#include "util/macros.h"
#include "common/v3d_debug.h"
#include "common/v3d_device_info.h"
+#include "common/v3d_limits.h"
#include "compiler/nir/nir.h"
#include "util/list.h"
#include "util/u_math.h"
#include "qpu/qpu_instr.h"
#include "pipe/p_state.h"
-#define V3D_MAX_TEXTURE_SAMPLERS 32
-#define V3D_MAX_SAMPLES 4
-#define V3D_MAX_FS_INPUTS 64
-#define V3D_MAX_VS_INPUTS 64
-
struct nir_builder;
struct v3d_fs_inputs {
* or physical registers later.
*/
QFILE_TEMP,
- QFILE_UNIF,
- QFILE_TLB,
- QFILE_TLBU,
/**
* VPM reads use this with an index value to say what part of the VPM
return (struct qreg){file, index};
}
+static inline struct qreg vir_magic_reg(uint32_t index)
+{
+ return (struct qreg){QFILE_MAGIC, index};
+}
+
+static inline struct qreg vir_nop_reg(void)
+{
+ return (struct qreg){QFILE_NULL, 0};
+}
+
/**
* A reference to an actual register at the QPU level, for register
* allocation.
/* Pre-register-allocation references to src/dst registers */
struct qreg dst;
struct qreg src[3];
- bool cond_is_exec_mask;
- bool has_implicit_uniform;
bool is_last_thrsw;
- /* After vir_to_qpu.c: If instr reads a uniform, which uniform from
- * the uncompiled stream it is.
+ /* If the instruction reads a uniform (other than through src[i].file
+ * == QFILE_UNIF), that uniform's index in c->uniform_contents. ~0
+ * otherwise.
*/
int uniform;
};
QUNIFORM_TMU_CONFIG_P0,
QUNIFORM_TMU_CONFIG_P1,
+ QUNIFORM_IMAGE_TMU_CONFIG_P0,
+
QUNIFORM_TEXTURE_FIRST_LEVEL,
QUNIFORM_TEXTURE_WIDTH,
QUNIFORM_TEXRECT_SCALE_X,
QUNIFORM_TEXRECT_SCALE_Y,
+ /* Returns the base offset of the SSBO given by the data value. */
+ QUNIFORM_SSBO_OFFSET,
+
+ /* Returns the size of the SSBO given by the data value. */
+ QUNIFORM_GET_BUFFER_SIZE,
+
+ /* Sizes (in pixels) of a shader image given by the data value. */
+ QUNIFORM_IMAGE_WIDTH,
+ QUNIFORM_IMAGE_HEIGHT,
+ QUNIFORM_IMAGE_DEPTH,
+ QUNIFORM_IMAGE_ARRAY_SIZE,
+
QUNIFORM_ALPHA_REF,
+ /* Number of workgroups passed to glDispatchCompute in the dimension
+ * selected by the data value.
+ */
+ QUNIFORM_NUM_WORK_GROUPS,
+
/**
* Returns the the offset of the scratch buffer for register spilling.
*/
QUNIFORM_SPILL_OFFSET,
QUNIFORM_SPILL_SIZE_PER_THREAD,
+
+ /**
+ * Returns the offset of the shared memory for compute shaders.
+ *
+ * This will be accessed using TMU general memory operations, so the
+ * L2T cache will effectively be the shared memory area.
+ */
+ QUNIFORM_SHARED_OFFSET,
};
-static inline uint32_t v3d_tmu_config_data_create(uint32_t unit, uint32_t value)
+static inline uint32_t v3d_unit_data_create(uint32_t unit, uint32_t value)
{
+ assert(value < (1 << 24));
return unit << 24 | value;
}
-static inline uint32_t v3d_tmu_config_data_get_unit(uint32_t data)
+static inline uint32_t v3d_unit_data_get_unit(uint32_t data)
{
return data >> 24;
}
-static inline uint32_t v3d_tmu_config_data_get_value(uint32_t data)
+static inline uint32_t v3d_unit_data_get_offset(uint32_t data)
{
return data & 0xffffff;
}
return slot.slot_and_component & 3;
}
-struct v3d_ubo_range {
- /**
- * offset in bytes from the start of the ubo where this range is
- * uploaded.
