#include "bifrost.h"
#include "compiler/nir/nir.h"
+#include "panfrost/util/pan_ir.h"
/* Bifrost opcodes are tricky -- the same op may exist on both FMA and
* ADD with two completely different opcodes, and opcodes can be varying
BI_CMP,
BI_BLEND,
BI_BITWISE,
+ BI_COMBINE,
BI_CONVERT,
BI_CSEL,
BI_DISCARD,
- BI_EXTRACT,
BI_FMA,
+ BI_FMOV,
BI_FREXP,
+ BI_IMATH,
BI_LOAD,
+ BI_LOAD_UNIFORM,
BI_LOAD_ATTR,
BI_LOAD_VAR,
BI_LOAD_VAR_ADDRESS,
- BI_MAKE_VEC,
BI_MINMAX,
BI_MOV,
- BI_SHIFT,
+ BI_REDUCE_FMA,
+ BI_SELECT,
BI_STORE,
BI_STORE_VAR,
- BI_SPECIAL, /* _FAST, _TABLE on supported GPUs */
- BI_SWIZZLE,
+ BI_SPECIAL, /* _FAST on supported GPUs */
+ BI_TABLE,
BI_TEX,
BI_ROUND,
+ BI_IMUL,
BI_NUM_CLASSES
};
/* abs/neg/outmod valid for a float op */
#define BI_MODS (1 << 0)
-/* Generic enough that little class-specific information is required. In other
- * words, it acts as a "normal" ALU op, even if the encoding ends up being
- * irregular enough to warrant a separate class */
-#define BI_GENERIC (1 << 1)
+/* Accepts a bi_cond */
+#define BI_CONDITIONAL (1 << 1)
/* Accepts a bifrost_roundmode */
#define BI_ROUNDMODE (1 << 2)
/* For scheduling purposes this is a high latency instruction and must be at
* the end of a clause. Implies ADD */
-#define BI_SCHED_HI_LATENCY ((1 << 7) | BI_SCHED_ADD)
+#define BI_SCHED_HI_LATENCY (1 << 7)
-/* It can't get any worse than csel4... can it? */
-#define BIR_SRC_COUNT 4
+/* Intrinsic is vectorized and acts with `vector_channels` components */
+#define BI_VECTOR (1 << 8)
-/* Class-specific data for BI_LD_ATTR, BI_LD_VAR_ADDR */
-struct bi_load {
- /* Note: no indirects here */
- unsigned location;
+/* Use a data register for src0/dest respectively, bypassing the usual
+ * register accessor. Mutually exclusive. */
+#define BI_DATA_REG_SRC (1 << 9)
+#define BI_DATA_REG_DEST (1 << 10)
- /* Only for BI_LD_ATTR. But number of vector channels */
- unsigned channels;
-};
+/* Quirk: cannot encode multiple abs on FMA in fp16 mode */
+#define BI_NO_ABS_ABS_FP16_FMA (1 << 11)
+
+/* It can't get any worse than csel4... can it? */
+#define BIR_SRC_COUNT 4
/* BI_LD_VARY */
struct bi_load_vary {
- /* All parameters used here. Indirect location specified in
- * src1 and ignoring location, if present. */
- struct bi_load load;
-
enum bifrost_interp_mode interp_mode;
bool reuse;
bool flat;
BI_COND_NE,
};
-struct bi_branch {
- /* Types are specified in src_types and must be compatible (either both
- * int, or both float, 16/32, and same size or 32/16 if float. Types
- * ignored if BI_COND_ALWAYS is set for an unconditional branch. */
-
- enum bi_cond cond;
- struct bi_block *target;
-};
-
/* Opcodes within a class */
enum bi_minmax_op {
BI_MINMAX_MIN,
BI_BITWISE_XOR
};
-enum bi_round_op {
- BI_ROUND_MODE, /* use round mode */
- BI_ROUND_ROUND /* i.e.: fround() */
+enum bi_imath_op {
+ BI_IMATH_ADD,
+ BI_IMATH_SUB,
+};
+
+enum bi_imul_op {
+ BI_IMUL_IMUL,
+};
+
+enum bi_table_op {
+ /* fp32 log2() with low precision, suitable for GL or half_log2() in
+ * CL. In the first argument, takes x. Letting u be such that x =
+ * 2^{-m} u with m integer and 0.75 <= u < 1.5, returns
+ * log2(u) / (u - 1). */
+
+ BI_TABLE_LOG2_U_OVER_U_1_LOW,
+};
+
+enum bi_reduce_op {
+ /* Takes two fp32 arguments and returns x + frexp(y). Used in
+ * low-precision log2 argument reduction on newer models. */
+
+ BI_REDUCE_ADD_FREXPM,
+};
+
+enum bi_frexp_op {
+ BI_FREXPE_LOG,
+};
+
+enum bi_special_op {
+ BI_SPECIAL_FRCP,
