From b645e8c96dc1e3b153cf882c8931f10e0c006f04 Mon Sep 17 00:00:00 2001 From: Nicolai Haehnle Date: Sun, 18 Mar 2007 18:32:32 +0100 Subject: [PATCH] r300: Streamlined fragment program LIT implementation Fix a bug in the LIT implementation (clamp exponent to 128, not 0.5) and change the implementation around. In theory, the new implementation needs as little as 5 instruction slots. Unfortunately, the dependency analysis in find_and_replace_slot is not strong enough to look at individual components of a register yet. --- src/mesa/drivers/dri/r300/r300_fragprog.c | 163 ++++++++++++++-------- 1 file changed, 101 insertions(+), 62 deletions(-) diff --git a/src/mesa/drivers/dri/r300/r300_fragprog.c b/src/mesa/drivers/dri/r300/r300_fragprog.c index 89e9f6531a9..b0681e28083 100644 --- a/src/mesa/drivers/dri/r300/r300_fragprog.c +++ b/src/mesa/drivers/dri/r300/r300_fragprog.c @@ -492,7 +492,7 @@ static GLuint emit_param4fv(struct r300_fragment_program *rp, return r; } -static GLuint emit_const4fv(struct r300_fragment_program *rp, GLfloat *cp) +static GLuint emit_const4fv(struct r300_fragment_program *rp, const GLfloat* cp) { GLuint r = undef; GLuint index; @@ -1405,15 +1405,112 @@ static void make_sin_const(struct r300_fragment_program *rp) } } +/** + * Emit a LIT instruction. + * \p flags may be PFS_FLAG_SAT + * + * Definition of LIT (from ARB_fragment_program): + * tmp = VectorLoad(op0); + * if (tmp.x < 0) tmp.x = 0; + * if (tmp.y < 0) tmp.y = 0; + * if (tmp.w < -(128.0-epsilon)) tmp.w = -(128.0-epsilon); + * else if (tmp.w > 128-epsilon) tmp.w = 128-epsilon; + * result.x = 1.0; + * result.y = tmp.x; + * result.z = (tmp.x > 0) ? RoughApproxPower(tmp.y, tmp.w) : 0.0; + * result.w = 1.0; + * + * The longest path of computation is the one leading to result.z, + * consisting of 5 operations. This implementation of LIT takes + * 5 slots. So unless there's some special undocumented opcode, + * this implementation is potentially optimal. Unfortunately, + * emit_arith is a bit too conservative because it doesn't understand + * partial writes to the vector component. + */ +static void emit_lit(struct r300_fragment_program *rp, + GLuint dest, + int mask, + GLuint src, + int flags) +{ + COMPILE_STATE; + static const GLfloat cnstv[4] = { 127.999999, 127.999999, 127.999999, -127.999999 }; + GLuint cnst; + int needTemporary; + GLuint temp; + + cnst = emit_const4fv(rp, cnstv); + + needTemporary = 0; + if ((mask & WRITEMASK_XYZW) != WRITEMASK_XYZW) { + needTemporary = 1; + } else if (REG_GET_TYPE(dest) == REG_TYPE_OUTPUT) { + // LIT is typically followed by DP3/DP4, so there's no point + // in creating special code for this case + needTemporary = 1; + } + + if (needTemporary) { + temp = keep(get_temp_reg(rp)); + } else { + temp = keep(dest); + } + + // Npte: The order of emit_arith inside the slots is relevant, + // because emit_arith only looks at scalar vs. vector when resolving + // dependencies, and it does not consider individual vector components, + // so swizzling between the two parts can create fake dependencies. + + // First slot + emit_arith(rp, PFS_OP_MAX, temp, WRITEMASK_XY, + keep(src), pfs_zero, undef, 0); + emit_arith(rp, PFS_OP_MAX, temp, WRITEMASK_W, + src, cnst, undef, 0); + + // Second slot + emit_arith(rp, PFS_OP_MIN, temp, WRITEMASK_Z, + swizzle(temp, W, W, W, W), cnst, undef, 0); + emit_arith(rp, PFS_OP_LG2, temp, WRITEMASK_W, + swizzle(temp, Y, Y, Y, Y), undef, undef, 0); + + // Third slot + // If desired, we saturate the y result here. + // This