/* * Copyright (C) 2019 Connor Abbott * Copyright (C) 2019 Lyude Paul * Copyright (C) 2019 Ryan Houdek * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the next * paragraph) shall be included in all copies or substantial portions of the * Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ #include #include #include #include #include #include #include "bifrost.h" #include "disassemble.h" #include "util/macros.h" // return bits (high, lo] static uint64_t bits(uint32_t word, unsigned lo, unsigned high) { if (high == 32) return word >> lo; return (word & ((1 << high) - 1)) >> lo; } // each of these structs represents an instruction that's dispatched in one // cycle. Note that these instructions are packed in funny ways within the // clause, hence the need for a separate struct. struct bifrost_alu_inst { uint32_t fma_bits; uint32_t add_bits; uint64_t reg_bits; }; struct bifrost_regs { unsigned uniform_const : 8; unsigned reg2 : 6; unsigned reg3 : 6; unsigned reg0 : 5; unsigned reg1 : 6; unsigned ctrl : 4; }; static unsigned get_reg0(struct bifrost_regs regs) { if (regs.ctrl == 0) return regs.reg0 | ((regs.reg1 & 0x1) << 5); return regs.reg0 <= regs.reg1 ? regs.reg0 : 63 - regs.reg0; } static unsigned get_reg1(struct bifrost_regs regs) { return regs.reg0 <= regs.reg1 ? regs.reg1 : 63 - regs.reg1; } enum bifrost_reg_write_unit { REG_WRITE_NONE = 0, // don't write REG_WRITE_TWO, // write using reg2 REG_WRITE_THREE, // write using reg3 }; // this represents the decoded version of the ctrl register field. struct bifrost_reg_ctrl{ bool read_reg0; bool read_reg1; bool read_reg3; enum bifrost_reg_write_unit fma_write_unit; enum bifrost_reg_write_unit add_write_unit; bool clause_start; }; enum fma_src_type { FMA_ONE_SRC, FMA_TWO_SRC, FMA_FADD, FMA_FMINMAX, FMA_FADD16, FMA_FMINMAX16, FMA_FCMP, FMA_FCMP16, FMA_THREE_SRC, FMA_FMA, FMA_FMA16, FMA_FOUR_SRC, FMA_FMA_MSCALE, FMA_SHIFT_ADD64, }; struct fma_op_info { unsigned op; char name[30]; enum fma_src_type src_type; }; enum add_src_type { ADD_ONE_SRC, ADD_TWO_SRC, ADD_FADD, ADD_FMINMAX, ADD_FADD16, ADD_FMINMAX16, ADD_THREE_SRC, ADD_FADDMscale, ADD_FCMP, ADD_FCMP16, ADD_TEX_COMPACT, // texture instruction with embedded sampler ADD_TEX, // texture instruction with sampler/etc. in uniform port ADD_VARYING_INTERP, ADD_BLENDING, ADD_LOAD_ATTR, ADD_VARYING_ADDRESS, ADD_BRANCH, }; struct add_op_info { unsigned op; char name[30]; enum add_src_type src_type; bool has_data_reg; }; 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; }; 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; }; enum 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 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 branch_code { BR_ALWAYS = 63, }; void dump_header(struct bifrost_header header, bool verbose); void dump_instr(const struct bifrost_alu_inst *instr, struct bifrost_regs next_regs, uint64_t *consts, unsigned data_reg, unsigned offset, bool verbose); bool dump_clause(uint32_t *words, unsigned *size, unsigned offset, bool verbose); void dump_header(struct bifrost_header header, bool verbose) { if (header.clause_type != 0) { printf("id(%du) ", header.scoreboard_index); } if (header.scoreboard_deps != 0) { printf("next-wait("); bool first = true; for (unsigned i = 0; i < 8; i++) { if (header.scoreboard_deps & (1 << i)) { if (!first) { printf(", "); } printf("%d", i); first = false; } } printf(") "); } if (header.datareg_writebarrier) printf("data-reg-barrier "); if (!header.no_end_of_shader) printf("eos "); if (!header.back_to_back) { printf("nbb "); if (header.branch_cond) printf("branch-cond "); else printf("branch-uncond "); } if (header.elide_writes) printf("we "); if (header.suppress_inf) printf("suppress-inf "); if (header.suppress_nan) printf("suppress-nan "); if (header.unk0) printf("unk0 "); if (header.unk1) printf("unk1 "); if (header.unk2) printf("unk2 "); if (header.unk3) printf("unk3 "); if (header.unk4) printf("unk4 "); printf("\n"); if (verbose) { printf("# clause type %d, next clause type %d\n", header.clause_type, header.next_clause_type); } } static struct bifrost_reg_ctrl DecodeRegCtrl(struct bifrost_regs regs) { struct bifrost_reg_ctrl decoded = {}; unsigned ctrl; if (regs.ctrl == 0) { ctrl = regs.reg1 >> 2; decoded.read_reg0 = !(regs.reg1 & 0x2); decoded.read_reg1 = false; } else { ctrl = regs.ctrl; decoded.read_reg0 = decoded.read_reg1 = true; } switch (ctrl) { case 1: decoded.fma_write_unit = REG_WRITE_TWO; break; case 3: decoded.fma_write_unit = REG_WRITE_TWO; decoded.read_reg3 = true; break; case 4: decoded.read_reg3 = true; break; case 5: decoded.add_write_unit = REG_WRITE_TWO; break; case 6: decoded.add_write_unit = REG_WRITE_TWO; decoded.read_reg3 = true; break; case 8: decoded.clause_start = true; break; case 9: decoded.fma_write_unit = REG_WRITE_TWO; decoded.clause_start = true; break; case 11: break; case 12: decoded.read_reg3 = true; decoded.clause_start = true; break; case 13: decoded.add_write_unit = REG_WRITE_TWO; decoded.clause_start = true; break; case 15: decoded.fma_write_unit = REG_WRITE_THREE; decoded.add_write_unit = REG_WRITE_TWO; break; default: printf("# unknown reg ctrl %d\n", ctrl); } return decoded; } // Pass in the add_write_unit or fma_write_unit, and this returns which register // the ADD/FMA units are writing to static unsigned GetRegToWrite(enum bifrost_reg_write_unit unit, struct bifrost_regs regs) { switch (unit) { case REG_WRITE_TWO: return regs.reg2; case REG_WRITE_THREE: return regs.reg3; default: /* REG_WRITE_NONE */ assert(0); return 0; } } static void dump_regs(struct bifrost_regs srcs) { struct bifrost_reg_ctrl ctrl = DecodeRegCtrl(srcs); printf("# "); if (ctrl.read_reg0) printf("port 0: R%d ", get_reg0(srcs)); if (ctrl.read_reg1) printf("port 1: R%d ", get_reg1(srcs)); if (ctrl.fma_write_unit == REG_WRITE_TWO) printf("port 2: R%d (write FMA) ", srcs.reg2); else if (ctrl.add_write_unit == REG_WRITE_TWO) printf("port 2: R%d (write ADD) ", srcs.reg2); if (ctrl.fma_write_unit == REG_WRITE_THREE) printf("port 3: R%d (write FMA) ", srcs.reg3); else if (ctrl.add_write_unit == REG_WRITE_THREE) printf("port 3: R%d (write ADD) ", srcs.reg3); else if (ctrl.read_reg3) printf("port 3: R%d (read) ", srcs.reg3); if (srcs.uniform_const) { if (srcs.uniform_const & 0x80) { printf("uniform: U%d", (srcs.uniform_const & 0x7f) * 2); } } printf("\n"); } static void dump_const_imm(uint32_t imm) { union { float f; uint32_t i; } fi; fi.i = imm; printf("0x%08x /* %f */", imm, fi.f); } static uint64_t get_const(uint64_t *consts, struct bifrost_regs srcs) { unsigned low_bits = srcs.uniform_const & 0xf; uint64_t imm; switch (srcs.uniform_const >> 4) { case 4: imm = consts[0]; break; case 5: imm = consts[1]; break; case 6: imm = consts[2]; break; case 7: imm = consts[3]; break; case 2: imm = consts[4]; break; case 3: imm = consts[5]; break; default: assert(0); break; } return imm | low_bits; } static void dump_uniform_const_src(struct bifrost_regs srcs, uint64_t *consts, bool high32) { if (srcs.uniform_const & 0x80) { unsigned uniform = (srcs.uniform_const & 0x7f) * 2; printf("U%d", uniform + (high32 ? 