#include "broadcom/common/v3d_device_info.h"
#include "qpu_instr.h"
-#ifndef QPU_MASK
-#define QPU_MASK(high, low) ((((uint64_t)1<<((high)-(low)+1))-1)<<(low))
-/* Using the GNU statement expression extension */
-#define QPU_SET_FIELD(value, field) \
- ({ \
- uint64_t fieldval = (uint64_t)(value) << field ## _SHIFT; \
- assert((fieldval & ~ field ## _MASK) == 0); \
- fieldval & field ## _MASK; \
- })
-
-#define QPU_GET_FIELD(word, field) ((uint32_t)(((word) & field ## _MASK) >> field ## _SHIFT))
-
-#define QPU_UPDATE_FIELD(inst, value, field) \
- (((inst) & ~(field ## _MASK)) | QPU_SET_FIELD(value, field))
-#endif /* QPU_MASK */
-
-#define VC5_QPU_OP_MUL_SHIFT 58
-#define VC5_QPU_OP_MUL_MASK QPU_MASK(63, 58)
-
-#define VC5_QPU_SIG_SHIFT 53
-#define VC5_QPU_SIG_MASK QPU_MASK(57, 53)
-
-#define VC5_QPU_COND_SHIFT 46
-#define VC5_QPU_COND_MASK QPU_MASK(52, 46)
-
-#define VC5_QPU_COND_IFA 0
-#define VC5_QPU_COND_IFB 1
-#define VC5_QPU_COND_IFNA 2
-#define VC5_QPU_COND_IFNB 3
-
-#define VC5_QPU_MM QPU_MASK(45, 45)
-#define VC5_QPU_MA QPU_MASK(44, 44)
-
-#define V3D_QPU_WADDR_M_SHIFT 38
-#define V3D_QPU_WADDR_M_MASK QPU_MASK(43, 38)
-
-#define VC5_QPU_BRANCH_ADDR_LOW_SHIFT 35
-#define VC5_QPU_BRANCH_ADDR_LOW_MASK QPU_MASK(55, 35)
-
-#define V3D_QPU_WADDR_A_SHIFT 32
-#define V3D_QPU_WADDR_A_MASK QPU_MASK(37, 32)
-
-#define VC5_QPU_BRANCH_COND_SHIFT 32
-#define VC5_QPU_BRANCH_COND_MASK QPU_MASK(34, 32)
-
-#define VC5_QPU_BRANCH_ADDR_HIGH_SHIFT 24
-#define VC5_QPU_BRANCH_ADDR_HIGH_MASK QPU_MASK(31, 24)
-
-#define VC5_QPU_OP_ADD_SHIFT 24
-#define VC5_QPU_OP_ADD_MASK QPU_MASK(31, 24)
-
-#define VC5_QPU_MUL_B_SHIFT 21
-#define VC5_QPU_MUL_B_MASK QPU_MASK(23, 21)
-
-#define VC5_QPU_BRANCH_MSFIGN_SHIFT 21
-#define VC5_QPU_BRANCH_MSFIGN_MASK QPU_MASK(22, 21)
-
-#define VC5_QPU_MUL_A_SHIFT 18
-#define VC5_QPU_MUL_A_MASK QPU_MASK(20, 18)
-
-#define VC5_QPU_ADD_B_SHIFT 15
-#define VC5_QPU_ADD_B_MASK QPU_MASK(17, 15)
-
-#define VC5_QPU_BRANCH_BDU_SHIFT 15
-#define VC5_QPU_BRANCH_BDU_MASK QPU_MASK(17, 15)
-
-#define VC5_QPU_BRANCH_UB QPU_MASK(14, 14)
-
-#define VC5_QPU_ADD_A_SHIFT 12
-#define VC5_QPU_ADD_A_MASK QPU_MASK(14, 12)
-
-#define VC5_QPU_BRANCH_BDI_SHIFT 12
