diff options
Diffstat (limited to 'src/gallium/auxiliary')
32 files changed, 2204 insertions, 414 deletions
diff --git a/src/gallium/auxiliary/cso_cache/cso_hash.c b/src/gallium/auxiliary/cso_cache/cso_hash.c index 7f0044c5a7..4e7664f9bf 100644 --- a/src/gallium/auxiliary/cso_cache/cso_hash.c +++ b/src/gallium/auxiliary/cso_cache/cso_hash.c @@ -431,3 +431,9 @@ struct cso_hash_iter cso_hash_erase(struct cso_hash *hash, struct cso_hash_iter --hash->data.d->size; return ret; } + +boolean cso_hash_contains(struct cso_hash *hash, unsigned key) +{ + struct cso_node **node = cso_hash_find_node(hash, key); + return (*node != hash->data.e); +} diff --git a/src/gallium/auxiliary/cso_cache/cso_hash.h b/src/gallium/auxiliary/cso_cache/cso_hash.h index 85f3e276c6..5891c325fa 100644 --- a/src/gallium/auxiliary/cso_cache/cso_hash.h +++ b/src/gallium/auxiliary/cso_cache/cso_hash.h @@ -44,6 +44,7 @@ #ifndef CSO_HASH_H #define CSO_HASH_H +#include "pipe/p_compiler.h" #ifdef __cplusplus extern "C" { @@ -95,6 +96,11 @@ struct cso_hash_iter cso_hash_first_node(struct cso_hash *hash); */ struct cso_hash_iter cso_hash_find(struct cso_hash *hash, unsigned key); +/** + * Returns true if a value with the given key exists in the hash + */ +boolean cso_hash_contains(struct cso_hash *hash, unsigned key); + int cso_hash_iter_is_null(struct cso_hash_iter iter); unsigned cso_hash_iter_key(struct cso_hash_iter iter); diff --git a/src/gallium/auxiliary/draw/draw_private.h b/src/gallium/auxiliary/draw/draw_private.h index 37c4c87f87..5d531146c5 100644 --- a/src/gallium/auxiliary/draw/draw_private.h +++ b/src/gallium/auxiliary/draw/draw_private.h @@ -196,7 +196,7 @@ struct draw_context const float (*aligned_constants)[4]; - float (*aligned_constant_storage)[4]; + const float (*aligned_constant_storage)[4]; unsigned const_storage_size; diff --git a/src/gallium/auxiliary/draw/draw_vs_aos_io.c b/src/gallium/auxiliary/draw/draw_vs_aos_io.c index 39f75b50b7..dd79bc799a 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos_io.c +++ b/src/gallium/auxiliary/draw/draw_vs_aos_io.c @@ -338,7 +338,7 @@ static void emit_store_R32G32B32A32( struct aos_compilation *cp, struct x86_reg dst_ptr, struct x86_reg dataXMM ) { - sse_movups(cp->func, dst_ptr, dataXMM); + sse_movaps(cp->func, dst_ptr, dataXMM); } static void emit_store_R32G32B32( struct aos_compilation *cp, diff --git a/src/gallium/auxiliary/draw/draw_vs_exec.c b/src/gallium/auxiliary/draw/draw_vs_exec.c index 44563803f9..79a19d6be2 100644 --- a/src/gallium/auxiliary/draw/draw_vs_exec.c +++ b/src/gallium/auxiliary/draw/draw_vs_exec.c @@ -46,6 +46,7 @@ struct exec_vertex_shader { struct draw_vertex_shader base; struct tgsi_exec_machine *machine; + const struct tgsi_token *machine_tokens; }; static struct exec_vertex_shader *exec_vertex_shader( struct draw_vertex_shader *vs ) @@ -62,12 +63,16 @@ vs_exec_prepare( struct draw_vertex_shader *shader, { struct exec_vertex_shader *evs = exec_vertex_shader(shader); - /* specify the vertex program to interpret/execute */ - tgsi_exec_machine_bind_shader(evs->machine, - shader->state.tokens, - PIPE_MAX_SAMPLERS, - NULL /*samplers*/ ); - + /* Specify the vertex program to interpret/execute. + * Avoid rebinding when possible. + */ + if (evs->machine_tokens != shader->state.tokens) { + tgsi_exec_machine_bind_shader(evs->machine, + shader->state.tokens, + PIPE_MAX_SAMPLERS, + NULL /*samplers*/ ); + evs->machine_tokens = shader->state.tokens; + } } diff --git a/src/gallium/auxiliary/draw/draw_vs_sse.c b/src/gallium/auxiliary/draw/draw_vs_sse.c index 0efabd9de8..b11ae31662 100644 --- a/src/gallium/auxiliary/draw/draw_vs_sse.c +++ b/src/gallium/auxiliary/draw/draw_vs_sse.c @@ -37,7 +37,7 @@ #include "draw_vs.h" -#if defined(PIPE_ARCH_X86) +#if defined(PIPE_ARCH_X86) && defined(PIPE_ARCH_SSE) #include "pipe/p_shader_tokens.h" diff --git a/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp b/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp index 3a4a41e544..3a2f2878a3 100644 --- a/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp +++ b/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp @@ -158,8 +158,8 @@ void gallivm_cpu_jit_compile(struct gallivm_cpu_engine *cpu, struct gallivm_prog llvm::ExistingModuleProvider *mp = new llvm::ExistingModuleProvider(mod); llvm::ExecutionEngine *ee = cpu->engine; assert(ee); - /*FIXME : remove */ - ee->DisableLazyCompilation(); + /*FIXME : why was this disabled ? we need it for pow/sqrt/... */ + ee->DisableLazyCompilation(false); ee->addModuleProvider(mp); llvm::Function *func = func_for_shader(prog); @@ -202,7 +202,6 @@ int gallivm_cpu_vs_exec(struct gallivm_prog *prog, unsigned int i, j; unsigned slot; vertex_shader_runner runner = reinterpret_cast<vertex_shader_runner>(prog->function); - assert(runner); for (i = 0; i < count; i += MAX_TGSI_VERTICES) { diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.cpp b/src/gallium/auxiliary/gallivm/instructionssoa.cpp index a658072551..d5600fd22d 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.cpp +++ b/src/gallium/auxiliary/gallivm/instructionssoa.cpp @@ -90,68 +90,11 @@ llvm::Value * InstructionsSoa::vectorFromVals(llvm::Value *x, llvm::Value *y, return res; } -std::vector<llvm::Value*> InstructionsSoa::arl(const std::vector<llvm::Value*> in) -{ - std::vector<llvm::Value*> res(4); - - //Extract x's - llvm::Value *x1 = m_builder.CreateExtractElement(in[0], - m_storage->constantInt(0), - name("extractX")); - //cast it to an unsigned int - x1 = m_builder.CreateFPToUI(x1, IntegerType::get(32), name("x1IntCast")); - - res[0] = x1;//vectorFromVals(x1, x2, x3, x4); - //only x is valid. the others shouldn't be necessary - /* - res[1] = Constant::getNullValue(m_floatVecType); - res[2] = Constant::getNullValue(m_floatVecType); - res[3] = Constant::getNullValue(m_floatVecType); - */ - - return res; -} - - -std::vector<llvm::Value*> InstructionsSoa::add(const std::vector<llvm::Value*> in1, - const std::vector<llvm::Value*> in2) -{ - std::vector<llvm::Value*> res(4); - - res[0] = m_builder.CreateAdd(in1[0], in2[0], name("addx")); - res[1] = m_builder.CreateAdd(in1[1], in2[1], name("addy")); - res[2] = m_builder.CreateAdd(in1[2], in2[2], name("addz")); - res[3] = m_builder.CreateAdd(in1[3], in2[3], name("addw")); - - return res; -} - -std::vector<llvm::Value*> InstructionsSoa::mul(const std::vector<llvm::Value*> in1, - const std::vector<llvm::Value*> in2) -{ - std::vector<llvm::Value*> res(4); - - res[0] = m_builder.CreateMul(in1[0], in2[0], name("mulx")); - res[1] = m_builder.CreateMul(in1[1], in2[1], name("muly")); - res[2] = m_builder.CreateMul(in1[2], in2[2], name("mulz")); - res[3] = m_builder.CreateMul(in1[3], in2[3], name("mulw")); - - return res; -} - void InstructionsSoa::end() { m_builder.CreateRetVoid(); } -std::vector<llvm::Value*> InstructionsSoa::madd(const std::vector<llvm::Value*> in1, - const std::vector<llvm::Value*> in2, - const std::vector<llvm::Value*> in3) -{ - std::vector<llvm::Value*> res = mul(in1, in2); - return add(res, in3); -} - std::vector<llvm::Value*> InstructionsSoa::extractVector(llvm::Value *vector) { std::vector<llvm::Value*> res(4); @@ -171,6 +114,11 @@ std::vector<llvm::Value*> InstructionsSoa::extractVector(llvm::Value *vector) return res; } +llvm::IRBuilder<>* InstructionsSoa::getIRBuilder() +{ + return &m_builder; +} + void InstructionsSoa::createFunctionMap() { m_functionsMap[TGSI_OPCODE_ABS] = "abs"; @@ -274,6 +222,41 @@ std::vector<llvm::Value*> InstructionsSoa::abs(const std::vector<llvm::Value*> i return callBuiltin(func, in1); } +std::vector<llvm::Value*> InstructionsSoa::add(const std::vector<llvm::Value*> in1, + const std::vector<llvm::Value*> in2) +{ + std::vector<llvm::Value*> res(4); + + res[0] = m_builder.CreateAdd(in1[0], in2[0], name("addx")); + res[1] = m_builder.CreateAdd(in1[1], in2[1], name("addy")); + res[2] = m_builder.CreateAdd(in1[2], in2[2], name("addz")); + res[3] = m_builder.CreateAdd(in1[3], in2[3], name("addw")); + + return res; +} + +std::vector<llvm::Value*> InstructionsSoa::arl(const std::vector<llvm::Value*> in) +{ + std::vector<llvm::Value*> res(4); + + //Extract x's + llvm::Value *x1 = m_builder.CreateExtractElement(in[0], + m_storage->constantInt(0), + name("extractX")); + //cast it to an unsigned int + x1 = m_builder.CreateFPToUI(x1, IntegerType::get(32), name("x1IntCast")); + + res[0] = x1;//vectorFromVals(x1, x2, x3, x4); + //only x is valid. the others shouldn't be necessary + /* + res[1] = Constant::getNullValue(m_floatVecType); + res[2] = Constant::getNullValue(m_floatVecType); + res[3] = Constant::getNullValue(m_floatVecType); + */ + + return res; +} + std::vector<llvm::Value*> InstructionsSoa::dp3(const std::vector<llvm::Value*> in1, const std::vector<llvm::Value*> in2) { @@ -281,6 +264,59 @@ std::vector<llvm::Value*> InstructionsSoa::dp3(const std::vector<llvm::Value*> i return callBuiltin(func, in1, in2); } +std::vector<llvm::Value*> InstructionsSoa::lit(const std::vector<llvm::Value*> in) +{ + llvm::Function *func = function(TGSI_OPCODE_LIT); + return callBuiltin(func, in); +} + +std::vector<llvm::Value*> InstructionsSoa::madd(const std::vector<llvm::Value*> in1, + const std::vector<llvm::Value*> in2, + const std::vector<llvm::Value*> in3) +{ + std::vector<llvm::Value*> res = mul(in1, in2); + return add(res, in3); +} + +std::vector<llvm::Value*> InstructionsSoa::max(const std::vector<llvm::Value*> in1, + const std::vector<llvm::Value*> in2) +{ + llvm::Function *func = function(TGSI_OPCODE_MAX); + return callBuiltin(func, in1, in2); +} + +std::vector<llvm::Value*> InstructionsSoa::min(const std::vector<llvm::Value*> in1, + const std::vector<llvm::Value*> in2) +{ + llvm::Function *func = function(TGSI_OPCODE_MIN); + return callBuiltin(func, in1, in2); +} + +std::vector<llvm::Value*> InstructionsSoa::mul(const std::vector<llvm::Value*> in1, + const std::vector<llvm::Value*> in2) +{ + std::vector<llvm::Value*> res(4); + + res[0] = m_builder.CreateMul(in1[0], in2[0], name("mulx")); + res[1] = m_builder.CreateMul(in1[1], in2[1], name("muly")); + res[2] = m_builder.CreateMul(in1[2], in2[2], name("mulz")); + res[3] = m_builder.CreateMul(in1[3], in2[3], name("mulw")); + + return res; +} + +std::vector<llvm::Value*> InstructionsSoa::pow(const std::vector<llvm::Value*> in1, + const std::vector<llvm::Value*> in2) +{ + llvm::Function *func = function(TGSI_OPCODE_POWER); + return callBuiltin(func, in1, in2); +} + +std::vector<llvm::Value*> InstructionsSoa::rsq(const std::vector<llvm::Value*> in) +{ + llvm::Function *func = function(TGSI_OPCODE_RSQ); + return callBuiltin(func, in); +} std::vector<llvm::Value*> InstructionsSoa::slt(const std::vector<llvm::Value*> in1, const std::vector<llvm::Value*> in2) @@ -289,6 +325,37 @@ std::vector<llvm::Value*> InstructionsSoa::slt(const std::vector<llvm::Value*> i return callBuiltin(func, in1, in2); } +std::vector<llvm::Value*> InstructionsSoa::sub(const std::vector<llvm::Value*> in1, + const std::vector<llvm::Value*> in2) +{ + std::vector<llvm::Value*> res(4); + + res[0] = m_builder.CreateSub(in1[0], in2[0], name("subx")); + res[1] = m_builder.CreateSub(in1[1], in2[1], name("suby")); + res[2] = m_builder.CreateSub(in1[2], in2[2], name("subz")); + res[3] = m_builder.CreateSub(in1[3], in2[3], name("subw")); + + return res; +} + +void checkFunction(Function *func) +{ + for (Function::const_iterator BI = func->begin(), BE = func->end(); + BI != BE; ++BI) { + const BasicBlock &BB = *BI; + for (BasicBlock::const_iterator II = BB.begin(), IE = BB.end(); + II != IE; ++II) { + const Instruction &I = *II; + std::cout<< "Instr = "<<I; + for (unsigned op = 0, E = I.getNumOperands(); op != E; ++op) { + const Value *Op = I.getOperand(op); + std::cout<< "\top = "<<Op<<"("<<op<<")"<<std::endl; + //I->setOperand(op, V); + } + } + } +} + llvm::Value * InstructionsSoa::allocaTemp() { VectorType *vector = VectorType::get(Type::FloatTy, 4); @@ -408,46 +475,6 @@ std::vector<Value*> InstructionsSoa::callBuiltin(llvm::Function *func, const std return allocaToResult(allocaPtr); } -std::vector<llvm::Value*> InstructionsSoa::pow(const std::vector<llvm::Value*> in1, - const std::vector<llvm::Value*> in2) -{ - llvm::Function *func = function(TGSI_OPCODE_POWER); - return callBuiltin(func, in1, in2); -} - -std::vector<llvm::Value*> InstructionsSoa::min(const std::vector<llvm::Value*> in1, - const std::vector<llvm::Value*> in2) -{ - llvm::Function *func = function(TGSI_OPCODE_MIN); - return callBuiltin(func, in1, in2); -} - - -std::vector<llvm::Value*> InstructionsSoa::max(const std::vector<llvm::Value*> in1, - const std::vector<llvm::Value*> 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(); - BI != BE; ++BI) { - const BasicBlock &BB = *BI; - for (BasicBlock::const_iterator II = BB.begin(), IE = BB.end(); - II != IE; ++II) { - const Instruction &I = *II; - std::cout<< "Instr = "<<I; - for (unsigned op = 0, E = I.getNumOperands(); op != E; ++op) { - const Value *Op = I.getOperand(op); - std::cout<< "\top = "<<Op<<"("<<op<<")"<<std::endl; - //I->setOperand(op, V); - } - } - } -} - void InstructionsSoa::injectFunction(llvm::Function *originalFunc, int op) { assert(originalFunc); @@ -492,28 +519,4 @@ void InstructionsSoa::injectFunction(llvm::Function *originalFunc, int op) } } -std::vector<llvm::Value*> InstructionsSoa::sub(const std::vector<llvm::Value*> in1, - const std::vector<llvm::Value*> in2) -{ - std::vector<llvm::Value*> res(4); - - res[0] = m_builder.CreateSub(in1[0], in2[0], name("subx")); - res[1] = m_builder.CreateSub(in1[1], in2[1], name("suby")); - res[2] = m_builder.CreateSub(in1[2], in2[2], name("subz")); - res[3] = m_builder.CreateSub(in1[3], in2[3], name("subw")); - - return res; -} - -std::vector<llvm::Value*> InstructionsSoa::lit(const std::vector<llvm::Value*> in) -{ - llvm::Function *func = function(TGSI_OPCODE_LIT); - return callBuiltin(func, in); -} - -std::vector<llvm::Value*> InstructionsSoa::rsq(const std::vector<llvm::Value*> 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 3817fdc904..d6831e0a6b 100644 --- a/src/gallium/auxiliary/gallivm/instructionssoa.h +++ b/src/gallium/auxiliary/gallivm/instructionssoa.h @@ -76,6 +76,7 @@ public: void end(); std::vector<llvm::Value*> extractVector(llvm::Value *vector); + llvm::IRBuilder<>* getIRBuilder(); private: const char * name(const char *prefix) const; llvm::Value *vectorFromVals(llvm::Value *x, llvm::Value *y, diff --git a/src/gallium/auxiliary/gallivm/storagesoa.cpp b/src/gallium/auxiliary/gallivm/storagesoa.cpp index 78d754371f..4fc075cf6d 100644 --- a/src/gallium/auxiliary/gallivm/storagesoa.cpp +++ b/src/gallium/auxiliary/gallivm/storagesoa.cpp @@ -93,7 +93,7 @@ void StorageSoa::declareImmediates() std::vector<float> vals(4); std::vector<Constant*> channelArray; - vals[0] = vec[0]; vals[1] = vec[0]; vals[2] = vec[0]; vals[3] = vec[0]; + vals[0] = vec[0]; vals[1] = vec[1]; vals[2] = vec[2]; vals[3] = vec[3]; llvm::Constant *xChannel = createConstGlobalVector(vals); vals[0] = vec[1]; vals[1] = vec[1]; vals[2] = vec[1]; vals[3] = vec[1]; @@ -144,22 +144,43 @@ std::vector<llvm::Value*> StorageSoa::inputElement(llvm::Value *idx) return res; } -std::vector<llvm::Value*> StorageSoa::constElement(llvm::Value *idx) +llvm::Value* StorageSoa::unpackConstElement(llvm::IRBuilder<>* m_builder, llvm::Value* vector, int cc) { - std::vector<llvm::Value*> res(4); + std::vector<llvm::Value*> x(4); + x[0] = m_builder->CreateExtractElement(vector, + constantInt(cc), + name("x")); + + VectorType *vectorType = VectorType::get(Type::FloatTy, 4); + Constant *constVector = Constant::getNullValue(vectorType); + Value *res = m_builder->CreateInsertElement(constVector, x[0], + constantInt(0), + name("vecx")); + res = m_builder->CreateInsertElement(res, x[0], constantInt(1), + name("vecxx")); + res = m_builder->CreateInsertElement(res, x[0], constantInt(2), + name("vecxxx")); + res = m_builder->CreateInsertElement(res, x[0], constantInt(3), + name("vecxxxx")); + return res; +} + +std::vector<llvm::Value*> StorageSoa::constElement(llvm::IRBuilder<>* m_builder, llvm::Value *idx) +{ + llvm::Value* res; + std::vector<llvm::Value*> res2(4); llvm::Value *xChannel, *yChannel, *zChannel, *wChannel; xChannel = elementPointer(m_consts, idx, 0); - yChannel = elementPointer(m_consts, idx, 1); - zChannel = elementPointer(m_consts, idx, 2); - wChannel = elementPointer(m_consts, idx, 3); - res[0] = alignedArrayLoad(xChannel); - res[1] = alignedArrayLoad(yChannel); - res[2] = alignedArrayLoad(zChannel); - res[3] = alignedArrayLoad(wChannel); + res = alignedArrayLoad(xChannel); - return res; + res2[0]=unpackConstElement(m_builder, res,0); + res2[1]=unpackConstElement(m_builder, res,1); + res2[2]=unpackConstElement(m_builder, res,2); + res2[3]=unpackConstElement(m_builder, res,3); + + return res2; } std::vector<llvm::Value*> StorageSoa::outputElement(llvm::Value *idx) @@ -260,6 +281,12 @@ llvm::Module * StorageSoa::currentModule() const return m_block->getParent()->getParent(); } +llvm::Constant * StorageSoa::createConstGlobalFloat(const float val) +{ + Constant*c = ConstantFP::get(APFloat(val)); + return c; +} + llvm::Constant * StorageSoa::createConstGlobalVector(const std::vector<float> &vec) { VectorType *vectorType = VectorType::get(Type::FloatTy, 4); @@ -278,7 +305,7 @@ llvm::Constant * StorageSoa::createConstGlobalVector(const std::vector<float> &v } std::vector<llvm::Value*> StorageSoa::load(enum tgsi_file_type type, int idx, int swizzle, - llvm::Value *indIdx) + llvm::IRBuilder<>* m_builder,llvm::Value *indIdx) { std::vector<llvm::Value*> val(4); @@ -302,7 +329,7 @@ std::vector<llvm::Value*> StorageSoa::load(enum tgsi_file_type type, int idx, in val = tempElement(realIndex); break; case TGSI_FILE_CONSTANT: - val = constElement(realIndex); + val = constElement(m_builder, realIndex); break; case TGSI_FILE_IMMEDIATE: val = immediateElement(realIndex); diff --git a/src/gallium/auxiliary/gallivm/storagesoa.h b/src/gallium/auxiliary/gallivm/storagesoa.h index ae2fc7c6ae..f21ca6ec43 100644 --- a/src/gallium/auxiliary/gallivm/storagesoa.h +++ b/src/gallium/auxiliary/gallivm/storagesoa.h @@ -29,6 +29,7 @@ #define STORAGESOA_H #include <pipe/p_shader_tokens.h> +#include <llvm/Support/IRBuilder.h> #include <vector> #include <list> @@ -56,7 +57,7 @@ public: std::vector<llvm::Value*> load(enum tgsi_file_type type, int idx, int swizzle, - llvm::Value *indIdx =0); + llvm::IRBuilder<>* m_builder, llvm::Value *indIdx =0); void store(enum tgsi_file_type type, int idx, const std::vector<llvm::Value*> &val, int mask); @@ -76,10 +77,12 @@ private: const char *name(const char *prefix) const; llvm::Value *alignedArrayLoad(llvm::Value *val); llvm::Module *currentModule() const; + llvm::Constant *createConstGlobalFloat(const float val); llvm::Constant *createConstGlobalVector(const std::vector<float> &vec); std::vector<llvm::Value*> inputElement(llvm::Value *indIdx); - std::vector<llvm::Value*> constElement(llvm::Value *indIdx); + llvm::Value* unpackConstElement(llvm::IRBuilder<>* m_builder, llvm::Value *indIdx, int cc); + std::vector<llvm::Value*> constElement(llvm::IRBuilder<>* m_builder, llvm::Value *indIdx); std::vector<llvm::Value*> outputElement(llvm::Value *indIdx); std::vector<llvm::Value*> tempElement(llvm::Value *indIdx); std::vector<llvm::Value*> immediateElement(llvm::Value *indIdx); diff --git a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp index 7292c0e366..1191a6cae9 100644 --- a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp +++ b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp @@ -52,7 +52,7 @@ static inline FunctionType *vertexShaderFunctionType() // pass are castable to the following: // [4 x <4 x float>] inputs, // [4 x <4 x float>] output, - // [4 x [4 x float]] consts, + // [4 x [1 x float]] consts, // [4 x <4 x float>] temps std::vector<const Type*> funcArgs; @@ -61,7 +61,7 @@ static inline FunctionType *vertexShaderFunctionType() PointerType *vectorArrayPtr = PointerType::get(vectorArray, 0); ArrayType *floatArray = ArrayType::get(Type::FloatTy, 4); - ArrayType *constsArray = ArrayType::get(floatArray, 4); + ArrayType *constsArray = ArrayType::get(floatArray, 1); PointerType *constsArrayPtr = PointerType::get(constsArray, 0); funcArgs.push_back(vectorArrayPtr);//inputs @@ -246,6 +246,7 @@ translate_instruction(llvm::Module *module, val = storage->constElement(src->SrcRegister.Index, indIdx); } else if (src->SrcRegister.File == TGSI_FILE_INPUT) { val = storage->inputElement(src->SrcRegister.Index, indIdx); + // FIXME we should not be generating elements for temporaries, this creates useless memory writes } else if (src->SrcRegister.File == TGSI_FILE_TEMPORARY) { val = storage->tempElement(src->SrcRegister.Index); } else if (src->SrcRegister.File == TGSI_FILE_OUTPUT) { @@ -676,6 +677,7 @@ translate_instruction(llvm::Module *module, if (dst->DstRegister.File == TGSI_FILE_OUTPUT) { storage->setOutputElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask); + // FIXME we should not be generating elements for temporaries, this creates useless memory writes } else if (dst->DstRegister.File == TGSI_FILE_TEMPORARY) { storage->setTempElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask); } else if (dst->DstRegister.File == TGSI_FILE_ADDRESS) { @@ -707,9 +709,8 @@ translate_instructionir(llvm::Module *module, if (src->SrcRegister.Indirect) { indIdx = storage->addrElement(src->SrcRegisterInd.Index); } - val = storage->load((enum tgsi_file_type)src->SrcRegister.File, - src->SrcRegister.Index, swizzle, indIdx); + src->SrcRegister.Index, swizzle, instr->getIRBuilder(), indIdx); inputs[i] = val; } @@ -1025,7 +1026,6 @@ translate_instructionir(llvm::Module *module, /* store results */ for (int i = 0; i < inst->Instruction.NumDstRegs; ++i) { struct tgsi_full_dst_register *dst = &inst->FullDstRegisters[i]; - storage->store((enum tgsi_file_type)dst->DstRegister.File, dst->DstRegister.Index, out, dst->DstRegister.WriteMask); } diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc.c b/src/gallium/auxiliary/rtasm/rtasm_ppc.c new file mode 100644 index 0000000000..534a23568d --- /dev/null +++ b/src/gallium/auxiliary/rtasm/rtasm_ppc.c @@ -0,0 +1,365 @@ +/************************************************************************** + * + * Copyright (C) 2008 Tungsten Graphics, Inc. All Rights Reserved. + * + * 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 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 + * BRIAN PAUL 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. + * + **************************************************************************/ + +/** + * PPC code generation. + * \author Brian Paul + */ + + +#include "util/u_memory.h" +#include "pipe/p_debug.h" +#include "rtasm_ppc.h" + + +void +ppc_init_func(struct ppc_function *p, unsigned max_inst) +{ + p->store = align_malloc(max_inst * PPC_INST_SIZE, 16); + p->num_inst = 0; + p->max_inst = max_inst; + p->vec_used = ~0; +} + + +void +ppc_release_func(struct ppc_function *p) +{ + assert(p->num_inst <= p->max_inst); + if (p->store != NULL) { + align_free(p->store); + } + p->store = NULL; +} + + +/** + * Alloate a vector register. + * \return register index or -1 if none left. + */ +int +ppc_allocate_vec_register(struct ppc_function *p, int reg) +{ + unsigned i; + for (i = 0; i < PPC_NUM_VEC_REGS; i++) { + const uint64_t mask = 1 << i; + if ((p->vec_used & mask) != 0) { + p->vec_used &= ~mask; + return i; + } + } + + return -1; +} + + +/** + * Mark the given vector register as "unallocated". + */ +void +ppc_release_vec_register(struct ppc_function *p, int reg) +{ + assert(reg < PPC_NUM_VEC_REGS); + assert((p->vec_used & (1 << reg)) == 0); + + p->vec_used |= (1 << reg); +} + + + +union vx_inst { + uint32_t bits; + struct { + unsigned op:6; + unsigned vD:5; + unsigned vA:5; + unsigned vB:5; + unsigned op2:11; + } inst; +}; + +union vxr_inst { + uint32_t bits; + struct { + unsigned op:6; + unsigned vD:5; + unsigned vA:5; + unsigned vB:5; + unsigned rC:1; + unsigned op2:10; + } inst; +}; + +union va_inst { + uint32_t bits; + struct { + unsigned op:6; + unsigned vD:5; + unsigned vA:5; + unsigned vB:5; + unsigned vC:5; + unsigned op2:6; + } inst; +}; + + +static inline void +emit_vx(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB) +{ + union vx_inst inst; + inst.inst.op = 4; + inst.inst.vD = vD; + inst.inst.vA = vA; + inst.inst.vB = vB; + inst.inst.op2 = op2; + p->store[p->num_inst++] = inst.bits; + assert(p->num_inst <= p->max_inst); +}; + +static inline void +emit_vxr(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB) +{ + union vxr_inst inst; + inst.inst.op = 4; + inst.inst.vD = vD; + inst.inst.vA = vA; + inst.inst.vB = vB; + inst.inst.rC = 0; + inst.inst.op2 = op2; + p->store[p->num_inst++] = inst.bits; + assert(p->num_inst <= p->max_inst); +}; + +static inline void +emit_va(struct ppc_function *p, uint op2, uint vD, uint vA, uint vB, uint vC) +{ + union va_inst inst; + inst.inst.op = 4; + inst.inst.vD = vD; + inst.inst.vA = vA; + inst.inst.vB = vB; + inst.inst.vC = vC; + inst.inst.op2 = op2; + p->store[p->num_inst++] = inst.bits; + assert(p->num_inst <= p->max_inst); +}; + + + +/** + ** float vector arithmetic + **/ + +/** vector float add */ +void +ppc_vaddfp(struct ppc_function *p,uint vD, uint vA, uint vB) +{ + emit_vx(p, 10, vD, vA, vB); +} + +/** vector float substract */ +void +ppc_vsubfp(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vx(p, 74, vD, vA, vB); +} + +/** vector float