1 : 0)); } else if (srcs.uniform_const >= 0x20) { uint64_t imm = get_const(consts, srcs); if (high32) dump_const_imm(imm >> 32); else dump_const_imm(imm); } else { switch (srcs.uniform_const) { case 0: printf("0"); break; case 5: printf("atest-data"); break; case 6: printf("sample-ptr"); break; case 8: case 9: case 10: case 11: case 12: case 13: case 14: case 15: printf("blend-descriptor%u", (unsigned) srcs.uniform_const - 8); break; default: printf("unkConst%u", (unsigned) srcs.uniform_const); break; } if (high32) printf(".y"); else printf(".x"); } } static void dump_src(unsigned src, struct bifrost_regs srcs, uint64_t *consts, bool isFMA) { switch (src) { case 0: printf("R%d", get_reg0(srcs)); break; case 1: printf("R%d", get_reg1(srcs)); break; case 2: printf("R%d", srcs.reg3); break; case 3: if (isFMA) printf("0"); else printf("T"); // i.e. the output of FMA this cycle break; case 4: dump_uniform_const_src(srcs, consts, false); break; case 5: dump_uniform_const_src(srcs, consts, true); break; case 6: printf("T0"); break; case 7: printf("T1"); break; } } static void dump_output_mod(unsigned mod) { switch (mod) { case 0: break; case 1: printf(".clamp_0_inf"); break; // max(out, 0) case 2: printf(".clamp_m1_1"); break; // clamp(out, -1, 1) case 3: printf(".clamp_0_1"); break; // clamp(out, 0, 1) default: break; } } static void dump_minmax_mode(unsigned mod) { switch (mod) { case 0: /* 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. */ break; case 1: /* 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. */ printf(".nan_wins"); break; case 2: /* 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. */ printf(".src1_wins"); break; case 3: /* For max, implement src0 < src1 ? src1 : src0 * For min, implement src0 > src1 ? src1 : src0 */ printf(".src0_wins"); break; default: break; } } static void dump_round_mode(unsigned mod) { switch (mod) { case 0: /* roundTiesToEven, the IEEE default. */ break; case 1: /* roundTowardPositive in the IEEE spec. */ printf(".round_pos"); break; case 2: /* roundTowardNegative in the IEEE spec. */ printf(".round_neg"); break; case 3: /* roundTowardZero in the IEEE spec. */ printf(".round_zero"); break; default: break; } } static const struct fma_op_info FMAOpInfos[] = { { 0x00000, "FMA.f32", FMA_FMA }, { 0x40000, "MAX.f32", FMA_FMINMAX }, { 0x44000, "MIN.f32", FMA_FMINMAX }, { 0x48000, "FCMP.GL", FMA_FCMP }, { 0x4c000, "FCMP.D3D", FMA_FCMP }, { 0x4ff98, "ADD.i32", FMA_TWO_SRC }, { 0x4ffd8, "SUB.i32", FMA_TWO_SRC }, { 0x4fff0, "SUBB.i32", FMA_TWO_SRC }, { 0x50000, "FMA_MSCALE", FMA_FMA_MSCALE }, { 0x58000, "ADD.f32", FMA_FADD }, { 0x5c000, "CSEL.FEQ.f32", FMA_FOUR_SRC }, { 0x5c200, "CSEL.FGT.f32", FMA_FOUR_SRC }, { 0x5c400, "CSEL.FGE.f32", FMA_FOUR_SRC }, { 0x5c600, "CSEL.IEQ.f32", FMA_FOUR_SRC }, { 0x5c800, "CSEL.IGT.i32", FMA_FOUR_SRC }, { 0x5ca00, "CSEL.IGE.i32", FMA_FOUR_SRC }, { 0x5cc00, "CSEL.UGT.i32", FMA_FOUR_SRC }, { 0x5ce00, "CSEL.UGE.i32", FMA_FOUR_SRC }, { 0x5d8d0, "ICMP.D3D.GT.v2i16", FMA_TWO_SRC }, { 0x5d9d0, "UCMP.D3D.GT.v2i16", FMA_TWO_SRC }, { 0x5dad0, "ICMP.D3D.GE.v2i16", FMA_TWO_SRC }, { 0x5dbd0, "UCMP.D3D.GE.v2i16", FMA_TWO_SRC }, { 0x5dcd0, "ICMP.D3D.EQ.v2i16", FMA_TWO_SRC }, { 0x5de40, "ICMP.GL.GT.i32", FMA_TWO_SRC }, // src0 > src1 ? 1 : 0 { 0x5de48, "ICMP.GL.GE.i32", FMA_TWO_SRC }, { 0x5de50, "UCMP.GL.GT.i32", FMA_TWO_SRC }, { 0x5de58, "UCMP.GL.GE.i32", FMA_TWO_SRC }, { 0x5de60, "ICMP.GL.EQ.i32", FMA_TWO_SRC }, { 0x5dec0, "ICMP.D3D.GT.i32", FMA_TWO_SRC }, // src0 > src1 ? ~0 : 0 { 0x5dec8, "ICMP.D3D.GE.i32", FMA_TWO_SRC }, { 0x5ded0, "UCMP.D3D.GT.i32", FMA_TWO_SRC }, { 0x5ded8, "UCMP.D3D.GE.i32", FMA_TWO_SRC }, { 0x5dee0, "ICMP.D3D.EQ.i32", FMA_TWO_SRC }, { 0x60200, "RSHIFT_NAND.i32", FMA_THREE_SRC }, { 0x603c0, "RSHIFT_NAND.v2i16", FMA_THREE_SRC }, { 0x60e00, "RSHIFT_OR.i32", FMA_THREE_SRC }, { 0x60fc0, "RSHIFT_OR.v2i16", FMA_THREE_SRC }, { 0x61200, "RSHIFT_AND.i32", FMA_THREE_SRC }, { 0x613c0, "RSHIFT_AND.v2i16", FMA_THREE_SRC }, { 0x61e00, "RSHIFT_NOR.i32", FMA_THREE_SRC }, // ~((src0 << src2) | src1) { 0x61fc0, "RSHIFT_NOR.v2i16", FMA_THREE_SRC }, // ~((src0 << src2) | src1) { 0x62200, "LSHIFT_NAND.i32", FMA_THREE_SRC }, { 0x623c0, "LSHIFT_NAND.v2i16", FMA_THREE_SRC }, { 0x62e00, "LSHIFT_OR.i32", FMA_THREE_SRC }, // (src0 << src2) | src1 { 0x62fc0, "LSHIFT_OR.v2i16", FMA_THREE_SRC }, // (src0 << src2) | src1 { 0x63200, "LSHIFT_AND.i32", FMA_THREE_SRC }, // (src0 << src2) & src1 { 0x633c0, "LSHIFT_AND.v2i16", FMA_THREE_SRC }, { 0x63e00, "LSHIFT_NOR.i32", FMA_THREE_SRC }, { 0x63fc0, "LSHIFT_NOR.v2i16", FMA_THREE_SRC }, { 0x64200, "RSHIFT_XOR.i32", FMA_THREE_SRC }, { 0x643c0, "RSHIFT_XOR.v2i16", FMA_THREE_SRC }, { 0x64600, "RSHIFT_XNOR.i32", FMA_THREE_SRC }, // ~((src0 >> src2) ^ src1) { 0x647c0, "RSHIFT_XNOR.v2i16", FMA_THREE_SRC }, // ~((src0 >> src2) ^ src1) { 0x64a00, "LSHIFT_XOR.i32", FMA_THREE_SRC }, { 0x64bc0, "LSHIFT_XOR.v2i16", FMA_THREE_SRC }, { 0x64e00, "LSHIFT_XNOR.i32", FMA_THREE_SRC }, // ~((src0 >> src2) ^ src1) { 0x64fc0, "LSHIFT_XNOR.v2i16", FMA_THREE_SRC }, // ~((src0 >> src2) ^ src1) { 0x65200, "LSHIFT_ADD.i32", FMA_THREE_SRC }, { 0x65600, "LSHIFT_SUB.i32", FMA_THREE_SRC }, // (src0 << src2) - src1 { 0x65a00, "LSHIFT_RSUB.i32", FMA_THREE_SRC }, // src1 - (src0 << src2) { 0x65e00, "RSHIFT_ADD.i32", FMA_THREE_SRC }, { 0x66200, "RSHIFT_SUB.i32", FMA_THREE_SRC }, { 0x66600, "RSHIFT_RSUB.i32", FMA_THREE_SRC }, { 0x66a00, "ARSHIFT_ADD.i32", FMA_THREE_SRC }, { 0x66e00, "ARSHIFT_SUB.i32", FMA_THREE_SRC }, { 0x67200, "ARSHIFT_RSUB.i32", FMA_THREE_SRC }, { 0x80000, "FMA.v2f16", FMA_FMA16 }, { 0xc0000, "MAX.v2f16", FMA_FMINMAX16 }, { 0xc4000, "MIN.v2f16", FMA_FMINMAX16 }, { 0xc8000, "FCMP.GL", FMA_FCMP16 }, { 0xcc000, "FCMP.D3D", FMA_FCMP16 }, { 0xcf900, "ADD.v2i16", FMA_TWO_SRC }, { 0xcfc10, "ADDC.i32", FMA_TWO_SRC }, { 0xcfd80, "ADD.i32.i16.X", FMA_TWO_SRC }, { 0xcfd90, "ADD.i32.u16.X", FMA_TWO_SRC }, { 0xcfdc0, "ADD.i32.i16.Y", FMA_TWO_SRC }, { 0xcfdd0, "ADD.i32.u16.Y", FMA_TWO_SRC }, { 0xd8000, "ADD.v2f16", FMA_FADD16 }, { 0xdc000, "CSEL.FEQ.v2f16", FMA_FOUR_SRC }, { 0xdc200, "CSEL.FGT.v2f16", FMA_FOUR_SRC }, { 0xdc400, "CSEL.FGE.v2f16", FMA_FOUR_SRC }, { 0xdc600, "CSEL.IEQ.v2f16", FMA_FOUR_SRC }, { 0xdc800, "CSEL.IGT.v2i16", FMA_FOUR_SRC }, { 0xdca00, "CSEL.IGE.v2i16", FMA_FOUR_SRC }, { 0xdcc00, "CSEL.UGT.v2i16", FMA_FOUR_SRC }, { 0xdce00, "CSEL.UGE.v2i16", FMA_FOUR_SRC }, { 0xdd000, "F32_TO_F16", FMA_TWO_SRC }, { 0xe0046, "F16_TO_I16.XX", FMA_ONE_SRC }, { 0xe0047, "F16_TO_U16.XX", FMA_ONE_SRC }, { 0xe004e, "F16_TO_I16.YX", FMA_ONE_SRC }, { 0xe004f, "F16_TO_U16.YX", FMA_ONE_SRC }, { 0xe0056, "F16_TO_I16.XY", FMA_ONE_SRC }, { 0xe0057, "F16_TO_U16.XY", FMA_ONE_SRC }, { 0xe005e, "F16_TO_I16.YY", FMA_ONE_SRC }, { 0xe005f, "F16_TO_U16.YY", FMA_ONE_SRC }, { 0xe00c0, "I16_TO_F16.XX", FMA_ONE_SRC }, { 0xe00c1, "U16_TO_F16.XX", FMA_ONE_SRC }, { 0xe00c8, "I16_TO_F16.YX", FMA_ONE_SRC }, { 0xe00c9, "U16_TO_F16.YX", FMA_ONE_SRC }, { 0xe00d0, "I16_TO_F16.XY", FMA_ONE_SRC }, { 0xe00d1, "U16_TO_F16.XY", FMA_ONE_SRC }, { 0xe00d8, "I16_TO_F16.YY", FMA_ONE_SRC }, { 0xe00d9, "U16_TO_F16.YY", FMA_ONE_SRC }, { 0xe0136, "F32_TO_I32", FMA_ONE_SRC }, { 0xe0137, "F32_TO_U32", FMA_ONE_SRC }, { 