+ BI_SPECIAL_FRSQ,
+
+ /* fp32 exp2() with low precision, suitable for half_exp2() in CL or
+ * exp2() in GL. In the first argument, it takes f2i_rte(x * 2^24). In
+ * the second, it takes x itself. */
+ BI_SPECIAL_EXP2_LOW,
+ BI_SPECIAL_IABS,
+};
+
+enum bi_tex_op {
+ BI_TEX_NORMAL,
+ BI_TEX_COMPACT,
+ BI_TEX_DUAL
+};
+
+struct bi_bitwise {
+ bool src_invert[2];
+ bool rshift; /* false for lshift */
+};
+
+struct bi_texture {
+ /* Constant indices. Indirect would need to be in src[..] like normal,
+ * we can reserve some sentinels there for that for future. */
+ unsigned texture_index, sampler_index;
};
typedef struct {
struct list_head link; /* Must be first */
enum bi_class type;
- /* Indices, see bir_ssa_index etc. Note zero is special cased
+ /* Indices, see pan_ssa_index etc. Note zero is special cased
* to "no argument" */
unsigned dest;
unsigned src[BIR_SRC_COUNT];
- /* If one of the sources has BIR_INDEX_CONSTANT... Also, for
- * BI_EXTRACT, the component index is stored here. */
+ /* 32-bit word offset for destination, added to the register number in
+ * RA when lowering combines */
+ unsigned dest_offset;
+
+ /* If one of the sources has BIR_INDEX_CONSTANT */
union {
uint64_t u64;
uint32_t u32;
/* Source types if required by the class */
nir_alu_type src_types[BIR_SRC_COUNT];
- /* If the source type is 8-bit or 16-bit such that SIMD is possible, and
- * the class has BI_SWIZZLABLE, this is a swizzle for the input. Swizzles
- * in practice only occur with one-source arguments (conversions,
- * dedicated swizzle ops) and as component selection on two-sources
- * where it is unambiguous which is which. Bounds are 32/type_size. */
- unsigned swizzle[4];
+ /* If the source type is 8-bit or 16-bit such that SIMD is possible,
+ * and the class has BI_SWIZZLABLE, this is a swizzle in the usual
+ * sense. On non-SIMD instructions, it can be used for component
+ * selection, so we don't have to special case extraction. */
+ uint8_t swizzle[BIR_SRC_COUNT][NIR_MAX_VEC_COMPONENTS];
+
+ /* For VECTOR ops, how many channels are written? */
+ unsigned vector_channels;
+
+ /* The comparison op. BI_COND_ALWAYS may not be valid. */
+ enum bi_cond cond;
/* A class-specific op from which the actual opcode can be derived
* (along with the above information) */
union {
enum bi_minmax_op minmax;
enum bi_bitwise_op bitwise;
- enum bi_round_op round;
+ enum bi_special_op special;
+ enum bi_reduce_op reduce;
+ enum bi_table_op table;
+ enum bi_frexp_op frexp;
+ enum bi_tex_op texture;
+ enum bi_imath_op imath;
+ enum bi_imul_op imul;
+
+ /* For FMA/ADD, should we add a biased exponent? */
+ bool mscale;
} op;
/* Union for class-specific information */
union {
enum bifrost_minmax_mode minmax;
- struct bi_load load;
struct bi_load_vary load_vary;
- struct bi_branch branch;
+ struct bi_block *branch_target;
+
+ /* For BLEND -- the location 0-7 */
+ unsigned blend_location;
- /* For CSEL, the comparison op. BI_COND_ALWAYS doesn't make
- * sense here but you can always just use a move for that */
- enum bi_cond csel_cond;
+ struct bi_bitwise bitwise;
+ struct bi_texture texture;
};
} bi_instruction;
-/* Scheduling takes place in two steps. Step 1 groups instructions within a
- * block into distinct clauses (bi_clause). Step 2 schedules instructions
- * within a clause into FMA/ADD pairs (bi_bundle).
- *
- * A bi_bundle contains two paired instruction pointers. If a slot is unfilled,
- * leave it NULL; the emitter will fill in a nop.
