unsigned unk3 : 1; // part of clauseType?
unsigned next_clause_type: 4;
unsigned unk4 : 1; // part of nextClauseType?
+} __attribute__((packed));
+
+enum bifrost_packed_src {
+ BIFROST_SRC_PORT0 = 0,
+ BIFROST_SRC_PORT1 = 1,
+ BIFROST_SRC_PORT3 = 2,
+ BIFROST_SRC_STAGE = 3,
+ BIFROST_SRC_CONST_LO = 4,
+ BIFROST_SRC_CONST_HI = 5,
+ BIFROST_SRC_PASS_FMA = 6,
+ BIFROST_SRC_PASS_ADD = 7,
};
struct bifrost_fma_inst {
unsigned src0 : 3;
unsigned op : 20;
-};
+} __attribute__((packed));
struct bifrost_add_inst {
unsigned src0 : 3;
unsigned op : 17;
+} __attribute__((packed));
+
+enum bifrost_outmod {
+ BIFROST_NONE = 0x0,
+ BIFROST_POS = 0x1,
+ BIFROST_SAT_SIGNED = 0x2,
+ BIFROST_SAT = 0x3,
+};
+
+enum bifrost_roundmode {
+ BIFROST_RTE = 0x0, /* round to even */
+ BIFROST_RTP = 0x1, /* round to positive */
+ BIFROST_RTN = 0x2, /* round to negative */
+ BIFROST_RTZ = 0x3 /* round to zero */
+};
+
+/* NONE: Same as fmax() and fmin() -- return the other
+ * number if any number is NaN. Also always return +0 if
+ * one argument is +0 and the other is -0.
+ *
+ * NAN_WINS: Instead of never returning a NaN, always return
+ * one. The "greater"/"lesser" NaN is always returned, first
+ * by checking the sign and then the mantissa bits.
+ *
+ * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
+ * For min, implement src0 < src1 ? src0 : src1. This
+ * includes handling NaN's and signedness of 0 differently
+ * from above, since +0 and -0 compare equal and comparisons
+ * always return false for NaN's. As a result, this mode is
+ * *not* commutative.
+ *
+ * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
+ * For min, implement src0 > src1 ? src1 : src0
+ */
+
+
+enum bifrost_minmax_mode {
+ BIFROST_MINMAX_NONE = 0x0,
+ BIFROST_NAN_WINS = 0x1,
+ BIFROST_SRC1_WINS = 0x2,
+ BIFROST_SRC0_WINS = 0x3,
};
+#define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
+
+struct bifrost_fma_add {
+ unsigned src0 : 3;
+ unsigned src1 : 3;
+ unsigned src1_abs : 1;
+ unsigned src0_neg : 1;
+ unsigned src1_neg : 1;
+ unsigned unk : 3;
+ unsigned src0_abs : 1;
+ enum bifrost_outmod outmod : 2;
+ enum bifrost_roundmode roundmode : 2;
+ unsigned op : 6;
+} __attribute__((packed));
+
+#define BIFROST_FMA_OP_FMA (0x00)
+
+struct bifrost_fma_fma {
+ unsigned src0 : 3;
+ unsigned src1 : 3;
+ unsigned src2 : 3;
+ unsigned src_expand : 3;
+ unsigned src0_abs : 1;
+ unsigned unk : 4;
+ unsigned src0_neg : 1;
+ unsigned src2_neg : 1;
+ unsigned src1_abs : 1;
+ unsigned src2_abs : 1;
+ unsigned op : 5;
+} __attribute__((packed));
+
enum bifrost_csel_cond {
BIFROST_FEQ_F = 0x0,
BIFROST_FGT_F = 0x1,
unsigned src3 : 3;
enum bifrost_csel_cond cond : 3;
unsigned op : 8;
-};
+} __attribute__((packed));
struct bifrost_shift_fma {
unsigned src0 : 3;
unsigned src1 : 3;
unsigned src2 : 3;
- unsigned half : 3; /* 000 for i32, 111 for v2i16 */
+ unsigned half : 3; /* 000 for i32, 100 for i8, 111 for v2i16 */
unsigned unk : 1; /* always set? */
unsigned invert_1 : 1; /* Inverts sources to combining op */
/* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
unsigned invert_2 : 1;
unsigned op : 8;
+} __attribute__((packed));
+
+struct bifrost_shift_add {
+ unsigned src0 : 3;
+ unsigned src1 : 3;
+ unsigned src2 : 3;
+ unsigned zero : 2;
+
+ unsigned invert_1 : 1;
+ unsigned invert_2 : 1;
+
+ unsigned op : 7;
+} __attribute__((packed));
+
+enum bifrost_ldst_type {
+ BIFROST_LDST_F16 = 0,
+ BIFROST_LDST_F32 = 1,
+ BIFROST_LDST_I32 = 2,
+ BIFROST_LDST_U32 = 3
+};
+
+struct bifrost_ld_var_addr {
+ unsigned src0 : 3;
+ unsigned src1 : 3;
+ unsigned location : 5;
+ enum bifrost_ldst_type type : 2;
+ unsigned op : 7;
+} __attribute__((packed));
+
+struct bifrost_ld_attr {
+ unsigned src0 : 3;
+ unsigned src1 : 3;
+ unsigned location : 5;
+ unsigned channels : 2; /* MALI_POSITIVE */
+ enum bifrost_ldst_type type : 2;
+ unsigned op : 5;
+} __attribute__((packed));
+
+enum bifrost_interp_mode {
+ BIFROST_INTERP_PER_FRAG = 0x0,
+ BIFROST_INTERP_CENTROID = 0x1,
+ BIFROST_INTERP_DEFAULT = 0x2,
+ BIFROST_INTERP_EXPLICIT = 0x3
+};
+
+#define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
+#define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
+
+struct bifrost_ld_var {
+ unsigned src0 : 3;
+
+ /* If top two bits set, indirect with src in bottom three */
+ unsigned addr : 5;
+
+ unsigned channels : 2; /* MALI_POSITIVE */
+ enum bifrost_interp_mode interp_mode : 2;
+ unsigned reuse : 1;
+ unsigned flat : 1;
+ unsigned op : 6;
+} __attribute__((packed));
+
+struct bifrost_tex_ctrl {
+ unsigned sampler_index : 4; // also used to signal indirects
+ unsigned tex_index : 7;
+ bool no_merge_index : 1; // whether to merge (direct) sampler & texture indices
+ bool filter : 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
+ unsigned unk0 : 2;
+ bool texel_offset : 1; // *Offset()
+ bool is_shadow : 1;
+ bool is_array : 1;
+ unsigned tex_type : 2; // 2D, 3D, Cube, Buffer
+ bool compute_lod : 1; // 0 for *Lod()
+ bool not_supply_lod : 1; // 0 for *Lod() or when a bias is applied
+ bool calc_gradients : 1; // 0 for *Grad()
+ unsigned unk1 : 1;
+ unsigned result_type : 4; // integer, unsigned, float TODO: why is this 4 bits?
+ unsigned unk2 : 4;
+} __attribute__((packed));
+
+struct bifrost_dual_tex_ctrl {
+ unsigned sampler_index0 : 2;
+ unsigned unk0 : 2;
+ unsigned tex_index0 : 2;
+ unsigned sampler_index1 : 2;
+ unsigned tex_index1 : 2;
+ unsigned unk1 : 22;
+} __attribute__((packed));
+
+enum branch_bit_size {
+ BR_SIZE_32 = 0,
+ BR_SIZE_16XX = 1,
+ BR_SIZE_16YY = 2,
+ // For the above combinations of bitsize and location, an extra bit is
+ // encoded via comparing the sources. The only possible source of ambiguity
+ // would be if the sources were the same, but then the branch condition
+ // would be always true or always false anyways, so we can ignore it. But
+ // this no longer works when comparing the y component to the x component,
+ // since it's valid to compare the y component of a source against its own
+ // x component. Instead, the extra bit is encoded via an extra bitsize.