0xe0178, "I32_TO_F32", FMA_ONE_SRC }, { 0xe0179, "U32_TO_F32", FMA_ONE_SRC }, { 0xe0198, "I16_TO_I32.X", FMA_ONE_SRC }, { 0xe0199, "U16_TO_U32.X", FMA_ONE_SRC }, { 0xe019a, "I16_TO_I32.Y", FMA_ONE_SRC }, { 0xe019b, "U16_TO_U32.Y", FMA_ONE_SRC }, { 0xe019c, "I16_TO_F32.X", FMA_ONE_SRC }, { 0xe019d, "U16_TO_F32.X", FMA_ONE_SRC }, { 0xe019e, "I16_TO_F32.Y", FMA_ONE_SRC }, { 0xe019f, "U16_TO_F32.Y", FMA_ONE_SRC }, { 0xe01a2, "F16_TO_F32.X", FMA_ONE_SRC }, { 0xe01a3, "F16_TO_F32.Y", FMA_ONE_SRC }, { 0xe032c, "NOP", FMA_ONE_SRC }, { 0xe032d, "MOV", FMA_ONE_SRC }, { 0xe032f, "SWZ.YY.v2i16", FMA_ONE_SRC }, // From the ARM patent US20160364209A1: // "Decompose v (the input) into numbers x1 and s such that v = x1 * 2^s, // and x1 is a floating point value in a predetermined range where the // value 1 is within the range and not at one extremity of the range (e.g. // choose a range where 1 is towards middle of range)." // // This computes x1. { 0xe0345, "LOG_FREXPM", FMA_ONE_SRC }, // Given a floating point number m * 2^e, returns m * 2^{-1}. This is // exactly the same as the mantissa part of frexp(). { 0xe0365, "FRCP_FREXPM", FMA_ONE_SRC }, // Given a floating point number m * 2^e, returns m * 2^{-2} if e is even, // and m * 2^{-1} if e is odd. In other words, scales by powers of 4 until // within the range [0.25, 1). Used for square-root and reciprocal // square-root. { 0xe0375, "FSQRT_FREXPM", FMA_ONE_SRC }, // Given a floating point number m * 2^e, computes -e - 1 as an integer. // Zero and infinity/NaN return 0. { 0xe038d, "FRCP_FREXPE", FMA_ONE_SRC }, // Computes floor(e/2) + 1. { 0xe03a5, "FSQRT_FREXPE", FMA_ONE_SRC }, // Given a floating point number m * 2^e, computes -floor(e/2) - 1 as an // integer. { 0xe03ad, "FRSQ_FREXPE", FMA_ONE_SRC }, { 0xe03c5, "LOG_FREXPE", FMA_ONE_SRC }, { 0xe0b80, "IMAX3", FMA_THREE_SRC }, { 0xe0bc0, "UMAX3", FMA_THREE_SRC }, { 0xe0c00, "IMIN3", FMA_THREE_SRC }, { 0xe0c40, "UMIN3", FMA_THREE_SRC }, { 0xe0f40, "CSEL", FMA_THREE_SRC }, // src2 != 0 ? src1 : src0 { 0xe0fc0, "MUX.i32", FMA_THREE_SRC }, // see ADD comment { 0xe1845, "CEIL", FMA_ONE_SRC }, { 0xe1885, "FLOOR", FMA_ONE_SRC }, { 0xe19b0, "ATAN_LDEXP.Y.f32", FMA_TWO_SRC }, { 0xe19b8, "ATAN_LDEXP.X.f32", FMA_TWO_SRC }, // These instructions in the FMA slot, together with LSHIFT_ADD_HIGH32.i32 // in the ADD slot, allow one to do a 64-bit addition with an extra small // shift on one of the sources. There are three possible scenarios: // // 1) Full 64-bit addition. Do: // out.x = LSHIFT_ADD_LOW32.i64 src1.x, src2.x, shift // out.y = LSHIFT_ADD_HIGH32.i32 src1.y, src2.y // // The shift amount is applied to src2 before adding. The shift amount, and // any extra bits from src2 plus the overflow bit, are sent directly from // FMA to ADD instead of being passed explicitly. Hence, these two must be // bundled together into the same instruction. // // 2) Add a 64-bit value src1 to a zero-extended 32-bit value src2. Do: // out.x = LSHIFT_ADD_LOW32.u32 src1.x, src2, shift // out.y = LSHIFT_ADD_HIGH32.i32 src1.x, 0 // // Note that in this case, the second argument to LSHIFT_ADD_HIGH32 is // ignored, so it can actually be anything. As before, the shift is applied // to src2 before adding. // // 3) Add a 64-bit value to a sign-extended 32-bit value src2. Do: // out.x = LSHIFT_ADD_LOW32.i32 src1.x, src2, shift // out.y = LSHIFT_ADD_HIGH32.i32 src1.x, 0 // // The only difference is the .i32 instead of .u32. Otherwise, this is // exactly the same as before. // // In all these instructions, the shift amount is stored where the third // source would be, so the shift has to be a small immediate from 0 to 7. // This is fine for the expected use-case of these instructions, which is // manipulating 64-bit pointers. // // These instructions can also be combined with various load/store // instructions which normally take a 64-bit pointer in order to add a // 32-bit or 64-bit offset to the pointer before doing the operation, // optionally shifting the offset. The load/store op implicity does // LSHIFT_ADD_HIGH32.i32 internally. Letting ptr be the pointer, and offset // the desired offset, the cases go as follows: // // 1) Add a 64-bit offset: // LSHIFT_ADD_LOW32.i64 ptr.x, offset.x, shift // ld_st_op ptr.y, offset.y, ... // // Note that the output of LSHIFT_ADD_LOW32.i64 is not used, instead being // implicitly sent to the load/store op to serve as the low 32 bits of the // pointer. // // 2) Add a 32-bit unsigned offset: // temp = LSHIFT_ADD_LOW32.u32 ptr.x, offset, shift // ld_st_op temp, ptr.y, ... // // Now, the low 32 bits of offset << shift + ptr are passed explicitly to // the ld_st_op, to match the case where there is no offset and ld_st_op is // called directly. // // 3) Add a 32-bit signed offset: // temp = LSHIFT_ADD_LOW32.i32 ptr.x, offset, shift // ld_st_op temp, ptr.y, ... // // Again, the same as the unsigned case except for the offset. { 0xe1c80, "LSHIFT_ADD_LOW32.u32", FMA_SHIFT_ADD64 }, { 0xe1cc0, "LSHIFT_ADD_LOW32.i64", FMA_SHIFT_ADD64 }, { 0xe1d80, "LSHIFT_ADD_LOW32.i32", FMA_SHIFT_ADD64 }, { 0xe1e00, "SEL.XX.i16", FMA_TWO_SRC }, { 0xe1e08, "SEL.YX.i16", FMA_TWO_SRC }, { 0xe1e10, "SEL.XY.i16", FMA_TWO_SRC }, { 0xe1e18, "SEL.YY.i16", FMA_TWO_SRC }, { 0xe7800, "IMAD", FMA_THREE_SRC }, { 0xe78db, "POPCNT", FMA_ONE_SRC }, }; static struct fma_op_info find_fma_op_info(unsigned op) { for (unsigned i = 0; i < ARRAY_SIZE(FMAOpInfos); i++) { unsigned opCmp = ~0; switch (FMAOpInfos[i].src_type) { case FMA_ONE_SRC: opCmp = op; break; case FMA_TWO_SRC: opCmp = op & ~0x7; break; case FMA_FCMP: case FMA_FCMP16: opCmp = op & ~0x1fff; break; case FMA_THREE_SRC: case FMA_SHIFT_ADD64: opCmp = op & ~0x3f; break; case FMA_FADD: case FMA_FMINMAX: case FMA_FADD16: case FMA_FMINMAX16: opCmp = op & ~0x3fff; break; case FMA_FMA: case FMA_FMA16: opCmp = op & ~0x3ffff; break; case FMA_FOUR_SRC: opCmp = op & ~0x1ff; break; case FMA_FMA_MSCALE: opCmp = op & ~0x7fff; break; default: opCmp = ~0; break; } if (FMAOpInfos[i].op == opCmp) return FMAOpInfos[i]; } struct fma_op_info info; snprintf(info.name, sizeof(info.name), "op%04x", op); info.op = op; info.src_type = FMA_THREE_SRC; return info; } static void dump_fcmp(unsigned op) { switch (op) { case 0: printf(".OEQ"); break; case 1: printf(".OGT"); break; case 2: printf(".OGE"); break; case 3: printf(".UNE"); break; case 4: printf(".OLT"); break; case 5: printf(".OLE"); break; default: printf(".unk%d", op); break; } } static void dump_16swizzle(unsigned swiz) { if (swiz == 2) return; printf(".