does not affect the use as a condition variable in the CMP later + emit_arith(rp, PFS_OP_MAD, temp, WRITEMASK_W, + temp, swizzle(temp, Z, Z, Z, Z), pfs_zero, 0); + emit_arith(rp, PFS_OP_MAD, temp, WRITEMASK_Y, + swizzle(temp, X, X, X, X), pfs_one, pfs_zero, flags); + + // Fourth slot + emit_arith(rp, PFS_OP_MAD, temp, WRITEMASK_X, + pfs_one, pfs_one, pfs_zero, 0); + emit_arith(rp, PFS_OP_EX2, temp, WRITEMASK_W, + temp, undef, undef, 0); + + // Fifth slot + emit_arith(rp, PFS_OP_CMP, temp, WRITEMASK_Z, + swizzle(temp, W, W, W, W), pfs_zero, swizzle(temp, Y, Y, Y, Y), flags); + emit_arith(rp, PFS_OP_MAD, temp, WRITEMASK_W, + pfs_one, pfs_one, pfs_zero, 0); + + if (needTemporary) { + emit_arith(rp, PFS_OP_MAD, dest, mask, + temp, pfs_one, pfs_zero, flags); + free_temp(rp, temp); + } else { + // Decrease refcount of the destination + t_hw_dst(rp, dest, GL_FALSE, cs->nrslots); + } +} + + static GLboolean parse_program(struct r300_fragment_program *rp) { struct gl_fragment_program *mp = &rp->mesa_program; const struct prog_instruction *inst = mp->Base.Instructions; struct prog_instruction *fpi; GLuint src[3], dest, temp[2]; - GLuint cnst; int flags, mask = 0; - GLfloat cnstv[4] = {0.0, 0.0, 0.0, 0.0}; if (!inst || inst[0].Opcode == OPCODE_END) { ERROR("empty program?\n"); @@ -1612,66 +1709,8 @@ static GLboolean parse_program(struct r300_fragment_program *rp) flags); break; case OPCODE_LIT: - /* LIT - * if (s.x < 0) t.x = 0; else t.x = s.x; - * if (s.y < 0) t.y = 0; else t.y = s.y; - * if (s.w > 128.0) t.w = 128.0; else t.w = s.w; - * if (s.w < -128.0) t.w = -128.0; else t.w = s.w; - * r.x = 1.0 - * if (t.x > 0) r.y = pow(t.y, t.w); else r.y = 0; - * Also r.y = 0 if t.y < 0 - * For the t.x > 0 FGLRX use the CMPH opcode which - * change the compare to (t.x + 0.5) > 0.5 we may - * save one instruction by doing CMP -t.x - */ - cnstv[0] = cnstv[1] = cnstv[2] = cnstv[3] = 0.50001; src[0] = t_src(rp, fpi->SrcReg[0]); - temp[0] = get_temp_reg(rp); - cnst = emit_const4fv(rp, cnstv); - emit_arith(rp, PFS_OP_CMP, temp[0], - WRITEMASK_X | WRITEMASK_Y, - src[0], pfs_zero, src[0], flags); - emit_arith(rp, PFS_OP_MIN, temp[0], WRITEMASK_Z, - swizzle(keep(src[0]), W, W, W, W), - cnst, undef, flags); - emit_arith(rp, PFS_OP_LG2, temp[0], WRITEMASK_W, - swizzle(temp[0], Y, Y, Y, Y), - undef, undef, flags); - emit_arith(rp, PFS_OP_MAX, temp[0], WRITEMASK_Z, - temp[0], negate(cnst), undef, flags); - emit_arith(rp, PFS_OP_MAD, temp[0], WRITEMASK_W, - temp[0], swizzle(temp[0], Z, Z, Z, Z), - pfs_zero, flags); - emit_arith(rp, PFS_OP_EX2, temp[0], WRITEMASK_W, - temp[0], undef, undef, flags); - emit_arith(rp, PFS_OP_MAD, dest, WRITEMASK_Y, - swizzle(keep(temp[0]), X, X, X, X), - pfs_one, pfs_zero, flags); -#if 0 - emit_arith(rp, PFS_OP_MAD, temp[0], WRITEMASK_X, - temp[0], pfs_one, pfs_half, flags); - emit_arith(rp, PFS_OP_CMPH, temp[0], WRITEMASK_Z, - swizzle(keep(temp[0]), W, W, W, W), - pfs_zero, swizzle(keep(temp[0]), X, X, X, X), - flags); -#else - emit_arith(rp, PFS_OP_CMP, temp[0], WRITEMASK_Z, - pfs_zero, - swizzle(keep(temp[0]), W, W, W, W), - negate(swizzle(keep(temp[0]), X, X, X, X)), - flags); -#endif - emit_arith(rp, PFS_OP_CMP, dest, WRITEMASK_Z, - pfs_zero, temp[0], - negate(swizzle(keep(temp[0]), Y, Y, Y, Y)), - flags); - emit_arith(rp, PFS_OP_MAD, dest, - WRITEMASK_X | WRITEMASK_W, - pfs_one, - pfs_one, - pfs_zero, - flags); - free_temp(rp, temp[0]); + emit_lit(rp, dest, mask, src[0], flags); break; case OPCODE_LRP: src[0] = t_src(rp, fpi->SrcReg[0]); -- 2.30.2