From 59766ac273c426557b7790b0fcb566c8095fd820 Mon Sep 17 00:00:00 2001 From: Zack Rusin Date: Thu, 15 May 2008 17:46:20 -0400 Subject: llvm: implement sub and abs --- src/gallium/auxiliary/gallivm/soabuiltins.c | 38 ++++++++++++++++++++--------- 1 file changed, 27 insertions(+), 11 deletions(-) (limited to 'src/gallium/auxiliary/gallivm/soabuiltins.c') diff --git a/src/gallium/auxiliary/gallivm/soabuiltins.c b/src/gallium/auxiliary/gallivm/soabuiltins.c index 40addebd8c..f04e4c974d 100644 --- a/src/gallium/auxiliary/gallivm/soabuiltins.c +++ b/src/gallium/auxiliary/gallivm/soabuiltins.c @@ -33,6 +33,33 @@ */ typedef __attribute__(( ext_vector_type(4) )) float float4; + +extern float fabsf(float val); + +void abs(float4 *res, + float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w) +{ + res[0].x = fabsf(tmp0x.x); + res[0].y = fabsf(tmp0x.y); + res[0].z = fabsf(tmp0x.z); + res[0].w = fabsf(tmp0x.w); + + res[1].x = fabsf(tmp0y.x); + res[1].y = fabsf(tmp0y.y); + res[1].z = fabsf(tmp0y.z); + res[1].w = fabsf(tmp0y.w); + + res[2].x = fabsf(tmp0z.x); + res[2].y = fabsf(tmp0z.y); + res[2].z = fabsf(tmp0z.z); + res[2].w = fabsf(tmp0z.w); + + res[3].x = fabsf(tmp0w.x); + res[3].y = fabsf(tmp0w.y); + res[3].z = fabsf(tmp0w.z); + res[3].w = fabsf(tmp0w.w); +} + void dp3(float4 *res, float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w, float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w) @@ -77,14 +104,3 @@ void pow(float4 *res, res[2] = p; res[3] = p; } - -#if 0 -void yo(float4 *out, float4 *in) -{ - float4 res[4]; - - dp3(res, in[0], in[1], in[2], in[3], - in[4], in[5], in[6], in[7]); - out[1] = res[1]; -} -#endif -- cgit v1.2.3 From ea1a607292ef31df70cda8c6476755e0224c9f7d Mon Sep 17 00:00:00 2001 From: Zack Rusin Date: Fri, 16 May 2008 14:54:40 -0400 Subject: implement min/max and abstract ops on vectors --- src/gallium/auxiliary/gallivm/instructionssoa.cpp | 30 ++++++++- src/gallium/auxiliary/gallivm/instructionssoa.h | 4 ++ src/gallium/auxiliary/gallivm/soabuiltins.c | 75 ++++++++++++++++++++--- src/gallium/auxiliary/gallivm/tgsitollvm.cpp | 2 + 4 files changed, 100 insertions(+), 11 deletions(-) (limited to 'src/gallium/auxiliary/gallivm/soabuiltins.c') diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.cpp b/src/gallium/auxiliary/gallivm/instructionssoa.cpp index 4520559ba2..55fdda2791 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.cpp +++ b/src/gallium/auxiliary/gallivm/instructionssoa.cpp @@ -176,14 +176,17 @@ void InstructionsSoa::createFunctionMap() m_functionsMap[TGSI_OPCODE_ABS] = "abs"; m_functionsMap[TGSI_OPCODE_DP3] = "dp3"; m_functionsMap[TGSI_OPCODE_DP4] = "dp4"; + m_functionsMap[TGSI_OPCODE_MIN] = "min"; + m_functionsMap[TGSI_OPCODE_MAX] = "max"; m_functionsMap[TGSI_OPCODE_POWER] = "pow"; } void InstructionsSoa::createDependencies() { { - std::vector powDeps(1); + std::vector powDeps(2); powDeps[0] = "powf"; + powDeps[1] = "powvec"; m_builtinDependencies["pow"] = powDeps; } { @@ -191,6 +194,16 @@ void InstructionsSoa::createDependencies() absDeps[0] = "fabsf"; m_builtinDependencies["abs"] = absDeps; } + { + std::vector