+/* Represents the assignment of ports for a given bi_bundle */
+
+typedef struct {
+ /* Register to assign to each port */
+ unsigned port[4];
+
+ /* Read ports can be disabled */
+ bool enabled[2];
+
+ /* Should we write FMA? what about ADD? If only a single port is
+ * enabled it is in port 2, else ADD/FMA is 2/3 respectively */
+ bool write_fma, write_add;
+
+ /* Should we read with port 3? */
+ bool read_port3;
+
+ /* Packed uniform/constant */
+ uint8_t uniform_constant;
+
+ /* Whether writes are actually for the last instruction */
+ bool first_instruction;
+} bi_registers;
+
+/* A bi_bundle contains two paired instruction pointers. If a slot is unfilled,
+ * leave it NULL; the emitter will fill in a nop. Instructions reference
+ * registers via ports which are assigned per bundle.
*/
typedef struct {
+ bi_registers regs;
bi_instruction *fma;
bi_instruction *add;
} bi_bundle;
+struct bi_block;
+
typedef struct {
struct list_head link;
+ /* Link back up for branch calculations */
+ struct bi_block *block;
+
/* A clause can have 8 instructions in bundled FMA/ADD sense, so there
- * can be 8 bundles. But each bundle can have both an FMA and an ADD,
- * so a clause can have up to 16 bi_instructions. Whether bundles or
- * instructions are used depends on where in scheduling we are. */
+ * can be 8 bundles. */
- unsigned instruction_count;
unsigned bundle_count;
-
- union {
- bi_instruction *instructions[16];
- bi_bundle bundles[8];
- };
+ bi_bundle bundles[8];
/* For scoreboarding -- the clause ID (this is not globally unique!)
* and its dependencies in terms of other clauses, computed during
bool back_to_back;
bool branch_conditional;
+ /* Assigned data register */
+ unsigned data_register;
+
/* Corresponds to the usual bit but shifted by a clause */
bool data_register_write_barrier;
- /* Constants read by this clause. ISA limit. */
+ /* Constants read by this clause. ISA limit. Must satisfy:
+ *
+ * constant_count + bundle_count <= 13
+ *
+ * Also implicitly constant_count <= bundle_count since a bundle only
+ * reads a single constant.
+ */
uint64_t constants[8];
unsigned constant_count;
+
+ /* Branches encode a constant offset relative to the program counter
+ * with some magic flags. By convention, if there is a branch, its
+ * constant will be last. Set this flag to indicate this is required.
+ */
+ bool branch_constant;
+
+ /* What type of high latency instruction is here, basically */
+ unsigned clause_type;
} bi_clause;
typedef struct bi_block {
- struct list_head link; /* must be first */
- unsigned name; /* Just for pretty-printing */
+ pan_block base; /* must be first */
/* If true, uses clauses; if false, uses instructions */
bool scheduled;
-
- union {
- struct list_head instructions; /* pre-schedule, list of bi_instructions */
- struct list_head clauses; /* list of bi_clause */
- };
-
- /* Control flow graph */
- struct set *predecessors;
- struct bi_block *successors[2];
+ struct list_head clauses; /* list of bi_clause */
} bi_block;
typedef struct {
nir_shader *nir;
gl_shader_stage stage;
struct list_head blocks; /* list of bi_block */
+ struct panfrost_sysvals sysvals;
uint32_t quirks;
-} bi_context;
-/* So we can distinguish between SSA/reg/sentinel quickly */
-#define BIR_NO_ARG (0)
-#define BIR_IS_REG (1)
+ /* During NIR->BIR */
+ nir_function_impl *impl;
+ bi_block *current_block;
+ bi_block *after_block;
+ bi_block *break_block;
+ bi_block *continue_block;
+ bool emitted_atest;
+ nir_alu_type *blend_types;
+
+ /* For creating temporaries */
+ unsigned temp_alloc;
+
+ /* Analysis results */
+ bool has_liveness;
+
+ /* Stats for shader-db */
+ unsigned instruction_count;
+ unsigned loop_count;
+} bi_context;
+
+static inline bi_instruction *
+bi_emit(bi_context *ctx, bi_instruction ins)
+{
+ bi_instruction *u = rzalloc(ctx, bi_instruction);
+ memcpy(u, &ins, sizeof(ins));
+ list_addtail(&u->link, &ctx->current_block->base.instructions);
+ return u;
+}
+
+static inline bi_instruction *
+bi_emit_before(bi_context *ctx, bi_instruction *tag, bi_instruction ins)
+{
+ bi_instruction *u = rzalloc(ctx, bi_instruction);
+ memcpy(u, &ins, sizeof(ins));
+ list_addtail(&u->link, &tag->link);
+ return u;
+}
+
+static inline void
+bi_remove_instruction(bi_instruction *ins)
+{
+ list_del(&ins->link);
+}
/* If high bits are set, instead of SSA/registers, we have specials indexed by
* the low bits if necessary.