min */ +void +ppc_vminfp(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vx(p, 1098, vD, vA, vB); +} + +/** vector float max */ +void +ppc_vmaxfp(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vx(p, 1034, vD, vA, vB); +} + +/** vector float mult add */ +void +ppc_vmaddfp(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC) +{ + emit_va(p, 46, vD, vA, vB, vC); +} + +/** vector float compare greater than */ +void +ppc_vcmpgtfpx(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vxr(p, 710, vD, vA, vB); +} + +/** vector float compare greater than or equal to */ +void +ppc_vcmpgefpx(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vxr(p, 454, vD, vA, vB); +} + +/** vector float compare equal */ +void +ppc_vcmpeqfpx(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vxr(p, 198, vD, vA, vB); +} + +/** vector float 2^x */ +void +ppc_vexptefp(struct ppc_function *p, uint vD, uint vB) +{ + emit_vx(p, 394, vD, 0, vB); +} + +/** vector float log2(x) */ +void +ppc_vlogefp(struct ppc_function *p, uint vD, uint vB) +{ + emit_vx(p, 458, vD, 0, vB); +} + +/** vector float reciprocol */ +void +ppc_vrefp(struct ppc_function *p, uint vD, uint vB) +{ + emit_vx(p, 266, vD, 0, vB); +} + +/** vector float reciprocol sqrt estimate */ +void +ppc_vrsqrtefp(struct ppc_function *p, uint vD, uint vB) +{ + emit_vx(p, 330, vD, 0, vB); +} + +/** vector float round to negative infinity */ +void +ppc_vrfim(struct ppc_function *p, uint vD, uint vB) +{ + emit_vx(p, 714, vD, 0, vB); +} + +/** vector float round to positive infinity */ +void +ppc_vrfip(struct ppc_function *p, uint vD, uint vB) +{ + emit_vx(p, 650, vD, 0, vB); +} + +/** vector float round to nearest int */ +void +ppc_vrfin(struct ppc_function *p, uint vD, uint vB) +{ + emit_vx(p, 522, vD, 0, vB); +} + +/** vector float round to int toward zero */ +void +ppc_vrfiz(struct ppc_function *p, uint vD, uint vB) +{ + emit_vx(p, 586, vD, 0, vB); +} + + + +/** + ** bitwise operations + **/ + + +/** vector and */ +void +ppc_vand(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vx(p, 1028, vD, vA, vB); +} + +/** vector and complement */ +void +ppc_vandc(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vx(p, 1092, vD, vA, vB); +} + +/** vector or */ +void +ppc_vor(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vx(p, 1156, vD, vA, vB); +} + +/** vector nor */ +void +ppc_vnor(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vx(p, 1284, vD, vA, vB); +} + +/** vector xor */ +void +ppc_vxor(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vx(p, 1220, vD, vA, vB); +} + + +/** + ** Vector shuffle / select / splat / etc + **/ + +/** vector permute */ +void +ppc_vperm(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC) +{ + emit_va(p, 43, vD, vA, vB, vC); +} + +/** vector select */ +void +ppc_vsel(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC) +{ + emit_va(p, 42, vD, vA, vB, vC); +} + +/** vector splat byte */ +void +ppc_vspltb(struct ppc_function *p, uint vD, uint vB, uint imm) +{ + emit_vx(p, 42, vD, imm, vB); +} + +/** vector splat half word */ +void +ppc_vsplthw(struct ppc_function *p, uint vD, uint vB, uint imm) +{ + emit_vx(p, 588, vD, imm, vB); +} + +/** vector splat word */ +void +ppc_vspltw(struct ppc_function *p, uint vD, uint vB, uint imm) +{ + emit_vx(p, 652, vD, imm, vB); +} diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc.h b/src/gallium/auxiliary/rtasm/rtasm_ppc.h new file mode 100644 index 0000000000..ed14e943df --- /dev/null +++ b/src/gallium/auxiliary/rtasm/rtasm_ppc.h @@ -0,0 +1,181 @@ +/************************************************************************** + * + * Copyright (C) 2008 Tungsten Graphics, Inc. All Rights Reserved. + * + * 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 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 + * BRIAN PAUL 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. + * + **************************************************************************/ + +/** + * PPC code generation. + * \author Brian Paul + */ + + +#ifndef RTASM_PPC_H +#define RTASM_PPC_H + + +#include "pipe/p_compiler.h" + + +#define PPC_INST_SIZE 4 /**< 4 bytes / instruction */ + +#define PPC_NUM_VEC_REGS 32 + + +struct ppc_function +{ + uint32_t *store; /**< instruction buffer */ + uint num_inst; + uint max_inst; + uint32_t vec_used; /** used/free vector registers bitmask */ + uint32_t reg_used; /** used/free general-purpose registers bitmask */ +}; + + + +extern void ppc_init_func(struct ppc_function *p, unsigned max_inst); +extern void ppc_release_func(struct ppc_function *p); + +extern int ppc_allocate_vec_register(struct ppc_function *p, int reg); +extern void ppc_release_vec_register(struct ppc_function *p, int reg); + + +/** + ** float vector arithmetic + **/ + +/** vector float add */ +extern void +ppc_vaddfp(struct ppc_function *p,uint vD, uint vA, uint vB); + +/** vector float substract */ +extern void +ppc_vsubfp(struct ppc_function *p, uint vD, uint vA, uint vB); + +/** vector float min */ +extern void +ppc_vminfp(struct ppc_function *p, uint vD, uint vA, uint vB); + +/** vector float max */ +extern void +ppc_vmaxfp(struct ppc_function *p, uint vD, uint vA, uint vB); + +/** vector float mult add */ +extern void +ppc_vmaddfp(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC); + +/** vector float compare greater than */ +extern void +ppc_vcmpgtfpx(struct ppc_function *p, uint vD, uint vA, uint vB); + +/** vector float compare greater than or equal to */ +extern void +ppc_vcmpgefpx(struct ppc_function *p, uint vD, uint vA, uint vB); + +/** vector float compare equal */ +extern void +ppc_vcmpeqfpx(struct ppc_function *p, uint vD, uint vA, uint vB); + +/** vector float 2^x */ +extern void +ppc_vexptefp(struct ppc_function *p, uint vD, uint vB); + +/** vector float log2(x) */ +extern void +ppc_vlogefp(struct ppc_function *p, uint vD, uint vB); + +/** vector float reciprocol */ +extern void +ppc_vrefp(struct ppc_function *p, uint vD, uint vB); + +/** vector float reciprocol sqrt estimate */ +extern void +ppc_vrsqrtefp(struct ppc_function *p, uint vD, uint vB); + +/** vector float round to negative infinity */ +extern void +ppc_vrfim(struct ppc_function *p, uint vD, uint vB); + +/** vector float round to positive infinity */ +extern void +ppc_vrfip(struct ppc_function *p, uint vD, uint vB); + +/** vector float round to nearest int */ +extern void +ppc_vrfin(struct ppc_function *p, uint vD, uint vB); + +/** vector float round to int toward zero */ +extern void +ppc_vrfiz(struct ppc_function *p, uint vD, uint vB); + + + +/** + ** bitwise operations + **/ + + +/** vector and */ +extern void +ppc_vand(struct ppc_function *p, uint vD, uint vA, uint vB); + +/** vector and complement */ +extern void +ppc_vandc(struct ppc_function *p, uint vD, uint vA, uint vB); + +/** vector or */ +extern void +ppc_vor(struct ppc_function *p, uint vD, uint vA, uint vB); + +/** vector nor */ +extern void +ppc_vnor(struct ppc_function *p, uint vD, uint vA, uint vB); + +/** vector xor */ +extern void +ppc_vxor(struct ppc_function *p, uint vD, uint vA, uint vB); + + +/** + ** Vector shuffle / select / splat / etc + **/ + +/** vector permute */ +extern void +ppc_vperm(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC); + +/** vector select */ +extern void +ppc_vsel(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC); + +/** vector splat byte */ +extern void +ppc_vspltb(struct ppc_function *p, uint vD, uint vB, uint imm); + +/** vector splat half word */ +extern void +ppc_vsplthw(struct ppc_function *p, uint vD, uint vB, uint imm); + +/** vector splat word */ +extern void +ppc_vspltw(struct ppc_function *p, uint vD, uint vB, uint imm); + + +#endif /* RTASM_PPC_H */ diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c index a04cc6c4ff..cc35f0ba5b 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c +++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c @@ -27,12 +27,16 @@ * Real-time assembly generation interface for Cell B.E. SPEs. * * \author Ian Romanick <idr@us.ibm.com> + * \author Brian Paul */ + +#include <stdio.h> #include "pipe/p_compiler.h" #include "util/u_memory.h" #include "rtasm_ppc_spe.h" + #ifdef GALLIUM_CELL /** * SPE instruction types @@ -143,8 +147,46 @@ union spe_inst_RI18 { /*@}*/ +static void +indent(const struct spe_function *p) +{ + int i; + for (i = 0; i < p->indent; i++) { + putchar(' '); + } +} + + +static const char * +rem_prefix(const char *longname) +{ + return longname + 4; +} + + +static const char * +reg_name(int reg) +{ + switch (reg) { + case SPE_REG_SP: + return "$sp"; + case SPE_REG_RA: + return "$lr"; + default: + { + /* cycle through four buffers to handle multiple calls per printf */ + static char buf[4][10]; + static int b = 0; + b = (b + 1) % 4; + sprintf(buf[b], "$%d", reg); + return buf[b]; + } + } +} + + static void emit_RR(struct spe_function *p, unsigned op, unsigned rT, - unsigned rA, unsigned rB) + unsigned rA, unsigned rB, const char *name) { union spe_inst_RR inst; inst.inst.op = op; @@ -153,11 +195,16 @@ static void emit_RR(struct spe_function *p, unsigned op, unsigned rT, inst.inst.rT = rT; p->store[p->num_inst++] = inst.bits; assert(p->num_inst <= p->max_inst); + if (p->print) { + indent(p); + printf("%s\t%s, %s, %s\n", + rem_prefix(name), reg_name(rT), reg_name(rA), reg_name(rB)); + } } static void emit_RRR(struct spe_function *p, unsigned op, unsigned rT, - unsigned rA, unsigned rB, unsigned rC) + unsigned rA, unsigned rB, unsigned rC, const char *name) { union spe_inst_RRR inst; inst.inst.op = op; @@ -167,11 +214,16 @@ static void emit_RRR(struct spe_function *p, unsigned op, unsigned rT, inst.inst.rC = rC; p->store[p->num_inst++] = inst.bits; assert(p->num_inst <= p->max_inst); + if (p->print) { + indent(p); + printf("%s\t%s, %s, %s, %s\n", rem_prefix(name), reg_name(rT), + reg_name(rA), reg_name(rB), reg_name(rC)); + } } static void emit_RI7(struct spe_function *p, unsigned op, unsigned rT, - unsigned rA, int imm) + unsigned rA, int imm, const char *name) { union spe_inst_RI7 inst; inst.inst.op = op; @@ -180,12 +232,17 @@ static void emit_RI7(struct spe_function *p, unsigned op, unsigned rT, inst.inst.rT = rT; p->store[p->num_inst++] = inst.bits; assert(p->num_inst <= p->max_inst); + if (p->print) { + indent(p); + printf("%s\t%s, %s, 0x%x\n", + rem_prefix(name), reg_name(rT), reg_name(rA), imm); + } } static void emit_RI8(struct spe_function *p, unsigned op, unsigned rT, - unsigned rA, int imm) + unsigned rA, int imm, const char *name) { union spe_inst_RI8 inst; inst.inst.op = op; @@ -194,12 +251,17 @@ static void emit_RI8(struct spe_function *p, unsigned op, unsigned rT, inst.inst.rT = rT; p->store[p->num_inst++] = inst.bits; assert(p->num_inst <= p->max_inst); + if (p->print) { + indent(p); + printf("%s\t%s, %s, 0x%x\n", + rem_prefix(name), reg_name(rT), reg_name(rA), imm); + } } static void emit_RI10(struct spe_function *p, unsigned op, unsigned rT, - unsigned rA, int imm) + unsigned rA, int imm, const char *name) { union spe_inst_RI10 inst; inst.inst.op = op; @@ -208,11 +270,16 @@ static void emit_RI10(struct spe_function *p, unsigned op, unsigned rT, inst.inst.rT = rT; p->store[p->num_inst++] = inst.bits; assert(p->num_inst <= p->max_inst); + if (p->print) { + indent(p); + printf("%s\t%s, %s, 0x%x\n", + rem_prefix(name), reg_name(rT), reg_name(rA), imm); + } } static void emit_RI16(struct spe_function *p, unsigned op, unsigned rT, - int imm) + int imm, const char *name) { union spe_inst_RI16 inst; inst.inst.op = op; @@ -220,11 +287,15 @@ static void emit_RI16(struct spe_function *p, unsigned op, unsigned rT, inst.inst.rT = rT; p->store[p->num_inst++] = inst.bits; assert(p->num_inst <= p->max_inst); + if (p->print) { + indent(p); + printf("%s\t%s, 0x%x\n", rem_prefix(name), reg_name(rT), imm); + } } static void emit_RI18(struct spe_function *p, unsigned op, unsigned rT, - int imm) + int imm, const char *name) { union spe_inst_RI18 inst; inst.inst.op = op; @@ -232,6 +303,10 @@ static void emit_RI18(struct spe_function *p, unsigned op, unsigned rT, inst.inst.rT = rT; p->store[p->num_inst++] = inst.bits; assert(p->num_inst <= p->max_inst); + if (p->print) { + indent(p); + printf("%s\t%s, 0x%x\n", rem_prefix(name), reg_name(rT), imm); + } } @@ -240,80 +315,91 @@ static void emit_RI18(struct spe_function *p, unsigned op, unsigned rT, #define EMIT_(_name, _op) \ void _name (struct spe_function *p, unsigned rT) \ { \ - emit_RR(p, _op, rT, 0, 0); \ + emit_RR(p, _op, rT, 0, 0, __FUNCTION__); \ } #define EMIT_R(_name, _op) \ void _name (struct spe_function *p, unsigned rT, unsigned rA) \ { \ - emit_RR(p, _op, rT, rA, 0); \ + emit_RR(p, _op, rT, rA, 0, __FUNCTION__); \ } #define EMIT_RR(_name, _op) \ void _name (struct