-#define VC5_QPU_BRANCH_BDI_MASK QPU_MASK(13, 12)
-
-#define VC5_QPU_RADDR_A_SHIFT 6
-#define VC5_QPU_RADDR_A_MASK QPU_MASK(11, 6)
-
-#define VC5_QPU_RADDR_B_SHIFT 0
-#define VC5_QPU_RADDR_B_MASK QPU_MASK(5, 0)
-
const char *
v3d_qpu_magic_waddr_name(enum v3d_qpu_waddr waddr)
{
[V3D_QPU_WADDR_VPMU] = "vpmu",
[V3D_QPU_WADDR_SYNC] = "sync",
[V3D_QPU_WADDR_SYNCU] = "syncu",
+ [V3D_QPU_WADDR_SYNCB] = "syncb",
[V3D_QPU_WADDR_RECIP] = "recip",
[V3D_QPU_WADDR_RSQRT] = "rsqrt",
[V3D_QPU_WADDR_EXP] = "exp",
[V3D_QPU_WADDR_LOG] = "log",
[V3D_QPU_WADDR_SIN] = "sin",
[V3D_QPU_WADDR_RSQRT2] = "rsqrt2",
+ [V3D_QPU_WADDR_TMUC] = "tmuc",
+ [V3D_QPU_WADDR_TMUS] = "tmus",
+ [V3D_QPU_WADDR_TMUT] = "tmut",
+ [V3D_QPU_WADDR_TMUR] = "tmur",
+ [V3D_QPU_WADDR_TMUI] = "tmui",
+ [V3D_QPU_WADDR_TMUB] = "tmub",
+ [V3D_QPU_WADDR_TMUDREF] = "tmudref",
+ [V3D_QPU_WADDR_TMUOFF] = "tmuoff",
+ [V3D_QPU_WADDR_TMUSCM] = "tmuscm",
+ [V3D_QPU_WADDR_TMUSF] = "tmusf",
+ [V3D_QPU_WADDR_TMUSLOD] = "tmuslod",
+ [V3D_QPU_WADDR_TMUHS] = "tmuhs",
+ [V3D_QPU_WADDR_TMUHSCM] = "tmuscm",
+ [V3D_QPU_WADDR_TMUHSF] = "tmuhsf",
+ [V3D_QPU_WADDR_TMUHSLOD] = "tmuhslod",
+ [V3D_QPU_WADDR_R5REP] = "r5rep",
};
return waddr_magic[waddr];
[V3D_QPU_A_NEG] = "neg",
[V3D_QPU_A_FLAPUSH] = "flapush",
[V3D_QPU_A_FLBPUSH] = "flbpush",
- [V3D_QPU_A_FLBPOP] = "flbpop",
+ [V3D_QPU_A_FLPOP] = "flpop",
+ [V3D_QPU_A_RECIP] = "recip",
[V3D_QPU_A_SETMSF] = "setmsf",
[V3D_QPU_A_SETREVF] = "setrevf",
[V3D_QPU_A_NOP] = "nop",
[V3D_QPU_A_VDWWT] = "vdwwt",
[V3D_QPU_A_IID] = "iid",
[V3D_QPU_A_SAMPID] = "sampid",
- [V3D_QPU_A_PATCHID] = "patchid",
+ [V3D_QPU_A_BARRIERID] = "barrierid",
[V3D_QPU_A_TMUWT] = "tmuwt",
[V3D_QPU_A_VPMSETUP] = "vpmsetup",
[V3D_QPU_A_VPMWT] = "vpmwt",
- [V3D_QPU_A_LDVPMV] = "ldvpmv",
- [V3D_QPU_A_LDVPMD] = "ldvpmd",
+ [V3D_QPU_A_LDVPMV_IN] = "ldvpmv_in",
+ [V3D_QPU_A_LDVPMV_OUT] = "ldvpmv_out",
+ [V3D_QPU_A_LDVPMD_IN] = "ldvpmd_in",
+ [V3D_QPU_A_LDVPMD_OUT] = "ldvpmd_out",
[V3D_QPU_A_LDVPMP] = "ldvpmp",
- [V3D_QPU_A_LDVPMG] = "ldvpmg",
+ [V3D_QPU_A_RSQRT] = "rsqrt",
+ [V3D_QPU_A_EXP] = "exp",
+ [V3D_QPU_A_LOG] = "log",
+ [V3D_QPU_A_SIN] = "sin",
+ [V3D_QPU_A_RSQRT2] = "rsqrt2",
+ [V3D_QPU_A_LDVPMG_IN] = "ldvpmg_in",
+ [V3D_QPU_A_LDVPMG_OUT] = "ldvpmg_out",