maxDeps(1); + maxDeps[0] = "maxvec"; + m_builtinDependencies["max"] = maxDeps; + } + { + std::vector minDeps(1); + minDeps[0] = "minvec"; + m_builtinDependencies["min"] = minDeps; + } } llvm::Function * InstructionsSoa::function(int op) @@ -374,6 +387,21 @@ std::vector InstructionsSoa::pow(const std::vector i return callBuiltin(func, in1, in2); } +std::vector InstructionsSoa::min(const std::vector in1, + const std::vector in2) +{ + llvm::Function *func = function(TGSI_OPCODE_MIN); + return callBuiltin(func, in1, in2); +} + + +std::vector InstructionsSoa::max(const std::vector in1, + const std::vector in2) +{ + llvm::Function *func = function(TGSI_OPCODE_MAX); + return callBuiltin(func, in1, in2); +} + void checkFunction(Function *func) { for (Function::const_iterator BI = func->begin(), BE = func->end(); diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.h b/src/gallium/auxiliary/gallivm/instructionssoa.h index 02e5fab51f..5a7f8fdf72 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.h +++ b/src/gallium/auxiliary/gallivm/instructionssoa.h @@ -59,6 +59,10 @@ public: std::vector madd(const std::vector in1, const std::vector in2, const std::vector in3); + std::vector max(const std::vector in1, + const std::vector in2); + std::vector min(const std::vector in1, + const std::vector in2); std::vector mul(const std::vector in1, const std::vector in2); std::vector pow(const std::vector in1, diff --git a/src/gallium/auxiliary/gallivm/soabuiltins.c b/src/gallium/auxiliary/gallivm/soabuiltins.c index f04e4c974d..935283f962 100644 --- a/src/gallium/auxiliary/gallivm/soabuiltins.c +++ b/src/gallium/auxiliary/gallivm/soabuiltins.c @@ -89,18 +89,73 @@ void dp4(float4 *res, extern float powf(float num, float p); +float4 powvec(float4 vec, float4 q) +{ + float4 p; + p.x = powf(vec.x, q.x); + p.y = powf(vec.y, q.y); + p.z = powf(vec.z, q.z); + p.w = powf(vec.w, q.w); + return p; +} + void pow(float4 *res, float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w, float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w) { - float4 p; - p.x = powf(tmp0x.x, tmp1x.x); - p.y = powf(tmp0x.y, tmp1x.y); - p.z = powf(tmp0x.z, tmp1x.z); - p.w = powf(tmp0x.w, tmp1x.w); - - res[0] = p; - res[1] = p; - res[2] = p; - res[3] = p; + res[0] = powvec(tmp0x, tmp1x); + res[1] = res[0]; + res[2] = res[0]; + res[3] = res[0]; +} + +void lit(float4 *res, + float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w, + float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w) +{ + const float4 zerovec = (float4) {0, 0, 0, 0}; + float4 tmpx = tmp0x; + float4 tmpy = tmp0y; + + res[0] = (float4){1.0, 1.0, 1.0, 1.0}; + res[1] = tmpx; + res[2] = tmpy; + res[3] = (float4){1.0, 1.0, 1.0, 1.0}; +} + +float4 minvec(float4 a, float4 b) +{ + return (float4){(a.x < b.x) ? a.x : b.x, + (a.y < b.y) ? a.y : b.y, + (a.z < b.z) ? a.z : b.z, + (a.w < b.w) ? a.w : b.w}; +} + +void min(float4 *res, + float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w, + float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w) +{ + res[0] = minvec(tmp0x, tmp1x); + res[1] = minvec(tmp0y, tmp1y); + res[2] = minvec(tmp0z, tmp1z); + res[3] = minvec(tmp0w, tmp1w); +} + + +float4 maxvec(float4 a, float4 b) +{ + return (float4){(a.x > b.x) ? a.x : b.x, + (a.y > b.y) ? a.y : b.y, + (a.z > b.z) ? a.z : b.z, + (a.w > b.w) ? a.w : b.w}; +} + +void max(float4 *res, + float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w, + float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w) +{ + res[0] = maxvec(tmp0x, tmp1x); + res[1] = maxvec(tmp0y, tmp1y); + res[2] = maxvec(tmp0z, tmp1z); + res[3] = maxvec(tmp0w, tmp1w); } diff --git a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp index 5465d3a95e..007b5c169a 100644 --- a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp +++ b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp @@ -724,9 +724,11 @@ translate_instructionir(llvm::Module *module, } break; case TGSI_OPCODE_MIN: { + out = instr->min(inputs[0], inputs[1]); } break; case TGSI_OPCODE_MAX: { + out = instr->max(inputs[0], inputs[1]); } break; case TGSI_OPCODE_SLT: { -- cgit v1.2.3 From 1d1cf8edf6a0409caf9aa7d44e186eb51f51fa1f Mon Sep 17 00:00:00 2001 From: Zack Rusin Date: Fri, 16 May 2008 16:06:59 -0400 Subject: do the lit (some artifacts present) --- src/gallium/auxiliary/gallivm/instructionssoa.cpp | 15 +++++++++ src/gallium/auxiliary/gallivm/instructionssoa.h | 1 + src/gallium/auxiliary/gallivm/soabuiltins.c | 37 ++++++++++++++--------- src/gallium/auxiliary/gallivm/tgsitollvm.cpp | 1 + 4 files changed, 40 insertions(+), 14 deletions(-) (limited to 'src/gallium/auxiliary/gallivm/soabuiltins.c') diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.cpp b/src/gallium/auxiliary/gallivm/instructionssoa.cpp index 55fdda2791..074dd0ecd6 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.cpp +++ b/src/gallium/auxiliary/gallivm/instructionssoa.cpp @@ -179,6 +179,7 @@ void InstructionsSoa::createFunctionMap() m_functionsMap[TGSI_OPCODE_MIN] = "min"; m_functionsMap[TGSI_OPCODE_MAX] = "max"; m_functionsMap[TGSI_OPCODE_POWER] = "pow"; + m_functionsMap[TGSI_OPCODE_LIT] = "lit"; } void InstructionsSoa::createDependencies() @@ -204,6 +205,14 @@ void InstructionsSoa::createDependencies() minDeps[0] = "minvec"; m_builtinDependencies["min"] = minDeps; } + { + std::vector litDeps(4); + litDeps[0] = "minvec"; + litDeps[1] = "maxvec"; + litDeps[2] = "powf"; + litDeps[3] = "powvec"; + m_builtinDependencies["lit"] = litDeps; + } } llvm::Function * InstructionsSoa::function(int op) @@ -475,3 +484,9 @@ std::vector InstructionsSoa::sub(const std::vector i return res; } +std::vector InstructionsSoa::lit(const std::vector in) +{ + llvm::Function *func = function(TGSI_OPCODE_LIT); + return callBuiltin(func, in); +} + diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.h b/src/gallium/auxiliary/gallivm/instructionssoa.h index 5a7f8fdf72..477ef4a157 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.h +++ b/src/gallium/auxiliary/gallivm/instructionssoa.h @@ -56,6 +56,7 @@ public: const std::vector in2); std::vector dp4(const std::vector