#define PACKED __attribute__((__packed__))
#include <stdint.h>
+#include <stdbool.h>
#include <assert.h>
+/* size of largest OPC field of all the instruction categories: */
+#define NOPC_BITS 6
+
+#define _OPC(cat, opc) (((cat) << NOPC_BITS) | opc)
+
typedef enum {
/* category 0: */
- OPC_NOP = 0,
- OPC_BR = 1,
- OPC_JUMP = 2,
- OPC_CALL = 3,
- OPC_RET = 4,
- OPC_KILL = 5,
- OPC_END = 6,
- OPC_EMIT = 7,
- OPC_CUT = 8,
- OPC_CHMASK = 9,
- OPC_CHSH = 10,
- OPC_FLOW_REV = 11,
+ OPC_NOP = _OPC(0, 0),
+ OPC_BR = _OPC(0, 1),
+ OPC_JUMP = _OPC(0, 2),
+ OPC_CALL = _OPC(0, 3),
+ OPC_RET = _OPC(0, 4),
+ OPC_KILL = _OPC(0, 5),
+ OPC_END = _OPC(0, 6),
+ OPC_EMIT = _OPC(0, 7),
+ OPC_CUT = _OPC(0, 8),
+ OPC_CHMASK = _OPC(0, 9),
+ OPC_CHSH = _OPC(0, 10),
+ OPC_FLOW_REV = _OPC(0, 11),
/* category 1: */
- /* no opc.. all category 1 are variants of mov */
+ OPC_MOV = _OPC(1, 0),
/* category 2: */
- OPC_ADD_F = 0,
- OPC_MIN_F = 1,
- OPC_MAX_F = 2,
- OPC_MUL_F = 3,
- OPC_SIGN_F = 4,
- OPC_CMPS_F = 5,
- OPC_ABSNEG_F = 6,
- OPC_CMPV_F = 7,
+ OPC_ADD_F = _OPC(2, 0),
+ OPC_MIN_F = _OPC(2, 1),
+ OPC_MAX_F = _OPC(2, 2),
+ OPC_MUL_F = _OPC(2, 3),
+ OPC_SIGN_F = _OPC(2, 4),
+ OPC_CMPS_F = _OPC(2, 5),
+ OPC_ABSNEG_F = _OPC(2, 6),
+ OPC_CMPV_F = _OPC(2, 7),
/* 8 - invalid */
- OPC_FLOOR_F = 9,
- OPC_CEIL_F = 10,
- OPC_RNDNE_F = 11,
- OPC_RNDAZ_F = 12,
- OPC_TRUNC_F = 13,
+ OPC_FLOOR_F = _OPC(2, 9),
+ OPC_CEIL_F = _OPC(2, 10),
+ OPC_RNDNE_F = _OPC(2, 11),
+ OPC_RNDAZ_F = _OPC(2, 12),
+ OPC_TRUNC_F = _OPC(2, 13),
/* 14-15 - invalid */
- OPC_ADD_U = 16,
- OPC_ADD_S = 17,
- OPC_SUB_U = 18,
- OPC_SUB_S = 19,
- OPC_CMPS_U = 20,
- OPC_CMPS_S = 21,
- OPC_MIN_U = 22,
- OPC_MIN_S = 23,
- OPC_MAX_U = 24,
- OPC_MAX_S = 25,
- OPC_ABSNEG_S = 26,
+ OPC_ADD_U = _OPC(2, 16),
+ OPC_ADD_S = _OPC(2, 17),
+ OPC_SUB_U = _OPC(2, 18),
+ OPC_SUB_S = _OPC(2, 19),
+ OPC_CMPS_U = _OPC(2, 20),
+ OPC_CMPS_S = _OPC(2, 21),
+ OPC_MIN_U = _OPC(2, 22),
+ OPC_MIN_S = _OPC(2, 23),
+ OPC_MAX_U = _OPC(2, 24),
+ OPC_MAX_S = _OPC(2, 25),
+ OPC_ABSNEG_S = _OPC(2, 26),
/* 27 - invalid */
- OPC_AND_B = 28,
- OPC_OR_B = 29,
- OPC_NOT_B = 30,
- OPC_XOR_B = 31,
+ OPC_AND_B = _OPC(2, 28),
+ OPC_OR_B = _OPC(2, 29),
+ OPC_NOT_B = _OPC(2, 30),
+ OPC_XOR_B = _OPC(2, 31),
/* 32 - invalid */
- OPC_CMPV_U = 33,
- OPC_CMPV_S = 34,
+ OPC_CMPV_U = _OPC(2, 33),
+ OPC_CMPV_S = _OPC(2, 34),
/* 35-47 - invalid */