[V3D_QPU_A_FCMP] = "fcmp",
[V3D_QPU_A_VFMAX] = "vfmax",
[V3D_QPU_A_FROUND] = "fround",
[V3D_QPU_A_NEG] = D | A,
[V3D_QPU_A_FLAPUSH] = D | A,
[V3D_QPU_A_FLBPUSH] = D | A,
- [V3D_QPU_A_FLBPOP] = D | A,
+ [V3D_QPU_A_FLPOP] = D | A,
+ [V3D_QPU_A_RECIP] = D | A,
[V3D_QPU_A_SETMSF] = D | A,
[V3D_QPU_A_SETREVF] = D | A,
[V3D_QPU_A_NOP] = 0,
[V3D_QPU_A_VDWWT] = D,
[V3D_QPU_A_IID] = D,
[V3D_QPU_A_SAMPID] = D,
- [V3D_QPU_A_PATCHID] = D,
+ [V3D_QPU_A_BARRIERID] = D,
[V3D_QPU_A_TMUWT] = D,
[V3D_QPU_A_VPMWT] = D,
[V3D_QPU_A_VPMSETUP] = D | A,
- [V3D_QPU_A_LDVPMV] = D | A,
- [V3D_QPU_A_LDVPMD] = D | A,
+ [V3D_QPU_A_LDVPMV_IN] = D | A,
+ [V3D_QPU_A_LDVPMV_OUT] = D | A,
+ [V3D_QPU_A_LDVPMD_IN] = D | A,
+ [V3D_QPU_A_LDVPMD_OUT] = D | A,
[V3D_QPU_A_LDVPMP] = D | A,
- [V3D_QPU_A_LDVPMG] = D | A | B,
+ [V3D_QPU_A_RSQRT] = D | A,
+ [V3D_QPU_A_EXP] = D | A,
+ [V3D_QPU_A_LOG] = D | A,
+ [V3D_QPU_A_SIN] = D | A,
+ [V3D_QPU_A_RSQRT2] = D | A,
+ [V3D_QPU_A_LDVPMG_IN] = D | A | B,
+ [V3D_QPU_A_LDVPMG_OUT] = D | A | B,
/* FIXME: MOVABSNEG */
return 0;
}
+enum v3d_qpu_cond
+v3d_qpu_cond_invert(enum v3d_qpu_cond cond)
+{
+ switch (cond) {
+ case V3D_QPU_COND_IFA:
+ return V3D_QPU_COND_IFNA;
+ case V3D_QPU_COND_IFNA:
+ return V3D_QPU_COND_IFA;
+ case V3D_QPU_COND_IFB:
+ return V3D_QPU_COND_IFNB;
+ case V3D_QPU_COND_IFNB:
+ return V3D_QPU_COND_IFB;
+ default:
+ unreachable("Non-invertible cond");
+ }
+}
+
bool
v3d_qpu_magic_waddr_is_sfu(enum v3d_qpu_waddr waddr)
{
bool
v3d_qpu_magic_waddr_is_tmu(enum v3d_qpu_waddr waddr)
{
- switch (waddr) {
- case V3D_QPU_WADDR_TMU:
- case V3D_QPU_WADDR_TMUL:
- case V3D_QPU_WADDR_TMUD:
- case V3D_QPU_WADDR_TMUA:
- case V3D_QPU_WADDR_TMUAU:
- return true;
- default:
- return false;
- }
+ /* XXX: WADDR_TMU changed to UNIFA on 4.x */
+ return ((waddr >= V3D_QPU_WADDR_TMU &&
+ waddr <= V3D_QPU_WADDR_TMUAU) ||
+ (waddr >= V3D_QPU_WADDR_TMUC &&
+ waddr <= V3D_QPU_WADDR_TMUHSLOD));
+}
+
+bool
+v3d_qpu_waits_on_tmu(const struct v3d_qpu_instr *inst)
+{
+ return (inst->sig.ldtmu ||
+ (inst->type == V3D_QPU_INSTR_TYPE_ALU &&
+ inst->alu.add.op == V3D_QPU_A_TMUWT));
}
bool
v3d_qpu_magic_waddr_is_tsy(enum v3d_qpu_waddr waddr)
{
return (waddr == V3D_QPU_WADDR_SYNC ||
