BINDING(t) (t)
-
-/**
- * Stride in bytes between shader_time entries.
- *
- * We separate entries by a cacheline to reduce traffic between EUs writing to
- * different entries.
- */
-#define SHADER_TIME_STRIDE 64
-
-struct brw_cache_item {
- /**
- * Effectively part of the key, cache_id identifies what kind of state
- * buffer is involved, and also which dirty flag should set.
- */
- enum brw_cache_id cache_id;
- /** 32-bit hash of the key data */
- GLuint hash;
- GLuint key_size; /* for variable-sized keys */
- GLuint aux_size;
- const void *key;
-
- uint32_t offset;
- uint32_t size;
-
- struct brw_cache_item *next;
-};
-
-
struct brw_cache {
struct brw_context *brw;
GLuint step_rate;
};
struct brw_vertex_element {
- const struct gl_client_array *glarray;
+ const struct gl_vertex_array *glarray;
int buffer;
-
+ bool is_dual_slot;
/** Offset of the first element within the buffer object */
unsigned int offset;
};
} saved;
};
-#define MAX_GS_INPUT_VERTICES 6
-
#define BRW_MAX_XFB_STREAMS 4
struct brw_transform_feedback_object {
/** The most recent primitive mode (GL_TRIANGLES/GL_POINTS/GL_LINES). */
GLenum primitive_mode;
+ /**
+ * The maximum number of vertices that we can write without overflowing
+ * any of the buffers currently being used for transform feedback.
+ */
+ unsigned max_index;
+
/**
* Count of primitives generated during this transform feedback operation.
* @{
struct gen_l3_config;
+enum brw_query_kind {
+ PIPELINE_STATS
+};
+
+struct brw_perf_query_info
+{
+ enum brw_query_kind kind;
+ const char *name;
+ struct brw_perf_query_counter *counters;
+ int n_counters;
+ size_t data_size;
+};
+
/**
* brw_context is derived from gl_context.
*/
bool use_rep_send;
bool use_resource_streamer;
- /**
- * Whether LRI can be used to write register values from the batch buffer.
- */
- bool can_do_pipelined_register_writes;
-
/**
* Some versions of Gen hardware don't do centroid interpolation correctly
* on unlit pixels, causing incorrect values for derivatives near triangle
} predicate;
struct {
- /** A map from pipeline statistics counter IDs to MMIO addresses. */
- const int *statistics_registers;
-
- /** The number of active monitors using OA counters. */
- unsigned oa_users;
-
- /**
- * A buffer object storing OA counter snapshots taken at the start and
- * end of each batch (creating "bookends" around the batch).
- */
- drm_intel_bo *bookend_bo;
+ struct brw_perf_query_info *queries;
+ int n_queries;
- /** The number of snapshots written to bookend_bo. */
- int bookend_snapshots;
-
- /**
- * An array of monitors whose results haven't yet been assembled based on
- * the data in buffer objects.
- *
- * These may be active, or have already ended. However, the results
- * have not been requested.
- */
- struct brw_perf_monitor_object **unresolved;
- int unresolved_elements;
- int unresolved_array_size;
-
- /**
- * Mapping from a uint32_t offset within an OA snapshot to the ID of
- * the counter which MI_REPORT_PERF_COUNT stores there.
- */
- const int *oa_snapshot_layout;
-
- /** Number of 32-bit entries in a hardware counter snapshot. */
- int entries_per_oa_snapshot;
- } perfmon;
+ int n_active_pipeline_stats_queries;
+ } perfquery;
int num_atoms[BRW_NUM_PIPELINES];
const struct brw_tracked_state render_atoms[76];
struct intel_screen *screen;
};
-/*======================================================================
- * brw_vtbl.c
- */
-void brwInitVtbl( struct brw_context *brw );
-
/* brw_clear.c */
extern void intelInitClearFuncs(struct dd_function_table *functions);
/*======================================================================
* brw_misc_state.c
*/
-void brw_get_depthstencil_tile_masks(struct intel_mipmap_tree *depth_mt,
- uint32_t depth_level,
- uint32_t depth_layer,
- struct intel_mipmap_tree *stencil_mt,
- uint32_t *out_tile_mask_x,
- uint32_t *out_tile_mask_y);
void brw_workaround_depthstencil_alignment(struct brw_context *brw,
GLbitfield clear_mask);
void brw_write_depth_count(struct brw_context *brw, drm_intel_bo *bo, int idx);
/** hsw_queryobj.c */
+void hsw_overflow_result_to_gpr0(struct brw_context *brw,
+ struct brw_query_object *query,
+ int count);
void hsw_init_queryobj_functions(struct dd_function_table *functions);
/** brw_conditional_render.c */
void brwInitFragProgFuncs( struct dd_function_table *functions );
-/* Per-thread scratch space is a power-of-two multiple of 1KB. */
-static inline int
-brw_get_scratch_size(int size)
-{
- return MAX2(1024, util_next_power_of_two(size));
-}
void brw_get_scratch_bo(struct brw_context *brw,
drm_intel_bo **scratch_bo, int size);
void brw_alloc_stage_scratch(struct brw_context *brw,
unsigned thread_count);
void brw_init_shader_time(struct brw_context *brw);
int brw_get_shader_time_index(struct brw_context *brw,
- struct gl_shader_program *shader_prog,
struct gl_program *prog,
- enum shader_time_shader_type type);
+ enum shader_time_shader_type type,
+ bool is_glsl_sh);