- OPC_MUL_U = 48,
- OPC_MUL_S = 49,
- OPC_MULL_U = 50,
- OPC_BFREV_B = 51,
- OPC_CLZ_S = 52,
- OPC_CLZ_B = 53,
- OPC_SHL_B = 54,
- OPC_SHR_B = 55,
- OPC_ASHR_B = 56,
- OPC_BARY_F = 57,
- OPC_MGEN_B = 58,
- OPC_GETBIT_B = 59,
- OPC_SETRM = 60,
- OPC_CBITS_B = 61,
- OPC_SHB = 62,
- OPC_MSAD = 63,
+ OPC_MUL_U = _OPC(2, 48),
+ OPC_MUL_S = _OPC(2, 49),
+ OPC_MULL_U = _OPC(2, 50),
+ OPC_BFREV_B = _OPC(2, 51),
+ OPC_CLZ_S = _OPC(2, 52),
+ OPC_CLZ_B = _OPC(2, 53),
+ OPC_SHL_B = _OPC(2, 54),
+ OPC_SHR_B = _OPC(2, 55),
+ OPC_ASHR_B = _OPC(2, 56),
+ OPC_BARY_F = _OPC(2, 57),
+ OPC_MGEN_B = _OPC(2, 58),
+ OPC_GETBIT_B = _OPC(2, 59),
+ OPC_SETRM = _OPC(2, 60),
+ OPC_CBITS_B = _OPC(2, 61),
+ OPC_SHB = _OPC(2, 62),
+ OPC_MSAD = _OPC(2, 63),
/* category 3: */
- OPC_MAD_U16 = 0,
- OPC_MADSH_U16 = 1,
- OPC_MAD_S16 = 2,
- OPC_MADSH_M16 = 3, /* should this be .s16? */
- OPC_MAD_U24 = 4,
- OPC_MAD_S24 = 5,
- OPC_MAD_F16 = 6,
- OPC_MAD_F32 = 7,
- OPC_SEL_B16 = 8,
- OPC_SEL_B32 = 9,
- OPC_SEL_S16 = 10,
- OPC_SEL_S32 = 11,
- OPC_SEL_F16 = 12,
- OPC_SEL_F32 = 13,
- OPC_SAD_S16 = 14,
- OPC_SAD_S32 = 15,
+ OPC_MAD_U16 = _OPC(3, 0),
+ OPC_MADSH_U16 = _OPC(3, 1),
+ OPC_MAD_S16 = _OPC(3, 2),
+ OPC_MADSH_M16 = _OPC(3, 3), /* should this be .s16? */
+ OPC_MAD_U24 = _OPC(3, 4),
+ OPC_MAD_S24 = _OPC(3, 5),
+ OPC_MAD_F16 = _OPC(3, 6),
+ OPC_MAD_F32 = _OPC(3, 7),
+ OPC_SEL_B16 = _OPC(3, 8),
+ OPC_SEL_B32 = _OPC(3, 9),
+ OPC_SEL_S16 = _OPC(3, 10),
+ OPC_SEL_S32 = _OPC(3, 11),
+ OPC_SEL_F16 = _OPC(3, 12),
+ OPC_SEL_F32 = _OPC(3, 13),
+ OPC_SAD_S16 = _OPC(3, 14),
+ OPC_SAD_S32 = _OPC(3, 15),
/* category 4: */
- OPC_RCP = 0,
- OPC_RSQ = 1,
- OPC_LOG2 = 2,
- OPC_EXP2 = 3,
- OPC_SIN = 4,
- OPC_COS = 5,
- OPC_SQRT = 6,
+ OPC_RCP = _OPC(4, 0),
+ OPC_RSQ = _OPC(4, 1),
+ OPC_LOG2 = _OPC(4, 2),
+ OPC_EXP2 = _OPC(4, 3),
+ OPC_SIN = _OPC(4, 4),
+ OPC_COS = _OPC(4, 5),
+ OPC_SQRT = _OPC(4, 6),
// 7-63 - invalid
/* category 5: */
- OPC_ISAM = 0,
- OPC_ISAML = 1,
- OPC_ISAMM = 2,
- OPC_SAM = 3,
- OPC_SAMB = 4,
- OPC_SAML = 5,
- OPC_SAMGQ = 6,
- OPC_GETLOD = 7,
- OPC_CONV = 8,
- OPC_CONVM = 9,
- OPC_GETSIZE = 10,
- OPC_GETBUF = 11,
- OPC_GETPOS = 12,
- OPC_GETINFO = 13,
- OPC_DSX = 14,
- OPC_DSY = 15,
- OPC_GATHER4R = 16,
- OPC_GATHER4G = 17,
- OPC_GATHER4B = 18,
- OPC_GATHER4A = 19,
- OPC_SAMGP0 = 20,
- OPC_SAMGP1 = 21,
- OPC_SAMGP2 = 22,
- OPC_SAMGP3 = 23,
- OPC_DSXPP_1 = 24,
