/* * Copyright 2008 Nicolai Haehnle. * SPDX-License-Identifier: MIT */ /** * @file * * Shareable transformations that transform "special" ALU instructions * into ALU instructions that are supported by hardware. * */ #include "radeon_program_alu.h" #include "radeon_compiler.h" #include "radeon_compiler_util.h" #include "radeon_dataflow.h" #include "util/log.h" static struct rc_instruction * emit1(struct radeon_compiler *c, struct rc_instruction *after, rc_opcode Opcode, struct rc_sub_instruction *base, struct rc_dst_register DstReg, struct rc_src_register SrcReg) { struct rc_instruction *fpi = rc_insert_new_instruction(c, after); if (base) { memcpy(&fpi->U.I, base, sizeof(struct rc_sub_instruction)); } fpi->U.I.Opcode = Opcode; fpi->U.I.DstReg = DstReg; fpi->U.I.SrcReg[0] = SrcReg; return fpi; } static struct rc_instruction * emit2(struct radeon_compiler *c, struct rc_instruction *after, rc_opcode Opcode, struct rc_sub_instruction *base, struct rc_dst_register DstReg, struct rc_src_register SrcReg0, struct rc_src_register SrcReg1) { struct rc_instruction *fpi = rc_insert_new_instruction(c, after); if (base) { memcpy(&fpi->U.I, base, sizeof(struct rc_sub_instruction)); } fpi->U.I.Opcode = Opcode; fpi->U.I.DstReg = DstReg; fpi->U.I.SrcReg[0] = SrcReg0; fpi->U.I.SrcReg[1] = SrcReg1; return fpi; } static struct rc_instruction * emit3(struct radeon_compiler *c, struct rc_instruction *after, rc_opcode Opcode, struct rc_sub_instruction *base, struct rc_dst_register DstReg, struct rc_src_register SrcReg0, struct rc_src_register SrcReg1, struct rc_src_register SrcReg2) { struct rc_instruction *fpi = rc_insert_new_instruction(c, after); if (base) { memcpy(&fpi->U.I, base, sizeof(struct rc_sub_instruction)); } fpi->U.I.Opcode = Opcode; fpi->U.I.DstReg = DstReg; fpi->U.I.SrcReg[0] = SrcReg0; fpi->U.I.SrcReg[1] = SrcReg1; fpi->U.I.SrcReg[2] = SrcReg2; return fpi; } static struct rc_dst_register dstregtmpmask(int index, int mask) { struct rc_dst_register dst = {0, 0, 0}; dst.File = RC_FILE_TEMPORARY; dst.Index = index; dst.WriteMask = mask; return dst; } static const struct rc_src_register builtin_one = { .File = RC_FILE_NONE, .Index = 0, .Swizzle = RC_SWIZZLE_1111}; static const struct rc_src_register srcreg_undefined = { .File = RC_FILE_NONE, .Index = 0, .Swizzle = RC_SWIZZLE_XYZW}; static struct rc_src_register srcreg(int file, int index) { struct rc_src_register src = srcreg_undefined; src.File = file; src.Index = index; return src; } static struct rc_src_register srcregswz(int file, int index, int swz) { struct rc_src_register src = srcreg_undefined; src.File = file; src.Index = index; src.Swizzle = swz; return src; } static struct rc_src_register absolute(struct rc_src_register reg) { struct rc_src_register newreg = reg; newreg.Abs = 1; newreg.Negate = RC_MASK_NONE; return newreg; } static struct rc_src_register negate(struct rc_src_register reg) { struct rc_src_register newreg = reg; newreg.Negate = newreg.Negate ^ RC_MASK_XYZW; return newreg; } static struct rc_src_register swizzle(struct rc_src_register reg, rc_swizzle x, rc_swizzle y, rc_swizzle z, rc_swizzle w) { struct rc_src_register swizzled = reg; swizzled.Swizzle = combine_swizzles4(reg.Swizzle, x, y, z, w); return swizzled; } static struct rc_src_register swizzle_smear(struct rc_src_register reg, rc_swizzle x) { return swizzle(reg, x, x, x, x); } static struct rc_src_register swizzle_xxxx(struct rc_src_register reg) { return swizzle_smear(reg, RC_SWIZZLE_X); } static struct rc_src_register swizzle_yyyy(struct rc_src_register reg) { return swizzle_smear(reg, RC_SWIZZLE_Y); } static struct rc_src_register swizzle_zzzz(struct rc_src_register reg) { return swizzle_smear(reg, RC_SWIZZLE_Z); } static struct rc_src_register swizzle_wwww(struct rc_src_register reg) { return swizzle_smear(reg, RC_SWIZZLE_W); } static