+ waddr == V3D_QPU_WADDR_SYNCB ||
waddr == V3D_QPU_WADDR_SYNCU);
}
+bool
+v3d_qpu_magic_waddr_loads_unif(enum v3d_qpu_waddr waddr)
+{
+ switch (waddr) {
+ case V3D_QPU_WADDR_VPMU:
+ case V3D_QPU_WADDR_TLBU:
+ case V3D_QPU_WADDR_TMUAU:
+ case V3D_QPU_WADDR_SYNCU:
+ return true;
+ default:
+ return false;
+ }
+}
+
+static bool
+v3d_qpu_add_op_reads_vpm(enum v3d_qpu_add_op op)
+{
+ switch (op) {
+ case V3D_QPU_A_VPMSETUP:
+ case V3D_QPU_A_LDVPMV_IN:
+ case V3D_QPU_A_LDVPMV_OUT:
+ case V3D_QPU_A_LDVPMD_IN:
+ case V3D_QPU_A_LDVPMD_OUT:
+ case V3D_QPU_A_LDVPMP:
+ case V3D_QPU_A_LDVPMG_IN:
+ case V3D_QPU_A_LDVPMG_OUT:
+ return true;
+ default:
+ return false;
+ }
+}
+
+static bool
+v3d_qpu_add_op_writes_vpm(enum v3d_qpu_add_op op)
+{
+ switch (op) {
+ case V3D_QPU_A_VPMSETUP:
+ case V3D_QPU_A_STVPMV:
+ case V3D_QPU_A_STVPMD:
+ case V3D_QPU_A_STVPMP:
+ return true;
+ default:
+ return false;
+ }
+}
+
+bool
+v3d_qpu_uses_tlb(const struct v3d_qpu_instr *inst)
+{
+ if (inst->sig.ldtlb ||
+ inst->sig.ldtlbu)
+ return true;
+
+ if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
+ if (inst->alu.add.magic_write &&
+ v3d_qpu_magic_waddr_is_tlb(inst->alu.add.waddr)) {
+ return true;
+ }
+
+ if (inst->alu.mul.magic_write &&
+ v3d_qpu_magic_waddr_is_tlb(inst->alu.mul.waddr)) {
+ return true;
+ }
+ }
+
+ return false;
+}
+
+bool
+v3d_qpu_uses_sfu(const struct v3d_qpu_instr *inst)
+{
+ if (v3d_qpu_instr_is_sfu(inst))
+ return true;
+
+ if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
+ if (inst->alu.add.magic_write &&
+ v3d_qpu_magic_waddr_is_sfu(inst->alu.add.waddr)) {
+ return true;
+ }
+
+ if (inst->alu.mul.magic_write &&
+ v3d_qpu_magic_waddr_is_sfu(inst->alu.mul.waddr)) {
+ return true;
+ }
+ }
+
+ return false;
+}
+
+bool
+v3d_qpu_instr_is_sfu(const struct v3d_qpu_instr *inst)
+{
+ if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
+ switch (inst->alu.add.op) {
+ case V3D_QPU_A_RECIP:
+ case V3D_QPU_A_RSQRT:
+ case V3D_QPU_A_EXP:
+ case V3D_QPU_A_LOG:
+ case V3D_QPU_A_SIN:
+ case V3D_QPU_A_RSQRT2:
+ return true;
+ default:
+ return false;
+ }
+ }
+ return false;
+}
+
+bool
+v3d_qpu_writes_tmu(const struct v3d_qpu_instr *inst)
+{
+ return (inst->type == V3D_QPU_INSTR_TYPE_ALU &&
