#include <string.h>
#include "util/macros.h"
-#include "glsl/nir/nir.h"
+#include "compiler/nir/nir.h"
#include "util/list.h"
#include "util/u_math.h"
#include "vc4_screen.h"
+#include "vc4_qpu_defines.h"
+#include "vc4_qpu.h"
+#include "kernel/vc4_packet.h"
#include "pipe/p_state.h"
struct nir_builder;
QFILE_VARY,
QFILE_UNIF,
QFILE_VPM,
+ QFILE_TLB_COLOR_WRITE,
+ QFILE_TLB_COLOR_WRITE_MS,
+ QFILE_TLB_Z_WRITE,
+ QFILE_TLB_STENCIL_SETUP,
+
+ /* Payload registers that aren't in the physical register file, so we
+ * can just use the corresponding qpu_reg at qpu_emit time.
+ */
+ QFILE_FRAG_X,
+ QFILE_FRAG_Y,
+ QFILE_FRAG_REV_FLAG,
+
+ /**
+ * Stores an immediate value in the index field that will be used
+ * directly by qpu_load_imm().
+ */
+ QFILE_LOAD_IMM,
/**
* Stores an immediate value in the index field that can be turned
struct qreg {
enum qfile file;
uint32_t index;
+ int pack;
};
+static inline struct qreg qir_reg(enum qfile file, uint32_t index)
+{
+ return (struct qreg){file, index};
+}
+
enum qop {
QOP_UNDEF,
QOP_MOV,
+ QOP_FMOV,
+ QOP_MMOV,
QOP_FADD,
QOP_FSUB,
QOP_FMUL,
+ QOP_V8MULD,
+ QOP_V8MIN,
+ QOP_V8MAX,
+ QOP_V8ADDS,
+ QOP_V8SUBS,
QOP_MUL24,
QOP_FMIN,
QOP_FMAX,
QOP_XOR,
QOP_NOT,
- /* Note: Orderings of these compares must be the same as in
- * qpu_defines.h. Selects the src[0] if the ns flag bit is set,
- * otherwise 0. */
- QOP_SEL_X_0_ZS,
- QOP_SEL_X_0_ZC,
- QOP_SEL_X_0_NS,
- QOP_SEL_X_0_NC,
- /* Selects the src[0] if the ns flag bit is set, otherwise src[1]. */
- QOP_SEL_X_Y_ZS,
- QOP_SEL_X_Y_ZC,
- QOP_SEL_X_Y_NS,
- QOP_SEL_X_Y_NC,
-
QOP_FTOI,
QOP_ITOF,
QOP_RCP,
QOP_LOG2,
QOP_VW_SETUP,
QOP_VR_SETUP,
- QOP_PACK_SCALED,
- QOP_PACK_8888_F,
- QOP_PACK_8A_F,
- QOP_PACK_8B_F,
- QOP_PACK_8C_F,
- QOP_PACK_8D_F,
- QOP_TLB_DISCARD_SETUP,
- QOP_TLB_STENCIL_SETUP,
- QOP_TLB_Z_WRITE,
- QOP_TLB_COLOR_WRITE,
QOP_TLB_COLOR_READ,
+ QOP_MS_MASK,
QOP_VARY_ADD_C,
- QOP_FRAG_X,
- QOP_FRAG_Y,
QOP_FRAG_Z,
QOP_FRAG_W,
- QOP_FRAG_REV_FLAG,
-
- QOP_UNPACK_8A_F,
- QOP_UNPACK_8B_F,
- QOP_UNPACK_8C_F,
- QOP_UNPACK_8D_F,
- QOP_UNPACK_16A_F,
- QOP_UNPACK_16B_F,
-
- QOP_UNPACK_8A_I,
- QOP_UNPACK_8B_I,
- QOP_UNPACK_8C_I,
- QOP_UNPACK_8D_I,
- QOP_UNPACK_16A_I,
- QOP_UNPACK_16B_I,
/** Texture x coordinate parameter write */
QOP_TEX_S,
* the destination
*/
QOP_TEX_RESULT,
+
+ QOP_LOAD_IMM,
+
+ /* Jumps to block->successor[0] if the qinst->cond (as a
+ * QPU_COND_BRANCH_*) passes, or block->successor[1] if not. Note
+ * that block->successor[1] may be unset if the condition is ALWAYS.