%c%c", "xy"[swiz & 1], "xy"[(swiz >> 1) & 1]); } static void dump_fma_expand_src0(unsigned ctrl) { switch (ctrl) { case 3: case 4: case 6: printf(".x"); break; case 5: case 7: printf(".y"); break; case 0: case 1: case 2: break; default: printf(".unk"); break; } } static void dump_fma_expand_src1(unsigned ctrl) { switch (ctrl) { case 1: case 3: printf(".x"); break; case 2: case 4: case 5: printf(".y"); break; case 0: case 6: case 7: break; default: printf(".unk"); break; } } static void dump_fma(uint64_t word, struct bifrost_regs regs, struct bifrost_regs next_regs, uint64_t *consts, bool verbose) { if (verbose) { printf("# FMA: %016" PRIx64 "\n", word); } struct bifrost_fma_inst FMA; memcpy((char *) &FMA, (char *) &word, sizeof(struct bifrost_fma_inst)); struct fma_op_info info = find_fma_op_info(FMA.op); printf("%s", info.name); if (info.src_type == FMA_FADD || info.src_type == FMA_FMINMAX || info.src_type == FMA_FMA || info.src_type == FMA_FADD16 || info.src_type == FMA_FMINMAX16 || info.src_type == FMA_FMA16) { dump_output_mod(bits(FMA.op, 12, 14)); switch (info.src_type) { case FMA_FADD: case FMA_FMA: case FMA_FADD16: case FMA_FMA16: dump_round_mode(bits(FMA.op, 10, 12)); break; case FMA_FMINMAX: case FMA_FMINMAX16: dump_minmax_mode(bits(FMA.op, 10, 12)); break; default: assert(0); } } else if (info.src_type == FMA_FCMP || info.src_type == FMA_FCMP16) { dump_fcmp(bits(FMA.op, 10, 13)); if (info.src_type == FMA_FCMP) printf(".f32"); else printf(".v2f16"); } else if (info.src_type == FMA_FMA_MSCALE) { if (FMA.op & (1 << 11)) { switch ((FMA.op >> 9) & 0x3) { case 0: /* This mode seems to do a few things: * - Makes 0 * infinity (and incidentally 0 * nan) return 0, * since generating a nan would poison the result of * 1/infinity and 1/0. * - Fiddles with which nan is returned in nan * nan, * presumably to make sure that the same exact nan is * returned for 1/nan. */ printf(".rcp_mode"); break; case 3: /* Similar to the above, but src0 always wins when multiplying * 0 by infinity. */ printf(".sqrt_mode"); break; default: printf(".unk%d_mode", (int) (FMA.op >> 9) & 0x3); } } else { dump_output_mod(bits(FMA.op, 9, 11)); } } printf(" "); struct bifrost_reg_ctrl next_ctrl = DecodeRegCtrl(next_regs); if (next_ctrl.fma_write_unit != REG_WRITE_NONE) { printf("{R%d, T0}, ", GetRegToWrite(next_ctrl.fma_write_unit, next_regs)); } else { printf("T0, "); } switch (info.src_type) { case FMA_ONE_SRC: dump_src(FMA.src0, regs, consts, true); break; case FMA_TWO_SRC: dump_src(FMA.src0, regs, consts, true); printf(", "); dump_src(FMA.op & 0x7, regs, consts, true); break; case FMA_FADD: case FMA_FMINMAX: if (FMA.op & 0x10) printf("-"); if (FMA.op & 0x200) printf("abs("); dump_src(FMA.src0, regs, consts, true); dump_fma_expand_src0((FMA.op >> 6) & 0x7); if (FMA.op & 0x200) printf(")"); printf(", "); if (FMA.op & 0x20) printf("-"); if (FMA.op & 0x8) printf("abs("); dump_src(FMA.op & 0x7, regs, consts, true); dump_fma_expand_src1((FMA.op >> 6) & 0x7); if (FMA.op & 0x8) printf(")"); break; case FMA_FADD16: case FMA_FMINMAX16: { bool abs1 = FMA.op & 0x8; bool abs2 = (FMA.op & 0x7) < FMA.src0; if (FMA.op & 0x10) printf("-"); if (abs1 || abs2) printf("abs("); dump_src(FMA.src0, regs, consts, true); dump_16swizzle((FMA.op >> 6) & 0x3); if (abs1 || abs2) printf(")"); printf(", "); if (FMA.op & 0x20) printf("-"); if (abs1 && abs2) printf("abs("); dump_src(FMA.op & 0x7, regs, consts, true); dump_16swizzle((FMA.op >> 8) & 0x3); if (abs1 && abs2) printf(")"); break; } case FMA_FCMP: if (FMA.op & 0x200) printf("abs("); dump_src(FMA.src0, regs, consts, true); dump_fma_expand_src0((FMA.op >> 6) & 0x7); if (FMA.op & 0x200) printf(")"); printf(", "); if (FMA.op & 0x20) printf("-"); if (FMA.op & 0x8) printf("abs("); dump_src(FMA.op & 0x7, regs, consts, true); dump_fma_expand_src1((FMA.op >> 6) & 0x7); if (FMA.op & 0x8) printf(")"); break; case FMA_FCMP16: dump_src(FMA.src0, regs, consts, true); // Note: this is kinda a guess, I haven't seen the blob set this to // anything other than the identity, but it matches FMA_TWO_SRCFmod16 dump_16swizzle((FMA.op >> 6) & 0x3); printf(", "); dump_src(FMA.op & 0x7, regs, consts, true); dump_16swizzle((FMA.op >> 8) & 0x3); break; case FMA_SHIFT_ADD64: dump_src(FMA.src0, regs, consts, true); printf(", "); dump_src(FMA.op & 0x7, regs, consts, true); printf(", "); printf("shift:%u", (FMA.op >> 3) & 0x7); break; case FMA_THREE_SRC: dump_src(FMA.src0, regs, consts, true); printf(", "); dump_src(FMA.op & 0x7, regs, consts, true); printf(", "); dump_src((FMA.op >> 3) & 0x7, regs, consts, true); break; case FMA_FMA: if (FMA.op & (1 << 14)) printf("-"); if (FMA.op & (1 << 9)) printf("abs("); dump_src(FMA.src0, regs, consts, true); dump_fma_expand_src0((FMA.op >> 6) & 0x7); if (FMA.op & (1 << 9)) printf(")"); printf(", "); if (FMA.op & (1 << 16)) printf("abs("); dump_src(FMA.op & 0x7, regs, consts, true); dump_fma_expand_src1((FMA.op >> 6) & 0x7); if (FMA.op & (1 << 16)) printf(")"); printf(", "); if (FMA.op & (1 << 15)) printf("-"); if (FMA.op & (1 << 17)) printf("abs("); dump_src((FMA.op >> 3) & 0x7, regs, consts, true); if (FMA.op & (1 << 17)) printf(")"); break; case FMA_FMA16: if (FMA.op & (1 << 14)) printf("-"); dump_src(FMA.src0, regs, consts, true); dump_16swizzle((FMA.op >> 6) & 0x3); printf(", "); dump_src(FMA.op & 0x7, regs, consts, true); dump_16swizzle((FMA.op >> 8) & 0x3); printf(", "); if (FMA.op & (1 << 15)) printf("-"); dump_src((FMA.op >> 3) & 0x7, regs, consts, true); dump_16swizzle((FMA.op >> 16) & 0x3); break; case FMA_FOUR_SRC: dump_src(FMA.src0, regs, consts, true); printf(", "); dump_src(FMA.op & 0x7, regs, consts, true); printf(", "); dump_src((FMA.op >> 3) & 0x7, regs, consts, true); printf(", "); dump_src((FMA.op >> 6) & 0x7, regs, consts, true); break; case FMA_FMA_MSCALE: if (FMA.op & (1 << 12)) printf("abs("); dump_src(FMA.src0, regs, consts, true); if (FMA.op & (1 << 12)) printf(")"); printf(", "); if (FMA.op & (1 << 13)) printf("-"); dump_src(FMA.op & 0x7, regs, consts, true); printf(", "); if (FMA.op & (1 << 14)) printf("-"); dump_src((FMA.op >> 3) & 0x7, regs, consts, true); printf(", "); dump_src((FMA.op >> 6) & 0x7, regs, consts, true); break; } printf("\n"); } static const struct add_op_info add_op_infos[] = { { 0x00000, "MAX.f32", ADD_FMINMAX }, { 0x02000, "MIN.f32", ADD_FMINMAX }, { 0x04000, "ADD.f32", ADD_FADD }, { 0x06000, "FCMP.GL", ADD_FCMP }, { 0x07000, "FCMP.D3D", ADD_FCMP }, { 0x07856, "F16_TO_I16", ADD_ONE_SRC }, { 0x07857, "F16_TO_U16", ADD_ONE_SRC }, { 0x078c0, "I16_TO_F16.XX", ADD_ONE_SRC }, { 0x078c1, "U16_TO_F16.XX", ADD_ONE_SRC }, { 0x078c8, "I16_TO_F16.YX", ADD_ONE_SRC }, { 0x078c9, "U16_TO_F16.YX", ADD_ONE_SRC }, { 0x078d0, "I16_TO_F16.XY", ADD_ONE_SRC }, { 0x078d1, "U16_TO_F16.XY", ADD_ONE_SRC }, { 0x078d8, "I16_TO_F16.YY", ADD_ONE_SRC }, { 0x078d9, "U16_TO_F16.YY", ADD_ONE_SRC }, { 0x07936, "F32_TO_I32", ADD_ONE_SRC }, { 0x07937, "F32_TO_U32", ADD_ONE_SRC }, { 0x07978, "I32_TO_F32", ADD_ONE_SRC }, { 0x07979, "U32_TO_F32", ADD_ONE_SRC }, { 0x07998, "I16_TO_I32.X", ADD_ONE_SRC }, { 0x07999, "U16_TO_U32.X", ADD_ONE_SRC }, { 0x0799a, "I16_TO_I32.Y", ADD_ONE_SRC }, { 0x0799b, "U16_TO_U32.Y", ADD_ONE_SRC }, { 0x0799c, "I16_TO_F32.X", ADD_ONE_SRC }, { 0x0799d, "U16_TO_F32.X", ADD_ONE_SRC }, { 0x0799e, "I16_TO_F32.Y", ADD_ONE_SRC }, { 0x0799f, "U16_TO_F32.Y", ADD_ONE_SRC }, // take the low 16 bits, and expand it to a 32-bit float { 0x079a2, "F16_TO_F32.X", ADD_ONE_SRC }, // take the high 16 bits, ... { 0x079a3, "F16_TO_F32.Y", ADD_ONE_SRC }, { 0x07b2b, "SWZ.YX.v2i16", ADD_ONE_SRC }, { 0x07b2c, "NOP", ADD_ONE_SRC }, { 0x07b29, "SWZ.XX.v2i16", ADD_ONE_SRC }, // Logically, this should be SWZ.XY, but that's equivalent to a move, and // this seems to be the canonical way the blob generates a MOV. { 