diff options
Diffstat (limited to 'src/gallium/auxiliary/gallivm')
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_arit.c | 369 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_conv.c | 14 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_format_aos.c | 4 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_format_soa.c | 4 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_init.c | 2 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_logic.c | 26 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_pack.c | 21 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_quad.c | 4 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c | 24 | ||||
-rw-r--r-- | src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c | 14 |
10 files changed, 343 insertions, 139 deletions
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.c b/src/gallium/auxiliary/gallivm/lp_bld_arit.c index f5f2623e46..7b35dd4bb4 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_arit.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.c @@ -1,6 +1,6 @@ /************************************************************************** * - * Copyright 2009 VMware, Inc. + * Copyright 2009-2010 VMware, Inc. * All Rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a @@ -59,6 +59,19 @@ #include "lp_bld_arit.h" +/* + * XXX: Increasing eliminates some artifacts, but adds others, most + * noticeably corruption in the Earth halo in Google Earth. + */ +#define RCP_NEWTON_STEPS 0 + +#define RSQRT_NEWTON_STEPS 0 + +#define EXP_POLY_DEGREE 3 + +#define LOG_POLY_DEGREE 5 + + /** * Generate min(a, b) * No checks for special case values of a or b = 1 or 0 are done. @@ -72,6 +85,9 @@ lp_build_min_simple(struct lp_build_context *bld, const char *intrinsic = NULL; LLVMValueRef cond; + assert(lp_check_value(type, a)); + assert(lp_check_value(type, b)); + /* TODO: optimize the constant case */ if(type.width * type.length == 128) { @@ -118,6 +134,9 @@ lp_build_max_simple(struct lp_build_context *bld, const char *intrinsic = NULL; LLVMValueRef cond; + assert(lp_check_value(type, a)); + assert(lp_check_value(type, b)); + /* TODO: optimize the constant case */ if(type.width * type.length == 128) { @@ -160,6 +179,8 @@ lp_build_comp(struct lp_build_context *bld, { const struct lp_type type = bld->type; + assert(lp_check_value(type, a)); + if(a == bld->one) return bld->zero; if(a == bld->zero) @@ -173,9 +194,15 @@ lp_build_comp(struct lp_build_context *bld, } if(LLVMIsConstant(a)) - return LLVMConstSub(bld->one, a); + if (type.floating) + return LLVMConstFSub(bld->one, a); + else + return LLVMConstSub(bld->one, a); else - return LLVMBuildSub(bld->builder, bld->one, a, ""); + if (type.floating) + return LLVMBuildFSub(bld->builder, bld->one, a, ""); + else + return LLVMBuildSub(bld->builder, bld->one, a, ""); } @@ -190,6 +217,9 @@ lp_build_add(struct lp_build_context *bld, const struct lp_type type = bld->type; LLVMValueRef res; + assert(lp_check_value(type, a)); + assert(lp_check_value(type, b)); + if(a == bld->zero) return b; if(b == bld->zero) @@ -217,9 +247,15 @@ lp_build_add(struct lp_build_context *bld, } if(LLVMIsConstant(a) && LLVMIsConstant(b)) - res = LLVMConstAdd(a, b); + if (type.floating) + res = LLVMConstFAdd(a, b); + else + res = LLVMConstAdd(a, b); else - res = LLVMBuildAdd(bld->builder, a, b, ""); + if (type.floating) + res = LLVMBuildFAdd(bld->builder, a, b, ""); + else + res = LLVMBuildAdd(bld->builder, a, b, ""); /* clamp to ceiling of 1.0 */ if(bld->type.norm && (bld->type.floating || bld->type.fixed)) @@ -240,6 +276,8 @@ lp_build_sum_vector(struct lp_build_context *bld, LLVMValueRef index, res; unsigned i; + assert(lp_check_value(type, a)); + if (a == bld->zero) return bld->zero; if (a == bld->undef) @@ -253,9 +291,16 @@ lp_build_sum_vector(struct lp_build_context *bld, for (i = 1; i < type.length; i++) { index = LLVMConstInt(LLVMInt32Type(), i, 0); - res = LLVMBuildAdd(bld->builder, res, - LLVMBuildExtractElement(bld->builder, a, index, ""), - ""); + if (type.floating) + res = LLVMBuildFAdd(bld->builder, res, + LLVMBuildExtractElement(bld->builder, + a, index, ""), + ""); + else + res = LLVMBuildAdd(bld->builder, res, + LLVMBuildExtractElement(bld->builder, + a, index, ""), + ""); } return res; @@ -273,6 +318,9 @@ lp_build_sub(struct lp_build_context *bld, const struct lp_type type = bld->type; LLVMValueRef res; + assert(lp_check_value(type, a)); + assert(lp_check_value(type, b)); + if(b == bld->zero) return a; if(a == bld->undef || b == bld->undef) @@ -300,9 +348,15 @@ lp_build_sub(struct lp_build_context *bld, } if(LLVMIsConstant(a) && LLVMIsConstant(b)) - res = LLVMConstSub(a, b); + if (type.floating) + res = LLVMConstFSub(a, b); + else + res = LLVMConstSub(a, b); else - res = LLVMBuildSub(bld->builder, a, b, ""); + if (type.floating) + res = LLVMBuildFSub(bld->builder, a, b, ""); + else + res = LLVMBuildSub(bld->builder, a, b, ""); if(bld->type.norm && (bld->type.floating || bld->type.fixed)) res = lp_build_max_simple(bld, res, bld->zero); @@ -360,6 +414,10 @@ lp_build_mul_u8n(LLVMBuilderRef builder, LLVMValueRef c8; LLVMValueRef ab; + assert(!i16_type.floating); + assert(lp_check_value(i16_type, a)); + assert(lp_check_value(i16_type, b)); + c8 = lp_build_const_int_vec(i16_type, 8); #if 0 @@ -395,6 +453,9 @@ lp_build_mul(struct lp_build_context *bld, LLVMValueRef shift; LLVMValueRef res; + assert(lp_check_value(type, a)); + assert(lp_check_value(type, b)); + if(a == bld->zero) return bld->zero; if(a == bld->one) @@ -433,7 +494,10 @@ lp_build_mul(struct lp_build_context *bld, shift = NULL; if(LLVMIsConstant(a) && LLVMIsConstant(b)) { - res = LLVMConstMul(a, b); + if (type.floating) + res = LLVMConstFMul(a, b); + else + res = LLVMConstMul(a, b); if(shift) { if(type.sign) res = LLVMConstAShr(res, shift); @@ -442,7 +506,10 @@ lp_build_mul(struct lp_build_context *bld, } } else { - res = LLVMBuildMul(bld->builder, a, b, ""); + if (type.floating) + res = LLVMBuildFMul(bld->builder, a, b, ""); + else + res = LLVMBuildMul(bld->builder, a, b, ""); if(shift) { if(type.sign) res = LLVMBuildAShr(bld->builder, res, shift, ""); @@ -465,6 +532,8 @@ lp_build_mul_imm(struct lp_build_context *bld, { LLVMValueRef factor; + assert(lp_check_value(bld->type, a)); + if(b == 0) return bld->zero; @@ -472,7 +541,7 @@ lp_build_mul_imm(struct lp_build_context *bld, return a; if(b == -1) - return LLVMBuildNeg(bld->builder, a, ""); + return lp_build_negate(bld, a); if(b == 2 && bld->type.floating) return lp_build_add(bld, a, a); @@ -518,6 +587,9 @@ lp_build_div(struct lp_build_context *bld, { const struct lp_type type = bld->type; + assert(lp_check_value(type, a)); + assert(lp_check_value(type, b)); + if(a == bld->zero) return bld->zero; if(a == bld->one) @@ -529,13 +601,24 @@ lp_build_div(struct lp_build_context *bld, if(a == bld->undef || b == bld->undef) return bld->undef; - if(LLVMIsConstant(a) && LLVMIsConstant(b)) - return LLVMConstFDiv(a, b); + if(LLVMIsConstant(a) && LLVMIsConstant(b)) { + if (type.floating) + return LLVMConstFDiv(a, b); + else if (type.sign) + return LLVMConstSDiv(a, b); + else + return LLVMConstUDiv(a, b); + } if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4) return lp_build_mul(bld, a, lp_build_rcp(bld, b)); - return LLVMBuildFDiv(bld->builder, a, b, ""); + if (type.floating) + return LLVMBuildFDiv(bld->builder, a, b, ""); + else if (type.sign) + return LLVMBuildSDiv(bld->builder, a, b, ""); + else + return LLVMBuildUDiv(bld->builder, a, b, ""); } @@ -555,6 +638,10 @@ lp_build_lerp(struct lp_build_context *bld, LLVMValueRef delta; LLVMValueRef res; + assert(lp_check_value(bld->type, x)); + assert(lp_check_value(bld->type, v0)); + assert(lp_check_value(bld->type, v1)); + delta = lp_build_sub(bld, v1, v0); res = lp_build_mul(bld, x, delta); @@ -596,6 +683,9 @@ lp_build_min(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b) { + assert(lp_check_value(bld->type, a)); + assert(lp_check_value(bld->type, b)); + if(a == bld->undef || b == bld->undef) return bld->undef; @@ -624,6 +714,9 @@ lp_build_max(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b) { + assert(lp_check_value(bld->type, a)); + assert(lp_check_value(bld->type, b)); + if(a == bld->undef || b == bld->undef) return bld->undef; @@ -653,6 +746,10 @@ lp_build_clamp(struct lp_build_context *bld, LLVMValueRef min, LLVMValueRef max) { + assert(lp_check_value(bld->type, a)); + assert(lp_check_value(bld->type, min)); + assert(lp_check_value(bld->type, max)); + a = lp_build_min(bld, a, max); a = lp_build_max(bld, a, min); return a; @@ -669,6 +766,8 @@ lp_build_abs(struct lp_build_context *bld, const struct lp_type type = bld->type; LLVMTypeRef vec_type = lp_build_vec_type(type); + assert(lp_check_value(type, a)); + if(!type.sign) return a; @@ -702,7 +801,16 @@ LLVMValueRef lp_build_negate(struct lp_build_context *bld, LLVMValueRef a) { - return LLVMBuildNeg(bld->builder, a, ""); + assert(lp_check_value(bld->type, a)); + +#if HAVE_LLVM >= 0x0207 + if (bld->type.floating) + a = LLVMBuildFNeg(bld->builder, a, ""); + else +#endif + a = LLVMBuildNeg(bld->builder, a, ""); + + return a; } @@ -715,6 +823,8 @@ lp_build_sgn(struct lp_build_context *bld, LLVMValueRef cond; LLVMValueRef res; + assert(lp_check_value(type, a)); + /* Handle non-zero case */ if(!type.sign) { /* if not zero then sign must be positive */ @@ -773,6 +883,7 @@ lp_build_set_sign(struct lp_build_context *bld, LLVMValueRef val, res; assert(type.floating); + assert(lp_check_value(type, a)); /* val = reinterpret_cast<int>(a) */ val = LLVMBuildBitCast(bld->builder, a, int_vec_type, ""); @@ -1021,7 +1132,7 @@ lp_build_iround(struct lp_build_context *bld, half = LLVMBuildOr(bld->builder, sign, half, ""); half = LLVMBuildBitCast(bld->builder, half, vec_type, ""); - res = LLVMBuildAdd(bld->builder, a, half, ""); + res = LLVMBuildFAdd(bld->builder, a, half, ""); } res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, ""); @@ -1070,7 +1181,7 @@ lp_build_ifloor(struct lp_build_context *bld, offset = LLVMBuildAnd(bld->builder, offset, sign, ""); offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "ifloor.offset"); - res = LLVMBuildAdd(bld->builder, a, offset, "ifloor.res"); + res = LLVMBuildFAdd(bld->builder, a, offset, "ifloor.res"); } /* round to nearest (toward zero) */ @@ -1120,7 +1231,7 @@ lp_build_iceil(struct lp_build_context *bld, offset = LLVMBuildAnd(bld->builder, offset, sign, ""); offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "iceil.offset"); - res = LLVMBuildAdd(bld->builder, a, offset, "iceil.res"); + res = LLVMBuildFAdd(bld->builder, a, offset, "iceil.res"); } /* round to nearest (toward zero) */ @@ -1138,6 +1249,8 @@ lp_build_sqrt(struct lp_build_context *bld, LLVMTypeRef vec_type = lp_build_vec_type(type); char intrinsic[32]; + assert(lp_check_value(type, a)); + /* TODO: optimize the constant case */ /* TODO: optimize the constant case */ @@ -1148,12 +1261,39 @@ lp_build_sqrt(struct lp_build_context *bld, } +/** + * Do one Newton-Raphson step to improve reciprocate precision: + * + * x_{i+1} = x_i * (2 - a * x_i) + * + * See also: + * - http://en.wikipedia.org/wiki/Division_(digital)#Newton.E2.80.93Raphson_division + * - http://softwarecommunity.intel.com/articles/eng/1818.htm + */ +static INLINE LLVMValueRef +lp_build_rcp_refine(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef rcp_a) +{ + LLVMValueRef two = lp_build_const_vec(bld->type, 2.0); + LLVMValueRef res; + + res = LLVMBuildFMul(bld->builder, a, rcp_a, ""); + res = LLVMBuildFSub(bld->builder, two, res, ""); + res = LLVMBuildFMul(bld->builder, rcp_a, res, ""); + + return res; +} + + LLVMValueRef lp_build_rcp(struct lp_build_context *bld, LLVMValueRef a) { const struct lp_type type = bld->type; + assert(lp_check_value(type, a)); + if(a == bld->zero) return bld->undef; if(a == bld->one) @@ -1167,32 +1307,16 @@ lp_build_rcp(struct lp_build_context *bld, return LLVMConstFDiv(bld->one, a); if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4) { - /* - * XXX: Added precision is not always necessary, so only enable this - * when we have a better system in place to track minimum precision. - */ - -#if 0 - /* - * Do one Newton-Raphson step to improve precision: - * - * x1 = (2 - a * rcp(a)) * rcp(a) - */ - - LLVMValueRef two = lp_build_const_vec(bld->type, 2.0); - LLVMValueRef rcp_a; LLVMValueRef res; + unsigned i; - rcp_a = lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rcp.ps", lp_build_vec_type(type), a); + res = lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rcp.ps", bld->vec_type, a); - res = LLVMBuildMul(bld->builder, a, rcp_a, ""); - res = LLVMBuildSub(bld->builder, two, res, ""); - res = LLVMBuildMul(bld->builder, res, rcp_a, ""); + for (i = 0; i < RCP_NEWTON_STEPS; ++i) { + res = lp_build_rcp_refine(bld, a, res); + } - return rcp_a; -#else - return lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rcp.ps", lp_build_vec_type(type), a); -#endif + return res; } return LLVMBuildFDiv(bld->builder, bld->one, a, ""); @@ -1200,6 +1324,33 @@ lp_build_rcp(struct lp_build_context *bld, /** + * Do one Newton-Raphson step to improve rsqrt precision: + * + * x_{i+1} = 0.5 * x_i * (3.0 - a * x_i * x_i) + * + * See also: + * - http://softwarecommunity.intel.com/articles/eng/1818.htm + */ +static INLINE LLVMValueRef +lp_build_rsqrt_refine(struct lp_build_context *bld, + LLVMValueRef a, + LLVMValueRef rsqrt_a) +{ + LLVMValueRef half = lp_build_const_vec(bld->type, 0.5); + LLVMValueRef three = lp_build_const_vec(bld->type, 3.0); + LLVMValueRef res; + + res = LLVMBuildFMul(bld->builder, rsqrt_a, rsqrt_a, ""); + res = LLVMBuildFMul(bld->builder, a, res, ""); + res = LLVMBuildFSub(bld->builder, three, res, ""); + res = LLVMBuildFMul(bld->builder, rsqrt_a, res, ""); + res = LLVMBuildFMul(bld->builder, half, res, ""); + + return res; +} + + +/** * Generate 1/sqrt(a) */ LLVMValueRef @@ -1208,10 +1359,22 @@ lp_build_rsqrt(struct lp_build_context *bld, { const struct lp_type type = bld->type; + assert(lp_check_value(type, a)); + assert(type.floating); - if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4) - return lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rsqrt.ps", lp_build_vec_type(type), a); + if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4) { + LLVMValueRef res; + unsigned i; + + res = lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rsqrt.ps", bld->vec_type, a); + + for (i = 0; i < RSQRT_NEWTON_STEPS; ++i) { + res = lp_build_rsqrt_refine(bld, a, res); + } + + return res; + } return lp_build_rcp(bld, lp_build_sqrt(bld, a)); } @@ -1270,7 +1433,7 @@ lp_build_sin(struct lp_build_context *bld, */ LLVMValueRef FOPi = lp_build_const_v4sf(1.27323954473516); - LLVMValueRef scale_y = LLVMBuildMul(b, x_abs, FOPi, "scale_y"); + LLVMValueRef scale_y = LLVMBuildFMul(b, x_abs, FOPi, "scale_y"); /* * store the integer part of y in mm0 @@ -1344,9 +1507,9 @@ lp_build_sin(struct lp_build_context *bld, * xmm2 = _mm_mul_ps(y, xmm2); * xmm3 = _mm_mul_ps(y, xmm3); */ - LLVMValueRef xmm1 = LLVMBuildMul(b, y_2, DP1, "xmm1"); - LLVMValueRef xmm2 = LLVMBuildMul(b, y_2, DP2, "xmm2"); - LLVMValueRef xmm3 = LLVMBuildMul(b, y_2, DP3, "xmm3"); + LLVMValueRef xmm1 = LLVMBuildFMul(b, y_2, DP1, "xmm1"); + LLVMValueRef xmm2 = LLVMBuildFMul(b, y_2, DP2, "xmm2"); + LLVMValueRef xmm3 = LLVMBuildFMul(b, y_2, DP3, "xmm3"); /* * x = _mm_add_ps(x, xmm1); @@ -1354,16 +1517,16 @@ lp_build_sin(struct lp_build_context *bld, * x = _mm_add_ps(x, xmm3); */ - LLVMValueRef x_1 = LLVMBuildAdd(b, x_abs, xmm1, "x_1"); - LLVMValueRef x_2 = LLVMBuildAdd(b, x_1, xmm2, "x_2"); - LLVMValueRef x_3 = LLVMBuildAdd(b, x_2, xmm3, "x_3"); + LLVMValueRef x_1 = LLVMBuildFAdd(b, x_abs, xmm1, "x_1"); + LLVMValueRef x_2 = LLVMBuildFAdd(b, x_1, xmm2, "x_2"); + LLVMValueRef x_3 = LLVMBuildFAdd(b, x_2, xmm3, "x_3"); /* * Evaluate the first polynom (0 <= x <= Pi/4) * * z = _mm_mul_ps(x,x); */ - LLVMValueRef z = LLVMBuildMul(b, x_3, x_3, "z"); + LLVMValueRef z = LLVMBuildFMul(b, x_3, x_3, "z"); /* * _PS_CONST(coscof_p0, 2.443315711809948E-005); @@ -1378,12 +1541,12 @@ lp_build_sin(struct lp_build_context *bld, * y = *(v4sf*)_ps_coscof_p0; * y = _mm_mul_ps(y, z); */ - LLVMValueRef y_3 = LLVMBuildMul(b, z, coscof_p0, "y_3"); - LLVMValueRef y_4 = LLVMBuildAdd(b, y_3, coscof_p1, "y_4"); - LLVMValueRef y_5 = LLVMBuildMul(b, y_4, z, "y_5"); - LLVMValueRef y_6 = LLVMBuildAdd(b, y_5, coscof_p2, "y_6"); - LLVMValueRef y_7 = LLVMBuildMul(b, y_6, z, "y_7"); - LLVMValueRef y_8 = LLVMBuildMul(b, y_7, z, "y_8"); + LLVMValueRef y_3 = LLVMBuildFMul(b, z, coscof_p0, "y_3"); + LLVMValueRef y_4 = LLVMBuildFAdd(b, y_3, coscof_p1, "y_4"); + LLVMValueRef y_5 = LLVMBuildFMul(b, y_4, z, "y_5"); + LLVMValueRef y_6 = LLVMBuildFAdd(b, y_5, coscof_p2, "y_6"); + LLVMValueRef y_7 = LLVMBuildFMul(b, y_6, z, "y_7"); + LLVMValueRef y_8 = LLVMBuildFMul(b, y_7, z, "y_8"); /* @@ -1392,10 +1555,10 @@ lp_build_sin(struct lp_build_context *bld, * y = _mm_add_ps(y, *(v4sf*)_ps_1); */ LLVMValueRef half = lp_build_const_v4sf(0.5); - LLVMValueRef tmp = LLVMBuildMul(b, z, half, "tmp"); - LLVMValueRef y_9 = LLVMBuildSub(b, y_8, tmp, "y_8"); + LLVMValueRef tmp = LLVMBuildFMul(b, z, half, "tmp"); + LLVMValueRef y_9 = LLVMBuildFSub(b, y_8, tmp, "y_8"); LLVMValueRef one = lp_build_const_v4sf(1.0); - LLVMValueRef y_10 = LLVMBuildAdd(b, y_9, one, "y_9"); + LLVMValueRef y_10 = LLVMBuildFAdd(b, y_9, one, "y_9"); /* * _PS_CONST(sincof_p0, -1.9515295891E-4); @@ -1419,13 +1582,13 @@ lp_build_sin(struct lp_build_context *bld, * y2 = _mm_add_ps(y2, x); */ - LLVMValueRef y2_3 = LLVMBuildMul(b, z, sincof_p0, "y2_3"); - LLVMValueRef y2_4 = LLVMBuildAdd(b, y2_3, sincof_p1, "y2_4"); - LLVMValueRef y2_5 = LLVMBuildMul(b, y2_4, z, "y2_5"); - LLVMValueRef y2_6 = LLVMBuildAdd(b, y2_5, sincof_p2, "y2_6"); - LLVMValueRef y2_7 = LLVMBuildMul(b, y2_6, z, "y2_7"); - LLVMValueRef y2_8 = LLVMBuildMul(b, y2_7, x_3, "y2_8"); - LLVMValueRef y2_9 = LLVMBuildAdd(b, y2_8, x_3, "y2_9"); + LLVMValueRef y2_3 = LLVMBuildFMul(b, z, sincof_p0, "y2_3"); + LLVMValueRef y2_4 = LLVMBuildFAdd(b, y2_3, sincof_p1, "y2_4"); + LLVMValueRef y2_5 = LLVMBuildFMul(b, y2_4, z, "y2_5"); + LLVMValueRef y2_6 = LLVMBuildFAdd(b, y2_5, sincof_p2, "y2_6"); + LLVMValueRef y2_7 = LLVMBuildFMul(b, y2_6, z, "y2_7"); + LLVMValueRef y2_8 = LLVMBuildFMul(b, y2_7, x_3, "y2_8"); + LLVMValueRef y2_9 = LLVMBuildFAdd(b, y2_8, x_3, "y2_9"); /* * select the correct result from the two polynoms @@ -1481,7 +1644,7 @@ lp_build_cos(struct lp_build_context *bld, */ LLVMValueRef FOPi = lp_build_const_v4sf(1.27323954473516); - LLVMValueRef scale_y = LLVMBuildMul(b, x_abs, FOPi, "scale_y"); + LLVMValueRef scale_y = LLVMBuildFMul(b, x_abs, FOPi, "scale_y"); /* * store the integer part of y in mm0 @@ -1561,9 +1724,9 @@ lp_build_cos(struct lp_build_context *bld, * xmm2 = _mm_mul_ps(y, xmm2); * xmm3 = _mm_mul_ps(y, xmm3); */ - LLVMValueRef xmm1 = LLVMBuildMul(b, y_2, DP1, "xmm1"); - LLVMValueRef xmm2 = LLVMBuildMul(b, y_2, DP2, "xmm2"); - LLVMValueRef xmm3 = LLVMBuildMul(b, y_2, DP3, "xmm3"); + LLVMValueRef xmm1 = LLVMBuildFMul(b, y_2, DP1, "xmm1"); + LLVMValueRef xmm2 = LLVMBuildFMul(b, y_2, DP2, "xmm2"); + LLVMValueRef xmm3 = LLVMBuildFMul(b, y_2, DP3, "xmm3"); /* * x = _mm_add_ps(x, xmm1); @@ -1571,16 +1734,16 @@ lp_build_cos(struct lp_build_context *bld, * x = _mm_add_ps(x, xmm3); */ - LLVMValueRef x_1 = LLVMBuildAdd(b, x_abs, xmm1, "x_1"); - LLVMValueRef x_2 = LLVMBuildAdd(b, x_1, xmm2, "x_2"); - LLVMValueRef x_3 = LLVMBuildAdd(b, x_2, xmm3, "x_3"); + LLVMValueRef x_1 = LLVMBuildFAdd(b, x_abs, xmm1, "x_1"); + LLVMValueRef x_2 = LLVMBuildFAdd(b, x_1, xmm2, "x_2"); + LLVMValueRef x_3 = LLVMBuildFAdd(b, x_2, xmm3, "x_3"); /* * Evaluate the first polynom (0 <= x <= Pi/4) * * z = _mm_mul_ps(x,x); */ - LLVMValueRef z = LLVMBuildMul(b, x_3, x_3, "z"); + LLVMValueRef z = LLVMBuildFMul(b, x_3, x_3, "z"); /* * _PS_CONST(coscof_p0, 2.443315711809948E-005); @@ -1595,12 +1758,12 @@ lp_build_cos(struct lp_build_context *bld, * y = *(v4sf*)_ps_coscof_p0; * y = _mm_mul_ps(y, z); */ - LLVMValueRef y_3 = LLVMBuildMul(b, z, coscof_p0, "y_3"); - LLVMValueRef y_4 = LLVMBuildAdd(b, y_3, coscof_p1, "y_4"); - LLVMValueRef y_5 = LLVMBuildMul(b, y_4, z, "y_5"); - LLVMValueRef y_6 = LLVMBuildAdd(b, y_5, coscof_p2, "y_6"); - LLVMValueRef y_7 = LLVMBuildMul(b, y_6, z, "y_7"); - LLVMValueRef y_8 = LLVMBuildMul(b, y_7, z, "y_8"); + LLVMValueRef y_3 = LLVMBuildFMul(b, z, coscof_p0, "y_3"); + LLVMValueRef y_4 = LLVMBuildFAdd(b, y_3, coscof_p1, "y_4"); + LLVMValueRef y_5 = LLVMBuildFMul(b, y_4, z, "y_5"); + LLVMValueRef y_6 = LLVMBuildFAdd(b, y_5, coscof_p2, "y_6"); + LLVMValueRef y_7 = LLVMBuildFMul(b, y_6, z, "y_7"); + LLVMValueRef y_8 = LLVMBuildFMul(b, y_7, z, "y_8"); /* @@ -1609,10 +1772,10 @@ lp_build_cos(struct lp_build_context *bld, * y = _mm_add_ps(y, *(v4sf*)_ps_1); */ LLVMValueRef half = lp_build_const_v4sf(0.5); - LLVMValueRef tmp = LLVMBuildMul(b, z, half, "tmp"); - LLVMValueRef y_9 = LLVMBuildSub(b, y_8, tmp, "y_8"); + LLVMValueRef tmp = LLVMBuildFMul(b, z, half, "tmp"); + LLVMValueRef y_9 = LLVMBuildFSub(b, y_8, tmp, "y_8"); LLVMValueRef one = lp_build_const_v4sf(1.0); - LLVMValueRef y_10 = LLVMBuildAdd(b, y_9, one, "y_9"); + LLVMValueRef y_10 = LLVMBuildFAdd(b, y_9, one, "y_9"); /* * _PS_CONST(sincof_p0, -1.9515295891E-4); @@ -1636,13 +1799,13 @@ lp_build_cos(struct lp_build_context *bld, * y2 = _mm_add_ps(y2, x); */ - LLVMValueRef y2_3 = LLVMBuildMul(b, z, sincof_p0, "y2_3"); - LLVMValueRef y2_4 = LLVMBuildAdd(b, y2_3, sincof_p1, "y2_4"); - LLVMValueRef y2_5 = LLVMBuildMul(b, y2_4, z, "y2_5"); - LLVMValueRef y2_6 = LLVMBuildAdd(b, y2_5, sincof_p2, "y2_6"); - LLVMValueRef y2_7 = LLVMBuildMul(b, y2_6, z, "y2_7"); - LLVMValueRef y2_8 = LLVMBuildMul(b, y2_7, x_3, "y2_8"); - LLVMValueRef y2_9 = LLVMBuildAdd(b, y2_8, x_3, "y2_9"); + LLVMValueRef y2_3 = LLVMBuildFMul(b, z, sincof_p0, "y2_3"); + LLVMValueRef y2_4 = LLVMBuildFAdd(b, y2_3, sincof_p1, "y2_4"); + LLVMValueRef y2_5 = LLVMBuildFMul(b, y2_4, z, "y2_5"); + LLVMValueRef y2_6 = LLVMBuildFAdd(b, y2_5, sincof_p2, "y2_6"); + LLVMValueRef y2_7 = LLVMBuildFMul(b, y2_6, z, "y2_7"); + LLVMValueRef y2_8 = LLVMBuildFMul(b, y2_7, x_3, "y2_8"); + LLVMValueRef y2_9 = LLVMBuildFAdd(b, y2_8, x_3, "y2_9"); /* * select the correct result from the two polynoms @@ -1695,6 +1858,8 @@ lp_build_exp(struct lp_build_context *bld, /* log2(e) = 1/log(2) */ LLVMValueRef log2e = lp_build_const_vec(bld->type, 1.4426950408889634); + assert(lp_check_value(bld->type, x)); + return lp_build_mul(bld, log2e, lp_build_exp2(bld, x)); } @@ -1709,14 +1874,12 @@ lp_build_log(struct lp_build_context *bld, /* log(2) */ LLVMValueRef log2 = lp_build_const_vec(bld->type, 0.69314718055994529); + assert(lp_check_value(bld->type, x)); + return lp_build_mul(bld, log2, lp_build_exp2(bld, x)); } -#define EXP_POLY_DEGREE 3 -#define LOG_POLY_DEGREE 5 - - /** * Generate polynomial. * Ex: coeffs[0] + x * coeffs[1] + x^2 * coeffs[2]. @@ -1731,6 +1894,8 @@ lp_build_polynomial(struct lp_build_context *bld, LLVMValueRef res = NULL; unsigned i; + assert(lp_check_value(bld->type, x)); + /* TODO: optimize the constant case */ if(LLVMIsConstant(x)) debug_printf("%s: inefficient/imprecise constant arithmetic\n", @@ -1802,6 +1967,8 @@ lp_build_exp2_approx(struct lp_build_context *bld, LLVMValueRef expfpart = NULL; LLVMValueRef res = NULL; + assert(lp_check_value(bld->type, x)); + if(p_exp2_int_part || p_frac_part || p_exp2) { /* TODO: optimize the constant case */ if(LLVMIsConstant(x)) @@ -1817,7 +1984,7 @@ lp_build_exp2_approx(struct lp_build_context *bld, ipart = lp_build_floor(bld, x); /* fpart = x - ipart */ - fpart = LLVMBuildSub(bld->builder, x, ipart, ""); + fpart = LLVMBuildFSub(bld->builder, x, ipart, ""); } if(p_exp2_int_part || p_exp2) { @@ -1832,7 +1999,7 @@ lp_build_exp2_approx(struct lp_build_context *bld, expfpart = lp_build_polynomial(bld, fpart, lp_build_exp2_polynomial, Elements(lp_build_exp2_polynomial)); - res = LLVMBuildMul(bld->builder, expipart, expfpart, ""); + res = LLVMBuildFMul(bld->builder, expipart, expfpart, ""); } if(p_exp2_int_part) @@ -1915,6 +2082,8 @@ lp_build_log2_approx(struct lp_build_context *bld, LLVMValueRef logmant = NULL; LLVMValueRef res = NULL; + assert(lp_check_value(bld->type, x)); + if(p_exp || p_floor_log2 || p_log2) { /* TODO: optimize the constant case */ if(LLVMIsConstant(x)) @@ -1945,9 +2114,9 @@ lp_build_log2_approx(struct lp_build_context *bld, Elements(lp_build_log2_polynomial)); /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/ - logmant = LLVMBuildMul(bld->builder, logmant, LLVMBuildSub(bld->builder, mant, bld->one, ""), ""); + logmant = LLVMBuildFMul(bld->builder, logmant, LLVMBuildFSub(bld->builder, mant, bld->one, ""), ""); - res = LLVMBuildAdd(bld->builder, logmant, logexp, ""); + res = LLVMBuildFAdd(bld->builder, logmant, logexp, ""); } if(p_exp) { diff --git a/src/gallium/auxiliary/gallivm/lp_bld_conv.c b/src/gallium/auxiliary/gallivm/lp_bld_conv.c index 77012f1fac..8b477313d4 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_conv.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_conv.c @@ -117,8 +117,8 @@ lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder, scale = (double)mask/ubound; bias = (double)((unsigned long long)1 << (mantissa - n)); - res = LLVMBuildMul(builder, src, lp_build_const_vec(src_type, scale), ""); - res = LLVMBuildAdd(builder, res, lp_build_const_vec(src_type, bias), ""); + res = LLVMBuildFMul(builder, src, lp_build_const_vec(src_type, scale), ""); + res = LLVMBuildFAdd(builder, res, lp_build_const_vec(src_type, bias), ""); res = LLVMBuildBitCast(builder, res, int_vec_type, ""); if(dst_width > n) { @@ -175,6 +175,8 @@ lp_build_unsigned_norm_to_float(LLVMBuilderRef builder, double scale; double bias; + assert(dst_type.floating); + mantissa = lp_mantissa(dst_type); n = MIN2(mantissa, src_width); @@ -199,8 +201,8 @@ lp_build_unsigned_norm_to_float(LLVMBuilderRef builder, res = LLVMBuildBitCast(builder, res, vec_type, ""); - res = LLVMBuildSub(builder, res, bias_, ""); - res = LLVMBuildMul(builder, res, lp_build_const_vec(dst_type, scale), ""); + res = LLVMBuildFSub(builder, res, bias_, ""); + res = LLVMBuildFMul(builder, res, lp_build_const_vec(dst_type, scale), ""); return res; } @@ -296,7 +298,7 @@ lp_build_conv(LLVMBuilderRef builder, if (dst_scale != 1.0) { LLVMValueRef scale = lp_build_const_vec(tmp_type, dst_scale); for(i = 0; i < num_tmps; ++i) - tmp[i] = LLVMBuildMul(builder, tmp[i], scale, ""); + tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, ""); } /* Use an equally sized integer for intermediate computations */ @@ -391,7 +393,7 @@ lp_build_conv(LLVMBuilderRef builder, if (src_scale != 1.0) { LLVMValueRef scale = lp_build_const_vec(tmp_type, 1.0/src_scale); for(i = 0; i < num_tmps; ++i) - tmp[i] = LLVMBuildMul(builder, tmp[i], scale, ""); + tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, ""); } } } diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c b/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c index 0f01fc1d75..247cb83ce6 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c @@ -240,7 +240,7 @@ lp_build_unpack_arith_rgba_aos(LLVMBuilderRef builder, */ if (normalized) - scaled = LLVMBuildMul(builder, casted, LLVMConstVector(scales, 4), ""); + scaled = LLVMBuildFMul(builder, casted, LLVMConstVector(scales, 4), ""); else scaled = casted; @@ -322,7 +322,7 @@ lp_build_pack_rgba_aos(LLVMBuilderRef builder, } if (normalized) - scaled = LLVMBuildMul(builder, unswizzled, LLVMConstVector(scales, 4), ""); + scaled = LLVMBuildFMul(builder, unswizzled, LLVMConstVector(scales, 4), ""); else scaled = unswizzled; diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c index 9f405921b0..c724a4453e 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c @@ -197,7 +197,7 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder, if (format_desc->channel[chan].normalized) { double scale = 1.0 / ((1 << (format_desc->channel[chan].size - 1)) - 1); LLVMValueRef scale_val = lp_build_const_vec(type, scale); - input = LLVMBuildMul(builder, input, scale_val, ""); + input = LLVMBuildFMul(builder, input, scale_val, ""); } } else { @@ -227,7 +227,7 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder, double scale = 1.0 / ((1 << (format_desc->channel[chan].size/2)) - 1); LLVMValueRef scale_val = lp_build_const_vec(type, scale); input = LLVMBuildSIToFP(builder, input, lp_build_vec_type(type), ""); - input = LLVMBuildMul(builder, input, scale_val, ""); + input = LLVMBuildFMul(builder, input, scale_val, ""); } else { /* FIXME */ diff --git a/src/gallium/auxiliary/gallivm/lp_bld_init.c b/src/gallium/auxiliary/gallivm/lp_bld_init.c index ef0888079c..60d8bcfa55 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_init.