+ */
+ QOP_BRANCH,
};
struct queued_qpu_inst {
struct qreg dst;
struct qreg *src;
bool sf;
+ uint8_t cond;
};
enum qstage {
/** A reference to a texture config parameter 2 cubemap stride uniform */
QUNIFORM_TEXTURE_CONFIG_P2,
+ QUNIFORM_TEXTURE_MSAA_ADDR,
+
QUNIFORM_UBO_ADDR,
QUNIFORM_TEXRECT_SCALE_X,
QUNIFORM_BLEND_CONST_COLOR_Y,
QUNIFORM_BLEND_CONST_COLOR_Z,
QUNIFORM_BLEND_CONST_COLOR_W,
+ QUNIFORM_BLEND_CONST_COLOR_RGBA,
+ QUNIFORM_BLEND_CONST_COLOR_AAAA,
QUNIFORM_STENCIL,
QUNIFORM_ALPHA_REF,
+ QUNIFORM_SAMPLE_MASK,
};
-struct vc4_varying_semantic {
- uint8_t semantic;
- uint8_t index;
+struct vc4_varying_slot {
+ uint8_t slot;
uint8_t swizzle;
};
struct vc4_uncompiled_shader *shader_state;
struct {
enum pipe_format format;
- unsigned compare_mode:1;
- unsigned compare_func:3;
- unsigned wrap_s:3;
- unsigned wrap_t:3;
uint8_t swizzle[4];
+ union {
+ struct {
+ unsigned compare_mode:1;
+ unsigned compare_func:3;
+ unsigned wrap_s:3;
+ unsigned wrap_t:3;
+ };
+ struct {
+ uint16_t msaa_width, msaa_height;
+ };
+ };
} tex[VC4_MAX_TEXTURE_SAMPLERS];
uint8_t ucp_enables;
};
bool alpha_test;
bool point_coord_upper_left;
bool light_twoside;
+ bool msaa;
+ bool sample_coverage;
+ bool sample_alpha_to_coverage;
+ bool sample_alpha_to_one;
uint8_t alpha_test_func;
uint8_t logicop_func;
uint32_t point_sprite_mask;
enum pipe_format attr_formats[8];
bool is_coord;
bool per_vertex_point_size;
+ bool clamp_color;
+};
+
+/** A basic block of QIR intructions. */
+struct qblock {
+ struct list_head link;
+
+ struct list_head instructions;
+
+ struct set *predecessors;
+ struct qblock *successors[2];
+
+ int index;
+
+ /** @{ used by vc4_qir_live_variables.c */
+ BITSET_WORD *def;
+ BITSET_WORD *use;
+ BITSET_WORD *live_in;
+ BITSET_WORD *live_out;
+ int start_ip, end_ip;
+ /** @} */
};
struct vc4_compile {
*/
struct qreg *inputs;
struct qreg *outputs;
+ bool msaa_per_sample_output;
+ struct qreg color_reads[VC4_MAX_SAMPLES];
+ struct qreg sample_colors[VC4_MAX_SAMPLES];
uint32_t inputs_array_size;
uint32_t outputs_array_size;
uint32_t uniforms_array_size;
struct qreg line_x, point_x, point_y;
struct qreg discard;
+ struct qreg payload_FRAG_Z;
+ struct qreg payload_FRAG_W;
uint8_t vattr_sizes[8];
/**
- * Array of the TGSI semantics of all FS QFILE_VARY reads.
+ * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
*
* This includes those that aren't part of the VPM varyings, like
* point/line coordinates.
