diff options
Diffstat (limited to 'src/gallium/auxiliary')
51 files changed, 4810 insertions, 506 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/Makefile b/src/gallium/auxiliary/draw/Makefile index f2e36a89e9..bdbf5a08ed 100644 --- a/src/gallium/auxiliary/draw/Makefile +++ b/src/gallium/auxiliary/draw/Makefile @@ -40,6 +40,7 @@ C_SOURCES = \ draw_vs_aos_machine.c \ draw_vs_exec.c \ draw_vs_llvm.c \ + draw_vs_ppc.c \ draw_vs_sse.c diff --git a/src/gallium/auxiliary/draw/SConscript b/src/gallium/auxiliary/draw/SConscript index 544a04918b..5f05aa324a 100644 --- a/src/gallium/auxiliary/draw/SConscript +++ b/src/gallium/auxiliary/draw/SConscript @@ -38,6 +38,7 @@ draw = env.ConvenienceLibrary( 'draw_vs_aos_machine.c', 'draw_vs_exec.c', 'draw_vs_llvm.c', + 'draw_vs_ppc.c', 'draw_vs_sse.c', 'draw_vs_varient.c' ]) 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.c b/src/gallium/auxiliary/draw/draw_vs.c index 34adbd49b0..7f305304ff 100644 --- a/src/gallium/auxiliary/draw/draw_vs.c +++ b/src/gallium/auxiliary/draw/draw_vs.c @@ -85,7 +85,10 @@ draw_create_vertex_shader(struct draw_context *draw, if (!vs) { vs = draw_create_vs_sse( draw, shader ); if (!vs) { - vs = draw_create_vs_exec( draw, shader ); + vs = draw_create_vs_ppc( draw, shader ); + if (!vs) { + vs = draw_create_vs_exec( draw, shader ); + } } } diff --git a/src/gallium/auxiliary/draw/draw_vs.h b/src/gallium/auxiliary/draw/draw_vs.h index 68c24abad3..89ae158751 100644 --- a/src/gallium/auxiliary/draw/draw_vs.h +++ b/src/gallium/auxiliary/draw/draw_vs.h @@ -158,6 +158,10 @@ draw_create_vs_sse(struct draw_context *draw, const struct pipe_shader_state *templ); struct draw_vertex_shader * +draw_create_vs_ppc(struct draw_context *draw, + const struct pipe_shader_state *templ); + +struct draw_vertex_shader * draw_create_vs_llvm(struct draw_context *draw, const struct pipe_shader_state *templ); diff --git a/src/gallium/auxiliary/draw/draw_vs_exec.c b/src/gallium/auxiliary/draw/draw_vs_exec.c index 82d27d4493..80c3606657 100644 --- a/src/gallium/auxiliary/draw/draw_vs_exec.c +++ b/src/gallium/auxiliary/draw/draw_vs_exec.c @@ -62,12 +62,15 @@ 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*/ ); + } } diff --git a/src/gallium/auxiliary/draw/draw_vs_ppc.c b/src/gallium/auxiliary/draw/draw_vs_ppc.c new file mode 100644 index 0000000000..8b75136144 --- /dev/null +++ b/src/gallium/auxiliary/draw/draw_vs_ppc.c @@ -0,0 +1,239 @@ +/************************************************************************** + * + * 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. + * + **************************************************************************/ + + /* + * Authors: + * Keith Whitwell <keith@tungstengraphics.com> + * Brian Paul + */ + +#include "util/u_math.h" +#include "util/u_memory.h" +#include "pipe/p_config.h" + +#include "draw_vs.h" + +#if defined(PIPE_ARCH_PPC) + +#include "pipe/p_shader_tokens.h" + +#include "draw_private.h" +#include "draw_context.h" + +#include "rtasm/rtasm_cpu.h" +#include "rtasm/rtasm_ppc.h" +#include "tgsi/tgsi_ppc.h" +#include "tgsi/tgsi_parse.h" + + + +typedef void (PIPE_CDECL *codegen_function) (float (*inputs)[4][4], + float (*outputs)[4][4], + float (*temps)[4][4], + float (*immeds)[4], + float (*consts)[4], + const float *builtins); + + +struct draw_ppc_vertex_shader { + struct draw_vertex_shader base; + struct ppc_function ppc_program; + + codegen_function func; +}; + + +static void +vs_ppc_prepare( struct draw_vertex_shader *base, + struct draw_context *draw ) +{ + /* nothing */ +} + + +/** + * Simplified vertex shader interface for the pt paths. Given the + * complexity of code-generating all the above operations together, + * it's time to try doing all the other stuff separately. + */ +static void +vs_ppc_run_linear( struct draw_vertex_shader *base, + const float (*input)[4], + float (*output)[4], + const float (*constants)[4], + unsigned count, + unsigned input_stride, + unsigned output_stride ) +{ + struct draw_ppc_vertex_shader *shader = (struct draw_ppc_vertex_shader *)base; + unsigned int i; + +#define MAX_VERTICES 4 + + /* loop over verts */ + for (i = 0; i < count; i += MAX_VERTICES) { + const uint max_vertices = MIN2(MAX_VERTICES, count - i); + float inputs_soa[PIPE_MAX_SHADER_INPUTS][4][4] ALIGN16_ATTRIB; + float outputs_soa[PIPE_MAX_SHADER_OUTPUTS][4][4] ALIGN16_ATTRIB; + float temps_soa[TGSI_EXEC_NUM_TEMPS][4][4] ALIGN16_ATTRIB; + uint attr; + + /* convert (up to) four input verts to SoA format */ + for (attr = 0; attr < base->info.num_inputs; attr++) { + const float *vIn = (const float *) input; + uint vert; + for (vert = 0; vert < max_vertices; vert++) { +#if 0 + if (attr==0) + printf("Input v%d a%d: %f %f %f %f\n", + vert, attr, vIn[0], vIn[1], vIn[2], vIn[3]); +#endif + inputs_soa[attr][0][vert] = vIn[attr * 4 + 0]; + inputs_soa[attr][1][vert] = vIn[attr * 4 + 1]; + inputs_soa[attr][2][vert] = vIn[attr * 4 + 2]; + inputs_soa[attr][3][vert] = vIn[attr * 4 + 3]; + vIn += input_stride / 4; + } + } + + /* run compiled shader + */ + shader->func(inputs_soa, outputs_soa, temps_soa, + (float (*)[4]) shader->base.immediates, + (float (*)[4]) constants, + ppc_builtin_constants); + + /* convert (up to) four output verts from SoA back to AoS format */ + for (attr = 0; attr < base->info.num_outputs; attr++) { + float *vOut = (float *) output; + uint vert; + for (vert = 0; vert < max_vertices; vert++) { + vOut[attr * 4 + 0] = outputs_soa[attr][0][vert]; + vOut[attr * 4 + 1] = outputs_soa[attr][1][vert]; + vOut[attr * 4 + 2] = outputs_soa[attr][2][vert]; + vOut[attr * 4 + 3] = outputs_soa[attr][3][vert]; +#if 0 + if (attr==0) + printf("Output v%d a%d: %f %f %f %f\n", + vert, attr, vOut[0], vOut[1], vOut[2], vOut[3]); +#endif + vOut += output_stride / 4; + } + } + + /* advance to next group of four input/output verts */ + input = (const float (*)[4])((const char *)input + input_stride * max_vertices); + output = (float (*)[4])((char *)output + output_stride * max_vertices); + } +} + + +static void +vs_ppc_delete( struct draw_vertex_shader *base ) +{ + struct draw_ppc_vertex_shader *shader = (struct draw_ppc_vertex_shader *)base; + + ppc_release_func( &shader->ppc_program ); + + align_free( (void *) shader->base.immediates ); + + FREE( (void*) shader->base.state.tokens ); + FREE( shader ); +} + + +struct draw_vertex_shader * +draw_create_vs_ppc(struct draw_context *draw, + const struct pipe_shader_state *templ) +{ + struct draw_ppc_vertex_shader *vs; + + vs = CALLOC_STRUCT( draw_ppc_vertex_shader ); + if (vs == NULL) + return NULL; + + /* we make a private copy of the tokens */ + vs->base.state.tokens = tgsi_dup_tokens(templ->tokens); + if (!vs->base.state.tokens) + goto fail; + + tgsi_scan_shader(templ->tokens, &vs->base.info); + + vs->base.draw = draw; +#if 0 + if (1) + vs->base.create_varient = draw_vs_varient_aos_ppc; + else +#endif + vs->base.create_varient = draw_vs_varient_generic; + vs->base.prepare = vs_ppc_prepare; + vs->base.run_linear = vs_ppc_run_linear; + vs->base.delete = vs_ppc_delete; + + vs->base.immediates = align_malloc(TGSI_EXEC_NUM_IMMEDIATES * 4 * + sizeof(float), 16); + + ppc_init_func( &vs->ppc_program ); + + if (!tgsi_emit_ppc( (struct tgsi_token *) vs->base.state.tokens, + &vs->ppc_program, + (float (*)[4]) vs->base.immediates, + TRUE )) + goto fail; + + vs->func = (codegen_function) ppc_get_func( &vs->ppc_program ); + if (!vs->func) { + goto fail; + } + + return &vs->base; + +fail: + /* + debug_error("tgsi_emit_ppc() failed, falling back to interpreter\n"); + */ + + ppc_release_func( &vs->ppc_program ); + + FREE(vs); + return NULL; +} + + + +#else /* PIPE_ARCH_PPC */ + + +struct draw_vertex_shader * +draw_create_vs_ppc( struct draw_context *draw, + const struct pipe_shader_state *templ ) +{ + return (void *) 0; +} + + +#endif /* PIPE_ARCH_PPC */ diff --git a/src/gallium/auxiliary/draw/draw_vs_sse.c b/src/gallium/auxiliary/draw/draw_vs_sse.c index 77ba5152f9..0e2036f12a 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..93a9748bdb 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); @@ -179,8 +179,7 @@ struct gallivm_cpu_engine * gallivm_global_cpu_engine() typedef void (*vertex_shader_runner)(void *ainputs, void *dests, - float (*aconsts)[4], - void *temps); + float (*aconsts)[4]); #define MAX_TGSI_VERTICES 4 /*! @@ -202,7 +201,6 @@ 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) { @@ -224,8 +222,7 @@ int gallivm_cpu_vs_exec(struct gallivm_prog *prog, /* run shader */ runner(machine->Inputs, machine->Outputs, - (float (*)[4]) constants, - machine->Temps); + (float (*)[4]) constants); /* Unswizzle all output results */ 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..e1e5cabcf5 100644 --- a/src/gallium/auxiliary/gallivm/storagesoa.cpp +++ b/src/gallium/auxiliary/gallivm/storagesoa.cpp @@ -48,13 +48,11 @@ using namespace llvm; StorageSoa::StorageSoa(llvm::BasicBlock *block, llvm::Value *input, llvm::Value *output, - llvm::Value *consts, - llvm::Value *temps) + llvm::Value *consts) : m_block(block), m_input(input), m_output(output), m_consts(consts), - m_temps(temps), m_immediates(0), m_idx(0) { @@ -93,7 +91,7 @@ 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 +142,43 @@ 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); - llvm::Value *xChannel, *yChannel, *zChannel, *wChannel; + 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; 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) @@ -174,14 +193,15 @@ std::vector<llvm::Value*> StorageSoa::outputElement(llvm::Value *idx) return res; } -std::vector<llvm::Value*> StorageSoa::tempElement(llvm::Value *idx) +std::vector<llvm::Value*> StorageSoa::tempElement(llvm::IRBuilder<>* m_builder, int idx) { std::vector<llvm::Value*> res(4); + llvm::Value *temp = m_temps[idx]; - res[0] = element(m_temps, idx, 0); - res[1] = element(m_temps, idx, 1); - res[2] = element(m_temps, idx, 2); - res[3] = element(m_temps, idx, 3); + res[0] = element(temp, constantInt(0), 0); + res[1] = element(temp, constantInt(0), 1); + res[2] = element(temp, constantInt(0), 2); + res[3] = element(temp, constantInt(0), 3); return res; } @@ -260,6 +280,12 @@ 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 +304,7 @@ 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); @@ -299,10 +325,10 @@ std::vector<llvm::Value*> StorageSoa::load(enum tgsi_file_type type, int idx, in val = outputElement(realIndex); break; case TGSI_FILE_TEMPORARY: - val = tempElement(realIndex); + val = tempElement(m_builder, idx); break; case TGSI_FILE_CONSTANT: - val = constElement(realIndex); + val = constElement(m_builder, realIndex); break; case TGSI_FILE_IMMEDIATE: val = immediateElement(realIndex); @@ -328,19 +354,39 @@ std::vector<llvm::Value*> StorageSoa::load(enum tgsi_file_type type, int idx, in return res; } +llvm::Value * StorageSoa::allocaTemp(llvm::IRBuilder<>* m_builder) +{ + VectorType *vector = VectorType::get(Type::FloatTy, 4); + ArrayType *vecArray = ArrayType::get(vector, 4); + AllocaInst *alloca = new AllocaInst(vecArray, "temp", + m_builder->GetInsertBlock()); + + return alloca; +} + + void StorageSoa::store(enum tgsi_file_type type, int idx, const std::vector<llvm::Value*> &val, - int mask) + int mask, llvm::IRBuilder<>* m_builder) { llvm::Value *out = 0; + llvm::Value *realIndex = 0; switch(type) { case TGSI_FILE_OUTPUT: out = m_output; + realIndex = constantInt(idx); break; case TGSI_FILE_TEMPORARY: - out = m_temps; + // if that temp doesn't