0x07b2d, "MOV", ADD_ONE_SRC }, { 0x07b2f, "SWZ.YY.v2i16", ADD_ONE_SRC }, // Given a floating point number m * 2^e, returns m ^ 2^{-1}. { 0x07b65, "FRCP_FREXPM", ADD_ONE_SRC }, { 0x07b75, "FSQRT_FREXPM", ADD_ONE_SRC }, { 0x07b8d, "FRCP_FREXPE", ADD_ONE_SRC }, { 0x07ba5, "FSQRT_FREXPE", ADD_ONE_SRC }, { 0x07bad, "FRSQ_FREXPE", ADD_ONE_SRC }, // From the ARM patent US20160364209A1: // "Decompose v (the input) into numbers x1 and s such that v = x1 * 2^s, // and x1 is a floating point value in a predetermined range where the // value 1 is within the range and not at one extremity of the range (e.g. // choose a range where 1 is towards middle of range)." // // This computes s. { 0x07bc5, "FLOG_FREXPE", ADD_ONE_SRC }, { 0x07d45, "CEIL", ADD_ONE_SRC }, { 0x07d85, "FLOOR", ADD_ONE_SRC }, { 0x07f18, "LSHIFT_ADD_HIGH32.i32", ADD_TWO_SRC }, { 0x08000, "LD_ATTR.f16", ADD_LOAD_ATTR, true }, { 0x08100, "LD_ATTR.v2f16", ADD_LOAD_ATTR, true }, { 0x08200, "LD_ATTR.v3f16", ADD_LOAD_ATTR, true }, { 0x08300, "LD_ATTR.v4f16", ADD_LOAD_ATTR, true }, { 0x08400, "LD_ATTR.f32", ADD_LOAD_ATTR, true }, { 0x08500, "LD_ATTR.v3f32", ADD_LOAD_ATTR, true }, { 0x08600, "LD_ATTR.v3f32", ADD_LOAD_ATTR, true }, { 0x08700, "LD_ATTR.v4f32", ADD_LOAD_ATTR, true }, { 0x08800, "LD_ATTR.i32", ADD_LOAD_ATTR, true }, { 0x08900, "LD_ATTR.v3i32", ADD_LOAD_ATTR, true }, { 0x08a00, "LD_ATTR.v3i32", ADD_LOAD_ATTR, true }, { 0x08b00, "LD_ATTR.v4i32", ADD_LOAD_ATTR, true }, { 0x08c00, "LD_ATTR.u32", ADD_LOAD_ATTR, true }, { 0x08d00, "LD_ATTR.v3u32", ADD_LOAD_ATTR, true }, { 0x08e00, "LD_ATTR.v3u32", ADD_LOAD_ATTR, true }, { 0x08f00, "LD_ATTR.v4u32", ADD_LOAD_ATTR, true }, { 0x0a000, "LD_VAR.32", ADD_VARYING_INTERP, true }, { 0x0b000, "TEX", ADD_TEX_COMPACT, true }, { 0x0c188, "LOAD.i32", ADD_TWO_SRC, true }, { 0x0c1a0, "LD_UBO.i32", ADD_TWO_SRC, true }, { 0x0c1b8, "LD_SCRATCH.v2i32", ADD_TWO_SRC, true }, { 0x0c1c8, "LOAD.v2i32", ADD_TWO_SRC, true }, { 0x0c1e0, "LD_UBO.v2i32", ADD_TWO_SRC, true }, { 0x0c1f8, "LD_SCRATCH.v2i32", ADD_TWO_SRC, true }, { 0x0c208, "LOAD.v4i32", ADD_TWO_SRC, true }, // src0 = offset, src1 = binding { 0x0c220, "LD_UBO.v4i32", ADD_TWO_SRC, true }, { 0x0c238, "LD_SCRATCH.v4i32", ADD_TWO_SRC, true }, { 0x0c248, "STORE.v4i32", ADD_TWO_SRC, true }, { 0x0c278, "ST_SCRATCH.v4i32", ADD_TWO_SRC, true }, { 0x0c588, "STORE.i32", ADD_TWO_SRC, true }, { 0x0c5b8, "ST_SCRATCH.i32", ADD_TWO_SRC, true }, { 0x0c5c8, "STORE.v2i32", ADD_TWO_SRC, true }, { 0x0c5f8, "ST_SCRATCH.v2i32", ADD_TWO_SRC, true }, { 0x0c648, "LOAD.u16", ADD_TWO_SRC, true }, // zero-extends { 0x0ca88, "LOAD.v3i32", ADD_TWO_SRC, true }, { 0x0caa0, "LD_UBO.v3i32", ADD_TWO_SRC, true }, { 0x0cab8, "LD_SCRATCH.v3i32", ADD_TWO_SRC, true }, { 0x0cb88, "STORE.v3i32", ADD_TWO_SRC, true }, { 0x0cbb8, "ST_SCRATCH.v3i32", ADD_TWO_SRC, true }, // *_FAST does not exist on G71 (added to G51, G72, and everything after) { 0x0cc00, "FRCP_FAST.f32", ADD_ONE_SRC }, { 0x0cc20, "FRSQ_FAST.f32", ADD_ONE_SRC }, // Given a floating point number m * 2^e, produces a table-based // approximation of 2/m using the top 17 bits. Includes special cases for // infinity, NaN, and zero, and copies the sign bit. { 0x0ce00, "FRCP_TABLE", ADD_ONE_SRC }, // Exists on G71 { 0x0ce10, "FRCP_FAST.f16.X", ADD_ONE_SRC }, // A similar table for inverse square root, using the high 17 bits of the // mantissa as well as the low bit of the exponent. { 0x0ce20, "FRSQ_TABLE", ADD_ONE_SRC }, { 0x0ce30, "FRCP_FAST.f16.Y", ADD_ONE_SRC }, { 0x0ce50, "FRSQ_FAST.f16.X", ADD_ONE_SRC }, // Used in the argument reduction for log. Given a floating-point number // m * 2^e, uses the top 4 bits of m to produce an approximation to 1/m // with the exponent forced to 0 and only the top 5 bits are nonzero. 0, // infinity, and NaN all return 1.0. // See the ARM patent for more information. { 0x0ce60, "FRCP_APPROX", ADD_ONE_SRC }, { 0x0ce70, "FRSQ_FAST.f16.Y", ADD_ONE_SRC }, { 0x0cf40, "ATAN_ASSIST", ADD_TWO_SRC }, { 0x0cf48, "ATAN_TABLE", ADD_TWO_SRC }, { 0x0cf50, "SIN_TABLE", ADD_ONE_SRC }, { 0x0cf51, "COS_TABLE", ADD_ONE_SRC }, { 0x0cf58, "EXP_TABLE", ADD_ONE_SRC }, { 0x0cf60, "FLOG2_TABLE", ADD_ONE_SRC }, { 0x0cf64, "FLOGE_TABLE", ADD_ONE_SRC }, { 0x0d000, "BRANCH", ADD_BRANCH }, // For each bit i, return src2[i] ? src0[i] : src1[i]. In other words, this // is the same as (src2 & src0) | (~src2 & src1). { 0x0e8c0, "MUX", ADD_THREE_SRC }, { 0x0e9b0, "ATAN_LDEXP.Y.f32", ADD_TWO_SRC }, { 0x0e9b8, "ATAN_LDEXP.X.f32", ADD_TWO_SRC }, { 0x0ea60, "SEL.XX.i16", ADD_TWO_SRC }, { 0x0ea70, "SEL.XY.i16", ADD_TWO_SRC }, { 0x0ea68, "SEL.YX.i16", ADD_TWO_SRC }, { 0x0ea78, "SEL.YY.i16", ADD_TWO_SRC }, { 0x0ec00, "F32_TO_F16", ADD_TWO_SRC }, { 0x0f640, "ICMP.GL.GT", ADD_TWO_SRC }, // src0 > src1 ? 1 : 0 { 0x0f648, "ICMP.GL.GE", ADD_TWO_SRC }, { 0x0f650, "UCMP.GL.GT", ADD_TWO_SRC }, { 0x0f658, "UCMP.GL.GE", ADD_TWO_SRC }, { 0x0f660, "ICMP.GL.EQ", ADD_TWO_SRC }, { 0x0f6c0, "ICMP.D3D.GT", ADD_TWO_SRC }, // src0 > src1 ? ~0 : 0 { 0x0f6c8, "ICMP.D3D.GE", ADD_TWO_SRC }, { 0x0f6d0, "UCMP.D3D.GT", ADD_TWO_SRC }, { 0x0f6d8, "UCMP.D3D.GE", ADD_TWO_SRC }, { 0x0f6e0, "ICMP.D3D.EQ", ADD_TWO_SRC }, { 0x10000, "MAX.v2f16", ADD_FMINMAX16 }, { 0x11000, "ADD_MSCALE.f32", ADD_FADDMscale }, { 0x12000, "MIN.v2f16", ADD_FMINMAX16 }, { 0x14000, "ADD.v2f16", ADD_FADD16 }, { 0x17000, "FCMP.D3D", ADD_FCMP16 }, { 0x178c0, "ADD.i32", ADD_TWO_SRC }, { 0x17900, "ADD.v2i16", ADD_TWO_SRC }, { 0x17ac0, "SUB.i32", ADD_TWO_SRC }, { 0x17c10, "ADDC.i32", ADD_TWO_SRC }, // adds src0 to the bottom bit of src1 { 0x17d80, "ADD.i32.i16.X", ADD_TWO_SRC }, { 0x17d90, "ADD.i32.u16.X", ADD_TWO_SRC }, { 0x17dc0, "ADD.i32.i16.Y", ADD_TWO_SRC }, { 0x17dd0, "ADD.i32.u16.Y", ADD_TWO_SRC }, // Compute varying address and datatype (for storing in the vertex shader), // and store the vec3 result in the data register. The result is passed as // the 3 normal arguments to ST_VAR. { 0x18000, "LD_VAR_ADDR.f16", ADD_VARYING_ADDRESS, true }, { 0x18100, "LD_VAR_ADDR.f32", ADD_VARYING_ADDRESS, true }, { 0x18200, "LD_VAR_ADDR.i32", ADD_VARYING_ADDRESS, true }, { 0x18300, "LD_VAR_ADDR.u32", ADD_VARYING_ADDRESS, true }, // Implements alpha-to-coverage, as well as possibly the late depth and // stencil tests. The first source is the existing sample mask in R60 // (possibly modified by gl_SampleMask), and the second source is the alpha // value. The sample mask is written right away based on the // alpha-to-coverage result using the normal register write mechanism, // since that doesn't need to read from any memory, and then written again // later based on the result of the stencil and depth tests using the // special register. { 0x191e8, "ATEST.f32", ADD_TWO_SRC, true }, { 0x191f0, "ATEST.X.f16", ADD_TWO_SRC, true }, { 0x191f8, "ATEST.Y.f16", ADD_TWO_SRC, true }, // store a varying given the address and datatype from LD_VAR_ADDR { 0x19300, "ST_VAR.v1", ADD_THREE_SRC, true }, { 0x19340, "ST_VAR.v2", ADD_THREE_SRC, true }, { 0x19380, "ST_VAR.v3", ADD_THREE_SRC, true }, { 0x193c0, "ST_VAR.v4", ADD_THREE_SRC, true }, // This takes the sample coverage mask (computed by ATEST above) as a // regular argument, in addition to the vec4 color in the