*/
- struct vc4_varying_semantic *input_semantics;
- uint32_t num_input_semantics;
- uint32_t input_semantics_array_size;
+ struct vc4_varying_slot *input_slots;
+ uint32_t num_input_slots;
+ uint32_t input_slots_array_size;
/**
- * An entry per outputs[] in the VS indicating what the semantic of
- * the output is. Used to emit from the VS in the order that the FS
- * needs.
+ * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
+ * of the output is. Used to emit from the VS in the order that the
+ * FS needs.
*/
- struct vc4_varying_semantic *output_semantics;
+ struct vc4_varying_slot *output_slots;
struct pipe_shader_state *shader_state;
struct vc4_key *key;
struct vc4_fs_key *fs_key;
struct vc4_vs_key *vs_key;
+ /* Live ranges of temps. */
+ int *temp_start, *temp_end;
+
uint32_t *uniform_data;
enum quniform_contents *uniform_contents;
uint32_t uniform_array_size;
uint32_t num_outputs;
uint32_t num_texture_samples;
uint32_t output_position_index;
- uint32_t output_clipvertex_index;
uint32_t output_color_index;
uint32_t output_point_size_index;
+ uint32_t output_sample_mask_index;
struct qreg undef;
enum qstage stage;
uint32_t num_temps;
- struct list_head instructions;
- uint32_t immediates[1024];
+
+ struct list_head blocks;
+ int next_block_index;
+ struct qblock *cur_block;
struct list_head qpu_inst_list;
uint64_t *qpu_insts;
*/
#define VC4_NIR_TLB_COLOR_READ_INPUT 2000000000
+#define VC4_NIR_MS_MASK_OUTPUT 2000000000
+
/* Special offset for nir_load_uniform values to get a QUNIFORM_*
* state-dependent value.
*/
-#define VC4_NIR_STATE_UNIFORM_OFFSET 2000000000
+#define VC4_NIR_STATE_UNIFORM_OFFSET 1000000000
struct vc4_compile *qir_compile_init(void);
void qir_compile_destroy(struct vc4_compile *c);
+struct qblock *qir_new_block(struct vc4_compile *c);
+void qir_set_emit_block(struct vc4_compile *c, struct qblock *block);
+void qir_link_blocks(struct qblock *predecessor, struct qblock *successor);
+struct qblock *qir_entry_block(struct vc4_compile *c);
+struct qblock *qir_exit_block(struct vc4_compile *c);
struct qinst *qir_inst(enum qop op, struct qreg dst,
struct qreg src0, struct qreg src1);
struct qinst *qir_inst4(enum qop op, struct qreg dst,
struct qreg qir_uniform(struct vc4_compile *c,
enum quniform_contents contents,
uint32_t data);
+void qir_schedule_instructions(struct vc4_compile *c);
void qir_reorder_uniforms(struct vc4_compile *c);
-void qir_emit(struct vc4_compile *c, struct qinst *inst);
+
+struct qreg qir_emit_def(struct vc4_compile *c, struct qinst *inst);
+struct qinst *qir_emit_nondef(struct vc4_compile *c, struct qinst *inst);
+
struct qreg qir_get_temp(struct vc4_compile *c);
+void qir_calculate_live_intervals(struct vc4_compile *c);
int qir_get_op_nsrc(enum qop qop);
bool qir_reg_equals(struct qreg a, struct qreg b);
bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst);
-bool qir_is_multi_instruction(struct qinst *inst);
+bool qir_is_mul(struct qinst *inst);
+bool qir_is_raw_mov(struct qinst *inst);