- *
- * Only set once used is set.
- */
- uint32_t dst_offset;
-
- /**
- * offset in bytes from the start of the gallium uniforms where the
- * data comes from.
- */
- uint32_t src_offset;
-
- /** size in bytes of this ubo range */
- uint32_t size;
-};
-
struct v3d_key {
void *shader_state;
struct {
uint8_t swizzle[4];
uint8_t return_size;
uint8_t return_channels;
- unsigned compare_mode:1;
- unsigned compare_func:3;
bool clamp_s:1;
bool clamp_t:1;
bool clamp_r:1;
bool sample_alpha_to_one;
bool clamp_color;
bool shade_model_flat;
- uint8_t nr_cbufs;
+ /* Mask of which color render targets are present. */
+ uint8_t cbufs;
uint8_t swap_color_rb;
/* Mask of which render targets need to be written as 32-bit floats */
uint8_t f32_color_rb;
/** @{ used by v3d_vir_live_variables.c */
BITSET_WORD *def;
+ BITSET_WORD *defin;
+ BITSET_WORD *defout;
BITSET_WORD *use;
BITSET_WORD *live_in;
BITSET_WORD *live_out;
struct v3d_compiler {
const struct v3d_device_info *devinfo;
struct ra_regs *regs;
+ unsigned int reg_class_any[3];
+ unsigned int reg_class_r5[3];
unsigned int reg_class_phys[3];
unsigned int reg_class_phys_or_acc[3];
};
struct exec_list *cf_node_list;
const struct v3d_compiler *compiler;
+ void (*debug_output)(const char *msg,
+ void *debug_output_data);
+ void *debug_output_data;
+
/**
* Mapping from nir_register * or nir_ssa_def * to array of struct
* qreg for the values.
uint32_t centroid_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
bool uses_center_w;
-
- struct v3d_ubo_range *ubo_ranges;
- bool *ubo_range_used;
- uint32_t ubo_ranges_array_size;
- /** Number of uniform areas tracked in ubo_ranges. */
- uint32_t num_ubo_ranges;
- uint32_t next_ubo_dst_offset;
+ bool writes_z;
/* State for whether we're executing on each channel currently. 0 if
* yes, otherwise a block number + 1 that the channel jumped to.
*/
struct qreg execute;
+ bool in_control_flow;
struct qreg line_x, point_x, point_y;
/* Fragment shader payload regs. */
struct qreg payload_w, payload_w_centroid, payload_z;
- uint8_t vattr_sizes[V3D_MAX_VS_INPUTS];
- uint32_t num_vpm_writes;
+ struct qreg cs_payload[2];
+ struct qreg cs_shared_offset;
+ int local_invocation_index_bits;
+
+ uint8_t vattr_sizes[V3D_MAX_VS_INPUTS / 4];
+ uint32_t vpm_output_size;
/* Size in bytes of registers that have been spilled. This is how much
* space needs to be available in the spill BO per thread per QPU.
*/
uint32_t spill_size;
- /* Shader-db stats for register spilling. */
- uint32_t spills, fills;
+ /* Shader-db stats */
+ uint32_t spills, fills, loops;
/**
* Register spilling's per-thread base address, shared between each
* spill/fill's addressing calculations.
enum quniform_contents *uniform_contents;
uint32_t uniform_array_size;
uint32_t num_uniforms;
- uint32_t num_outputs;
uint32_t output_position_index;
nir_variable *output_color_var[4];
- uint32_t output_point_size_index;
uint32_t output_sample_mask_index;
struct qreg undef;
*/
uint32_t num_inputs;
- /**
- * Number of inputs from num_inputs remaining to be queued to the read
- * FIFO in the VS/CS.