+ ((inst->alu.add.magic_write &&
+ v3d_qpu_magic_waddr_is_tmu(inst->alu.add.waddr)) ||
+ (inst->alu.mul.magic_write &&
+ v3d_qpu_magic_waddr_is_tmu(inst->alu.mul.waddr))));
+}
+
+bool
+v3d_qpu_writes_tmu_not_tmuc(const struct v3d_qpu_instr *inst)
+{
+ return v3d_qpu_writes_tmu(inst) &&
+ (!inst->alu.add.magic_write ||
+ inst->alu.add.waddr != V3D_QPU_WADDR_TMUC) &&
+ (!inst->alu.mul.magic_write ||
+ inst->alu.mul.waddr != V3D_QPU_WADDR_TMUC);
+}
+
+bool
+v3d_qpu_reads_vpm(const struct v3d_qpu_instr *inst)
+{
+ if (inst->sig.ldvpm)
+ return true;
+
+ if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
+ if (v3d_qpu_add_op_reads_vpm(inst->alu.add.op))
+ return true;
+ }
+
+ return false;
+}
+
+bool
+v3d_qpu_writes_vpm(const struct v3d_qpu_instr *inst)
+{
+ if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
+ if (v3d_qpu_add_op_writes_vpm(inst->alu.add.op))
+ return true;
+
+ if (inst->alu.add.magic_write &&
+ v3d_qpu_magic_waddr_is_vpm(inst->alu.add.waddr)) {
+ return true;
+ }
+
+ if (inst->alu.mul.magic_write &&
+ v3d_qpu_magic_waddr_is_vpm(inst->alu.mul.waddr)) {
+ return true;
+ }
+ }
+
+ return false;
+}
+
+static bool
+v3d_qpu_waits_vpm(const struct v3d_qpu_instr *inst)
+{
+ return inst->type == V3D_QPU_INSTR_TYPE_ALU &&
+ inst->alu.add.op == V3D_QPU_A_VPMWT;
+}
+
+bool
+v3d_qpu_reads_or_writes_vpm(const struct v3d_qpu_instr *inst)
+{
+ return v3d_qpu_reads_vpm(inst) || v3d_qpu_writes_vpm(inst);
+}
+
+bool
+v3d_qpu_uses_vpm(const struct v3d_qpu_instr *inst)
+{
+ return v3d_qpu_reads_vpm(inst) ||
+ v3d_qpu_writes_vpm(inst) ||
+ v3d_qpu_waits_vpm(inst);
+}
+
bool
v3d_qpu_writes_r3(const struct v3d_device_info *devinfo,
const struct v3d_qpu_instr *inst)
v3d_qpu_writes_r4(const struct v3d_device_info *devinfo,
const struct v3d_qpu_instr *inst)
{
- if (inst->sig.ldtmu)
- return true;
-
if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
if (inst->alu.add.magic_write &&
(inst->alu.add.waddr == V3D_QPU_WADDR_R4 ||
}
}
- if (v3d_qpu_sig_writes_address(devinfo, &inst->sig) &&
- inst->sig_magic && inst->sig_addr == V3D_QPU_WADDR_R4) {
+ if (v3d_qpu_sig_writes_address(devinfo, &inst->sig)) {
+ if (inst->sig_magic && inst->sig_addr == V3D_QPU_WADDR_R4)