bool qir_is_tex(struct qinst *inst);
+bool qir_is_float_input(struct qinst *inst);
bool qir_depends_on_flags(struct qinst *inst);
bool qir_writes_r4(struct qinst *inst);
-bool qir_src_needs_a_file(struct qinst *inst);
struct qreg qir_follow_movs(struct vc4_compile *c, struct qreg reg);
+uint8_t qir_channels_written(struct qinst *inst);
void qir_dump(struct vc4_compile *c);
void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
const char *qir_get_stage_name(enum qstage stage);
+void qir_validate(struct vc4_compile *c);
+
void qir_optimize(struct vc4_compile *c);
bool qir_opt_algebraic(struct vc4_compile *c);
bool qir_opt_constant_folding(struct vc4_compile *c);
bool qir_opt_copy_propagation(struct vc4_compile *c);
-bool qir_opt_cse(struct vc4_compile *c);
bool qir_opt_dead_code(struct vc4_compile *c);
+bool qir_opt_peephole_sf(struct vc4_compile *c);
bool qir_opt_small_immediates(struct vc4_compile *c);
-bool qir_opt_vpm_writes(struct vc4_compile *c);
-void vc4_nir_lower_blend(struct vc4_compile *c);
-void vc4_nir_lower_io(struct vc4_compile *c);
+bool qir_opt_vpm(struct vc4_compile *c);
+void vc4_nir_lower_blend(nir_shader *s, struct vc4_compile *c);
+void vc4_nir_lower_io(nir_shader *s, struct vc4_compile *c);
nir_ssa_def *vc4_nir_get_state_uniform(struct nir_builder *b,
enum quniform_contents contents);
nir_ssa_def *vc4_nir_get_swizzled_channel(struct nir_builder *b,
nir_ssa_def **srcs, int swiz);
+void vc4_nir_lower_txf_ms(nir_shader *s, struct vc4_compile *c);
void qir_lower_uniforms(struct vc4_compile *c);
-void qpu_schedule_instructions(struct vc4_compile *c);
+uint32_t qpu_schedule_instructions(struct vc4_compile *c);
void qir_SF(struct vc4_compile *c, struct qreg src);
static inline struct qreg \
qir_##name(struct vc4_compile *c) \
{ \
- struct qreg t = qir_get_temp(c); \
- qir_emit(c, qir_inst(QOP_##name, t, c->undef, c->undef)); \
- return t; \
+ return qir_emit_def(c, qir_inst(QOP_##name, c->undef, \
+ c->undef, c->undef)); \
+} \
+static inline struct qinst * \
+qir_##name##_dest(struct vc4_compile *c, struct qreg dest) \
+{ \
+ return qir_emit_nondef(c, qir_inst(QOP_##name, dest, \
+ c->undef, c->undef)); \
}
#define QIR_ALU1(name) \
static inline struct qreg \
qir_##name(struct vc4_compile *c, struct qreg a) \
{ \
- struct qreg t = qir_get_temp(c); \
- qir_emit(c, qir_inst(QOP_##name, t, a, c->undef)); \
- return t; \
+ return qir_emit_def(c, qir_inst(QOP_##name, c->undef, \
+ a, c->undef)); \
+} \
+static inline struct qinst * \
+qir_##name##_dest(struct vc4_compile *c, struct qreg dest, \
+ struct qreg a) \
+{ \
+ return qir_emit_nondef(c, qir_inst(QOP_##name, dest, a, \
+ c->undef)); \
}
#define QIR_ALU2(name) \
static inline struct qreg \
qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b) \
{ \
- struct qreg t = qir_get_temp(c); \
- qir_emit(c, qir_inst(QOP_##name, t, a, b)); \
- return t; \
+ return qir_emit_def(c, qir_inst(QOP_##name, c->undef, a, b)); \
+} \
+static inline struct qinst * \