+ BR_SIZE_16YX0 = 3,
+ BR_SIZE_16YX1 = 4,
+ BR_SIZE_32_AND_16X = 5,
+ BR_SIZE_32_AND_16Y = 6,
+ // Used for comparisons with zero and always-true, see below. I think this
+ // only works for integer comparisons.
+ BR_SIZE_ZERO = 7,
+};
+
+enum bifrost_reg_write_unit {
+ REG_WRITE_NONE = 0, // don't write
+ REG_WRITE_TWO, // write using reg2
+ REG_WRITE_THREE, // write using reg3
+};
+
+struct bifrost_regs {
+ unsigned uniform_const : 8;
+ unsigned reg2 : 6;
+ unsigned reg3 : 6;
+ unsigned reg0 : 5;
+ unsigned reg1 : 6;
+ unsigned ctrl : 4;
+} __attribute__((packed));
+
+enum bifrost_branch_cond {
+ BR_COND_LT = 0,
+ BR_COND_LE = 1,
+ BR_COND_GE = 2,
+ BR_COND_GT = 3,
+ // Equal vs. not-equal determined by src0/src1 comparison
+ BR_COND_EQ = 4,
+ // floating-point comparisons
+ // Becomes UNE when you flip the arguments
+ BR_COND_OEQ = 5,
+ // TODO what happens when you flip the arguments?
+ BR_COND_OGT = 6,
+ BR_COND_OLT = 7,
+};
+
+enum bifrost_branch_code {
+ BR_ALWAYS = 63,
+};
+
+struct bifrost_branch {
+ unsigned src0 : 3;
+
+ /* For BR_SIZE_ZERO, upper two bits become ctrl */
+ unsigned src1 : 3;
+
+ /* Offset source -- always uniform/const but
+ * theoretically could support indirect jumps? */
+ unsigned src2 : 3;
+
+ enum bifrost_branch_cond cond : 3;
+ enum branch_bit_size size : 3;
+
+ unsigned op : 5;
+};
+
+/* Clause packing */
+
+#define BIFROST_FMA_NOP (0x701960)
+#define BIFROST_ADD_NOP (0x3D960)
+
+struct bifrost_fmt1 {
+ unsigned ins_0 : 3;
+ unsigned tag : 5;
+ uint64_t ins_1 : 64;
+ unsigned ins_2 : 11;
+ uint64_t header : 45;
+} __attribute__((packed));
+
+#define BIFROST_FMT1_INSTRUCTIONS 0b00101
+#define BIFROST_FMT1_FINAL 0b01001
+#define BIFROST_FMT1_CONSTANTS 0b00001
+
+enum bifrost_reg_control {
+ BIFROST_WRITE_FMA_P2 = 1,
+ BIFROST_WRITE_FMA_P2_READ_P3 = 2,
+ BIFROST_WRITE_FMA_P2_READ_P3_ALT = 3,
+ BIFROST_READ_P3 = 4,
+ BIFROST_WRITE_ADD_P2 = 5,
+ BIFROST_WRITE_ADD_P2_READ_P3 = 6,
+ BIFROST_WRITE_ADD_P2_FMA_P3 = 7,
+
+ BIFROST_FIRST_NONE = 8,
+ BIFROST_FIRST_WRITE_FMA_P2 = 9,
+ BIFROST_REG_NONE = 11,
+ BIFROST_FIRST_READ_P3 = 12,
+ BIFROST_FIRST_WRITE_ADD_P2 = 13,
+ BIFROST_FIRST_WRITE_ADD_P2_READ_P3 = 14,
+ BIFROST_FIRST_WRITE_ADD_P2_FMA_P3 = 15
};
#endif