in1, const std::vector in2); + std::vector lit(const std::vector in); std::vector madd(const std::vector in1, const std::vector in2, const std::vector in3); diff --git a/src/gallium/auxiliary/gallivm/soabuiltins.c b/src/gallium/auxiliary/gallivm/soabuiltins.c index 935283f962..b3bfebfe50 100644 --- a/src/gallium/auxiliary/gallivm/soabuiltins.c +++ b/src/gallium/auxiliary/gallivm/soabuiltins.c @@ -109,20 +109,6 @@ void pow(float4 *res, res[3] = res[0]; } -void lit(float4 *res, - float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w, - float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w) -{ - const float4 zerovec = (float4) {0, 0, 0, 0}; - float4 tmpx = tmp0x; - float4 tmpy = tmp0y; - - res[0] = (float4){1.0, 1.0, 1.0, 1.0}; - res[1] = tmpx; - res[2] = tmpy; - res[3] = (float4){1.0, 1.0, 1.0, 1.0}; -} - float4 minvec(float4 a, float4 b) { return (float4){(a.x < b.x) ? a.x : b.x, @@ -159,3 +145,26 @@ void max(float4 *res, res[2] = maxvec(tmp0z, tmp1z); res[3] = maxvec(tmp0w, tmp1w); } + + +void lit(float4 *res, + float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w) +{ + const float4 zerovec = (float4) {0, 0, 0, 0}; + const float4 min128 = (float4) {-128.f, -128.f, -128.f, -128.f}; + const float4 plus128 = (float4) {128.f, 128.f, 128.f, 128.f}; + + res[0] = (float4){1.0, 1.0, 1.0, 1.0}; + if (tmp0x.x > 0) { + float4 tmpx = maxvec(tmpx, zerovec); + float4 tmpy = maxvec(tmp0y, zerovec); + float4 tmpw = minvec(tmp0w, plus128); + tmpw = maxvec(tmpw, min128); + res[1] = tmpx; + res[2] = powvec(tmpy, tmpw); + } else { + res[1] = zerovec; + res[2] = zerovec; + } + res[3] = (float4){1.0, 1.0, 1.0, 1.0}; +} diff --git a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp index 007b5c169a..abcb240f46 100644 --- a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp +++ b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp @@ -692,6 +692,7 @@ translate_instructionir(llvm::Module *module, } break; case TGSI_OPCODE_LIT: { + out = instr->lit(inputs[0]); } break; case TGSI_OPCODE_RCP: { -- cgit v1.2.3 From 02e45b2dadd42c38247cb992a07eb520ac86519b Mon Sep 17 00:00:00 2001 From: Zack Rusin Date: Fri, 16 May 2008 17:10:52 -0400 Subject: fix abs and start on rsq --- src/gallium/auxiliary/gallivm/instructionssoa.cpp | 20 +++++++- src/gallium/auxiliary/gallivm/instructionssoa.h | 1 + src/gallium/auxiliary/gallivm/soabuiltins.c | 56 +++++++++++++++-------- src/gallium/auxiliary/gallivm/tgsitollvm.cpp | 1 + 4 files changed, 58 insertions(+), 20 deletions(-) (limited to 'src/gallium/auxiliary/gallivm/soabuiltins.c') diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.cpp b/src/gallium/auxiliary/gallivm/instructionssoa.cpp index 074dd0ecd6..76049ade7c 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.cpp +++ b/src/gallium/auxiliary/gallivm/instructionssoa.cpp @@ -180,6 +180,7 @@ void InstructionsSoa::createFunctionMap() m_functionsMap[TGSI_OPCODE_MAX] = "max"; m_functionsMap[TGSI_OPCODE_POWER] = "pow"; m_functionsMap[TGSI_OPCODE_LIT] = "lit"; + m_functionsMap[TGSI_OPCODE_RSQ] = "rsq"; } void