struct rc_dst_register new_dst_reg(struct radeon_compiler *c, struct rc_instruction *inst) { unsigned tmp = rc_find_free_temporary(c); return dstregtmpmask(tmp, inst->U.I.DstReg.WriteMask); } static void transform_DP2(struct radeon_compiler *c, struct rc_instruction *inst) { struct rc_src_register src0 = inst->U.I.SrcReg[0]; struct rc_src_register src1 = inst->U.I.SrcReg[1]; src0.Negate &= ~(RC_MASK_Z | RC_MASK_W); src0.Swizzle &= ~(63 << (3 * 2)); src0.Swizzle |= (RC_SWIZZLE_ZERO << (3 * 2)) | (RC_SWIZZLE_ZERO << (3 * 3)); src1.Negate &= ~(RC_MASK_Z | RC_MASK_W); src1.Swizzle &= ~(63 << (3 * 2)); src1.Swizzle |= (RC_SWIZZLE_ZERO << (3 * 2)) | (RC_SWIZZLE_ZERO << (3 * 3)); emit2(c, inst->Prev, RC_OPCODE_DP3, &inst->U.I, inst->U.I.DstReg, src0, src1); rc_remove_instruction(inst); } static void transform_RSQ(struct radeon_compiler *c, struct rc_instruction *inst) { inst->U.I.SrcReg[0] = absolute(inst->U.I.SrcReg[0]); } static void transform_KILP(struct radeon_compiler *c, struct rc_instruction *inst) { inst->U.I.SrcReg[0] = negate(builtin_one); inst->U.I.Opcode = RC_OPCODE_KIL; } /** * Can be used as a transformation for @ref radeonClauseLocalTransform, * no userData necessary. * * Transforms RSQ to Radeon's native RSQ by explicitly setting * absolute value. * * @note should be applicable to R300 and R500 fragment programs. */ int radeonTransformALU(struct radeon_compiler *c, struct rc_instruction *inst, void *unused) { switch (inst->U.I.Opcode) { case RC_OPCODE_DP2: transform_DP2(c, inst); return 1; case RC_OPCODE_KILP: transform_KILP(c, inst); return 1; case RC_OPCODE_RSQ: transform_RSQ(c, inst); return 1; case RC_OPCODE_SEQ: unreachable(); case RC_OPCODE_SGE: unreachable(); case RC_OPCODE_SLT: unreachable(); case RC_OPCODE_SNE: unreachable(); default: return 0; } } static void transform_r300_vertex_CMP(struct radeon_compiler *c, struct rc_instruction *inst) { /* R5xx has a CMP, but we can use it only if it reads from less than * three different temps. */ if (c->is_r500 && !rc_inst_has_three_diff_temp_srcs(inst)) return; unreachable(); } static void transform_r300_vertex_DP2(struct radeon_compiler *c, struct rc_instruction *inst) { struct rc_instruction *next_inst = inst->Next; transform_DP2(c, inst); next_inst->Prev->U.I.Opcode = RC_OPCODE_DP4; } static void transform_r300_vertex_DP3(struct radeon_compiler *c, struct rc_instruction *inst) { struct rc_src_register src0 = inst->U.I.SrcReg[0]; struct rc_src_register src1 = inst->U.I.SrcReg[1]; src0.Negate &= ~RC_MASK_W; src0.Swizzle &= ~(7 << (3 * 3)); src0.Swizzle |= RC_SWIZZLE_ZERO << (3 * 3); src1.Negate &= ~RC_MASK_W; src1.Swizzle &= ~(7 << (3 * 3)); src1.Swizzle |= RC_SWIZZLE_ZERO << (3 * 3); emit2(c, inst->Prev, RC_OPCODE_DP4, &inst->U.I, inst->U.I.DstReg, src0, src1); rc_remove_instruction(inst); } static void transform_r300_vertex_fix_LIT(struct radeon_compiler *c, struct rc_instruction *inst) { struct rc_dst_register dst = new_dst_reg(c, inst); unsigned constant_swizzle; int constant = rc_constants_add_immediate_scalar(&c->Program.Constants, 0.0000000000000000001, &constant_swizzle); /* MOV dst, src */ dst.WriteMask = RC_MASK_XYZW; emit1(c, inst->Prev, RC_OPCODE_MOV, NULL, dst, inst->U.I.SrcReg[0]); /* MAX dst.y, src, 0.00...001 */ emit2(c, inst->Prev, RC_OPCODE_MAX, NULL, dstregtmpmask(dst.Index, RC_MASK_Y), srcreg(RC_FILE_TEMPORARY, dst.Index), srcregswz(RC_FILE_CONSTANT, constant, constant_swizzle)); inst->U.I.SrcReg[0] = srcreg(RC_FILE_TEMPORARY, dst.Index); } static void transform_r300_vertex_SEQ(struct radeon_compiler *c, struct rc_instruction *inst) { /* x = y <==> x >= y && y >= x */ /* x <= y */ struct rc_dst_register dst0 = new_dst_reg(c, inst); emit2(c, inst->Prev, RC_OPCODE_SGE, NULL, dst0, inst->U.I.SrcReg[0], inst->U.I.SrcReg[1]); /* y <= x */ int tmp = rc_find_free_temporary(c); emit2(c, inst->Prev, RC_OPCODE_SGE, NULL, dstregtmpmask(tmp, inst->U.I.DstReg.WriteMask), inst->U.I.SrcReg[1], inst->U.I.SrcReg[0]); /* x && y = x * y */ emit2(c, inst->Prev, RC_OPCODE_MUL, NULL, inst->U.I.DstReg, srcreg(dst0.File, dst0.Index), srcreg(RC_FILE_TEMPORARY, tmp)); rc_remove_instruction(inst); } static void transform_r300_vertex_SNE(struct radeon_compiler *c, struct rc_instruction *inst) { /* x != y <==> x < y || y < x */ /* x < y */ struct rc_dst_register dst0 = new_dst_reg(c, inst); emit2(c, inst->Prev, RC_OPCODE_SLT, NULL, dst0, inst->U.I.SrcReg[0], inst->U.I.SrcReg[1]); /* y < x */ int tmp = rc_find_free_temporary(c); emit2(c, inst->Prev, RC_OPCODE_SLT, NULL, dstregtmpmask(tmp, inst->U.I.DstReg.WriteMask), inst->U.I.SrcReg[1], inst->U.I.SrcReg[0]); /* x || y = max(x, y) */ emit2(c, inst->Prev, RC_OPCODE_MAX, NULL, inst->U.I.DstReg, srcreg(dst0.File, dst0.Index), srcreg(RC_FILE_TEMPORARY, tmp)); rc_remove_instruction(inst); } /** * For use with rc_local_transform, this transforms non-native ALU * instructions of the r300 up to r500 vertex engine. */ int r300_transform_vertex_alu(struct radeon_compiler *c, struct rc_instruction *inst, void *unused) { switch (inst->U.I.Opcode) { case RC_OPCODE_CMP: transform_r300_vertex_CMP(c, inst); return 1; case RC_OPCODE_DP2: transform_r300_vertex_DP2(c, inst); return 1; case RC_OPCODE_DP3: transform_r300_vertex_DP3(c, inst); return 1; case RC_OPCODE_LIT: transform_r300_vertex_fix_LIT(c, inst); return 1; case RC_OPCODE_SEQ: if (!c->is_r500) { transform_r300_vertex_SEQ(c, inst); return 1; } return 0; case RC_OPCODE_SNE: if (!c->is_r500) { transform_r300_vertex_SNE(c, inst); return 1; } return 0; default: return 0; } } /** * Replaces DDX/DDY instructions with MOV 0 to avoid using dummy shaders on r300/r400. * * @warning This explicitly changes the form of DDX and DDY! */ int radeonStubDeriv(struct radeon_compiler *c, struct rc_instruction *inst, void *unused) { if (inst->U.I.Opcode != RC_OPCODE_DDX && inst->U.I.Opcode != RC_OPCODE_DDY) return 0; inst->U.I.Opcode = RC_OPCODE_MOV; inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_0000; mesa_logw_once("r300: WARNING: Shader is trying to use derivatives, " "but the hardware doesn't support it. " "Expect possible misrendering (it's not a bug, do not report it)."); return 1; } /** * Rewrite DDX/DDY instructions to properly work with r5xx shaders. * The r5xx MDH/MDV instruction provides per-quad partial derivatives. * It takes the form A*B+C. A and C are set by setting src0. B should be -1. * * @warning This explicitly changes the form of DDX and DDY! */ int radeonTransformDeriv(struct radeon_compiler *c, struct rc_instruction *inst, void *unused) { if (inst->U.I.Opcode != RC_OPCODE_DDX && inst->U.I.Opcode != RC_OPCODE_DDY) return 0; inst->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_1111; inst->U.I.SrcReg[1].Negate = RC_MASK_XYZW; return 1; } int rc_force_output_alpha_to_one(struct radeon_compiler *c, struct rc_instruction *inst, void *data) { struct r300_fragment_program_compiler *fragc = (struct r300_fragment_program_compiler *)c; const struct rc_opcode_info *info = rc_get_opcode_info(inst->U.I.Opcode); unsigned tmp; if (!info->HasDstReg || inst->U.I.DstReg.File != RC_FILE_OUTPUT || inst->U.I.DstReg.Index == fragc->OutputDepth) return 1; tmp = rc_find_free_temporary(c); /* Insert MOV after inst, set alpha to 1. */ emit1(c, inst, RC_OPCODE_MOV, NULL, inst->U.I.DstReg, srcregswz(RC_FILE_TEMPORARY, tmp, RC_SWIZZLE_XYZ1)); /* Re-route the destination of inst to the source of mov. */ inst->U.I.DstReg.File = RC_FILE_TEMPORARY; inst->U.I.DstReg.Index = tmp; /* Move the saturate output modifier to the MOV instruction * (for better copy propagation). */ inst->Next->U.I.SaturateMode = inst->U.I.SaturateMode; inst->U.I.SaturateMode = RC_SATURATE_NONE; return 1; }