+ return true;
+ } else if (inst->sig.ldtmu) {
return true;
}
sig->ldtlb ||
sig->ldtlbu);
}
+
+bool
+v3d_qpu_reads_flags(const struct v3d_qpu_instr *inst)
+{
+ if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH) {
+ return inst->branch.cond != V3D_QPU_BRANCH_COND_ALWAYS;
+ } else if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
+ if (inst->flags.ac != V3D_QPU_COND_NONE ||
+ inst->flags.mc != V3D_QPU_COND_NONE ||
+ inst->flags.auf != V3D_QPU_UF_NONE ||
+ inst->flags.muf != V3D_QPU_UF_NONE)
+ return true;
+
+ switch (inst->alu.add.op) {
+ case V3D_QPU_A_VFLA:
+ case V3D_QPU_A_VFLNA:
+ case V3D_QPU_A_VFLB:
+ case V3D_QPU_A_VFLNB:
+ case V3D_QPU_A_FLAPUSH:
+ case V3D_QPU_A_FLBPUSH:
+ return true;
+ default:
+ break;
+ }
+ }
+
+ return false;
+}
+
+bool
+v3d_qpu_writes_flags(const struct v3d_qpu_instr *inst)
+{
+ if (inst->flags.apf != V3D_QPU_PF_NONE ||
+ inst->flags.mpf != V3D_QPU_PF_NONE ||
+ inst->flags.auf != V3D_QPU_UF_NONE ||
+ inst->flags.muf != V3D_QPU_UF_NONE) {
+ return true;
+ }
+
+ return false;
+}
+
+bool
+v3d_qpu_unpacks_f32(const struct v3d_qpu_instr *inst)
+{
+ if (inst->type != V3D_QPU_INSTR_TYPE_ALU)
+ return false;
+
+ switch (inst->alu.add.op) {
+ case V3D_QPU_A_FADD:
+ case V3D_QPU_A_FADDNF:
+ case V3D_QPU_A_FSUB:
+ case V3D_QPU_A_FMIN:
+ case V3D_QPU_A_FMAX:
+ case V3D_QPU_A_FCMP:
+ case V3D_QPU_A_FROUND:
+ case V3D_QPU_A_FTRUNC:
+ case V3D_QPU_A_FFLOOR:
+ case V3D_QPU_A_FCEIL:
+ case V3D_QPU_A_FDX:
+ case V3D_QPU_A_FDY:
+ case V3D_QPU_A_FTOIN:
+ case V3D_QPU_A_FTOIZ:
+ case V3D_QPU_A_FTOUZ:
+ case V3D_QPU_A_FTOC:
+ case V3D_QPU_A_VFPACK:
+ return true;
+ break;
+ default:
+ break;
+ }
+
+ switch (inst->alu.mul.op) {
+ case V3D_QPU_M_FMOV:
+ case V3D_QPU_M_FMUL:
+ return true;
+ break;
+ default:
+ break;
+ }
+
+ return false;
+}
+bool
+v3d_qpu_unpacks_f16(const struct v3d_qpu_instr *inst)
+{
+ if (inst->type != V3D_QPU_INSTR_TYPE_ALU)
+ return false;
+
+ switch (inst->alu.add.op) {
+ case V3D_QPU_A_VFMIN:
+ case V3D_QPU_A_VFMAX:
+ return true;
+ break;
+ default:
+ break;
+ }
+
+ switch (inst->alu.mul.op) {
+ case V3D_QPU_M_VFMUL:
+ return true;
+ break;
+ default:
+ break;
+ }
+
+ return false;
+}
projects / mesa.git / blobdiff