*
* Fixed register: do not allocate register, do not collect $200.
* Uniform: access a uniform register given by low bits.
- * Constant: access the specified constant
+ * Constant: access the specified constant (specifies a bit offset / shift)
* Zero: special cased to avoid wasting a constant
+ * Passthrough: a bifrost_packed_src to passthrough T/T0/T1
*/
#define BIR_INDEX_REGISTER (1 << 31)
#define BIR_INDEX_UNIFORM (1 << 30)
#define BIR_INDEX_CONSTANT (1 << 29)
#define BIR_INDEX_ZERO (1 << 28)
+#define BIR_INDEX_PASS (1 << 27)
/* Keep me synced please so we can check src & BIR_SPECIAL */
#define BIR_SPECIAL ((BIR_INDEX_REGISTER | BIR_INDEX_UNIFORM) | \
- (BIR_INDEX_CONSTANT | BIR_INDEX_ZERO)
+ (BIR_INDEX_CONSTANT | BIR_INDEX_ZERO | BIR_INDEX_PASS))
static inline unsigned
-bir_ssa_index(nir_ssa_def *ssa)
+bi_max_temp(bi_context *ctx)
{
- /* Off-by-one ensures BIR_NO_ARG is skipped */
- return ((ssa->index + 1) << 1) | 0;
+ unsigned alloc = MAX2(ctx->impl->reg_alloc, ctx->impl->ssa_alloc);
+ return ((alloc + 2 + ctx->temp_alloc) << 1);
}
static inline unsigned
-bir_src_index(nir_src *src)
+bi_make_temp(bi_context *ctx)
{
- if (src->is_ssa)
- return bir_ssa_index(src->ssa);
- else {
- assert(!src->reg.indirect);
- return (src->reg.reg->index << 1) | BIR_IS_REG;
- }
+ return (ctx->impl->ssa_alloc + 1 + ctx->temp_alloc++) << 1;
}
static inline unsigned
-bir_dest_index(nir_dest *dst)
+bi_make_temp_reg(bi_context *ctx)
{
- if (dst->is_ssa)
- return bir_ssa_index(&dst->ssa);
- else {
- assert(!dst->reg.indirect);
- return (dst->reg.reg->index << 1) | BIR_IS_REG;
- }
+ return ((ctx->impl->reg_alloc + ctx->temp_alloc++) << 1) | PAN_IS_REG;
}
/* Iterators for Bifrost IR */
#define bi_foreach_block(ctx, v) \
- list_for_each_entry(bi_block, v, &ctx->blocks, link)
+ list_for_each_entry(pan_block, v, &ctx->blocks, link)
#define bi_foreach_block_from(ctx, from, v) \
- list_for_each_entry_from(bi_block, v, from, &ctx->blocks, link)
+ list_for_each_entry_from(pan_block, v, from, &ctx->blocks, link)
+
+#define bi_foreach_block_from_rev(ctx, from, v) \
+ list_for_each_entry_from_rev(pan_block, v, from, &ctx->blocks, link)
#define bi_foreach_instr_in_block(block, v) \
- list_for_each_entry(bi_instruction, v, &block->instructions, link)
+ list_for_each_entry(bi_instruction, v, &(block)->base.instructions, link)
#define bi_foreach_instr_in_block_rev(block, v) \
- list_for_each_entry_rev(bi_instruction, v, &block->instructions, link)
+ list_for_each_entry_rev(bi_instruction, v, &(block)->base.instructions, link)
#define bi_foreach_instr_in_block_safe(block, v) \
- list_for_each_entry_safe(bi_instruction, v, &block->instructions, link)
+ list_for_each_entry_safe(bi_instruction, v, &(block)->base.instructions, link)
#define bi_foreach_instr_in_block_safe_rev(block, v) \
- list_for_each_entry_safe_rev(bi_instruction, v, &block->instructions, link)
+ list_for_each_entry_safe_rev(bi_instruction, v, &(block)->base.instructions, link)
#define bi_foreach_instr_in_block_from(block, v, from) \
- list_for_each_entry_from(bi_instruction, v, from, &block->instructions, link)
+ list_for_each_entry_from(bi_instruction, v, from, &(block)->base.instructions, link)
#define bi_foreach_instr_in_block_from_rev(block, v, from) \
- list_for_each_entry_from_rev(bi_instruction, v, from, &block->instructions, link)
+ list_for_each_entry_from_rev(bi_instruction, v, from, &(block)->base.instructions, link)
#define bi_foreach_clause_in_block(block, v) \
- list_for_each_entry(bi_clause, v, &block->clauses, link)