special register. { 0x1952c, "BLEND", ADD_BLENDING, true }, { 0x1a000, "LD_VAR.16", ADD_VARYING_INTERP, true }, { 0x1ae60, "TEX", ADD_TEX, true }, { 0x1c000, "RSHIFT_NAND.i32", ADD_THREE_SRC }, { 0x1c300, "RSHIFT_OR.i32", ADD_THREE_SRC }, { 0x1c400, "RSHIFT_AND.i32", ADD_THREE_SRC }, { 0x1c700, "RSHIFT_NOR.i32", ADD_THREE_SRC }, { 0x1c800, "LSHIFT_NAND.i32", ADD_THREE_SRC }, { 0x1cb00, "LSHIFT_OR.i32", ADD_THREE_SRC }, { 0x1cc00, "LSHIFT_AND.i32", ADD_THREE_SRC }, { 0x1cf00, "LSHIFT_NOR.i32", ADD_THREE_SRC }, { 0x1d000, "RSHIFT_XOR.i32", ADD_THREE_SRC }, { 0x1d100, "RSHIFT_XNOR.i32", ADD_THREE_SRC }, { 0x1d200, "LSHIFT_XOR.i32", ADD_THREE_SRC }, { 0x1d300, "LSHIFT_XNOR.i32", ADD_THREE_SRC }, { 0x1d400, "LSHIFT_ADD.i32", ADD_THREE_SRC }, { 0x1d500, "LSHIFT_SUB.i32", ADD_THREE_SRC }, { 0x1d500, "LSHIFT_RSUB.i32", ADD_THREE_SRC }, { 0x1d700, "RSHIFT_ADD.i32", ADD_THREE_SRC }, { 0x1d800, "RSHIFT_SUB.i32", ADD_THREE_SRC }, { 0x1d900, "RSHIFT_RSUB.i32", ADD_THREE_SRC }, { 0x1da00, "ARSHIFT_ADD.i32", ADD_THREE_SRC }, { 0x1db00, "ARSHIFT_SUB.i32", ADD_THREE_SRC }, { 0x1dc00, "ARSHIFT_RSUB.i32", ADD_THREE_SRC }, { 0x1dd18, "OR.i32", ADD_TWO_SRC }, { 0x1dd20, "AND.i32", ADD_TWO_SRC }, { 0x1dd60, "LSHIFT.i32", ADD_TWO_SRC }, { 0x1dd50, "XOR.i32", ADD_TWO_SRC }, { 0x1dd80, "RSHIFT.i32", ADD_TWO_SRC }, { 0x1dda0, "ARSHIFT.i32", ADD_TWO_SRC }, }; static struct add_op_info find_add_op_info(unsigned op) { for (unsigned i = 0; i < ARRAY_SIZE(add_op_infos); i++) { unsigned opCmp = ~0; switch (add_op_infos[i].src_type) { case ADD_ONE_SRC: case ADD_BLENDING: opCmp = op; break; case ADD_TWO_SRC: opCmp = op & ~0x7; break; case ADD_THREE_SRC: opCmp = op & ~0x3f; break; case ADD_TEX: opCmp = op & ~0xf; break; case ADD_FADD: case ADD_FMINMAX: case ADD_FADD16: opCmp = op & ~0x1fff; break; case ADD_FMINMAX16: case ADD_FADDMscale: opCmp = op & ~0xfff; break; case ADD_FCMP: case ADD_FCMP16: opCmp = op & ~0x7ff; break; case ADD_TEX_COMPACT: opCmp = op & ~0x3ff; break; case ADD_VARYING_INTERP: opCmp = op & ~0x7ff; break; case ADD_VARYING_ADDRESS: opCmp = op & ~0xff; break; case ADD_LOAD_ATTR: opCmp = op & ~0x7f; break; case ADD_BRANCH: opCmp = op & ~0xfff; break; default: opCmp = ~0; break; } if (add_op_infos[i].op == opCmp) return add_op_infos[i]; } struct add_op_info info; snprintf(info.name, sizeof(info.name), "op%04x", op); info.op = op; info.src_type = ADD_TWO_SRC; info.has_data_reg = true; return info; } static void dump_add(uint64_t word, struct bifrost_regs regs, struct bifrost_regs next_regs, uint64_t *consts, unsigned data_reg, unsigned offset, bool verbose) { if (verbose) { printf("# ADD: %016" PRIx64 "\n", word); } struct bifrost_add_inst ADD; memcpy((char *) &ADD, (char *) &word, sizeof(ADD)); struct add_op_info info = find_add_op_info(ADD.op); printf("%s", info.name); // float16 seems like it doesn't support output modifiers if (info.src_type == ADD_FADD || info.src_type == ADD_FMINMAX) { // output modifiers dump_output_mod(bits(ADD.op, 8, 10)); if (info.src_type == ADD_FADD) dump_round_mode(bits(ADD.op, 10, 12)); else dump_minmax_mode(bits(ADD.op, 10, 12)); } else if (info.src_type == ADD_FCMP || info.src_type == ADD_FCMP16) { dump_fcmp(bits(ADD.op, 3, 6)); if (info.src_type == ADD_FCMP) printf(".f32"); else printf(".v2f16"); } else if (info.src_type == ADD_FADDMscale) { switch ((ADD.op >> 6) & 0x7) { case 0: break; // causes GPU hangs on G71 case 1: printf(".invalid"); break; // Same as usual outmod value. case 2: printf(".clamp_0_1"); break; // If src0 is infinite or NaN, flush it to zero so that the other // source is passed through unmodified. case 3: printf(".flush_src0_inf_nan"); break; // Vice versa. case 4: printf(".flush_src1_inf_nan"); break; // Every other case seems to behave the same as the above? default: printf(".unk%d", (ADD.op >> 6) & 0x7); break; } } else if (info.src_type == ADD_VARYING_INTERP) { if (ADD.op & 0x200) printf(".reuse"); if (ADD.op & 0x400) printf(".flat"); switch ((ADD.op >> 7) & 0x3) { case 0: printf(".per_frag"); break; case 1: printf(".centroid"); break; case 2: break; case 3: printf(".explicit"); break; } printf(".v%d", ((ADD.op >> 5) & 0x3) + 1); } else if (info.src_type == ADD_BRANCH) { enum branch_code branchCode = (enum branch_code) ((ADD.op >> 6) & 0x3f); if (branchCode == BR_ALWAYS) { // unconditional branch } else { enum branch_cond cond = (enum branch_cond) ((ADD.op >> 6) & 0x7); enum branch_bit_size size = (enum branch_bit_size) ((ADD.op >> 9) & 0x7); bool portSwapped = (ADD.op & 0x7) < ADD.src0; // See the comment in branch_bit_size if (size == BR_SIZE_16YX0) portSwapped = true; if (size == BR_SIZE_16YX1) portSwapped = false; // These sizes are only for floating point comparisons, so the // non-floating-point comparisons are reused to encode the flipped // versions. if (size == BR_SIZE_32_AND_16X || size == BR_SIZE_32_AND_16Y) portSwapped = false; // There's only one argument, so we reuse the extra argument to // encode this. if (size == BR_SIZE_ZERO) portSwapped = !(ADD.op & 1); switch (cond) { case BR_COND_LT: if (portSwapped) printf(".LT.u"); else printf(".LT.i"); break; case BR_COND_LE: if (size == BR_SIZE_32_AND_16X || size == BR_SIZE_32_AND_16Y) { printf(".UNE.f"); } else { if (portSwapped) printf(".LE.u"); else printf(".LE.i"); } break; case BR_COND_GT: if (portSwapped) printf(".GT.u"); else printf(".GT.i"); break; case BR_COND_GE: if (portSwapped) printf(".GE.u"); else printf(".GE.i"); break; case BR_COND_EQ: if (portSwapped) printf(".NE.i"); else printf(".EQ.i"); break; case BR_COND_OEQ: if (portSwapped) printf(".UNE.f"); else printf(".OEQ.f"); break; case BR_COND_OGT: if (portSwapped) printf(".OGT.unk.f"); else printf(".OGT.f"); break; case BR_COND_OLT: if (portSwapped) printf(".OLT.unk.f"); else printf(".OLT.f"); break; } switch (size) { case BR_SIZE_32: case BR_SIZE_32_AND_16X: case BR_SIZE_32_AND_16Y: printf("32"); break; case BR_SIZE_16XX: case BR_SIZE_16YY: case BR_SIZE_16YX0: case BR_SIZE_16YX1: printf("16"); break; case BR_SIZE_ZERO: { unsigned ctrl = (ADD.op >> 1) & 0x3; if (ctrl == 0) printf("32.Z"); else printf("16.Z"); break; } } } } printf(" "); struct bifrost_reg_ctrl next_ctrl = DecodeRegCtrl(next_regs); if (next_ctrl.add_write_unit != REG_WRITE_NONE) { printf("{R%d, T1}, ", GetRegToWrite(next_ctrl.add_write_unit, next_regs)); } else { printf("T1, "); } switch (info.src_type) { case ADD_BLENDING: // Note: in this case, regs.uniform_const == location | 0x8 // This probably means we can't load uniforms or immediates in the // same instruction. This re-uses the encoding that normally means // "disabled", where the low 4 bits are ignored. Perhaps the extra // 0x8 or'd in indicates this is happening. printf("location:%d, ", regs.uniform_const & 0x7); // fallthrough case ADD_ONE_SRC: dump_src(ADD.src0, regs, consts, false); break; case ADD_TEX: case ADD_TEX_COMPACT: { int tex_index; int sampler_index; bool dualTex = false; if (info.src_type == ADD_TEX_COMPACT) { tex_index = (ADD.op >> 3) & 0x7; sampler_index = (ADD.op >> 7) & 0x7; bool unknown = (ADD.op & 0x40); // TODO: figure out if the unknown bit is ever 0 if (!unknown) printf("unknown "); } else { uint64_t