InstructionsSoa::createDependencies() @@ -191,8 +192,9 @@ void InstructionsSoa::createDependencies() m_builtinDependencies["pow"] = powDeps; } { - std::vector absDeps(1); + std::vector absDeps(2); absDeps[0] = "fabsf"; + absDeps[1] = "absvec"; m_builtinDependencies["abs"] = absDeps; } { @@ -213,6 +215,14 @@ void InstructionsSoa::createDependencies() litDeps[3] = "powvec"; m_builtinDependencies["lit"] = litDeps; } + { + std::vector rsqDeps(4); + rsqDeps[0] = "sqrtf"; + rsqDeps[1] = "sqrtvec"; + rsqDeps[2] = "fabsf"; + rsqDeps[3] = "absvec"; + m_builtinDependencies["rsq"] = rsqDeps; + } } llvm::Function * InstructionsSoa::function(int op) @@ -453,7 +463,9 @@ void InstructionsSoa::injectFunction(llvm::Function *originalFunc, int op) currentModule()->dump(); } else { DenseMap val; + val[m_builtins->getFunction("fabsf")] = currentModule()->getFunction("fabsf"); val[m_builtins->getFunction("powf")] = currentModule()->getFunction("powf"); + val[m_builtins->getFunction("sqrtf")] = currentModule()->getFunction("sqrtf"); func = CloneFunction(originalFunc, val); #if 0 std::cout <<" replacing "<getFunction("powf") @@ -490,3 +502,9 @@ std::vector InstructionsSoa::lit(const std::vector i return callBuiltin(func, in); } +std::vector InstructionsSoa::rsq(const std::vector in) +{ + llvm::Function *func = function(TGSI_OPCODE_RSQ); + return callBuiltin(func, in); +} + diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.h b/src/gallium/auxiliary/gallivm/instructionssoa.h index 477ef4a157..3e20b652dd 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.h +++ b/src/gallium/auxiliary/gallivm/instructionssoa.h @@ -68,6 +68,7 @@ public: const std::vector in2); std::vector pow(const std::vector in1, const std::vector in2); + std::vector rsq(const std::vector in1); std::vector sub(const std::vector in1, const std::vector in2); void end(); diff --git a/src/gallium/auxiliary/gallivm/soabuiltins.c b/src/gallium/auxiliary/gallivm/soabuiltins.c index b3bfebfe50..64c02aa967 100644 --- a/src/gallium/auxiliary/gallivm/soabuiltins.c +++ b/src/gallium/auxiliary/gallivm/soabuiltins.c @@ -36,28 +36,24 @@ typedef __attribute__(( ext_vector_type(4) )) float float4; extern float fabsf(float val); +float4 absvec(float4 vec) +{ + float4 res; + res.x = fabsf(vec.x); + res.y = fabsf(vec.y); + res.z = fabsf(vec.z); + res.w = fabsf(vec.w); + + return res; +} + void abs(float4 *res, float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w) { - res[0].x = fabsf(tmp0x.x); - res[0].y = fabsf(tmp0x.y); - res[0].z = fabsf(tmp0x.z); - res[0].w = fabsf(tmp0x.w); - - res[1].x = fabsf(tmp0y.x); - res[1].y = fabsf(tmp0y.y); - res[1].z = fabsf(tmp0y.z); - res[1].w = fabsf(tmp0y.w); - - res[2].x = fabsf(tmp0z.x); - res[2].y = fabsf(tmp0z.y); - res[2].z = fabsf(tmp0z.z); - res[2].w = fabsf(tmp0z.w); - - res[3].x = fabsf(tmp0w.x); - res[3].y = fabsf(tmp0w.y); - res[3].z = fabsf(tmp0w.z); - res[3].w = fabsf(tmp0w.w); + res[0] = absvec(tmp0x); + res[1] = absvec(tmp0y); + res[2] = absvec(tmp0z); + res[3] = absvec(tmp0w); } void dp3(float4 *res, @@ -88,6 +84,7 @@ void dp4(float4 *res, } extern float powf(float num, float p); +extern float sqrtf(float x); float4 powvec(float4 vec, float4 q) { @@ -168,3 +165,24 @@ void lit(float4 *res, } res[3] = (float4){1.0, 1.0, 1.0, 1.0}; } + + +float4 sqrtvec(float4 vec) +{ + float4 p; + p.x = sqrtf(vec.x); + p.y = sqrtf(vec.y); + p.z = sqrtf(vec.z); + p.w = sqrtf(vec.w); + return p; +} + +void rsq(float4 *res, + float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w) +{ + const float4 onevec = (float4) {1., 1., 1., 1.}; + res[0] = onevec/sqrtvec(absvec(tmp0x)); + res[1] = res[0]; + res[2] = res[0]; + res[3] = res[0]; +} diff --git a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp index abcb240f46..9695358ab8 100644 --- a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp +++ b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp @@ -699,6 +699,7 @@ translate_instructionir(llvm::Module *module, } break; case TGSI_OPCODE_RSQ: { + out = instr->rsq(inputs[0]); } break; case TGSI_OPCODE_EXP: -- cgit v1.2.3 From a7449d4d840148ccd9261b59e68d45e9d0d2be53 Mon Sep 17 00:00:00 2001 From: Zack Rusin Date: Fri, 16 May 2008 17:35:47 -0400 Subject: fix rsq --- src/gallium/auxiliary/gallivm/soabuiltins.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'src/gallium/auxiliary/gallivm/soabuiltins.c') diff --git a/src/gallium/auxiliary/gallivm/soabuiltins.c b/src/gallium/auxiliary/gallivm/soabuiltins.c index 64c02aa967..62c75f18f4 100644 --- a/src/gallium/auxiliary/gallivm/soabuiltins.c +++ b/src/gallium/auxiliary/gallivm/soabuiltins.c @@ -182,7 +182,7 @@ void rsq(float4 *res, { const float4 onevec = (float4) {1., 1., 1., 1.}; res[0] = onevec/sqrtvec(absvec(tmp0x)); - res[1] = res[0]; - res[2] = res[0]; - res[3] = res[0]; + res[1] = onevec/sqrtvec(absvec(tmp0y)); + res[2] = onevec/sqrtvec(absvec(tmp0z)); + res[3] = onevec/sqrtvec(absvec(tmp0w)); } -- cgit v1.2.3 From 0bf82c0111c9d5e33ffc76be2fd0d22eea316952 Mon Sep 17 00:00:00 2001 From: Zack Rusin Date: Fri, 16 May 2008 17:56:38 -0400 Subject: cosmetic changes --- src/gallium/auxiliary/gallivm/soabuiltins.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) (limited to 'src/gallium/auxiliary/gallivm/soabuiltins.c') diff --git a/src/gallium/auxiliary/gallivm/soabuiltins.c b/src/gallium/auxiliary/gallivm/soabuiltins.c index 62c75f18f4..78f84510e2 100644 --- a/src/gallium/auxiliary/gallivm/soabuiltins.c +++ b/src/gallium/auxiliary/gallivm/soabuiltins.c @@ -147,17 +147,16 @@ void max(float4 *res, void lit(float4 *res, float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w) { - const float4 zerovec = (float4) {0, 0, 0, 0}; + const float4 zerovec = (float4) {0.0, 0.0, 0.0, 0.0}; const float4 min128 = (float4) {-128.f, -128.f, -128.f, -128.f}; const float4 plus128 = (float4) {128.f, 128.f, 128.f, 128.f}; res[0] = (float4){1.0, 1.0, 1.0, 1.0}; if (tmp0x.x > 0) { - float4 tmpx = maxvec(tmpx, zerovec); float4 tmpy = maxvec(tmp0y, zerovec); float4 tmpw = minvec(tmp0w, plus128); tmpw = maxvec(tmpw, min128); - res[1] = tmpx; + res[1] = tmp0x; res[2] = powvec(tmpy, tmpw); } else { res[1] = zerovec; -- cgit v1.2.3