void brw_collect_and_report_shader_time(struct brw_context *brw);
void brw_destroy_shader_time(struct brw_context *brw);
*/
void brw_upload_cs_urb_state(struct brw_context *brw);
-/* brw_fs_reg_allocate.cpp
- */
-void brw_fs_alloc_reg_sets(struct brw_compiler *compiler);
-
-/* brw_vec4_reg_allocate.cpp */
-void brw_vec4_alloc_reg_set(struct brw_compiler *compiler);
-
-/* brw_disasm.c */
-int brw_disassemble_inst(FILE *file, const struct gen_device_info *devinfo,
- struct brw_inst *inst, bool is_compacted);
-
/* brw_vs.c */
gl_clip_plane *brw_select_clip_planes(struct gl_context *ctx);
/* brw_draw_upload.c */
unsigned brw_get_vertex_surface_type(struct brw_context *brw,
- const struct gl_client_array *glarray);
+ const struct gl_vertex_array *glarray);
static inline unsigned
brw_get_index_type(GLenum type)
struct gl_buffer_object *buffer_obj,
uint32_t *out_offset, unsigned num_vector_components,
unsigned stride_dwords, unsigned offset_dwords);
-void brw_upload_ubo_surfaces(struct brw_context *brw,
- struct gl_linked_shader *shader,
+void brw_upload_ubo_surfaces(struct brw_context *brw, struct gl_program *prog,
struct brw_stage_state *stage_state,
struct brw_stage_prog_data *prog_data);
void brw_upload_abo_surfaces(struct brw_context *brw,
- struct gl_linked_shader *shader,
+ const struct gl_program *prog,
struct brw_stage_state *stage_state,
struct brw_stage_prog_data *prog_data);
void brw_upload_image_surfaces(struct brw_context *brw,
- struct gl_linked_shader *shader,
+ const struct gl_program *prog,
struct brw_stage_state *stage_state,
struct brw_stage_prog_data *prog_data);
struct gl_renderbuffer *rb);
uint32_t brw_depth_format(struct brw_context *brw, mesa_format format);
-/* brw_performance_monitor.c */
-void brw_init_performance_monitors(struct brw_context *brw);
-void brw_dump_perf_monitors(struct brw_context *brw);
-void brw_perf_monitor_new_batch(struct brw_context *brw);
-void brw_perf_monitor_finish_batch(struct brw_context *brw);
+/* brw_performance_query.c */
+void brw_init_performance_queries(struct brw_context *brw);
/* intel_buffer_objects.c */
int brw_bo_map(struct brw_context *brw, drm_intel_bo *bo, int write_enable,
void
brw_end_transform_feedback(struct gl_context *ctx,
struct gl_transform_feedback_object *obj);
+void
+brw_pause_transform_feedback(struct gl_context *ctx,
+ struct gl_transform_feedback_object *obj);
+void
+brw_resume_transform_feedback(struct gl_context *ctx,
+ struct gl_transform_feedback_object *obj);
+void
+brw_save_primitives_written_counters(struct brw_context *brw,
+ struct brw_transform_feedback_object *obj);
+void
+brw_compute_xfb_vertices_written(struct brw_context *brw,
+ struct brw_transform_feedback_object *obj);
GLsizei
brw_get_transform_feedback_vertex_count(struct gl_context *ctx,
struct gl_transform_feedback_object *obj,
return (const struct brw_program *) p;
}
-static inline struct brw_fragment_program *
-brw_fragment_program(struct gl_program *p)
-{
- return (struct brw_fragment_program *) p;
-}
-
-static inline const struct brw_fragment_program *
-brw_fragment_program_const(const struct gl_program *p)
-{
- return (const struct brw_fragment_program *) p;
-}
-
-static inline struct brw_compute_program *
-brw_compute_program(struct gl_program *p)
-{
- return (struct brw_compute_program *) p;
-}
-
-/**
- * Pre-gen6, the register file of the EUs was shared between threads,
- * and each thread used some subset allocated on a 16-register block
- * granularity. The unit states wanted these block counts.
- */
-static inline int
-brw_register_blocks(int reg_count)
-{
- return ALIGN(reg_count, 16) / 16 - 1;
-}
-
static inline uint32_t
brw_program_reloc(struct brw_context *brw, uint32_t state_offset,
uint32_t prog_offset)
return brw->cache.bo->offset64 + prog_offset;
}
-bool brw_do_cubemap_normalize(struct exec_list *instructions);
-bool brw_lower_texture_gradients(struct brw_context *brw,
- struct exec_list *instructions);
+static inline bool
+brw_depth_writes_enabled(const struct brw_context *brw)
+{
+ const struct gl_context *ctx = &brw->ctx;
-extern const char * const conditional_modifier[16];
-extern const char *const pred_ctrl_align16[16];
+ /* We consider depth writes disabled if the depth function is GL_EQUAL,
+ * because it would just overwrite the existing depth value with itself.
+ *
+ * These bonus depth writes not only use bandwidth, but they also can
+ * prevent early depth processing. For example, if the pixel shader
+ * discards, the hardware must invoke the to determine whether or not
+ * to do the depth write. If writes are disabled, we may still be able
+ * to do the depth test before the shader, and skip the shader execution.
+ *
+ * The Broadwell 3DSTATE_WM_DEPTH_STENCIL documentation also contains
+ * a programming note saying to disable depth writes for EQUAL.
+ */
+ return ctx->Depth.Test && ctx->Depth.Mask && ctx->Depth.Func != GL_EQUAL;
+}
void
brw_emit_depthbuffer(struct brw_context *brw);
projects / mesa.git / blobdiff