constVal = get_const(consts, regs); uint32_t controlBits = (ADD.op & 0x8) ? (constVal >> 32) : constVal; struct bifrost_tex_ctrl ctrl; memcpy((char *) &ctrl, (char *) &controlBits, sizeof(ctrl)); // TODO: figure out what actually triggers dual-tex if (ctrl.result_type == 9) { struct bifrost_dual_tex_ctrl dualCtrl; memcpy((char *) &dualCtrl, (char *) &controlBits, sizeof(ctrl)); printf("(dualtex) tex0:%d samp0:%d tex1:%d samp1:%d ", dualCtrl.tex_index0, dualCtrl.sampler_index0, dualCtrl.tex_index1, dualCtrl.sampler_index1); if (dualCtrl.unk0 != 3) printf("unk:%d ", dualCtrl.unk0); dualTex = true; } else { if (ctrl.no_merge_index) { tex_index = ctrl.tex_index; sampler_index = ctrl.sampler_index; } else { tex_index = sampler_index = ctrl.tex_index; unsigned unk = ctrl.sampler_index >> 2; if (unk != 3) printf("unk:%d ", unk); if (ctrl.sampler_index & 1) tex_index = -1; if (ctrl.sampler_index & 2) sampler_index = -1; } if (ctrl.unk0 != 3) printf("unk0:%d ", ctrl.unk0); if (ctrl.unk1) printf("unk1 "); if (ctrl.unk2 != 0xf) printf("unk2:%x ", ctrl.unk2); switch (ctrl.result_type) { case 0x4: printf("f32 "); break; case 0xe: printf("i32 "); break; case 0xf: printf("u32 "); break; default: printf("unktype(%x) ", ctrl.result_type); } switch (ctrl.tex_type) { case 0: printf("cube "); break; case 1: printf("buffer "); break; case 2: printf("2D "); break; case 3: printf("3D "); break; } if (ctrl.is_shadow) printf("shadow "); if (ctrl.is_array) printf("array "); if (!ctrl.filter) { if (ctrl.calc_gradients) { int comp = (controlBits >> 20) & 0x3; printf("txg comp:%d ", comp); } else { printf("txf "); } } else { if (!ctrl.not_supply_lod) { if (ctrl.compute_lod) printf("lod_bias "); else printf("lod "); } if (!ctrl.calc_gradients) printf("grad "); } if (ctrl.texel_offset) printf("offset "); } } if (!dualTex) { if (tex_index == -1) printf("tex:indirect "); else printf("tex:%d ", tex_index); if (sampler_index == -1) printf("samp:indirect "); else printf("samp:%d ", sampler_index); } break; } case ADD_VARYING_INTERP: { unsigned addr = ADD.op & 0x1f; if (addr < 0b10100) { // direct addr printf("%d", addr); } else if (addr < 0b11000) { if (addr == 22) printf("fragw"); else if (addr == 23) printf("fragz"); else printf("unk%d", addr); } else { dump_src(ADD.op & 0x7, regs, consts, false); } printf(", "); dump_src(ADD.src0, regs, consts, false); break; } case ADD_VARYING_ADDRESS: { dump_src(ADD.src0, regs, consts, false); printf(", "); dump_src(ADD.op & 0x7, regs, consts, false); printf(", "); unsigned location = (ADD.op >> 3) & 0x1f; if (location < 16) { printf("location:%d", location); } else if (location == 20) { printf("location:%u", (uint32_t) get_const(consts, regs)); } else if (location == 21) { printf("location:%u", (uint32_t) (get_const(consts, regs) >> 32)); } else { printf("location:%d(unk)", location); } break; } case ADD_LOAD_ATTR: printf("location:%d, ", (ADD.op >> 3) & 0xf); case ADD_TWO_SRC: dump_src(ADD.src0, regs, consts, false); printf(", "); dump_src(ADD.op & 0x7, regs, consts, false); break; case ADD_THREE_SRC: dump_src(ADD.src0, regs, consts, false); printf(", "); dump_src(ADD.op & 0x7, regs, consts, false); printf(", "); dump_src((ADD.op >> 3) & 0x7, regs, consts, false); break; case ADD_FADD: case ADD_FMINMAX: if (ADD.op & 0x10) printf("-"); if (ADD.op & 0x1000) printf("abs("); dump_src(ADD.src0, regs, consts, false); switch ((ADD.op >> 6) & 0x3) { case 3: printf(".x"); break; default: break; } if (ADD.op & 0x1000) printf(")"); printf(", "); if (ADD.op & 0x20) printf("-"); if (ADD.op & 0x8) printf("abs("); dump_src(ADD.op & 0x7, regs, consts, false); switch ((ADD.op >> 6) & 0x3) { case 1: case 3: printf(".x"); break; case 2: printf(".y"); break; case 0: break; default: printf(".unk"); break; } if (ADD.op & 0x8) printf(")"); break; case ADD_FADD16: if (ADD.op & 0x10) printf("-"); if (ADD.op & 0x1000) printf("abs("); dump_src(ADD.src0, regs, consts, false); if (ADD.op & 0x1000) printf(")"); dump_16swizzle((ADD.op >> 6) & 0x3); printf(", "); if (ADD.op & 0x20) printf("-"); if (ADD.op & 0x8) printf("abs("); dump_src(ADD.op & 0x7, regs, consts, false); dump_16swizzle((ADD.op >> 8) & 0x3); if (ADD.op & 0x8) printf(")"); break; case ADD_FMINMAX16: { bool abs1 = ADD.op & 0x8; bool abs2 = (ADD.op & 0x7) < ADD.src0; if (ADD.op & 0x10) printf("-"); if (abs1 || abs2) printf("abs("); dump_src(ADD.src0, regs, consts, false); dump_16swizzle((ADD.op >> 6) & 0x3); if (abs1 || abs2) printf(")"); printf(", "); if (ADD.op & 0x20) printf("-"); if (abs1 && abs2) printf("abs("); dump_src(ADD.op & 0x7, regs, consts, false); dump_16swizzle((ADD.op >> 8) & 0x3); if (abs1 && abs2) printf(")"); break; } case ADD_FADDMscale: { if (ADD.op & 0x400) printf("-"); if (ADD.op & 0x200) printf("abs("); dump_src(ADD.src0, regs, consts, false); if (ADD.op & 0x200) printf(")"); printf(", "); if (ADD.op & 0x800) printf("-"); dump_src(ADD.op & 0x7, regs, consts, false); printf(", "); dump_src((ADD.op >> 3) & 0x7, regs, consts, false); break; } case ADD_FCMP: if (ADD.op & 0x400) { printf("-"); } if (ADD.op & 0x100) { printf("abs("); } dump_src(ADD.src0, regs, consts, false); switch ((ADD.op >> 6) & 0x3) { case 3: printf(".x"); break; default: break; } if (ADD.op & 0x100) { printf(")"); } printf(", "); if (ADD.op & 0x200) { printf("abs("); } dump_src(ADD.op & 0x7, regs, consts, false); switch ((ADD.op >> 6) & 0x3) { case 1: case 3: printf(".x"); break; case 2: printf(".y"); break; case 0: break; default: printf(".unk"); break; } if (ADD.op & 0x200) { printf(")"); } break; case ADD_FCMP16: dump_src(ADD.src0, regs, consts, false); dump_16swizzle((ADD.op >> 6) & 0x3); printf(", "); dump_src(ADD.op & 0x7, regs, consts, false); dump_16swizzle((ADD.op >> 8) & 0x3); break; case ADD_BRANCH: { enum branch_code code = (enum branch_code) ((ADD.op >> 6) & 0x3f); enum branch_bit_size size = (enum branch_bit_size) ((ADD.op >> 9) & 0x7); if (code != BR_ALWAYS) { dump_src(ADD.src0, regs, consts, false); switch (size) { case BR_SIZE_16XX: printf(".x"); break; case BR_SIZE_16YY: case BR_SIZE_16YX0: case BR_SIZE_16YX1: printf(".y"); break; case BR_SIZE_ZERO: { unsigned ctrl = (ADD.op >> 1) & 0x3; switch (ctrl) { case 1: printf(".y"); break; case 2: printf(".x"); break; default: break; } } default: break; } printf(", "); } if (code != BR_ALWAYS && size != BR_SIZE_ZERO) { dump_src(ADD.op & 0x7, regs, consts, false); switch (size) { case BR_SIZE_16XX: case BR_SIZE_16YX0: case BR_SIZE_16YX1: case BR_SIZE_32_AND_16X: printf(".x"); break; case BR_SIZE_16YY: case BR_SIZE_32_AND_16Y: printf(".y"); break; default: break; } printf(", "); } // I haven't had the chance to test if this actually specifies the // branch offset, since I couldn't get it to produce values other // than 5 (uniform/const high), but these three bits are always // consistent across branch instructions, so it makes sense... int offsetSrc = (ADD.op >> 3) & 0x7; if (offsetSrc == 4 || offsetSrc == 5) { // If the offset is known/constant, we can decode it uint32_t raw_offset; if (offsetSrc == 4) raw_offset = get_const(consts, regs); else raw_offset = get_const(consts, regs) >> 32; // The high 4 bits are flags, while