spe_function *p, unsigned rT, unsigned rA, unsigned rB) \ { \ - emit_RR(p, _op, rT, rA, rB); \ + emit_RR(p, _op, rT, rA, rB, __FUNCTION__); \ } #define EMIT_RRR(_name, _op) \ void _name (struct spe_function *p, unsigned rT, unsigned rA, unsigned rB, unsigned rC) \ { \ - emit_RRR(p, _op, rT, rA, rB, rC); \ + emit_RRR(p, _op, rT, rA, rB, rC, __FUNCTION__); \ } #define EMIT_RI7(_name, _op) \ void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \ { \ - emit_RI7(p, _op, rT, rA, imm); \ + emit_RI7(p, _op, rT, rA, imm, __FUNCTION__); \ } #define EMIT_RI8(_name, _op, bias) \ void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \ { \ - emit_RI8(p, _op, rT, rA, bias - imm); \ + emit_RI8(p, _op, rT, rA, bias - imm, __FUNCTION__); \ } #define EMIT_RI10(_name, _op) \ void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \ { \ - emit_RI10(p, _op, rT, rA, imm); \ + emit_RI10(p, _op, rT, rA, imm, __FUNCTION__); \ } #define EMIT_RI16(_name, _op) \ void _name (struct spe_function *p, unsigned rT, int imm) \ { \ - emit_RI16(p, _op, rT, imm); \ + emit_RI16(p, _op, rT, imm, __FUNCTION__); \ } #define EMIT_RI18(_name, _op) \ void _name (struct spe_function *p, unsigned rT, int imm) \ { \ - emit_RI18(p, _op, rT, imm); \ + emit_RI18(p, _op, rT, imm, __FUNCTION__); \ } #define EMIT_I16(_name, _op) \ void _name (struct spe_function *p, int imm) \ { \ - emit_RI16(p, _op, 0, imm); \ + emit_RI16(p, _op, 0, imm, __FUNCTION__); \ } #include "rtasm_ppc_spe.h" + /** * Initialize an spe_function. * \param code_size size of instruction buffer to allocate, in bytes. */ void spe_init_func(struct spe_function *p, unsigned code_size) { + unsigned int i; + p->store = align_malloc(code_size, 16); p->num_inst = 0; p->max_inst = code_size / SPE_INST_SIZE; + p->set_count = 0; + memset(p->regs, 0, SPE_NUM_REGS * sizeof(p->regs[0])); + /* Conservatively treat R0 - R2 and R80 - R127 as non-volatile. */ - p->regs[0] = ~7; - p->regs[1] = (1U << (80 - 64)) - 1; + p->regs[0] = p->regs[1] = p->regs[2] = 1; + for (i = 80; i <= 127; i++) { + p->regs[i] = 1; + } + + p->print = false; + p->indent = 0; } @@ -327,20 +413,23 @@ void spe_release_func(struct spe_function *p) } +/** Return current code size in bytes. */ +unsigned spe_code_size(const struct spe_function *p) +{ + return p->num_inst * SPE_INST_SIZE; +} + + /** - * Alloate a SPE register. + * Allocate a SPE register. * \return register index or -1 if none left. */ int spe_allocate_available_register(struct spe_function *p) { unsigned i; for (i = 0; i < SPE_NUM_REGS; i++) { - const uint64_t mask = (1ULL << (i % 64)); - const unsigned idx = i / 64; - - assert(idx < 2); - if ((p->regs[idx] & mask) != 0) { - p->regs[idx] &= ~mask; + if (p->regs[i] == 0) { + p->regs[i] = 1; return i; } } @@ -354,31 +443,150 @@ int spe_allocate_available_register(struct spe_function *p) */ int spe_allocate_register(struct spe_function *p, int reg) { - const unsigned idx = reg / 64; - const unsigned bit = reg % 64; - assert(reg < SPE_NUM_REGS); - assert((p->regs[idx] & (1ULL << bit)) != 0); - - p->regs[idx] &= ~(1ULL << bit); + assert(p->regs[reg] == 0); + p->regs[reg] = 1; return reg; } /** - * Mark the given SPE register as "unallocated". + * Mark the given SPE register as "unallocated". Note that this should + * only be used on registers allocated in the current register set; an + * assertion will fail if an attempt is made to deallocate a register + * allocated in an earlier register set. */ void spe_release_register(struct spe_function *p, int reg) { - const unsigned idx = reg / 64; - const unsigned bit = reg % 64; + assert(reg < SPE_NUM_REGS); + assert(p->regs[reg] == 1); - assert(idx < 2); + p->regs[reg] = 0; +} - assert(reg < SPE_NUM_REGS); - assert((p->regs[idx] & (1ULL << bit)) == 0); +/** + * Start a new set of registers. This can be called if + * it will be difficult later to determine exactly what + * registers were actually allocated during a code generation + * sequence, and you really just want to deallocate all of them. + */ +void spe_allocate_register_set(struct spe_function *p) +{ + unsigned int i; + + /* Keep track of the set count. If it ever wraps around to 0, + * we're in trouble. + */ + p->set_count++; + assert(p->set_count > 0); + + /* Increment the allocation count of all registers currently + * allocated. Then any registers that are allocated in this set + * will be the only ones with a count of 1; they'll all be released + * when the register set is released. + */ + for (i = 0; i < SPE_NUM_REGS; i++) { + if (p->regs[i] > 0) + p->regs[i]++; + } +} + +void spe_release_register_set(struct spe_function *p) +{ + unsigned int i; + + /* If the set count drops below zero, we're in trouble. */ + assert(p->set_count > 0); + p->set_count--; + + /* Drop the allocation level of all registers. Any allocated + * during this register set will drop to 0 and then become + * available. + */ + for (i = 0; i < SPE_NUM_REGS; i++) { + if (p->regs[i] > 0) + p->regs[i]--; + } +} + + +unsigned +spe_get_registers_used(const struct spe_function *p, ubyte used[]) +{ + unsigned i, num = 0; + /* only count registers in the range available to callers */ + for (i = 2; i < 80; i++) { + if (p->regs[i]) { + used[num++] = i; + } + } + return num; +} + + +void +spe_print_code(struct spe_function *p, boolean enable) +{ + p->print = enable; +} + + +void +spe_indent(struct spe_function *p, int spaces) +{ + p->indent += spaces; +} + + +void +spe_comment(struct spe_function *p, int rel_indent, const char *s) +{ + if (p->print) { + p->indent += rel_indent; + indent(p); + p->indent -= rel_indent; + printf("# %s\n", s); + } +} + + +/** + * Load quad word. + * NOTE: imm is in bytes and the least significant 4 bits must be zero! + */ +void spe_lqd(struct spe_function *p, unsigned rT, unsigned rA, int offset) +{ + const boolean pSave = p->print; + + p->print = FALSE; + assert(offset % 4 == 0); + emit_RI10(p, 0x034, rT, rA, offset >> 4, "spe_lqd"); + p->print = pSave; + + if (p->print) { + indent(p); + printf("lqd\t%s, %d(%s)\n", reg_name(rT), offset, reg_name(rA)); + } +} + + +/** + * Store quad word. + * NOTE: imm is in bytes and the least significant 4 bits must be zero! + */ +void spe_stqd(struct spe_function *p, unsigned rT, unsigned rA, int offset) +{ + const boolean pSave = p->print; - p->regs[idx] |= (1ULL << bit); + p->print = FALSE; + assert(offset % 4 == 0); + emit_RI10(p, 0x024, rT, rA, offset >> 4, "spe_stqd"); + p->print = pSave; + + if (p->print) { + indent(p); + printf("stqd\t%s, %d(%s)\n", reg_name(rT), offset, reg_name(rA)); + } } @@ -392,51 +600,51 @@ void spe_release_register(struct spe_function *p, int reg) /** Branch Indirect to address in rA */ void spe_bi(struct spe_function *p, unsigned rA, int d, int e) { - emit_RI7(p, 0x1a8, 0, rA, (d << 5) | (e << 4)); + emit_RI7(p, 0x1a8, 0, rA, (d << 5) | (e << 4), __FUNCTION__); } /** Interupt Return */ void spe_iret(struct spe_function *p, unsigned rA, int d, int e) { - emit_RI7(p, 0x1aa, 0, rA, (d << 5) | (e << 4)); + emit_RI7(p, 0x1aa, 0, rA, (d << 5) | (e << 4), __FUNCTION__); } /** Branch indirect and set link on external data */ void spe_bisled(struct spe_function *p, unsigned rT, unsigned rA, int d, int e) { - emit_RI7(p, 0x1ab, rT, rA, (d << 5) | (e << 4)); + emit_RI7(p, 0x1ab, rT, rA, (d << 5) | (e << 4), __FUNCTION__); } /** Branch indirect and set link. Save PC in rT, jump to rA. */ void spe_bisl(struct spe_function *p, unsigned rT, unsigned rA, int d, int e) { - emit_RI7(p, 0x1a9, rT, rA, (d << 5) | (e << 4)); + emit_RI7(p, 0x1a9, rT, rA, (d << 5) | (e << 4), __FUNCTION__); } /** Branch indirect if zero word. If rT.word[0]==0, jump to rA. */ void spe_biz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e) { - emit_RI7(p, 0x128, rT, rA, (d << 5) | (e << 4)); + emit_RI7(p, 0x128, rT, rA, (d << 5) | (e << 4), __FUNCTION__); } /** Branch indirect if non-zero word. If rT.word[0]!=0, jump to rA. */ void spe_binz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e) { - emit_RI7(p, 0x129, rT, rA, (d << 5) | (e << 4)); + emit_RI7(p, 0x129, rT, rA, (d << 5) | (e << 4), __FUNCTION__); } /** Branch indirect if zero halfword. If rT.halfword[1]==0, jump to rA. */ void spe_bihz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e) { - emit_RI7(p, 0x12a, rT, rA, (d << 5) | (e << 4)); + emit_RI7(p, 0x12a, rT, rA, (d << 5) | (e << 4), __FUNCTION__); } /** Branch indirect if non-zero halfword. If rT.halfword[1]!=0, jump to rA. */ void spe_bihnz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e) { - emit_RI7(p, 0x12b, rT, rA, (d << 5) | (e << 4)); + emit_RI7(p, 0x12b, rT, rA, (d << 5) | (e << 4), __FUNCTION__); } @@ -505,30 +713,223 @@ spe_load_int(struct spe_function *p, unsigned rT, int i) } else { spe_ilhu(p, rT, i >> 16); - spe_iohl(p, rT, i & 0xffff); + if (i & 0xffff) + spe_iohl(p, rT, i & 0xffff); + } +} + +void spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui) +{ + /* If the whole value is in the lower 18 bits, use ila, which + * doesn't sign-extend. Otherwise, if the two halfwords of + * the constant are identical, use ilh. Otherwise, if every byte of + * the desired value is 0x00 or 0xff, we can use Form Select Mask for + * Bytes Immediate (fsmbi) to load the value in a single instruction. + * Otherwise, in the general case, we have to use ilhu followed by iohl. + */ + if ((ui & 0x3ffff) == ui) { + spe_ila(p, rT, ui); + } + else if ((ui >> 16) == (ui & 0xffff)) { + spe_ilh(p, rT, ui & 0xffff); + } + else if ( + ((ui & 0x000000ff) == 0 || (ui & 0x000000ff) == 0x000000ff) && + ((ui & 0x0000ff00) == 0 || (ui & 0x0000ff00) == 0x0000ff00) && + ((ui & 0x00ff0000) == 0 || (ui & 0x00ff0000) == 0x00ff0000) && + ((ui & 0xff000000) == 0 || (ui & 0xff000000) == 0xff000000) + ) { + unsigned int mask = 0; + /* fsmbi duplicates each bit in the given mask eight times, + * using a 16-bit value to initialize a 16-byte quadword. + * Each 4-bit nybble of the mask corresponds to a full word + * of the result; look at the value and figure out the mask + * (replicated for each word in the quadword), and then + * form the "select mask" to get the value. + */ + if ((ui & 0x000000ff) == 0x000000ff) mask |= 0x1111; + if ((ui & 0x0000ff00) == 0x0000ff00) mask |= 0x2222; + if ((ui & 0x00ff0000) == 0x00ff0000) mask |= 0x4444; + if ((ui & 0xff000000) == 0xff000000) mask |= 0x8888; + spe_fsmbi(p, rT, mask); + } + else { + /* The general case: this usually uses two instructions, but + * may use only one if the low-order 16 bits of each word are 0. + */ + spe_ilhu(p, rT, ui >> 16); + if (ui & 0xffff) + spe_iohl(p, rT, ui & 0xffff); + } +} + +/** + * This function is constructed identically to spe_sor_uint() below. + * Changes to one should be made in the other. + */ +void +spe_and_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If we can, emit a single instruction, either And Byte Immediate + * (which uses the same constant across each byte), And Halfword Immediate + * (which sign-extends a 10-bit immediate to 16 bits and uses that + * across each halfword), or And Word Immediate (which sign-extends + * a 10-bit immediate to 32 bits). + * + * Otherwise, we'll need to use a temporary register. + */ + unsigned int tmp; + + /* If the upper 23 bits are all 0s or all 1s, sign extension + * will work and we can use And Word Immediate + */ + tmp = ui & 0xfffffe00; + if (tmp == 0xfffffe00 || tmp == 0) { + spe_andi(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric along halfword boundaries and + * the upper 7 bits of each halfword are all 0s or 1s, we + * can use And Halfword Immediate + */ + tmp = ui & 0xfe00fe00; + if ((tmp == 0xfe00fe00 || tmp == 0) && ((ui >> 16) == (ui & 0x0000ffff))) { + spe_andhi(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric in each byte, then we can use + * the And Byte Immediate instruction. + */ + tmp = ui & 0x000000ff; + if ((ui >> 24) == tmp && ((ui >> 16) & 0xff) == tmp && ((ui >> 8) & 0xff) == tmp) { + spe_andbi(p, rT, rA, tmp); + return; + } + + /* Otherwise, we'll have to use a temporary register. */ + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_and(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); +} + + +/** + * This function is constructed identically to spe_and_uint() above. + * Changes to one should be made in the other. + */ +void +spe_xor_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If we can, emit a single instruction, either Exclusive Or Byte + * Immediate (which uses the same constant across each byte), Exclusive + * Or Halfword Immediate (which sign-extends a 10-bit immediate to + * 16 bits and uses that across each halfword), or Exclusive Or Word + * Immediate (which sign-extends a 10-bit immediate to 32 bits). + * + * Otherwise, we'll need to use a temporary register. + */ + unsigned int tmp; + + /* If the upper 23 bits are all 0s or all 1s, sign extension + * will work and we can use Exclusive Or Word Immediate + */ + tmp = ui & 0xfffffe00; + if (tmp == 0xfffffe00 || tmp == 0) { + spe_xori(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric along halfword boundaries and + * the upper 7 bits of each halfword are all 0s or 1s, we + * can use Exclusive Or Halfword Immediate + */ + tmp = ui & 0xfe00fe00; + if ((tmp == 0xfe00fe00 || tmp == 0) && ((ui >> 16) == (ui & 0x0000ffff))) { + spe_xorhi(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric in each byte, then we can use + * the Exclusive Or Byte Immediate instruction. + */ + tmp = ui & 0x000000ff; + if ((ui >> 24) == tmp && ((ui >> 16) & 0xff) == tmp && ((ui >> 8) & 0xff) == tmp) { + spe_xorbi(p, rT, rA, tmp); + return; + } + + /* Otherwise, we'll have to use a temporary register. */ + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_xor(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); +} + +void +spe_compare_equal_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If the comparison value is 9 bits or less, it fits inside a + * Compare Equal Word Immediate instruction. + */ + if ((ui & 0x000001ff) == ui) { + spe_ceqi(p, rT, rA, ui); + } + /* Otherwise, we're going to have to load a word first. */ + else { + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_ceq(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); } } +void +spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If the comparison value is 10 bits or less, it fits inside a + * Compare Logical Greater Than Word Immediate instruction. + */ + if ((ui & 0x000003ff) == ui) { + spe_clgti(p, rT, rA, ui); + } + /* Otherwise, we're going to have to load a word first. */ + else { + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_clgt(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); + } +} void spe_splat(struct spe_function *p, unsigned rT, unsigned rA) { - spe_ila(p, rT, 66051); + /* Duplicate bytes 0, 1, 2, and 3 across the whole register */ + spe_ila(p, rT, 0x00010203); spe_shufb(p, rT, rA, rA, rT); } void -spe_complement(struct spe_function *p, unsigned rT) +spe_complement(struct spe_function *p, unsigned rT, unsigned rA) { - spe_nor(p, rT, rT, rT); + spe_nor(p, rT, rA, rA); } void spe_move(struct spe_function *p, unsigned rT, unsigned rA) { - spe_ori(p, rT, rA, 0); + /* Use different instructions depending on the instruction address + * to take advantage of the dual pipelines. + */ + if (p->num_inst & 1) + spe_shlqbyi(p, rT, rA, 0); /* odd pipe */ + else + spe_ori(p, rT, rA, 0); /* even pipe */ } @@ -539,4 +940,70 @@ spe_zero(struct spe_function *p, unsigned rT) } +void +spe_splat_word(struct spe_function *p, unsigned rT, unsigned rA, int word) +{ + assert(word >= 0); + assert(word <= 3); + + if (word == 0) { + int tmp1 = rT; + spe_ila(p, tmp1, 66051); + spe_shufb(p, rT, rA, rA, tmp1); + } + else { + /* XXX review this, we may not need the rotqbyi instruction */ + int tmp1 = rT; + int tmp2 = spe_allocate_available_register(p); + + spe_ila(p, tmp1, 66051); + spe_rotqbyi(p, tmp2, rA, 4 * word); + spe_shufb(p, rT, tmp2, tmp2, tmp1); + + spe_release_register(p, tmp2); + } +} + +/** + * For each 32-bit float element of rA and rB, choose the smaller of the + * two, compositing them into the rT register. + * + * The Float Compare Greater Than (fcgt) instruction will put 1s into + * compare_reg where rA > rB, and 0s where rA <= rB. + * + * Then the Select Bits (selb) instruction will take bits from rA where + * compare_reg is 0, and from rB where compare_reg is 1; i.e., from rA + * where rA <= rB and from rB where rB > rA, which is exactly the + * "min" operation. + * + * The compare_reg could in many cases be the same as rT, unless + * rT == rA || rt == rB. But since this is common in constructions + * like "x = min(x, a)", we always allocate a new register to be safe. + */ +void +spe_float_min(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB) +{ + unsigned int compare_reg = spe_allocate_available_register(p); + spe_fcgt(p, compare_reg, rA, rB); + spe_selb(p, rT, rA, rB, compare_reg); + spe_release_register(p, compare_reg); +} + +/** + * For each 32-bit float element of rA and rB, choose the greater of the + * two, compositing them into the rT register. + * + * The logic is similar to that of spe_float_min() above; the only + * difference is that the registers on spe_selb() have been reversed, + * so that the larger of the two is selected instead of the smaller. + */ +void +spe_float_max(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB) +{ + unsigned int compare_reg = spe_allocate_available_register(p); + spe_fcgt(p, compare_reg, rA, rB); + spe_selb(p, rT, rB, rA, compare_reg); + spe_release_register(p, compare_reg); +} + #endif /* GALLIUM_CELL */ diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h index d95e5aace3..47dadb343c 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h +++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h @@ -28,6 +28,7 @@ * For details, see /opt/cell/sdk/docs/arch/SPU_ISA_v1.2_27Jan2007_pub.pdf * * \author Ian Romanick <idr@us.ibm.com> + * \author Brian Paul */ #ifndef RTASM_PPC_SPE_H @@ -39,10 +40,10 @@ /** number of general-purpose SIMD registers */ #define SPE_NUM_REGS 128 -/** Return Address register */ +/** Return Address register (aka $lr / Link Register) */ #define SPE_REG_RA 0 -/** Stack Pointer register */ +/** Stack Pointer register (aka $sp) */ #define SPE_REG_SP 1 @@ -52,25 +53,49 @@ struct spe_function uint num_inst; uint max_inst; - /** - * Mask of used / unused registers - * - * Each set bit corresponds to an available register. Each cleared bit - * corresponds to an allocated register. + /** + * The "set count" reflects the number of nested register sets + * are allowed. In the unlikely case that we exceed the set count, + * register allocation will start to be confused, which is critical + * enough that we check for it. + */ + unsigned char set_count; + + /** + * Flags for used and unused registers. Each byte corresponds to a + * register; a 0 in that byte means that the register is available. + * A value of 1 means that the register was allocated in the current + * register set. Any other value N means that the register was allocated + * N register sets ago. * * \sa * spe_allocate_register, spe_allocate_available_register, - * spe_release_register + * spe_allocate_register_set, spe_release_register_set, spe_release_register, */ - uint64_t regs[SPE_NUM_REGS / 64]; + unsigned char regs[SPE_NUM_REGS]; + + boolean print; /**< print/dump instructions as they're emitted? */ + int indent; /**< number of spaces to indent */ }; + extern void spe_init_func(struct spe_function *p, unsigned code_size); extern void spe_release_func(struct spe_function *p); +extern unsigned spe_code_size(const struct spe_function *p); extern int spe_allocate_available_register(struct spe_function *p); extern int spe_allocate_register(struct spe_function *p, int reg); extern void spe_release_register(struct spe_function *p, int reg); +extern void spe_allocate_register_set(struct spe_function *p); +extern void spe_release_register_set(struct spe_function *p); + +extern unsigned +spe_get_registers_used(const struct spe_function *p, ubyte used[]); + +extern void spe_print_code(struct spe_function *p, boolean enable); +extern void spe_indent(struct spe_function *p, int spaces); +extern void spe_comment(struct spe_function *p, int rel_indent, const char *s); + #endif /* RTASM_PPC_SPE_H */ @@ -106,11 +131,9 @@ extern void spe_release_register(struct spe_function *p, int reg); /* Memory load / store instructions */ -EMIT_RI10(spe_lqd, 0x034); EMIT_RR (spe_lqx, 0x1c4); EMIT_RI16(spe_lqa, 0x061); EMIT_RI16(spe_lqr, 0x067); -EMIT_RI10(spe_stqd, 0x024); EMIT_RR (spe_stqx, 0x144); EMIT_RI16(spe_stqa, 0x041); EMIT_RI16(spe_stqr, 0x047); @@ -268,6 +291,12 @@ EMIT_RI16(spe_brz, 0x040); EMIT_RI16(spe_brhnz, 0x046); EMIT_RI16(spe_brhz, 0x044); +extern void +spe_lqd(struct spe_function *p, unsigned rT, unsigned rA, int offset); + +extern void +spe_stqd(struct spe_function *p, unsigned rT, unsigned rA, int offset); + extern void spe_bi(struct spe_function *p, unsigned rA, int d, int e); extern void spe_iret(struct spe_function *p, unsigned rA, int d, int e); extern void spe_bisled(struct spe_function *p, unsigned rT, unsigned rA, @@ -292,13 +321,33 @@ spe_load_float(struct spe_function *p, unsigned rT, float x); extern void spe_load_int(struct spe_function *p, unsigned rT, int i); +/** Load/splat immediate unsigned int into rT. */ +extern void +spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui); + +/** And immediate value into rT. */ +extern void +spe_and_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + +/** Xor immediate value into rT. */ +extern void +spe_xor_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + +/** Compare equal with immediate value. */ +extern void +spe_compare_equal_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + +/** Compare greater with immediate value. */ +extern void +spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + /** Replicate word 0 of rA across rT. */ extern void spe_splat(struct spe_function *p, unsigned rT, unsigned rA); -/** Complement/invert all bits in rT. */ +/** rT = complement_all_bits(rA). */ extern void -spe_complement(struct spe_function *p, unsigned rT); +spe_complement(struct spe_function *p, unsigned rT, unsigned rA); /** rT = rA. */ extern void @@ -308,6 +357,18 @@ spe_move(struct spe_function *p, unsigned rT, unsigned rA); extern void spe_zero(struct spe_function *p, unsigned rT); +/** rT = splat(rA, word) */ +extern void +spe_splat_word(struct spe_function *p, unsigned rT, unsigned rA, int word); + +/** rT = float min(rA, rB) */ +extern void +spe_float_min(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB); + +/** rT = float max(rA, rB) */ +extern void +spe_float_max(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB); + /* Floating-point instructions */ diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c index ad9d8f8ced..99ee74cf14 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c +++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c @@ -240,7 +240,8 @@ static void emit_modrm( struct x86_function *p, /* Oh-oh we've stumbled into the SIB thing. */ if (regmem.file == file_REG32 && - regmem.idx == reg_SP) { + regmem.idx == reg_SP && + regmem.mod != mod_REG) { emit_1ub(p, 0x24); /* simplistic! */ } @@ -439,25 +440,70 @@ void x86_call( struct x86_function *p, struct x86_reg reg) } -/* michal: - * Temporary. As I need immediate operands, and dont want to mess with the codegen, - * I load the immediate into general purpose register and use it. - */ void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, int imm ) { DUMP_RI( dst, imm ); + assert(dst.file == file_REG32); assert(dst.mod == mod_REG); emit_1ub(p, 0xb8 + dst.idx); emit_1i(p, imm); } -void x86_add_reg_imm8( struct x86_function *p, struct x86_reg dst, ubyte imm ) +/** + * Immediate group 1 instructions. + */ +static INLINE void +x86_group1_imm( struct x86_function *p, + unsigned op, struct x86_reg dst, int imm ) { - DUMP_RI( dst, imm ); + assert(dst.file == file_REG32); assert(dst.mod == mod_REG); - emit_1ub(p, 0x80); - emit_modrm_noreg(p, 0, dst); - emit_1ub(p, imm); + if(-0x80 <= imm && imm < 0x80) { + emit_1ub(p, 0x83); + emit_modrm_noreg(p, op, dst); + emit_1b(p, (char)imm); + } + else { + emit_1ub(p, 0x81); + emit_modrm_noreg(p, op, dst); + emit_1i(p, imm); + } +} + +void x86_add_imm( struct x86_function *p, struct x86_reg dst, int imm ) +{ + DUMP_RI( dst, imm ); + x86_group1_imm(p, 0, dst, imm); +} + +void x86_or_imm( struct x86_function *p, struct x86_reg dst, int imm ) +{ + DUMP_RI( dst, imm ); + x86_group1_imm(p, 1, dst, imm); +} + +void x86_and_imm( struct x86_function *p, struct x86_reg dst, int imm ) +{ + DUMP_RI( dst, imm ); + x86_group1_imm(p, 4, dst, imm); +} + +void x86_sub_imm( struct x86_function *p, struct x86_reg dst, int imm ) +{ + DUMP_RI( dst, imm ); + x86_group1_imm(p, 5, dst, imm); +} + +void x86_xor_imm( struct x86_function *p, struct x86_reg dst, int imm ) +{ + DUMP_RI( dst, imm ); + x86_group1_imm(p, 6, dst, imm); +} + +void x86_cmp_imm( struct x86_function *p, struct x86_reg dst, int imm ) +{ + DUMP_RI( dst, imm ); + x86_group1_imm(p, 7, dst, imm); } diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.h b/src/gallium/auxiliary/rtasm/rtasm_x86sse.h index af79f07dd3..1b5eaaca85 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.h +++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.h @@ -152,12 +152,13 @@ void x86_jmp( struct x86_function *p, int label ); /* void x86_call( struct