diff options
| -rw-r--r-- | src/broadcom/qpu/qpu_pack.c | 34 | ||||
| -rw-r--r-- | src/broadcom/qpu/tests/qpu_disasm.c | 6 |
2 files changed, 40 insertions, 0 deletions
diff --git a/src/broadcom/qpu/qpu_pack.c b/src/broadcom/qpu/qpu_pack.c index 6630ed4b85f..02aa1b86aa4 100644 --- a/src/broadcom/qpu/qpu_pack.c +++ b/src/broadcom/qpu/qpu_pack.c @@ -743,6 +743,19 @@ v3d_qpu_mul_unpack(const struct v3d_device_info *devinfo, uint64_t packed_inst, } break; + + case V3D_QPU_M_VFMUL: + instr->alu.mul.output_pack = V3D_QPU_PACK_NONE; + + if (!v3d_qpu_float16_unpack_unpack(((op & 0x7) - 4) & 7, + &instr->alu.mul.a_unpack)) { + return false; + } + + instr->alu.mul.b_unpack = V3D_QPU_UNPACK_NONE; + + break; + default: instr->alu.mul.output_pack = V3D_QPU_PACK_NONE; instr->alu.mul.a_unpack = V3D_QPU_UNPACK_NONE; @@ -1002,6 +1015,27 @@ v3d_qpu_mul_pack(const struct v3d_device_info *devinfo, break; } + case V3D_QPU_M_VFMUL: { + uint32_t packed; + + if (instr->alu.mul.output_pack != V3D_QPU_PACK_NONE) + return false; + + if (!v3d_qpu_float16_unpack_pack(instr->alu.mul.a_unpack, + &packed)) { + return false; + } + if (instr->alu.mul.a_unpack == V3D_QPU_UNPACK_SWAP_16) + opcode = 8; + else + opcode |= (packed + 4) & 7; + + if (instr->alu.mul.b_unpack != V3D_QPU_UNPACK_NONE) + return false; + + break; + } + default: break; } diff --git a/src/broadcom/qpu/tests/qpu_disasm.c b/src/broadcom/qpu/tests/qpu_disasm.c index c7f6476def5..59668a86ecc 100644 --- a/src/broadcom/qpu/tests/qpu_disasm.c +++ b/src/broadcom/qpu/tests/qpu_disasm.c @@ -40,6 +40,12 @@ static const struct { { 33, 0x3c002380b6edb000ull, "or rf0, r3, r3 ; mov vpm, r3" }, { 33, 0x57403006bbb80000ull, "nop ; fmul r0, rf0, r5 ; ldvpm; ldunif" }, + /* vfmul input packing */ + { 33, 0x101e8b6e8aad4000ull, "fmax.nornn rf46, r4.l, r2.l; vfmul.ifnb rf45, r3, r5" }, + { 33, 0x1857d3c219825000ull, "faddnf.norc r2.l, r5.l, r4; vfmul.ifb rf15, r0.ll, r4; ldunif" }, + { 33, 0x1c0a0dfde2294000ull, "fcmp.ifna rf61.h, r4.abs, r2.l; vfmul rf55, r2.hh, r1" }, + { 33, 0x2011c89b402cc000ull, "fsub.norz rf27, r4.abs, r1.abs; vfmul.ifa rf34, r3.swp, r1" }, + /* branch conditions */ { 33, 0x02000006002034c0ull, "b.anyap rf19" }, { 33, 0x02679356b4201000ull, "b.anyap -1268280496" }, |