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_init.c @@ -46,7 +46,7 @@ static const struct debug_named_value lp_bld_debug_flags[] = { DEBUG_NAMED_VALUE_END }; -DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0); +DEBUG_GET_ONCE_FLAGS_OPTION(gallivm_debug, "GALLIVM_DEBUG", lp_bld_debug_flags, 0) #endif diff --git a/src/gallium/auxiliary/gallivm/lp_bld_logic.c b/src/gallium/auxiliary/gallivm/lp_bld_logic.c index ab4ddb81c4..7d7db3b0d9 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_logic.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_logic.c @@ -83,6 +83,8 @@ lp_build_compare(LLVMBuilderRef builder, assert(func >= PIPE_FUNC_NEVER); assert(func <= PIPE_FUNC_ALWAYS); + assert(lp_check_value(type, a)); + assert(lp_check_value(type, b)); if(func == PIPE_FUNC_NEVER) return zeros; @@ -374,6 +376,9 @@ lp_build_select_bitwise(struct lp_build_context *bld, struct lp_type type = bld->type; LLVMValueRef res; + assert(lp_check_value(type, a)); + assert(lp_check_value(type, b)); + if (a == b) { return a; } @@ -419,6 +424,9 @@ lp_build_select(struct lp_build_context *bld, struct lp_type type = bld->type; LLVMValueRef res; + assert(lp_check_value(type, a)); + assert(lp_check_value(type, b)); + if(a == b) return a; @@ -484,6 +492,9 @@ lp_build_select_aos(struct lp_build_context *bld, const unsigned n = type.length; unsigned i, j; + assert(lp_check_value(type, a)); + assert(lp_check_value(type, b)); + if(a == b) return a; if(cond[0] && cond[1] && cond[2] && cond[3]) @@ -539,7 +550,22 @@ lp_build_select_aos(struct lp_build_context *bld, LLVMValueRef lp_build_andc(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b) { + const struct lp_type type = bld->type; + + assert(lp_check_value(type, a)); + assert(lp_check_value(type, b)); + + /* can't do bitwise ops on floating-point values */ + if(type.floating) { + a = LLVMBuildBitCast(bld->builder, a, bld->int_vec_type, ""); + b = LLVMBuildBitCast(bld->builder, b, bld->int_vec_type, ""); + } + b = LLVMBuildNot(bld->builder, b, ""); b = LLVMBuildAnd(bld->builder, a, b, ""); + + if(type.floating) { + b = LLVMBuildBitCast(bld->builder, b, bld->vec_type, ""); + } return b; } diff --git a/src/gallium/auxiliary/gallivm/lp_bld_pack.c b/src/gallium/auxiliary/gallivm/lp_bld_pack.c index 7748f8f099..b7b630f2e8 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_pack.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_pack.c @@ -171,14 +171,13 @@ lp_build_unpack2(LLVMBuilderRef builder, msb = lp_build_zero(src_type); /* Interleave bits */ - if(util_cpu_caps.little_endian) { +#ifdef PIPE_ARCH_LITTLE_ENDIAN *dst_lo = lp_build_interleave2(builder, src_type, src, msb, 0); *dst_hi = lp_build_interleave2(builder, src_type, src, msb, 1); - } - else { +#else *dst_lo = lp_build_interleave2(builder, src_type, msb, src, 0); *dst_hi = lp_build_interleave2(builder, src_type, msb, src, 1); - } +#endif /* Cast the result into the new type (twice as wide) */ @@ -261,13 +260,14 @@ lp_build_pack2(LLVMBuilderRef builder, #endif LLVMTypeRef dst_vec_type = lp_build_vec_type(dst_type); LLVMValueRef shuffle; - LLVMValueRef res; + LLVMValueRef res = NULL; assert(!src_type.floating); assert(!dst_type.floating); assert(src_type.width == dst_type.width * 2); assert(src_type.length * 2 == dst_type.length); + /* Check for special cases first */ if(util_cpu_caps.has_sse2 && src_type.width * src_type.length == 128) { switch(src_type.width) { case 32: @@ -283,8 +283,8 @@ lp_build_pack2(LLVMBuilderRef builder, return lp_build_intrinsic_binary(builder, "llvm.x86.sse41.packusdw", dst_vec_type, lo, hi); } else { - assert(0); - return LLVMGetUndef(dst_vec_type); + /* use generic shuffle below */ + res = NULL; } } break; @@ -310,10 +310,13 @@ lp_build_pack2(LLVMBuilderRef builder, break; } - res = LLVMBuildBitCast(builder, res, dst_vec_type, ""); - return res; + if (res) { + res = LLVMBuildBitCast(builder, res, dst_vec_type, ""); + return res; + } } + /* generic shuffle */ lo = LLVMBuildBitCast(builder, lo, dst_vec_type, ""); hi = LLVMBuildBitCast(builder, hi, dst_vec_type, ""); diff --git a/src/gallium/auxiliary/gallivm/lp_bld_quad.c b/src/gallium/auxiliary/gallivm/lp_bld_quad.c index ca36046d22..7b1088939b 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_quad.