+ list_for_each_entry(bi_clause, v, &(block)->clauses, link)
+
+#define bi_foreach_clause_in_block_from(block, v, from) \
+ list_for_each_entry_from(bi_clause, v, from, &(block)->clauses, link)
+
+#define bi_foreach_clause_in_block_from_rev(block, v, from) \
+ list_for_each_entry_from_rev(bi_clause, v, from, &(block)->clauses, link)
#define bi_foreach_instr_global(ctx, v) \
bi_foreach_block(ctx, v_block) \
- bi_foreach_instr_in_block(v_block, v)
+ bi_foreach_instr_in_block((bi_block *) v_block, v)
#define bi_foreach_instr_global_safe(ctx, v) \
bi_foreach_block(ctx, v_block) \
- bi_foreach_instr_in_block_safe(v_block, v)
-
-#define bi_foreach_successor(blk, v) \
- bi_block *v; \
- bi_block **_v; \
- for (_v = &blk->successors[0], \
- v = *_v; \
- v != NULL && _v < &blk->successors[2]; \
- _v++, v = *_v) \
+ bi_foreach_instr_in_block_safe((bi_block *) v_block, v)
/* Based on set_foreach, expanded with automatic type casts */
#define bi_foreach_predecessor(blk, v) \
struct set_entry *_entry_##v; \
bi_block *v; \
- for (_entry_##v = _mesa_set_next_entry(blk->predecessors, NULL), \
+ for (_entry_##v = _mesa_set_next_entry(blk->base.predecessors, NULL), \
v = (bi_block *) (_entry_##v ? _entry_##v->key : NULL); \
_entry_##v != NULL; \
- _entry_##v = _mesa_set_next_entry(blk->predecessors, _entry_##v), \
+ _entry_##v = _mesa_set_next_entry(blk->base.predecessors, _entry_##v), \
v = (bi_block *) (_entry_##v ? _entry_##v->key : NULL))
#define bi_foreach_src(ins, v) \
for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
+static inline bi_instruction *
+bi_prev_op(bi_instruction *ins)
+{
+ return list_last_entry(&(ins->link), bi_instruction, link);
+}
+
+static inline bi_instruction *
+bi_next_op(bi_instruction *ins)
+{
+ return list_first_entry(&(ins->link), bi_instruction, link);
+}
+
+static inline pan_block *
+pan_next_block(pan_block *block)
+{
+ return list_first_entry(&(block->link), pan_block, link);
+}
+
+/* Special functions */
+
+void bi_emit_fexp2(bi_context *ctx, nir_alu_instr *instr);
+void bi_emit_flog2(bi_context *ctx, nir_alu_instr *instr);
+
/* BIR manipulation */
bool bi_has_outmod(bi_instruction *ins);
bool bi_has_source_mods(bi_instruction *ins);
bool bi_is_src_swizzled(bi_instruction *ins, unsigned s);
+bool bi_has_arg(bi_instruction *ins, unsigned arg);
+uint16_t bi_from_bytemask(uint16_t bytemask, unsigned bytes);
+unsigned bi_get_component_count(bi_instruction *ins, signed s);
+uint16_t bi_bytemask_of_read_components(bi_instruction *ins, unsigned node);
+uint64_t bi_get_immediate(bi_instruction *ins, unsigned index);
+bool bi_writes_component(bi_instruction *ins, unsigned comp);
+unsigned bi_writemask(bi_instruction *ins);
+
+/* BIR passes */
+
+void bi_lower_combine(bi_context *ctx, bi_block *block);
+bool bi_opt_dead_code_eliminate(bi_context *ctx, bi_block *block);
+void bi_schedule(bi_context *ctx);
+void bi_register_allocate(bi_context *ctx);
+
+/* Liveness */
+
+void bi_compute_liveness(bi_context *ctx);
+void bi_liveness_ins_update(uint16_t *live, bi_instruction *ins, unsigned max);
+void bi_invalidate_liveness(bi_context *ctx);
+bool bi_is_live_after(bi_context *ctx, bi_block *block, bi_instruction *start, int src);
+
+/* Layout */
+
+bool bi_can_insert_bundle(bi_clause *clause, bool constant);
+unsigned bi_clause_quadwords(bi_clause *clause);
+signed bi_block_offset(bi_context *ctx, bi_clause *start, bi_block *target);
+
+/* Code emit */
+
+void bi_pack(bi_context *ctx, struct util_dynarray *emission);
#endif