- */
- uint32_t num_inputs_remaining;
-
- /* Number of inputs currently in the read FIFO for the VS/CS */
- uint32_t num_inputs_in_fifo;
-
- /** Next offset in the VPM to read from in the VS/CS */
- uint32_t vpm_read_offset;
-
uint32_t program_id;
uint32_t variant_id;
struct v3d_prog_data {
struct v3d_uniform_list uniforms;
- struct v3d_ubo_range *ubo_ranges;
- uint32_t num_ubo_ranges;
- uint32_t ubo_size;
uint32_t spill_size;
- uint8_t num_inputs;
uint8_t threads;
/* For threads > 1, whether the program should be dispatched in the
bool uses_iid, uses_vid;
/* Number of components read from each vertex attribute. */
- uint8_t vattr_sizes[32];
+ uint8_t vattr_sizes[V3D_MAX_VS_INPUTS / 4];
/* Total number of components read, for the shader state record. */
uint32_t vpm_input_size;
uint32_t centroid_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
+ uint8_t num_inputs;
bool writes_z;
- bool discard;
+ bool disable_ez;
bool uses_center_w;
};
+struct v3d_compute_prog_data {
+ struct v3d_prog_data base;
+ /* Size in bytes of the workgroup's shared space. */
+ uint32_t shared_size;
+};
+
+static inline bool
+vir_has_uniform(struct qinst *inst)
+{
+ return inst->uniform != ~0;
+}
+
/* Special nir_load_input intrinsic index for loading the current TLB
* destination color.
*/
void v3d_compiler_free(const struct v3d_compiler *compiler);
void v3d_optimize_nir(struct nir_shader *s);
-uint64_t *v3d_compile_vs(const struct v3d_compiler *compiler,
- struct v3d_vs_key *key,
- struct v3d_vs_prog_data *prog_data,
- nir_shader *s,
- int program_id, int variant_id,
- uint32_t *final_assembly_size);
-
-uint64_t *v3d_compile_fs(const struct v3d_compiler *compiler,
- struct v3d_fs_key *key,
- struct v3d_fs_prog_data *prog_data,
- nir_shader *s,
- int program_id, int variant_id,
- uint32_t *final_assembly_size);
+uint64_t *v3d_compile(const struct v3d_compiler *compiler,
+ struct v3d_key *key,
+ struct v3d_prog_data **prog_data,
+ nir_shader *s,
+ void (*debug_output)(const char *msg,
+ void *debug_output_data),
+ void *debug_output_data,
+ int program_id, int variant_id,
+ uint32_t *final_assembly_size);
void v3d_nir_to_vir(struct v3d_compile *c);
struct qreg src0, struct qreg src1);
struct qinst *vir_mul_inst(enum v3d_qpu_mul_op op, struct qreg dst,
struct qreg src0, struct qreg src1);
-struct qinst *vir_branch_inst(enum v3d_qpu_branch_cond cond, struct qreg src0);
+struct qinst *vir_branch_inst(struct v3d_compile *c,
+ enum v3d_qpu_branch_cond cond);
void vir_remove_instruction(struct v3d_compile *c, struct qinst *qinst);
+uint32_t vir_get_uniform_index(struct v3d_compile *c,
+ enum quniform_contents contents,
+ uint32_t data);
struct qreg vir_uniform(struct v3d_compile *c,
enum quniform_contents contents,
uint32_t data);
void vir_schedule_instructions(struct v3d_compile *c);
+void v3d_setup_spill_base(struct v3d_compile *c);
struct v3d_qpu_instr v3d_qpu_nop(void);
struct qreg vir_emit_def(struct v3d_compile *c, struct qinst *inst);
struct qinst *vir_emit_nondef(struct v3d_compile *c, struct qinst *inst);
void vir_set_cond(struct qinst *inst, enum v3d_qpu_cond cond);
void vir_set_pf(struct qinst *inst, enum v3d_qpu_pf pf);
+void vir_set_uf(struct qinst *inst, enum v3d_qpu_uf uf);
void vir_set_unpack(struct qinst *inst, int src,