+qir_##name##_dest(struct vc4_compile *c, struct qreg dest, \
+ struct qreg a, struct qreg b) \
+{ \
+ return qir_emit_nondef(c, qir_inst(QOP_##name, dest, a, b)); \
}
#define QIR_NODST_1(name) \
-static inline void \
+static inline struct qinst * \
qir_##name(struct vc4_compile *c, struct qreg a) \
{ \
- qir_emit(c, qir_inst(QOP_##name, c->undef, a, c->undef)); \
+ return qir_emit_nondef(c, qir_inst(QOP_##name, c->undef, \
+ a, c->undef)); \
}
#define QIR_NODST_2(name) \
-static inline void \
+static inline struct qinst * \
qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b) \
{ \
- qir_emit(c, qir_inst(QOP_##name, c->undef, a, b)); \
+ return qir_emit_nondef(c, qir_inst(QOP_##name, c->undef, \
+ a, b)); \
+}
+
+#define QIR_PAYLOAD(name) \
+static inline struct qreg \
+qir_##name(struct vc4_compile *c) \
+{ \
+ struct qreg *payload = &c->payload_##name; \
+ if (payload->file != QFILE_NULL) \
+ return *payload; \
+ *payload = qir_get_temp(c); \
+ struct qinst *inst = qir_inst(QOP_##name, *payload, \
+ c->undef, c->undef); \
+ struct qblock *entry = qir_entry_block(c); \
+ list_add(&inst->link, &entry->instructions); \
+ c->defs[payload->index] = inst; \
+ return *payload; \
}
QIR_ALU1(MOV)
+QIR_ALU1(FMOV)
+QIR_ALU1(MMOV)
QIR_ALU2(FADD)
QIR_ALU2(FSUB)
QIR_ALU2(FMUL)
+QIR_ALU2(V8MULD)
+QIR_ALU2(V8MIN)
+QIR_ALU2(V8MAX)
+QIR_ALU2(V8ADDS)
+QIR_ALU2(V8SUBS)
QIR_ALU2(MUL24)
-QIR_ALU1(SEL_X_0_ZS)
-QIR_ALU1(SEL_X_0_ZC)
-QIR_ALU1(SEL_X_0_NS)
-QIR_ALU1(SEL_X_0_NC)
-QIR_ALU2(SEL_X_Y_ZS)
-QIR_ALU2(SEL_X_Y_ZC)
-QIR_ALU2(SEL_X_Y_NS)
-QIR_ALU2(SEL_X_Y_NC)
QIR_ALU2(FMIN)
QIR_ALU2(FMAX)
QIR_ALU2(FMINABS)
QIR_ALU1(RSQ)
QIR_ALU1(EXP2)
QIR_ALU1(LOG2)
-QIR_ALU2(PACK_SCALED)
-QIR_ALU1(PACK_8888_F)
-QIR_ALU2(PACK_8A_F)
-QIR_ALU2(PACK_8B_F)
-QIR_ALU2(PACK_8C_F)
-QIR_ALU2(PACK_8D_F)
QIR_ALU1(VARY_ADD_C)
QIR_NODST_2(TEX_S)
QIR_NODST_2(TEX_T)
QIR_NODST_2(TEX_R)
QIR_NODST_2(TEX_B)
QIR_NODST_2(TEX_DIRECT)
-QIR_ALU0(FRAG_X)
-QIR_ALU0(FRAG_Y)
-QIR_ALU0(FRAG_Z)
-QIR_ALU0(FRAG_W)
-QIR_ALU0(FRAG_REV_FLAG)
+QIR_PAYLOAD(FRAG_Z)
+QIR_PAYLOAD(FRAG_W)
QIR_ALU0(TEX_RESULT)
QIR_ALU0(TLB_COLOR_READ)
-QIR_NODST_1(TLB_COLOR_WRITE)
-QIR_NODST_1(TLB_Z_WRITE)
-QIR_NODST_1(TLB_DISCARD_SETUP)
-QIR_NODST_1(TLB_STENCIL_SETUP)
+QIR_NODST_1(MS_MASK)
static inline struct qreg
-qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
+qir_SEL(struct vc4_compile *c, uint8_t cond, struct qreg src0, struct qreg src1)
{
struct qreg t = qir_get_temp(c);
- qir_emit(c, qir_inst(QOP_UNPACK_8A_F + i, t, src, c->undef));
+ struct qinst *a = qir_MOV_dest(c, t, src0);
+ struct qinst *b = qir_MOV_dest(c, t, src1);
+ a->cond = cond;
+ b->cond = qpu_cond_complement(cond);
+ return t;
+}
+
+static inline struct qreg
+qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
+{
+ struct qreg t = qir_FMOV(c, src);
+ c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
return t;
}
static inline struct qreg
qir_UNPACK_8_I(struct vc4_compile *c, struct qreg src, int i)
{
- struct qreg t = qir_get_temp(c);