- OPC_DSYPP_1 = 25,
- OPC_RGETPOS = 26,
- OPC_RGETINFO = 27,
+ OPC_ISAM = _OPC(5, 0),
+ OPC_ISAML = _OPC(5, 1),
+ OPC_ISAMM = _OPC(5, 2),
+ OPC_SAM = _OPC(5, 3),
+ OPC_SAMB = _OPC(5, 4),
+ OPC_SAML = _OPC(5, 5),
+ OPC_SAMGQ = _OPC(5, 6),
+ OPC_GETLOD = _OPC(5, 7),
+ OPC_CONV = _OPC(5, 8),
+ OPC_CONVM = _OPC(5, 9),
+ OPC_GETSIZE = _OPC(5, 10),
+ OPC_GETBUF = _OPC(5, 11),
+ OPC_GETPOS = _OPC(5, 12),
+ OPC_GETINFO = _OPC(5, 13),
+ OPC_DSX = _OPC(5, 14),
+ OPC_DSY = _OPC(5, 15),
+ OPC_GATHER4R = _OPC(5, 16),
+ OPC_GATHER4G = _OPC(5, 17),
+ OPC_GATHER4B = _OPC(5, 18),
+ OPC_GATHER4A = _OPC(5, 19),
+ OPC_SAMGP0 = _OPC(5, 20),
+ OPC_SAMGP1 = _OPC(5, 21),
+ OPC_SAMGP2 = _OPC(5, 22),
+ OPC_SAMGP3 = _OPC(5, 23),
+ OPC_DSXPP_1 = _OPC(5, 24),
+ OPC_DSYPP_1 = _OPC(5, 25),
+ OPC_RGETPOS = _OPC(5, 26),
+ OPC_RGETINFO = _OPC(5, 27),
/* category 6: */
- OPC_LDG = 0, /* load-global */
- OPC_LDL = 1,
- OPC_LDP = 2,
- OPC_STG = 3, /* store-global */
- OPC_STL = 4,
- OPC_STP = 5,
- OPC_STI = 6,
- OPC_G2L = 7,
- OPC_L2G = 8,
- OPC_PREFETCH = 9,
- OPC_LDLW = 10,
- OPC_STLW = 11,
- OPC_RESFMT = 14,
- OPC_RESINFO = 15,
- OPC_ATOMIC_ADD = 16,
- OPC_ATOMIC_SUB = 17,
- OPC_ATOMIC_XCHG = 18,
- OPC_ATOMIC_INC = 19,
- OPC_ATOMIC_DEC = 20,
- OPC_ATOMIC_CMPXCHG = 21,
- OPC_ATOMIC_MIN = 22,
- OPC_ATOMIC_MAX = 23,
- OPC_ATOMIC_AND = 24,
- OPC_ATOMIC_OR = 25,
- OPC_ATOMIC_XOR = 26,
- OPC_LDGB_TYPED_4D = 27,
- OPC_STGB_4D_4 = 28,
- OPC_STIB = 29,
- OPC_LDC_4 = 30,
- OPC_LDLV = 31,
+ OPC_LDG = _OPC(6, 0), /* load-global */
+ OPC_LDL = _OPC(6, 1),
+ OPC_LDP = _OPC(6, 2),
+ OPC_STG = _OPC(6, 3), /* store-global */
+ OPC_STL = _OPC(6, 4),
+ OPC_STP = _OPC(6, 5),
+ OPC_STI = _OPC(6, 6),
+ OPC_G2L = _OPC(6, 7),
+ OPC_L2G = _OPC(6, 8),
+ OPC_PREFETCH = _OPC(6, 9),
+ OPC_LDLW = _OPC(6, 10),
+ OPC_STLW = _OPC(6, 11),
+ OPC_RESFMT = _OPC(6, 14),
+ OPC_RESINFO = _OPC(6, 15),
+ OPC_ATOMIC_ADD = _OPC(6, 16),
+ OPC_ATOMIC_SUB = _OPC(6, 17),
+ OPC_ATOMIC_XCHG = _OPC(6, 18),
+ OPC_ATOMIC_INC = _OPC(6, 19),
+ OPC_ATOMIC_DEC = _OPC(6, 20),
+ OPC_ATOMIC_CMPXCHG = _OPC(6, 21),
+ OPC_ATOMIC_MIN = _OPC(6, 22),
+ OPC_ATOMIC_MAX = _OPC(6, 23),
+ OPC_ATOMIC_AND = _OPC(6, 24),
+ OPC_ATOMIC_OR = _OPC(6, 25),
+ OPC_ATOMIC_XOR = _OPC(6, 26),
+ OPC_LDGB = _OPC(6, 27),
+ OPC_STGB = _OPC(6, 28),
+ OPC_STIB = _OPC(6, 29),
+ OPC_LDC = _OPC(6, 30),
+ OPC_LDLV = _OPC(6, 31),
+
+ /* category 7: */