enum v3d_qpu_input_unpack unpack);
struct qreg vir_get_temp(struct v3d_compile *c);
void vir_emit_last_thrsw(struct v3d_compile *c);
void vir_calculate_live_intervals(struct v3d_compile *c);
-bool vir_has_implicit_uniform(struct qinst *inst);
-int vir_get_implicit_uniform_src(struct qinst *inst);
-int vir_get_non_sideband_nsrc(struct qinst *inst);
int vir_get_nsrc(struct qinst *inst);
bool vir_has_side_effects(struct v3d_compile *c, struct qinst *inst);
bool vir_get_add_op(struct qinst *inst, enum v3d_qpu_add_op *op);
bool vir_is_tex(struct qinst *inst);
bool vir_is_add(struct qinst *inst);
bool vir_is_mul(struct qinst *inst);
-bool vir_is_float_input(struct qinst *inst);
-bool vir_depends_on_flags(struct qinst *inst);
bool vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst);
bool vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst);
struct qreg vir_follow_movs(struct v3d_compile *c, struct qreg reg);
bool vir_opt_copy_propagate(struct v3d_compile *c);
bool vir_opt_dead_code(struct v3d_compile *c);
bool vir_opt_peephole_sf(struct v3d_compile *c);
+bool vir_opt_redundant_flags(struct v3d_compile *c);
bool vir_opt_small_immediates(struct v3d_compile *c);
bool vir_opt_vpm(struct v3d_compile *c);
void v3d_nir_lower_blend(nir_shader *s, struct v3d_compile *c);
void v3d_nir_lower_io(nir_shader *s, struct v3d_compile *c);
+void v3d_nir_lower_scratch(nir_shader *s);
void v3d_nir_lower_txf_ms(nir_shader *s, struct v3d_compile *c);
+void v3d_nir_lower_image_load_store(nir_shader *s);
void vir_lower_uniforms(struct v3d_compile *c);
void v3d33_vir_vpm_read_setup(struct v3d_compile *c, int num_components);
void v3d33_vir_vpm_write_setup(struct v3d_compile *c);
void v3d33_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
void v3d40_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
+void v3d40_vir_emit_image_load_store(struct v3d_compile *c,
+ nir_intrinsic_instr *instr);
void v3d_vir_to_qpu(struct v3d_compile *c, struct qpu_reg *temp_registers);
uint32_t v3d_qpu_schedule_instructions(struct v3d_compile *c);
struct qpu_reg *v3d_register_allocate(struct v3d_compile *c, bool *spilled);
bool vir_init_reg_sets(struct v3d_compiler *compiler);
-void vir_PF(struct v3d_compile *c, struct qreg src, enum v3d_qpu_pf pf);
+bool v3d_gl_format_is_return_32(GLenum format);
static inline bool
quniform_contents_is_texture_p0(enum quniform_contents contents)
V3D_MAX_TEXTURE_SAMPLERS));
}
+static inline bool
+vir_in_nonuniform_control_flow(struct v3d_compile *c)
+{
+ return c->execute.file != QFILE_NULL;
+}
+
static inline struct qreg
vir_uniform_ui(struct v3d_compile *c, uint32_t ui)
{
VIR_A_ALU0(YCD)
VIR_A_ALU0(MSF)
VIR_A_ALU0(REVF)
+VIR_A_ALU0(BARRIERID)
VIR_A_NODST_1(VPMSETUP)
VIR_A_NODST_0(VPMWT)
VIR_A_ALU2(FCMP)
vir_BRANCH(struct v3d_compile *c, enum v3d_qpu_branch_cond cond)
{
/* The actual uniform_data value will be set at scheduling time */
- return vir_emit_nondef(c, vir_branch_inst(cond, vir_uniform_ui(c, 0)));
+ return vir_emit_nondef(c, vir_branch_inst(c, cond));
}
#define vir_for_each_block(block, c) \
vir_for_each_block(_block, c) \
vir_for_each_inst(inst, _block)
+#define vir_for_each_inst_inorder_safe(inst, c) \
+ vir_for_each_block(_block, c) \
+ vir_for_each_inst_safe(inst, _block)
+
#endif /* V3D_COMPILER_H */
projects / mesa.git / blobdiff