- qir_emit(c, qir_inst(QOP_UNPACK_8A_I + i, t, src, c->undef));
+ struct qreg t = qir_MOV(c, src);
+ c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
return t;
}
static inline struct qreg
qir_UNPACK_16_F(struct vc4_compile *c, struct qreg src, int i)
{
- struct qreg t = qir_get_temp(c);
- qir_emit(c, qir_inst(QOP_UNPACK_16A_F + i, t, src, c->undef));
+ struct qreg t = qir_FMOV(c, src);
+ c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
return t;
}
static inline struct qreg
qir_UNPACK_16_I(struct vc4_compile *c, struct qreg src, int i)
{
- struct qreg t = qir_get_temp(c);
- qir_emit(c, qir_inst(QOP_UNPACK_16A_I + i, t, src, c->undef));
+ struct qreg t = qir_MOV(c, src);
+ c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
return t;
}
+static inline void
+qir_PACK_8_F(struct vc4_compile *c, struct qreg dest, struct qreg val, int chan)
+{
+ assert(!dest.pack);
+ dest.pack = QPU_PACK_MUL_8A + chan;
+ qir_emit_nondef(c, qir_inst(QOP_MMOV, dest, val, c->undef));
+}
+
static inline struct qreg
-qir_PACK_8_F(struct vc4_compile *c, struct qreg rest, struct qreg val, int chan)
+qir_PACK_8888_F(struct vc4_compile *c, struct qreg val)
{
- struct qreg t = qir_get_temp(c);
- qir_emit(c, qir_inst(QOP_PACK_8A_F + chan, t, rest, val));
- return t;
+ struct qreg dest = qir_MMOV(c, val);
+ c->defs[dest.index]->dst.pack = QPU_PACK_MUL_8888;
+ return dest;
}
static inline struct qreg
static inline void
qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
{
- static const struct qreg vpm = { QFILE_VPM, 0 };
- qir_emit(c, qir_inst(QOP_MOV, vpm, val, c->undef));
+ qir_MOV_dest(c, qir_reg(QFILE_VPM, 0), val);
+}
+
+static inline struct qreg
+qir_LOAD_IMM(struct vc4_compile *c, uint32_t val)
+{
+ return qir_emit_def(c, qir_inst(QOP_LOAD_IMM, c->undef,
+ qir_reg(QFILE_LOAD_IMM, val), c->undef));
+}
+
+static inline struct qinst *
+qir_BRANCH(struct vc4_compile *c, uint8_t cond)
+{
+ struct qinst *inst = qir_inst(QOP_BRANCH, c->undef, c->undef, c->undef);
+ inst->cond = cond;
+ qir_emit_nondef(c, inst);
+ return inst;
}
+#define qir_for_each_block(block, c) \
+ list_for_each_entry(struct qblock, block, &c->blocks, link)
+
+#define qir_for_each_block_rev(block, c) \
+ list_for_each_entry_rev(struct qblock, block, &c->blocks, link)
+
+/* Loop over the non-NULL members of the successors array. */
+#define qir_for_each_successor(succ, block) \
+ for (struct qblock *succ = block->successors[0]; \
+ succ != NULL; \
+ succ = (succ == block->successors[1] ? NULL : \
+ block->successors[1]))
+
+#define qir_for_each_inst(inst, block) \
+ list_for_each_entry(struct qinst, inst, &block->instructions, link)
+
+#define qir_for_each_inst_rev(inst, block) \
+ list_for_each_entry_rev(struct qinst, inst, &block->instructions, link)
+
+#define qir_for_each_inst_safe(inst, block) \
+ list_for_each_entry_safe(struct qinst, inst, &block->instructions, link)
+
+#define qir_for_each_inst_inorder(inst, c) \
+ qir_for_each_block(_block, c) \
+ qir_for_each_inst(inst, _block)
+
#endif /* VC4_QIR_H */