+ OPC_BAR = _OPC(7, 0),
+ OPC_FENCE = _OPC(7, 1),
/* meta instructions (category -1): */
/* placeholder instr to mark shader inputs: */
- OPC_META_INPUT = 0,
- OPC_META_PHI = 1,
+ OPC_META_INPUT = _OPC(-1, 0),
/* The "fan-in" and "fan-out" instructions are used for keeping
* track of instructions that write to multiple dst registers
* (fan-out) like texture sample instructions, or read multiple
* consecutive scalar registers (fan-in) (bary.f, texture samp)
*/
- OPC_META_FO = 2,
- OPC_META_FI = 3,
+ OPC_META_FO = _OPC(-1, 2),
+ OPC_META_FI = _OPC(-1, 3),
} opc_t;
+#define opc_cat(opc) ((int)((opc) >> NOPC_BITS))
+#define opc_op(opc) ((unsigned)((opc) & ((1 << NOPC_BITS) - 1)))
+
typedef enum {
TYPE_F16 = 0,
TYPE_F32 = 1,
int32_t immed : 20;
uint32_t dummy1 : 12;
} a4xx;
+ struct PACKED {
+ int32_t immed : 32;
+ } a5xx;
};
/* dword1: */
/* dword1: */
uint32_t dst : 8;
- uint32_t repeat : 3;
+ uint32_t repeat : 2;
+ uint32_t sat : 1;
uint32_t src1_r : 1;
uint32_t ss : 1;
uint32_t ul : 1; /* dunno */
/* dword1: */
uint32_t dst : 8;
- uint32_t repeat : 3;
+ uint32_t repeat : 2;
+ uint32_t sat : 1;
uint32_t src1_r : 1;
uint32_t ss : 1;
uint32_t ul : 1;
static inline bool instr_cat3_full(instr_cat3_t *cat3)
{
- switch (cat3->opc) {
+ switch (_OPC(3, cat3->opc)) {
case OPC_MAD_F16:
case OPC_MAD_U16:
case OPC_MAD_S16:
/* dword1: */
uint32_t dst : 8;
- uint32_t repeat : 3;
+ uint32_t repeat : 2;
+ uint32_t sat : 1;
uint32_t src_r : 1;
uint32_t ss : 1;
uint32_t ul : 1;
uint32_t dst : 8;
uint32_t mustbe0 : 1;
- uint32_t pad0 : 23;
+ uint32_t idx : 8;
+ uint32_t pad0 : 15;
} instr_cat6d_t;
-/* I think some of the other cat6 instructions use additional
- * sub-encodings..
+/* ldgb and atomics..
+ *
+ * ldgb: pad0=0, pad3=1
+ * atomic .g: pad0=1, pad3=1
+ * .l: pad0=1, pad3=0
*/
+typedef struct PACKED {
+ /* dword0: */
+ uint32_t pad0 : 1;
+ uint32_t src3 : 8;
+ uint32_t d : 2;
+ uint32_t typed : 1;
+ uint32_t type_size : 2;
+ uint32_t src1 : 8;
+ uint32_t src1_im : 1;
+ uint32_t src2_im : 1;
+ uint32_t src2 : 8;
+
+ /* dword1: */
+ uint32_t dst : 8;
+ uint32_t mustbe0 : 1;
+ uint32_t src_ssbo : 8;
+ uint32_t pad2 : 3; // type
+ uint32_t g : 1;
+ uint32_t pad3 : 1;
+ uint32_t pad4 : 10; // opc/jmp_tgt/sync/opc_cat
+} instr_cat6ldgb_t;
+
+/* stgb, pad0=0, pad3=2
+ */
+typedef struct PACKED {
+ /* dword0: */
+ uint32_t mustbe1 : 1; // ???