x86_function *p, void (*label)() ); */ void x86_call( struct x86_function *p, struct x86_reg reg); -/* michal: - * Temporary. As I need immediate operands, and dont want to mess with the codegen, - * I load the immediate into general purpose register and use it. - */ void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, int imm ); -void x86_add_reg_imm8( struct x86_function *p, struct x86_reg dst, ubyte imm ); +void x86_add_imm( struct x86_function *p, struct x86_reg dst, int imm ); +void x86_or_imm( struct x86_function *p, struct x86_reg dst, int imm ); +void x86_and_imm( struct x86_function *p, struct x86_reg dst, int imm ); +void x86_sub_imm( struct x86_function *p, struct x86_reg dst, int imm ); +void x86_xor_imm( struct x86_function *p, struct x86_reg dst, int imm ); +void x86_cmp_imm( struct x86_function *p, struct x86_reg dst, int imm ); /* Macro for sse_shufps() and sse2_pshufd(): diff --git a/src/gallium/auxiliary/tgsi/tgsi_build.c b/src/gallium/auxiliary/tgsi/tgsi_build.c index 74614d3688..38fcaf8829 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_build.c +++ b/src/gallium/auxiliary/tgsi/tgsi_build.c @@ -793,10 +793,14 @@ tgsi_default_instruction_ext_nv( void ) return instruction_ext_nv; } -union token_u32 + +/** test for inequality of 32-bit values pointed to by a and b */ +static INLINE boolean +compare32(const void *a, const void *b) { - unsigned u32; -}; + return *((uint32_t *) a) != *((uint32_t *) b); +} + unsigned tgsi_compare_instruction_ext_nv( @@ -805,7 +809,7 @@ tgsi_compare_instruction_ext_nv( { a.Padding = b.Padding = 0; a.Extended = b.Extended = 0; - return ((union token_u32 *) &a)->u32 != ((union token_u32 *) &b)->u32; + return compare32(&a, &b); } struct tgsi_instruction_ext_nv @@ -864,7 +868,7 @@ tgsi_compare_instruction_ext_label( { a.Padding = b.Padding = 0; a.Extended = b.Extended = 0; - return ((union token_u32 *) &a)->u32 != ((union token_u32 *) &b)->u32; + return compare32(&a, &b); } struct tgsi_instruction_ext_label @@ -905,7 +909,7 @@ tgsi_compare_instruction_ext_texture( { a.Padding = b.Padding = 0; a.Extended = b.Extended = 0; - return ((union token_u32 *) &a)->u32 != ((union token_u32 *) &b)->u32; + return compare32(&a, &b); } struct tgsi_instruction_ext_texture @@ -1027,7 +1031,7 @@ tgsi_compare_src_register_ext_swz( { a.Padding = b.Padding = 0; a.Extended = b.Extended = 0; - return ((union token_u32 *) &a)->u32 != ((union token_u32 *) &b)->u32; + return compare32(&a, &b); } struct tgsi_src_register_ext_swz @@ -1095,7 +1099,7 @@ tgsi_compare_src_register_ext_mod( { a.Padding = b.Padding = 0; a.Extended = b.Extended = 0; - return ((union token_u32 *) &a)->u32 != ((union token_u32 *) &b)->u32; + return compare32(&a, &b); } struct tgsi_src_register_ext_mod @@ -1241,7 +1245,7 @@ tgsi_compare_dst_register_ext_concode( { a.Padding = b.Padding = 0; a.Extended = b.Extended = 0; - return ((union token_u32 *) &a)->u32 != ((union token_u32 *) &b)->u32; + return compare32(&a, &b); } struct tgsi_dst_register_ext_concode @@ -1299,7 +1303,7 @@ tgsi_compare_dst_register_ext_modulate( { a.Padding = b.Padding = 0; a.Extended = b.Extended = 0; - return ((union token_u32 *) &a)->u32 != ((union token_u32 *) &b)->u32; + return compare32(&a, &b); } struct tgsi_dst_register_ext_modulate diff --git a/src/gallium/auxiliary/tgsi/tgsi_exec.c b/src/gallium/auxiliary/tgsi/tgsi_exec.c index df002939c6..1a5294eabc 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_exec.c +++ b/src/gallium/auxiliary/tgsi/tgsi_exec.c @@ -1674,6 +1674,7 @@ exec_declaration( break; default: + eval = NULL; assert( 0 ); } diff --git a/src/gallium/auxiliary/tgsi/tgsi_parse.c b/src/gallium/auxiliary/tgsi/tgsi_parse.c index 3757486ba9..2cd56e413a 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_parse.c +++ b/src/gallium/auxiliary/tgsi/tgsi_parse.c @@ -88,16 +88,33 @@ tgsi_parse_end_of_tokens( 1 + ctx->FullHeader.Header.HeaderSize + ctx->FullHeader.Header.BodySize; } + +/** + * This function is used to avoid and work-around type punning/aliasing + * warnings. The warnings seem harmless on x86 but on PPC they cause + * real failures. + */ +static INLINE void +copy_token(void *dst, const void *src) +{ + memcpy(dst, src, 4); +} + + +/** + * Get next 4-byte token, return it at address specified by 'token' + */ static void next_token( struct tgsi_parse_context *ctx, void *token ) { assert( !tgsi_parse_end_of_tokens( ctx ) ); - - *(struct tgsi_token *) token = ctx->Tokens[ctx->Position++]; + copy_token(token, &ctx->Tokens[ctx->Position]); + ctx->Position++; } + void tgsi_parse_token( struct tgsi_parse_context *ctx ) @@ -116,7 +133,7 @@ tgsi_parse_token( struct tgsi_full_declaration *decl = &ctx->FullToken.FullDeclaration; *decl = tgsi_default_full_declaration(); - decl->Declaration = *(struct tgsi_declaration *) &token; + copy_token(&decl->Declaration, &token); next_token( ctx, &decl->DeclarationRange ); @@ -132,8 +149,7 @@ tgsi_parse_token( struct tgsi_full_immediate *imm = &ctx->FullToken.FullImmediate; *imm = tgsi_default_full_immediate(); - imm->Immediate = *(struct tgsi_immediate *) &token; - + copy_token(&imm->Immediate, &token); assert( !imm->Immediate.Extended ); switch (imm->Immediate.DataType) { @@ -158,8 +174,7 @@ tgsi_parse_token( unsigned extended; *inst = tgsi_default_full_instruction(); - inst->Instruction = *(struct tgsi_instruction *) &token; - + copy_token(&inst->Instruction, &token); extended = inst->Instruction.Extended; while( extended ) { @@ -169,18 +184,15 @@ tgsi_parse_token( switch( token.Type ) { case TGSI_INSTRUCTION_EXT_TYPE_NV: - inst->InstructionExtNv = - *(struct tgsi_instruction_ext_nv *) &token; + copy_token(&inst->InstructionExtNv, &token); break; case TGSI_INSTRUCTION_EXT_TYPE_LABEL: - inst->InstructionExtLabel = - *(struct tgsi_instruction_ext_label *) &token; + copy_token(&inst->InstructionExtLabel, &token); break; case TGSI_INSTRUCTION_EXT_TYPE_TEXTURE: - inst->InstructionExtTexture = - *(struct tgsi_instruction_ext_texture *) &token; + copy_token(&inst->InstructionExtTexture, &token); break; default: @@ -212,13 +224,13 @@ tgsi_parse_token( switch( token.Type ) { case TGSI_DST_REGISTER_EXT_TYPE_CONDCODE: - inst->FullDstRegisters[i].DstRegisterExtConcode = - *(struct tgsi_dst_register_ext_concode *) &token; + copy_token(&inst->FullDstRegisters[i].DstRegisterExtConcode, + &token); break; case TGSI_DST_REGISTER_EXT_TYPE_MODULATE: - inst->FullDstRegisters[i].DstRegisterExtModulate = - *(struct tgsi_dst_register_ext_modulate *) &token; + copy_token(&inst->FullDstRegisters[i].DstRegisterExtModulate, + &token); break; default: @@ -245,13 +257,13 @@ tgsi_parse_token( switch( token.Type ) { case TGSI_SRC_REGISTER_EXT_TYPE_SWZ: - inst->FullSrcRegisters[i].SrcRegisterExtSwz = - *(struct tgsi_src_register_ext_swz *) &token; + copy_token(&inst->FullSrcRegisters[i].SrcRegisterExtSwz, + &token); break; case TGSI_SRC_REGISTER_EXT_TYPE_MOD: - inst->FullSrcRegisters[i].SrcRegisterExtMod = - *(struct tgsi_src_register_ext_mod *) &token; + copy_token(&inst->FullSrcRegisters[i].SrcRegisterExtMod, + &token); break; default: diff --git a/src/gallium/auxiliary/tgsi/tgsi_sse2.c b/src/gallium/auxiliary/tgsi/tgsi_sse2.c index 4681b29f52..f79170b9d6 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_sse2.c +++ b/src/gallium/auxiliary/tgsi/tgsi_sse2.c @@ -25,9 +25,14 @@ * **************************************************************************/ +#include "pipe/p_config.h" + +#if defined(PIPE_ARCH_X86) && defined(PIPE_ARCH_SSE) + #include "pipe/p_debug.h" #include "pipe/p_shader_tokens.h" #include "util/u_math.h" +#include "util/u_sse.h" #include "tgsi/tgsi_parse.h" #include "tgsi/tgsi_util.h" #include "tgsi_exec.h" @@ -35,8 +40,6 @@ #include "rtasm/rtasm_x86sse.h" -#ifdef PIPE_ARCH_X86 - /* for 1/sqrt() * * This costs about 100fps (close to 10%) in gears: @@ -480,10 +483,31 @@ emit_coef_dady( * Function call helpers. */ +/** + * NOTE: In gcc, if the destination uses the SSE intrinsics, then it must be + * defined with __attribute__((force_align_arg_pointer)), as we do not guarantee + * that the stack pointer is 16 byte aligned, as expected. + */ static void -emit_push_gp( - struct x86_function *func ) +emit_func_call_dst( + struct x86_function *func, + unsigned xmm_save, + unsigned xmm_dst, + void (PIPE_CDECL *code)() ) { + struct x86_reg ecx = x86_make_reg( file_REG32, reg_CX ); + unsigned i, n, xmm; + unsigned xmm_mask; + + /* Bitmask of the xmm registers to save */ + xmm_mask = (1 << xmm_save) - 1; + xmm_mask &= ~(1 << xmm_dst); + + sse_movaps( + func, + get_temp( TEMP_R0, 0 ), + make_xmm( xmm_dst ) ); + x86_push( func, x86_make_reg( file_REG32, reg_AX) ); @@ -493,12 +517,49 @@ emit_push_gp( x86_push( func, x86_make_reg( file_REG32, reg_DX) ); -} + + for(i = 0, n = 0; i < 8; ++i) + if(xmm_mask & (1 << i)) + ++n; + + x86_sub_imm( + func, + x86_make_reg( file_REG32, reg_SP ), + n*16); + + for(i = 0, n = 0; i < 8; ++i) + if(xmm_mask & (1 << i)) { + sse_movups( + func, + x86_make_disp( x86_make_reg( file_REG32, reg_SP ), n*16 ), + make_xmm( xmm ) ); + ++n; + } + + x86_lea( + func, + ecx, + get_temp( TEMP_R0, 0 ) ); + + x86_push( func, ecx ); + x86_mov_reg_imm( func, ecx, (unsigned long) code ); + x86_call( func, ecx ); + x86_pop(func, ecx ); + + for(i = 0, n = 0; i < 8; ++i) + if(xmm_mask & (1 << i)) { + sse_movups( + func, + make_xmm( xmm ), + x86_make_disp( x86_make_reg( file_REG32, reg_SP ), n*16 ) ); + ++n; + } + + x86_add_imm( + func, + x86_make_reg( file_REG32, reg_SP ), + n*16); -static void -x86_pop_gp( - struct x86_function *func ) -{ /* Restore GP registers in a reverse order. */ x86_pop( @@ -510,39 +571,6 @@ x86_pop_gp( x86_pop( func, x86_make_reg( file_REG32, reg_AX) ); -} - -static void -emit_func_call_dst( - struct x86_function *func, - unsigned xmm_dst, - void (PIPE_CDECL *code)() ) -{ - sse_movaps( - func, - get_temp( TEMP_R0, 0 ), - make_xmm( xmm_dst ) ); - - emit_push_gp( - func ); - - { - struct x86_reg ecx = x86_make_reg( file_REG32, reg_CX ); - - x86_lea( - func, - ecx, - get_temp( TEMP_R0, 0 ) ); - - x86_push( func, ecx ); - x86_mov_reg_imm( func, ecx, (unsigned long) code ); - x86_call( func, ecx ); - x86_pop(func, ecx ); - } - - - x86_pop_gp( - func ); sse_movaps( func, @@ -553,6 +581,7 @@ emit_func_call_dst( static void emit_func_call_dst_src( struct x86_function *func, + unsigned xmm_save, unsigned xmm_dst, unsigned xmm_src, void (PIPE_CDECL *code)() ) @@ -564,10 +593,111 @@ emit_func_call_dst_src( emit_func_call_dst( func, + xmm_save, xmm_dst, code ); } +/* + * Fast SSE2 implementation of special math functions. + */ + +#define POLY0(x, c0) _mm_set1_ps(c0) +#define POLY1(x, c0, c1) _mm_add_ps(_mm_mul_ps(POLY0(x, c1), x), _mm_set1_ps(c0)) +#define POLY2(x, c0, c1, c2) _mm_add_ps(_mm_mul_ps(POLY1(x, c1, c2), x), _mm_set1_ps(c0)) +#define POLY3(x, c0, c1, c2, c3) _mm_add_ps(_mm_mul_ps(POLY2(x, c1, c2, c3), x), _mm_set1_ps(c0)) +#define POLY4(x, c0, c1, c2, c3, c4) _mm_add_ps(_mm_mul_ps(POLY3(x, c1, c2, c3, c4), x), _mm_set1_ps(c0)) +#define POLY5(x, c0, c1, c2, c3, c4, c5) _mm_add_ps(_mm_mul_ps(POLY4(x, c1, c2, c3, c4, c5), x), _mm_set1_ps(c0)) + +#define EXP_POLY_DEGREE 3 +#define LOG_POLY_DEGREE 5 + +/** + * See http://www.devmaster.net/forums/showthread.php?p=43580 + */ +static INLINE __m128 +exp2f4(__m128 x) +{ + __m128i ipart; + __m128 fpart, expipart, expfpart; + + x = _mm_min_ps(x, _mm_set1_ps( 129.00000f)); + x = _mm_max_ps(x, _mm_set1_ps(-126.99999f)); + + /* ipart = int(x - 0.5) */ + ipart = _mm_cvtps_epi32(_mm_sub_ps(x, _mm_set1_ps(0.5f))); + + /* fpart = x - ipart */ + fpart = _mm_sub_ps(x, _mm_cvtepi32_ps(ipart)); + + /* expipart = (float) (1 << ipart) */ + expipart = _mm_castsi128_ps(_mm_slli_epi32(_mm_add_epi32(ipart, _mm_set1_epi32(127)), 23)); + + /* minimax polynomial fit of 2**x, in range [-0.5, 0.5[ */ +#if EXP_POLY_DEGREE == 5 + expfpart = POLY5(fpart, 9.9999994e-1f, 6.9315308e-1f, 2.4015361e-1f, 5.5826318e-2f, 8.9893397e-3f, 1.8775767e-3f); +#elif EXP_POLY_DEGREE == 4 + expfpart = POLY4(fpart, 1.0000026f, 6.9300383e-1f, 2.4144275e-1f, 5.2011464e-2f, 1.3534167e-2f); +#elif EXP_POLY_DEGREE == 3 + expfpart = POLY3(fpart, 9.9992520e-1f, 6.9583356e-1f, 2.2606716e-1f, 7.8024521e-2f); +#elif EXP_POLY_DEGREE == 2 + expfpart = POLY2(fpart, 1.0017247f, 6.5763628e-1f, 3.3718944e-1f); +#else +#error +#endif + + return _mm_mul_ps(expipart, expfpart); +} + +/** + * See http://www.devmaster.net/forums/showthread.php?p=43580 + */ +static INLINE __m128 +log2f4(__m128 x) +{ + __m128i expmask = _mm_set1_epi32(0x7f800000); + __m128i mantmask = _mm_set1_epi32(0x007fffff); + __m128 one = _mm_set1_ps(1.0f); + + __m128i i = _mm_castps_si128(x); + + /* exp = (float) exponent(x) */ + __m128 exp = _mm_cvtepi32_ps(_mm_sub_epi32(_mm_srli_epi32(_mm_and_si128(i, expmask), 