the rest is the // twos-complement offset in bytes (here we convert to // clauses). int32_t branch_offset = ((int32_t) raw_offset << 4) >> 8; // If high4 is the high 4 bits of the last 64-bit constant, // this is calculated as (high4 + 4) & 0xf, or 0 if the branch // offset itself is the last constant. Not sure if this is // actually used, or just garbage in unused bits, but in any // case, we can just ignore it here since it's redundant. Note // that if there is any padding, this will be 4 since the // padding counts as the last constant. unsigned flags = raw_offset >> 28; (void) flags; // Note: the offset is in bytes, relative to the beginning of the // current clause, so a zero offset would be a loop back to the // same clause (annoyingly different from Midgard). printf("clause_%d", offset + branch_offset); } else { dump_src(offsetSrc, regs, consts, false); } } } if (info.has_data_reg) { printf(", R%d", data_reg); } printf("\n"); } void dump_instr(const struct bifrost_alu_inst *instr, struct bifrost_regs next_regs, uint64_t *consts, unsigned data_reg, unsigned offset, bool verbose) { struct bifrost_regs regs; memcpy((char *) ®s, (char *) &instr->reg_bits, sizeof(regs)); if (verbose) { printf("# regs: %016" PRIx64 "\n", instr->reg_bits); dump_regs(regs); } dump_fma(instr->fma_bits, regs, next_regs, consts, verbose); dump_add(instr->add_bits, regs, next_regs, consts, data_reg, offset, verbose); } bool dump_clause(uint32_t *words, unsigned *size, unsigned offset, bool verbose) { // State for a decoded clause struct bifrost_alu_inst instrs[8] = {}; uint64_t consts[6] = {}; unsigned num_instrs = 0; unsigned num_consts = 0; uint64_t header_bits = 0; bool stopbit = false; unsigned i; for (i = 0; ; i++, words += 4) { if (verbose) { printf("# "); for (int j = 0; j < 4; j++) printf("%08x ", words[3 - j]); // low bit on the right printf("\n"); } unsigned tag = bits(words[0], 0, 8); // speculatively decode some things that are common between many formats, so we can share some code struct bifrost_alu_inst main_instr = {}; // 20 bits main_instr.add_bits = bits(words[2], 2, 32 - 13); // 23 bits main_instr.fma_bits = bits(words[1], 11, 32) | bits(words[2], 0, 2) << (32 - 11); // 35 bits main_instr.reg_bits = ((uint64_t) bits(words[1], 0, 11)) << 24 | (uint64_t) bits(words[0], 8, 32); uint64_t const0 = bits(words[0], 8, 32) << 4 | (uint64_t) words[1] << 28 | bits(words[2], 0, 4) << 60; uint64_t const1 = bits(words[2], 4, 32) << 4 | (uint64_t) words[3] << 32; bool stop = tag & 0x40; if (verbose) { printf("# tag: 0x%02x\n", tag); } if (tag & 0x80) { unsigned idx = stop ? 5 : 2; main_instr.add_bits |= ((tag >> 3) & 0x7) << 17; instrs[idx + 1] = main_instr; instrs[idx].add_bits = bits(words[3], 0, 17) | ((tag & 0x7) << 17); instrs[idx].fma_bits |= bits(words[2], 19, 32) << 10; consts[0] = bits(words[3], 17, 32) << 4; } else { bool done = false; switch ((tag >> 3) & 0x7) { case 0x0: switch (tag & 0x7) { case 0x3: main_instr.add_bits |= bits(words[3], 29, 32) << 17; instrs[1] = main_instr; num_instrs = 2; done = stop; break; case 0x4: instrs[2].add_bits = bits(words[3], 0, 17) | bits(words[3], 29, 32) << 17; instrs[2].fma_bits |= bits(words[2], 19, 32) << 10; consts[0] = const0; num_instrs = 3; num_consts = 1; done = stop; break; case 0x1: case 0x5: instrs[2].add_bits = bits(words[3], 0, 17) | bits(words[3], 29, 32) << 17; instrs[2].fma_bits |= bits(words[2], 19, 32) << 10; main_instr.add_bits |= bits(words[3], 26, 29) << 17; instrs[3] = main_instr; if ((tag & 0x7) == 0x5) { num_instrs = 4; done = stop; } break; case 0x6: instrs[5].add_bits = bits(words[3], 0, 17) | bits(words[3], 29, 32) << 17; instrs[5].fma_bits |= bits(words[2], 19, 32) << 10; consts[0] = const0; num_instrs = 6; num_consts = 1; done = stop; break; case 0x7: instrs[5].add_bits = bits(words[3], 0, 17) | bits(words[3], 29, 32) << 17; instrs[5].fma_bits |= bits(words[2], 19, 32) << 10; main_instr.add_bits |= bits(words[3], 26, 29) << 17; instrs[6] = main_instr; num_instrs = 7; done = stop; break; default: printf("unknown tag bits 0x%02x\n", tag); } break; case 0x2: case 0x3: { unsigned idx = ((tag >> 3) & 0x7) == 2 ? 4 : 7; main_instr.add_bits |= (tag & 0x7) << 17; instrs[idx] = main_instr; consts[0] |= (bits(words[2], 19, 32) | ((uint64_t) words[3] << 13)) << 19; num_consts = 1; num_instrs = idx + 1; done = stop; break; } case 0x4: { unsigned idx = stop ? 4 : 1; main_instr.add_bits |= (tag & 0x7) << 17; instrs[idx] = main_instr; instrs[idx + 1].fma_bits |= bits(words[3], 22, 32); instrs[idx + 1].reg_bits = bits(words[2], 19, 32) | (bits(words[3], 0, 22) << (32 - 19)); break; } case 0x1: // only constants can come after this num_instrs = 1; done = stop; case 0x5: header_bits = bits(words[2], 19, 32) | ((uint64_t) words[3] << (32 - 19)); main_instr.add_bits |= (tag & 0x7) << 17; instrs[0] = main_instr; break; case 0x6: case 0x7: { unsigned pos = tag & 0xf; // note that `pos' encodes both the total number of // instructions and the position in the constant stream, // presumably because decoded constants and instructions // share a buffer in the decoder, but we only care about // the position in the constant stream; the total number of // instructions is redundant. unsigned const_idx = 7; switch (pos) { case 0: case 1: case 2: case 6: const_idx = 0; break; case 3: case 4: case 7: case 9: const_idx = 1; break; case 5: case 0xa: const_idx = 2; break; case 8: case 0xb: case 0xc: const_idx = 3; break; case 0xd: const_idx = 4; break; default: printf("# unknown pos 0x%x\n", pos); } if (num_consts < const_idx + 2) num_consts = const_idx + 2; consts[const_idx] = const0; consts[const_idx + 1] = const1; done = stop; break; } default: break; } if (done) break; } } *size = i + 1; if (verbose) { printf("# header: %012" PRIx64 "\n", header_bits); } struct bifrost_header header; memcpy((char *) &header, (char *) &header_bits, sizeof(struct bifrost_header)); dump_header(header, verbose); if (!header.no_end_of_shader) stopbit = true; printf("{\n"); for (i = 0; i < num_instrs; i++) { struct bifrost_regs next_regs; if (i + 1 == num_instrs) { memcpy((char *) &next_regs, (char *) &instrs[0].reg_bits, sizeof(next_regs)); } else { memcpy((char *) &next_regs, (char *) &instrs[i + 1].reg_bits, sizeof(next_regs)); } dump_instr(&instrs[i], next_regs, consts, header.datareg, offset, verbose); } printf("}\n"); if (verbose) { for (unsigned i = 0; i < num_consts; i++) { printf("# const%d: %08lx\n", 2 * i, consts[i] & 0xffffffff); printf("# const%d: %08" PRIx64 "\n", 2 * i + 1, consts[i] >> 32); } } return stopbit; } void disassemble_bifrost(uint8_t *code, size_t size, bool verbose) { uint32_t *words = (uint32_t *) code; uint32_t *words_end = words + (size / 4); // used for displaying branch targets unsigned offset = 0; while (words != words_end) { // we don't know what the program-end bit is quite yet, so for now just // assume that an all-0 quadword is padding uint32_t zero[4] = {}; if (memcmp(words, zero, 4 * sizeof(uint32_t)) == 0) break; printf("clause_%d:\n", offset); unsigned size; if (dump_clause(words, &size, offset, verbose) == true) { break; } words += size * 4; offset += size; } }