+ uint32_t src1 : 8;
+ uint32_t d : 2;
+ uint32_t typed : 1;
+ uint32_t type_size : 2;
+ uint32_t pad0 : 9;
+ uint32_t src2_im : 1;
+ uint32_t src2 : 8;
+
+ /* dword1: */
+ uint32_t src3 : 8;
+ uint32_t src3_im : 1;
+ uint32_t dst_ssbo : 8;
+ uint32_t pad2 : 3; // type
+ uint32_t pad3 : 2;
+ uint32_t pad4 : 10; // opc/jmp_tgt/sync/opc_cat
+} instr_cat6stgb_t;
typedef union PACKED {
instr_cat6a_t a;
instr_cat6b_t b;
instr_cat6c_t c;
instr_cat6d_t d;
+ instr_cat6ldgb_t ldgb;
+ instr_cat6stgb_t stgb;
struct PACKED {
/* dword0: */
uint32_t src_off : 1;
};
} instr_cat6_t;
+typedef struct PACKED {
+ /* dword0: */
+ uint32_t pad1 : 32;
+
+ /* dword1: */
+ uint32_t pad2 : 12;
+ uint32_t ss : 1; /* maybe in the encoding, but blob only uses (sy) */
+ uint32_t pad3 : 6;
+ uint32_t w : 1; /* write */
+ uint32_t r : 1; /* read */
+ uint32_t l : 1; /* local */
+ uint32_t g : 1; /* global */
+ uint32_t opc : 4; /* presumed, but only a couple known OPCs */
+ uint32_t jmp_tgt : 1; /* (jp) */
+ uint32_t sync : 1; /* (sy) */
+ uint32_t opc_cat : 3;
+} instr_cat7_t;
+
typedef union PACKED {
instr_cat0_t cat0;
instr_cat1_t cat1;
instr_cat4_t cat4;
instr_cat5_t cat5;
instr_cat6_t cat6;
+ instr_cat7_t cat7;
struct PACKED {
/* dword0: */
- uint64_t pad1 : 40;
- uint32_t repeat : 3; /* cat0-cat4 */
- uint32_t pad2 : 1;
- uint32_t ss : 1; /* cat1-cat4 (cat0??) */
+ uint32_t pad1 : 32;
+
+ /* dword1: */
+ uint32_t pad2 : 12;
+ uint32_t ss : 1; /* cat1-cat4 (cat0??) and cat7 (?) */
uint32_t ul : 1; /* cat2-cat4 (and cat1 in blob.. which may be bug??) */
uint32_t pad3 : 13;
uint32_t jmp_tgt : 1;
};
} instr_t;
+static inline uint32_t instr_repeat(instr_t *instr)
+{
+ switch (instr->opc_cat) {
+ case 0: return instr->cat0.repeat;
+ case 1: return instr->cat1.repeat;
+ case 2: return instr->cat2.repeat;
+ case 3: return instr->cat3.repeat;
+ case 4: return instr->cat4.repeat;
+ default: return 0;
+ }
+}
+
+static inline bool instr_sat(instr_t *instr)
+{
+ switch (instr->opc_cat) {
+ case 2: return instr->cat2.sat;
+ case 3: return instr->cat3.sat;
+ case 4: return instr->cat4.sat;
+ default: return false;
+ }
+}
+
static inline uint32_t instr_opc(instr_t *instr)
{
switch (instr->opc_cat) {
case 4: return instr->cat4.opc;
case 5: return instr->cat5.opc;
case 6: return instr->cat6.opc;
+ case 7: return instr->cat7.opc;
default: return 0;
}
}
}
}
+static inline bool is_atomic(opc_t opc)
+{
+ switch (opc) {
+ case OPC_ATOMIC_ADD:
+ case OPC_ATOMIC_SUB:
+ case OPC_ATOMIC_XCHG:
+ case OPC_ATOMIC_INC:
+ case OPC_ATOMIC_DEC:
+ case OPC_ATOMIC_CMPXCHG:
+ case OPC_ATOMIC_MIN:
+ case OPC_ATOMIC_MAX:
+ case OPC_ATOMIC_AND:
+ case OPC_ATOMIC_OR:
+ case OPC_ATOMIC_XOR:
+ return true;
+ default:
+ return false;
+ }
+}
+
+static inline bool is_ssbo(opc_t opc)
+{
+ switch (opc) {
+ case OPC_RESFMT:
+ case OPC_RESINFO:
+ case OPC_LDGB:
+ case OPC_STGB:
+ case OPC_STIB:
+ return true;
+ default:
+ return false;
+ }
+}
+
#endif /* INSTR_A3XX_H_ */
projects / mesa.git / blobdiff