23), _mm_set1_epi32(127))); + + /* mant = (float) mantissa(x) */ + __m128 mant = _mm_or_ps(_mm_castsi128_ps(_mm_and_si128(i, mantmask)), one); + + __m128 logmant; + + /* Minimax polynomial fit of log2(x)/(x - 1), for x in range [1, 2[ + * These coefficients can be generate with + * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html + */ +#if LOG_POLY_DEGREE == 6 + logmant = POLY5(mant, 3.11578814719469302614f, -3.32419399085241980044f, 2.59883907202499966007f, -1.23152682416275988241f, 0.318212422185251071475f, -0.0344359067839062357313f); +#elif LOG_POLY_DEGREE == 5 + logmant = POLY4(mant, 2.8882704548164776201f, -2.52074962577807006663f, 1.48116647521213171641f, -0.465725644288844778798f, 0.0596515482674574969533f); +#elif LOG_POLY_DEGREE == 4 + logmant = POLY3(mant, 2.61761038894603480148f, -1.75647175389045657003f, 0.688243882994381274313f, -0.107254423828329604454f); +#elif LOG_POLY_DEGREE == 3 + logmant = POLY2(mant, 2.28330284476918490682f, -1.04913055217340124191f, 0.204446009836232697516f); +#else +#error +#endif + + /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/ + logmant = _mm_mul_ps(logmant, _mm_sub_ps(mant, one)); + + return _mm_add_ps(logmant, exp); +} + +static INLINE __m128 +powf4(__m128 x, __m128 y) +{ + return exp2f4(_mm_mul_ps(log2f4(x), y)); +} + + /** * Low-level instruction translators. */ @@ -610,38 +740,35 @@ cos4f( static void emit_cos( struct x86_function *func, + unsigned xmm_save, unsigned xmm_dst ) { emit_func_call_dst( func, + xmm_save, xmm_dst, cos4f ); } static void PIPE_CDECL +#if defined(PIPE_CC_GCC) +__attribute__((force_align_arg_pointer)) +#endif ex24f( float *store ) { -#if FAST_MATH - store[0] = util_fast_exp2( store[0] ); - store[1] = util_fast_exp2( store[1] ); - store[2] = util_fast_exp2( store[2] ); - store[3] = util_fast_exp2( store[3] ); -#else - store[0] = powf( 2.0f, store[0] ); - store[1] = powf( 2.0f, store[1] ); - store[2] = powf( 2.0f, store[2] ); - store[3] = powf( 2.0f, store[3] ); -#endif + _mm_store_ps(&store[0], exp2f4( _mm_load_ps(&store[0]) )); } static void emit_ex2( struct x86_function *func, + unsigned xmm_save, unsigned xmm_dst ) { emit_func_call_dst( func, + xmm_save, xmm_dst, ex24f ); } @@ -670,10 +797,12 @@ flr4f( static void emit_flr( struct x86_function *func, + unsigned xmm_save, unsigned xmm_dst ) { emit_func_call_dst( func, + xmm_save, xmm_dst, flr4f ); } @@ -691,31 +820,35 @@ frc4f( static void emit_frc( struct x86_function *func, + unsigned xmm_save, unsigned xmm_dst ) { emit_func_call_dst( func, + xmm_save, xmm_dst, frc4f ); } static void PIPE_CDECL +#if defined(PIPE_CC_GCC) +__attribute__((force_align_arg_pointer)) +#endif lg24f( float *store ) { - store[0] = util_fast_log2( store[0] ); - store[1] = util_fast_log2( store[1] ); - store[2] = util_fast_log2( store[2] ); - store[3] = util_fast_log2( store[3] ); + _mm_store_ps(&store[0], log2f4( _mm_load_ps(&store[0]) )); } static void emit_lg2( struct x86_function *func, + unsigned xmm_save, unsigned xmm_dst ) { emit_func_call_dst( func, + xmm_save, xmm_dst, lg24f ); } @@ -757,14 +890,14 @@ emit_neg( } static void PIPE_CDECL +#if defined(PIPE_CC_GCC) +__attribute__((force_align_arg_pointer)) +#endif pow4f( float *store ) { -#if FAST_MATH - store[0] = util_fast_pow( store[0], store[4] ); - store[1] = util_fast_pow( store[1], store[5] ); - store[2] = util_fast_pow( store[2], store[6] ); - store[3] = util_fast_pow( store[3], store[7] ); +#if 1 + _mm_store_ps(&store[0], powf4( _mm_load_ps(&store[0]), _mm_load_ps(&store[4]) )); #else store[0] = powf( store[0], store[4] ); store[1] = powf( store[1], store[5] ); @@ -776,11 +909,13 @@ pow4f( static void emit_pow( struct x86_function *func, + unsigned xmm_save, unsigned xmm_dst, unsigned xmm_src ) { emit_func_call_dst_src( func, + xmm_save, xmm_dst, xmm_src, pow4f ); @@ -873,10 +1008,12 @@ sin4f( static void emit_sin (struct x86_function *func, + unsigned xmm_save, unsigned xmm_dst) { emit_func_call_dst( func, + xmm_save, xmm_dst, sin4f ); } @@ -1296,7 +1433,7 @@ emit_instruction( get_temp( TGSI_EXEC_TEMP_MINUS_128_I, TGSI_EXEC_TEMP_MINUS_128_C ) ); - emit_pow( func, 1, 2 ); + emit_pow( func, 3, 1, 2 ); FETCH( func, *inst, 0, 0, CHAN_X ); sse_xorps( func, @@ -1342,11 +1479,11 @@ emit_instruction( if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) || IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) { emit_MOV( func, 1, 0 ); - emit_flr( func, 1 ); + emit_flr( func, 2, 1 ); /* dst.x = ex2(floor(src.x)) */ if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X )) { emit_MOV( func, 2, 1 ); - emit_ex2( func, 2 ); + emit_ex2( func, 3, 2 ); STORE( func, *inst, 2, 0, CHAN_X ); } /* dst.y = src.x - floor(src.x) */ @@ -1358,7 +1495,7 @@ emit_instruction( } /* dst.z = ex2(src.x) */ if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) { - emit_ex2( func, 0 ); + emit_ex2( func, 3, 0 ); STORE( func, *inst, 0, 0, CHAN_Z ); } } @@ -1376,21 +1513,21 @@ emit_instruction( FETCH( func, *inst, 0, 0, CHAN_X ); emit_abs( func, 0 ); emit_MOV( func, 1, 0 ); - emit_lg2( func, 1 ); + emit_lg2( func, 2, 1 ); /* dst.z = lg2(abs(src.x)) */ if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) { STORE( func, *inst, 1, 0, CHAN_Z ); } if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) || IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) { - emit_flr( func, 1 ); + emit_flr( func, 2, 1 ); /* dst.x = floor(lg2(abs(src.x))) */ if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X )) { STORE( func, *inst, 1, 0, CHAN_X ); } /* dst.x = abs(src)/ex2(floor(lg2(abs(src.x)))) */ if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) { - emit_ex2( func, 1 ); + emit_ex2( func, 2, 1 ); emit_rcp( func, 1, 1 ); emit_mul( func, 0, 1 ); STORE( func, *inst, 0, 0, CHAN_Y ); @@ -1580,7 +1717,7 @@ emit_instruction( /* TGSI_OPCODE_FRC */ FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { FETCH( func, *inst, 0, 0, chan_index ); - emit_frc( func, 0 ); + emit_frc( func, 0, 0 ); STORE( func, *inst, 0, 0, chan_index ); } break; @@ -1593,7 +1730,7 @@ emit_instruction( /* TGSI_OPCODE_FLR */ FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { FETCH( func, *inst, 0, 0, chan_index ); - emit_flr( func, 0 ); + emit_flr( func, 0, 0 ); STORE( func, *inst, 0, 0, chan_index ); } break; @@ -1605,7 +1742,7 @@ emit_instruction( case TGSI_OPCODE_EXPBASE2: /* TGSI_OPCODE_EX2 */ FETCH( func, *inst, 0, 0, CHAN_X ); - emit_ex2( func, 0 ); + emit_ex2( func, 0, 0 ); FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { STORE( func, *inst, 0, 0, chan_index ); } @@ -1614,7 +1751,7 @@ emit_instruction( case TGSI_OPCODE_LOGBASE2: /* TGSI_OPCODE_LG2 */ FETCH( func, *inst, 0, 0, CHAN_X ); - emit_lg2( func, 0 ); + emit_lg2( func, 0, 0 ); FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { STORE( func, *inst, 0, 0, chan_index ); } @@ -1624,7 +1761,7 @@ emit_instruction( /* TGSI_OPCODE_POW */ FETCH( func, *inst, 0, 0, CHAN_X ); FETCH( func, *inst, 1, 1, CHAN_X ); - emit_pow( func, 0, 1 ); + emit_pow( func, 0, 0, 1 ); FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { STORE( func, *inst, 0, 0, chan_index ); } @@ -1715,7 +1852,7 @@ emit_instruction( case TGSI_OPCODE_COS: FETCH( func, *inst, 0, 0, CHAN_X ); - emit_cos( func, 0 ); + emit_cos( func, 0, 0 ); FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { STORE( func, *inst, 0, 0, chan_index ); } @@ -1774,7 +1911,7 @@ emit_instruction( case TGSI_OPCODE_SIN: FETCH( func, *inst, 0, 0, CHAN_X ); - emit_sin( func, 0 ); + emit_sin( func, 0, 0 ); FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { STORE( func, *inst, 0, 0, chan_index ); } @@ -1868,12 +2005,12 @@ emit_instruction( case TGSI_OPCODE_SCS: IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) { FETCH( func, *inst, 0, 0, CHAN_X ); - emit_cos( func, 0 ); + emit_cos( func, 0, 0 ); STORE( func, *inst, 0, 0, CHAN_X ); } IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) { FETCH( func, *inst, 0, 0, CHAN_X ); - emit_sin( func, 0 ); + emit_sin( func, 0, 0 ); STORE( func, *inst, 0, 0, CHAN_Y ); } IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) { diff --git a/src/gallium/auxiliary/util/Makefile b/src/gallium/auxiliary/util/Makefile index d3951e4e7d..b3d1045a8f 100644 --- a/src/gallium/auxiliary/util/Makefile +++ b/src/gallium/auxiliary/util/Makefile @@ -10,6 +10,7 @@ C_SOURCES = \ u_gen_mipmap.c \ u_handle_table.c \ u_hash_table.c \ + u_keymap.c \ u_math.c \ u_mm.c \ u_rect.c \ diff --git a/src/gallium/auxiliary/util/SConscript b/src/gallium/auxiliary/util/SConscript index e65c17b1cc..8a04955a16 100644 --- a/src/gallium/auxiliary/util/SConscript +++ b/src/gallium/auxiliary/util/SConscript @@ -11,13 +11,14 @@ util = env.ConvenienceLibrary( 'u_gen_mipmap.c', 'u_handle_table.c', 'u_hash_table.c', + 'u_keymap.c', 'u_math.c', 'u_mm.c', 'u_rect.c', 'u_simple_shaders.c', 'u_snprintf.c', - 'u_stream_stdc.c', - 'u_stream_wd.c', + 'u_stream_stdc.c', + 'u_stream_wd.c', 'u_tile.c', 'u_time.c', ]) diff --git a/src/gallium/auxiliary/util/p_debug.c b/src/gallium/auxiliary/util/p_debug.c index b6cff281e6..3ed8bdfdf3 100644 --- a/src/gallium/auxiliary/util/p_debug.c +++ b/src/gallium/auxiliary/util/p_debug.c @@ -306,6 +306,13 @@ debug_get_flags_option(const char *name, str = _debug_get_option(name); if(!str) result = dfault; + else if (!util_strcmp(str, "help")) { + result = dfault; + while (flags->name) { + debug_printf("%s: help for %s: %s [0x%lx]\n", __FUNCTION__, name, flags->name, flags->value); + flags++; + } + } else { result = 0; while( flags->name ) { @@ -315,7 +322,12 @@ debug_get_flags_option(const char *name, } } - debug_printf("%s: %s = 0x%lx\n", __FUNCTION__, name, result); + if (str) { + debug_printf("%s: %s = 0x%lx (%s)\n", __FUNCTION__, name, result, str); + } + else { + debug_printf("%s: %s = 0x%lx\n", __FUNCTION__, name, result); + } return result; } diff --git a/src/gallium/auxiliary/util/u_blit.c b/src/gallium/auxiliary/util/u_blit.c index 9adf72944e..d28201ac8d 100644 --- a/src/gallium/auxiliary/util/u_blit.c +++ b/src/gallium/auxiliary/util/u_blit.c @@ -104,6 +104,7 @@ util_create_blit(struct pipe_context *pipe, struct cso_context *cso) ctx->rasterizer.cull_mode = PIPE_WINDING_NONE; ctx->rasterizer.bypass_clipping = 1; /*ctx->rasterizer.bypass_vs = 1;*/ + ctx->rasterizer.gl_rasterization_rules = 1; /* samplers */ memset(&ctx->sampler, 0, sizeof(ctx->sampler)); diff --git a/src/gallium/auxiliary/util/u_gen_mipmap.c b/src/gallium/auxiliary/util/u_gen_mipmap.c index b19a649bbc..9d305ad763 100644 --- a/src/gallium/auxiliary/util/u_gen_mipmap.c +++ b/src/gallium/auxiliary/util/u_gen_mipmap.c @@ -725,6 +725,7 @@ util_create_gen_mipmap(struct pipe_context *pipe, ctx->rasterizer.cull_mode = PIPE_WINDING_NONE; ctx->rasterizer.bypass_clipping = 1; /*ctx->rasterizer.bypass_vs = 1;*/ + ctx->rasterizer.gl_rasterization_rules = 1; /* sampler state */ memset(&ctx->sampler, 0, sizeof(ctx->sampler)); diff --git a/src/gallium/auxiliary/util/u_keymap.c b/src/gallium/auxiliary/util/u_keymap.c new file mode 100644 index 0000000000..01b17ddb1b --- /dev/null +++ b/src/gallium/auxiliary/util/u_keymap.c @@ -0,0 +1,309 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * 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, sub license, 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 NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS 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. + * + **************************************************************************/ + +/** + * Key lookup/associative container. + * + * Like Jose's u_hash_table, based on CSO cache code for now. + * + * Author: Brian Paul + */ + + +#include "pipe/p_compiler.h" +#include "pipe/p_debug.h" +#include "pipe/p_error.h" + +#include "cso_cache/cso_hash.h" + +#include "util/u_memory.h" +#include "util/u_keymap.h" + + +struct keymap +{ + struct cso_hash *cso; + unsigned key_size; + unsigned max_entries; /* XXX not obeyed net */ + unsigned num_entries; + keymap_delete_func delete_func; +}; + + +struct keymap_item +{ + void *key, *value; +}; + + +/** + * This the default key-delete function used when the client doesn't + * provide one. + */ +static void +default_delete_func(const struct keymap *map, + const void *key, void *data, void *user) +{ + FREE((void*) data); +} + + +static INLINE struct keymap_item * +hash_table_item(struct cso_hash_iter iter) +{ + return (struct keymap_item *) cso_hash_iter_data(iter); +} + + +/** + * Return 4-byte hash key for a block of bytes. + */ +static unsigned +hash(const void *key, unsigned keySize) +{ + unsigned i, hash; + + keySize /= 4; /* convert from bytes to uints */ + + hash = 0; + for (i = 0; i < keySize; i++) { + hash ^= (i + 1) * ((const unsigned *) key)[i]; + } + + /*hash = hash ^ (hash >> 11) ^ (hash >> 22);*/ + + return hash; +} + + +/** + * Create a new map. + * \param keySize size of the keys in bytes + * \param maxEntries max number of entries to allow (~0 = infinity) + * \param deleteFunc optional callback to call when entries + * are deleted/replaced + */ +struct keymap * +util_new_keymap(unsigned keySize, unsigned maxEntries, + keymap_delete_func deleteFunc) +{ + struct keymap *map = MALLOC_STRUCT(keymap); + if (!map) + return NULL; + + map->cso = cso_hash_create(); + if (!map->cso) { + FREE(map); + return NULL; + } + + map->max_entries = maxEntries; + map->num_entries = 0; + map->key_size = keySize; + map->delete_func = deleteFunc ? deleteFunc : default_delete_func; + + return map; +} + + +/** + * Delete/free a keymap and all entries. The deleteFunc that was given at + * create time will be called for each entry. + * \param user user-provided pointer passed through to the delete callback + */ +void +util_delete_keymap(struct keymap *map, void *user) +{ + util_keymap_remove_all(map, user); + cso_hash_delete(map->cso); + FREE(map); +} + + +static INLINE struct cso_hash_iter +hash_table_find_iter(const struct keymap *map, const void *key, + unsigned key_hash) +{ + struct cso_hash_iter iter; + struct keymap_item *item; + + iter = cso_hash_find(map->cso, key_hash); + while (!cso_hash_iter_is_null(iter)) { + item = (struct keymap_item *) cso_hash_iter_data(iter); + if (!memcmp(item->key, key, map->key_size)) + break; + iter = cso_hash_iter_next(iter); + } + + return iter; +} + + +static INLINE struct keymap_item * +hash_table_find_item(const struct keymap *map, const void *key, + unsigned key_hash) +{ + struct cso_hash_iter iter = hash_table_find_iter(map, key, key_hash); + if (cso_hash_iter_is_null(iter)) { + return NULL; + } + else { + return hash_table_item(iter); + } +} + + +/** + * Insert a new key + data pointer into the table. + * Note: we create a copy of the key, but not the data! + * If the key is already present in the table, replace the existing + * entry (calling the delete callback on the previous entry). + * If the maximum capacity of the map is reached an old entry + * will be deleted (the delete callback will be called). + */ +boolean +util_keymap_insert(struct keymap *map, const void *key, + const void *data, void *user) +{ + unsigned key_hash; + struct keymap_item *item; + struct cso_hash_iter iter; + + assert(map); + + key_hash = hash(key, map->key_size); + + item = hash_table_find_item(map, key, key_hash); + if (item) { + /* call delete callback for old entry/item */ + map->delete_func(map, item->key, item->value, user); + item->value = (void *) data; + return TRUE; + } + + item = MALLOC_STRUCT(keymap_item); + if (!item) + return FALSE; + + item->key = mem_dup(key, map->key_size); + item->value = (void *) data; + + iter = cso_hash_insert(map->cso, key_hash, item); + if (cso_hash_iter_is_null(iter)) { + FREE(item); + return FALSE; + } + + map->num_entries++; + + return TRUE; +} + + +/** + * Look up a key in the map and return the associated data pointer. + */ +const void * +util_keymap_lookup(const struct keymap *map, const void *key) +{ + unsigned key_hash; + struct keymap_item *item; + + assert(map); + + key_hash = hash(key, map->key_size); + + item = hash_table_find_item(map, key, key_hash); + if (!item) + return NULL; + + return item->value; +} + + +/** + * Remove an entry from the map. + * The delete callback will be called if the given key/entry is found. + * \param user passed to the delete callback as the last param. + */ +void +util_keymap_remove(struct keymap *map, const void *key, void *user) +{ + unsigned key_hash; + struct cso_hash_iter iter; + struct keymap_item *item; + + assert(map); + + key_hash = hash(key, map->key_size); + + iter = hash_table_find_iter(map, key, key_hash); + if (cso_hash_iter_is_null(iter)) + return; + + item = hash_table_item(iter); + assert(item); + map->delete_func(map, item->key, item->value, user); + FREE(item->key); + FREE(item); + + map->num_entries--; + + cso_hash_erase(map->cso, iter); +} + + +/** + * Remove all entries from the map, calling the delete callback for each. + * \param user passed to the delete callback as the last param. + */ +void +util_keymap_remove_all(struct keymap *map, void *user) +{ + struct cso_hash_iter iter; + struct keymap_item *item; + + assert(map); + + iter = cso_hash_first_node(map->cso); + while (!cso_hash_iter_is_null(iter)) { + item = (struct keymap_item *) + cso_hash_take(map->cso, cso_hash_iter_key(iter)); + map->delete_func(map, item->key, item->value, user); + FREE(item->key); + FREE(item); + iter = cso_hash_first_node(map->cso); + } +} + + +extern void +util_keymap_info(const struct keymap *map) +{ + debug_printf("Keymap %p: %u of max %u entries\n", + (void *) map, map->num_entries, map->max_entries); +} diff --git a/src/gallium/auxiliary/util/u_keymap.h b/src/gallium/auxiliary/util/u_keymap.h new file mode 100644 index 0000000000..8d60a76fc3 --- /dev/null +++ b/src/gallium/auxiliary/util/u_keymap.h @@ -0,0 +1,68 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * 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, sub license, 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 NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS 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. + * + **************************************************************************/ + +#ifndef U_KEYMAP_H +#define U_KEYMAP_H + +#include "pipe/p_compiler.h" + + +/** opaque keymap type */ +struct keymap; + + +/** Delete/callback function type */ +typedef void (*keymap_delete_func)(const struct keymap *map, + const void *key, void *data, + void *user); + + +extern struct keymap * +util_new_keymap(unsigned keySize, unsigned maxEntries, + keymap_delete_func deleteFunc); + +extern void +util_delete_keymap(struct keymap *map, void *user); + +extern boolean +util_keymap_insert(struct keymap *map, const void *key, + const void *data, void *user); + +extern const void * +util_keymap_lookup(const struct keymap *map, const void *key); + +extern void +util_keymap_remove(struct keymap *map, const void *key, void *user); + +extern void +util_keymap_remove_all(struct keymap *map, void *user); + +extern void +util_keymap_info(const struct keymap *map); + + +#endif /* U_KEYMAP_H */ diff --git a/src/gallium/auxiliary/util/u_math.c b/src/gallium/auxiliary/util/u_math.c index 0729114d6a..5b3cab4642 100644 --- a/src/gallium/auxiliary/util/u_math.c +++ b/src/gallium/auxiliary/util/u_math.c @@ -30,7 +30,7 @@ #include "util/u_math.h" - +/** 2^x, for x in [-1.0, 1.0[ */ float pow2_table[POW2_TABLE_SIZE]; @@ -38,9 +38,21 @@ static void init_pow2_table(void) { int i; - for (i = 0; i < POW2_TABLE_SIZE; i++) { - pow2_table[i] = (float) pow(2.0, i / POW2_TABLE_SCALE); - } + for (i = 0; i < POW2_TABLE_SIZE; i++) + pow2_table[i] = (float) pow(2.0, (i - POW2_TABLE_OFFSET) / POW2_TABLE_SCALE); +} + + +/** log2(x), for x in [1.0, 2.0[ */ +float log2_table[LOG2_TABLE_SIZE]; + + +static void +init_log2_table(void) +{ + unsigned i; + for (i = 0; i < LOG2_TABLE_SIZE; i++) + log2_table[i] = (float) log2(1.0 + i * (1.0 / LOG2_TABLE_SIZE)); } @@ -53,6 +65,7 @@ util_init_math(void) static boolean initialized = FALSE; if (!initialized) { init_pow2_table(); + init_log2_table(); initialized = TRUE; } } diff --git a/src/gallium/auxiliary/util/u_math.h b/src/gallium/auxiliary/util/u_math.h index 196aeb28fa..be7303e550 100644 --- a/src/gallium/auxiliary/util/u_math.h +++ b/src/gallium/auxiliary/util/u_math.h @@ -174,8 +174,10 @@ static INLINE float logf( float f ) -#define POW2_TABLE_SIZE 256 -#define POW2_TABLE_SCALE ((float) (POW2_TABLE_SIZE-1)) +#define POW2_TABLE_SIZE_LOG2 9 +#define POW2_TABLE_SIZE (1 << POW2_TABLE_SIZE_LOG2) +#define POW2_TABLE_OFFSET (POW2_TABLE_SIZE/2) +#define POW2_TABLE_SCALE ((float)(POW2_TABLE_SIZE/2)) extern float pow2_table[POW2_TABLE_SIZE]; @@ -186,98 +188,78 @@ util_init_math(void); union fi { float f; - int i; - unsigned ui; + int32_t i; + uint32_t ui; }; /** - * Fast approximation to exp(x). - * Compute with base 2 exponents: exp(x) = exp2(log2(e) * x) - * Note: log2(e) is a constant, k = 1.44269 - * So, exp(x) = exp2(k * x); + * Fast version of 2^x * Identity: exp2(a + b) = exp2(a) * exp2(b) - * Let ipart = int(k*x) - * Let fpart = k*x - ipart; - * So, exp2(k*x) = exp2(ipart) * exp2(fpart) + * Let ipart = int(x) + * Let fpart = x - ipart; + * So, exp2(x) = exp2(ipart) * exp2(fpart) * Compute exp2(ipart) with i << ipart * Compute exp2(fpart) with lookup table. */ static INLINE float -util_fast_exp(float x) +util_fast_exp2(float x) { - if (x >= 0.0f) { - float k = 1.44269f; /* = log2(e) */ - float kx = k * x; - int ipart = (int) kx; - float fpart = kx - (float) ipart; - float y = (float) (1 << ipart) - * pow2_table[(int) (fpart * POW2_TABLE_SCALE)]; - return y; - } - else { - /* exp(-x) = 1.0 / exp(x) */ - float k = -1.44269f; - float kx = k * x; - int ipart = (int) kx; - float fpart = kx - (float) ipart; - float y = (float) (1 << ipart) - * pow2_table[(int) (fpart * POW2_TABLE_SCALE)]; - return 1.0f / y; - } + int32_t ipart; + float fpart, mpart; + union fi epart; + + if(x > 129.00000f) + return 3.402823466e+38f; + + if(x < -126.99999f) + return 0.0f; + + ipart = (int32_t) x; + fpart = x - (float) ipart; + + /* same as + * epart.f = (float) (1 << ipart) + * but faster and without integer overflow for ipart > 31 */ + epart.i = (ipart + 127 ) << 23; + + mpart = pow2_table[POW2_TABLE_OFFSET + (int)(fpart * POW2_TABLE_SCALE)]; + + return epart.f * mpart; } /** - * Fast version of 2^x - * XXX the above function could be implemented in terms of this one. + * Fast approximation to exp(x). */ static INLINE float -util_fast_exp2(float x) +util_fast_exp(float x) { - if (x >= 0.0f) { - int ipart = (int) x; - float fpart = x - (float) ipart; - float y = (float) (1 << ipart) - * pow2_table[(int) (fpart * POW2_TABLE_SCALE)]; - return y; - } - else { - /* exp(-x) = 1.0 / exp(x) */ - int ipart = (int) -x; - float fpart = -x - (float) ipart; - float y = (float) (1 << ipart) - * pow2_table[(int) (fpart * POW2_TABLE_SCALE)]; - return 1.0f / y; - } + const float k = 1.44269f; /* = log2(e) */ + return util_fast_exp2(k * x); } -/** - * Based on code from http://www.flipcode.com/totd/ - */ +#define LOG2_TABLE_SIZE_LOG2 8 +#define LOG2_TABLE_SIZE (1 << LOG2_TABLE_SIZE_LOG2) +extern float log2_table[LOG2_TABLE_SIZE]; + + static INLINE float -util_fast_log2(float val) +util_fast_log2(float x) { union fi num; - int log_2; - num.f = val; - log_2 = ((num.i >> 23) & 255) - 128; - num.i &= ~(255 << 23); - num.i += 127 << 23; - num.f = ((-1.0f/3) * num.f + 2) * num.f - 2.0f/3; - return num.f + log_2; + float epart, mpart; + num.f = x; + epart = (float)(((num.i & 0x7f800000) >> 23) - 127); + mpart = log2_table[(num.i & 0x007fffff) >> (23 - LOG2_TABLE_SIZE_LOG2)]; + return epart + mpart; } static INLINE float util_fast_pow(float x, float y) { - /* XXX these tests may need adjustment */ - if (y >= 3.0f && (-0.02f <= x && x <= 0.02f)) - return 0.0f; - if (y >= 50.0f && (-0.9f <= x && x <= 0.9f)) - return 0.0f; return util_fast_exp2(util_fast_log2(x) * y); } diff --git a/src/gallium/auxiliary/util/u_sse.h b/src/gallium/auxiliary/util/u_sse.h new file mode 100644 index 0000000000..e2a8491e62 --- /dev/null +++ b/src/gallium/auxiliary/util/u_sse.h @@ -0,0 +1,77 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * 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, sub license, 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 NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS 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. + * + **************************************************************************/ + +/** + * @file + * SSE intrinsics portability header. + * + * Although the SSE intrinsics are support by all modern x86 and x86-64 + * compilers, there are some intrisincs missing in some implementations + * (especially older MSVC versions). This header abstracts that away. + */ + +#ifndef U_SSE_H_ +#define U_SSE_H_ + +#include "pipe/p_config.h" + +#if defined(PIPE_ARCH_SSE) + +#include <xmmintrin.h> +#include <emmintrin.h> + + +/* MSVC before VC8 does not support the _mm_castxxx_yyy */ +#if defined(_MSC_VER) && _MSC_VER < 1500 + +union __declspec(align(16)) m128_types { + __m128 m128; + __m128i m128i; + __m128d m128d; +}; + +static __inline __m128 +_mm_castsi128_ps(__m128i a) +{ + union m128_types u; + u.m128i = a; + return u.m128; +} + +static __inline __m128i +_mm_castps_si128(__m128 a) +{ + union m128_types u; + u.m128 = a; + return u.m128i; +} + +#endif /* defined(_MSC_VER) && _MSC_VER < 1500 */ + +#endif /* PIPE_ARCH_X86 || PIPE_ARCH_X86_64 */ + +#endif /* U_SSE_H_ */ |