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_quad.c @@ -85,7 +85,7 @@ lp_build_scalar_ddx(struct lp_build_context *bld, LLVMValueRef idx_right = LLVMConstInt(LLVMInt32Type(), LP_BLD_QUAD_TOP_RIGHT, 0); LLVMValueRef a_left = LLVMBuildExtractElement(bld->builder, a, idx_left, ""); LLVMValueRef a_right = LLVMBuildExtractElement(bld->builder, a, idx_right, ""); - return LLVMBuildSub(bld->builder, a_right, a_left, ""); + return lp_build_sub(bld, a_right, a_left); } @@ -97,5 +97,5 @@ lp_build_scalar_ddy(struct lp_build_context *bld, LLVMValueRef idx_bottom = LLVMConstInt(LLVMInt32Type(), LP_BLD_QUAD_BOTTOM_LEFT, 0); LLVMValueRef a_top = LLVMBuildExtractElement(bld->builder, a, idx_top, ""); LLVMValueRef a_bottom = LLVMBuildExtractElement(bld->builder, a, idx_bottom, ""); - return LLVMBuildSub(bld->builder, a_bottom, a_top, ""); + return lp_build_sub(bld, a_bottom, a_top); } diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c index 1a20d74cac..806c7d56a8 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c @@ -40,7 +40,6 @@ #include "util/u_memory.h" #include "util/u_math.h" #include "util/u_format.h" -#include "util/u_cpu_detect.h" #include "lp_bld_debug.h" #include "lp_bld_type.h" #include "lp_bld_const.h" @@ -811,7 +810,7 @@ lp_build_minify(struct lp_build_sample_context *bld, LLVMValueRef base_size, LLVMValueRef level) { - LLVMValueRef size = LLVMBuildAShr(bld->builder, base_size, level, "minify"); + LLVMValueRef size = LLVMBuildLShr(bld->builder, base_size, level, "minify"); size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one); return size; } @@ -888,17 +887,17 @@ lp_build_lod_selector(struct lp_build_sample_context *bld, /* Compute rho = max of all partial derivatives scaled by texture size. * XXX this could be vectorized somewhat */ - rho = LLVMBuildMul(bld->builder, + rho = LLVMBuildFMul(bld->builder, lp_build_max(float_bld, dsdx, dsdy), lp_build_int_to_float(float_bld, width), ""); if (dims > 1) { LLVMValueRef max; - max = LLVMBuildMul(bld->builder, + max = LLVMBuildFMul(bld->builder, lp_build_max(float_bld, dtdx, dtdy), lp_build_int_to_float(float_bld, height), ""); rho = lp_build_max(float_bld, rho, max); if (dims > 2) { - max = LLVMBuildMul(bld->builder, + max = LLVMBuildFMul(bld->builder, lp_build_max(float_bld, drdx, drdy), lp_build_int_to_float(float_bld, depth), ""); rho = lp_build_max(float_bld, rho, max); @@ -912,12 +911,12 @@ lp_build_lod_selector(struct lp_build_sample_context *bld, if (lod_bias) { lod_bias = LLVMBuildExtractElement(bld->builder, lod_bias, index0, ""); - lod = LLVMBuildAdd(bld->builder, lod, lod_bias, "shader_lod_bias"); + lod = LLVMBuildFAdd(bld->builder, lod, lod_bias, "shader_lod_bias"); } } /* add sampler lod bias */ - lod = LLVMBuildAdd(bld->builder, lod, sampler_lod_bias, "sampler_lod_bias"); + lod = LLVMBuildFAdd(bld->builder, lod, sampler_lod_bias, "sampler_lod_bias"); /* clamp lod */ lod = lp_build_clamp(float_bld, lod, min_lod, max_lod); @@ -1219,8 +1218,7 @@ lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord) /* ima = -0.5 / abs(coord); */ LLVMValueRef negHalf = lp_build_const_vec(coord_bld->type, -0.5); LLVMValueRef absCoord = lp_build_abs(coord_bld, coord); - LLVMValueRef ima = lp_build_mul(coord_bld, negHalf, - lp_build_rcp(coord_bld, absCoord)); + LLVMValueRef ima = lp_build_div(coord_bld, negHalf, absCoord); return ima; } @@ -1841,7 +1839,11 @@ lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld, unsigned i, j; for(j = 0; j < h16.type.length; j += 4) { - unsigned subindex = util_cpu_caps.little_endian ? 0 : 1; +#ifdef PIPE_ARCH_LITTLE_ENDIAN + unsigned subindex = 0; +#else + unsigned subindex = 1; +#endif LLVMValueRef index; index = LLVMConstInt(elem_type, j/2 + subindex, 0); @@ -2029,6 +2031,8 @@ lp_build_sample_soa(LLVMBuilderRef builder, debug_printf("Sample from %s\n", util_format_name(fmt)); } + assert(type.floating); + /* Setup our build context */ memset(&bld, 0, sizeof bld); bld.builder = builder; diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c index 21236839fb..0aa64affac 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c @@ -489,7 +489,7 @@ get_indirect_offsets(struct lp_build_tgsi_soa_context *bld, int_vec_type, ""); /* addr_vec = addr_vec * 4 */ - addr_vec = lp_build_mul(&bld->base, addr_vec, vec4); + addr_vec = lp_build_mul(&bld->int_bld, addr_vec, vec4); return addr_vec; } @@ -533,7 +533,7 @@ emit_fetch( reg->Register.Index * 4 + swizzle); /* index_vec = index_vec + addr_vec */ - index_vec = lp_build_add(&bld->base, index_vec, addr_vec); + index_vec = lp_build_add(&bld->int_bld, index_vec, addr_vec); /* Gather values from the constant buffer */ res = build_gather(bld, bld->consts_ptr, index_vec); @@ -612,11 +612,9 @@ emit_fetch( case TGSI_UTIL_SIGN_SET: /* TODO: Use bitwese OR for floating point */ res = lp_build_abs( &bld->base, res ); - res = LLVMBuildNeg( bld->base.builder, res, "" ); - break; - + /* fall through */ case TGSI_UTIL_SIGN_TOGGLE: - res = LLVMBuildNeg( bld->base.builder, res, "" ); + res = lp_build_negate( &bld->base, res ); break; case TGSI_UTIL_SIGN_KEEP: @@ -773,7 +771,9 @@ emit_store( addr = LLVMBuildExtractElement(bld->base.builder, addr, LLVMConstInt(LLVMInt32Type(), 0, 0), ""); - addr = lp_build_mul(&bld->base, addr, LLVMConstInt(LLVMInt32Type(), 4, 0)); + addr = LLVMBuildMul(bld->base.builder, + addr, LLVMConstInt(LLVMInt32Type(), 4, 0), + ""); } switch( reg->Register.File ) { |