already exist, alloca it + if (m_temps.find(idx) == m_temps.end()) + m_temps[idx] = allocaTemp(m_builder); + + out = m_temps[idx]; + + realIndex = constantInt(0); break; case TGSI_FILE_INPUT: out = m_input; + realIndex = constantInt(idx); break; case TGSI_FILE_ADDRESS: { llvm::Value *addr = m_addresses[idx]; @@ -358,7 +404,6 @@ void StorageSoa::store(enum tgsi_file_type type, int idx, const std::vector<llvm assert(0); break; } - llvm::Value *realIndex = constantInt(idx); if ((mask & TGSI_WRITEMASK_X)) { llvm::Value *xChannel = elementPointer(out, realIndex, 0); new StoreInst(val[0], xChannel, false, m_block); diff --git a/src/gallium/auxiliary/gallivm/storagesoa.h b/src/gallium/auxiliary/gallivm/storagesoa.h index ae2fc7c6ae..56886f85e7 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> @@ -51,14 +52,13 @@ public: StorageSoa(llvm::BasicBlock *block, llvm::Value *input, llvm::Value *output, - llvm::Value *consts, - llvm::Value *temps); + llvm::Value *consts); 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); + int mask, llvm::IRBuilder<>* m_builder); void addImmediate(float *vec); void declareImmediates(); @@ -76,12 +76,14 @@ 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*> tempElement(llvm::IRBuilder<>* m_builder, int idx); std::vector<llvm::Value*> immediateElement(llvm::Value *indIdx); private: llvm::BasicBlock *m_block; @@ -89,12 +91,13 @@ private: llvm::Value *m_input; llvm::Value *m_output; llvm::Value *m_consts; - llvm::Value *m_temps; + std::map<int, llvm::Value*> m_temps; llvm::GlobalVariable *m_immediates; std::map<int, llvm::Value*> m_addresses; std::vector<std::vector<float> > m_immediatesToFlush; + llvm::Value * allocaTemp(llvm::IRBuilder<>* m_builder); mutable std::map<int, llvm::ConstantInt*> m_constInts; mutable char m_name[32]; diff --git a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp index 7292c0e366..c11b88af9e 100644 --- a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp +++ b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp @@ -52,8 +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 <4 x float>] temps + // [4 x [1 x float]] consts, std::vector<const Type*> funcArgs; VectorType *vectorType = VectorType::get(Type::FloatTy, 4); @@ -61,13 +60,12 @@ 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 funcArgs.push_back(vectorArrayPtr);//output funcArgs.push_back(constsArrayPtr);//consts - funcArgs.push_back(vectorArrayPtr);//temps FunctionType *functionType = FunctionType::get( /*Result=*/Type::VoidTy, @@ -707,9 +705,8 @@ 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; } @@ -1025,9 +1022,9 @@ 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); + dst->DstRegister.Index, out, dst->DstRegister.WriteMask, + instr->getIRBuilder() ); } } @@ -1122,8 +1119,6 @@ llvm::Module * tgsi_to_llvmir(struct gallivm_ir *ir, output->setName("outputs"); Value *consts = args++; consts->setName("consts"); - Value *temps = args++; - temps->setName("temps"); BasicBlock *label_entry = BasicBlock::Create("entry", shader, 0); @@ -1132,7 +1127,7 @@ llvm::Module * tgsi_to_llvmir(struct gallivm_ir *ir, fi = tgsi_default_full_instruction(); fd = tgsi_default_full_declaration(); - StorageSoa storage(label_entry, input, output, consts, temps); + StorageSoa storage(label_entry, input, output, consts); InstructionsSoa instr(mod, shader, label_entry, &storage); while(!tgsi_parse_end_of_tokens(&parse)) { diff --git a/src/gallium/auxiliary/pipebuffer/pb_buffer.h b/src/gallium/auxiliary/pipebuffer/pb_buffer.h index 8505d333bd..19db8a6a91 100644 --- a/src/gallium/auxiliary/pipebuffer/pb_buffer.h +++ b/src/gallium/auxiliary/pipebuffer/pb_buffer.h @@ -177,12 +177,16 @@ pb_get_base_buffer( struct pb_buffer *buf, } +/** + * Don't call this directly. Use pb_reference instead. + */ static INLINE void pb_destroy(struct pb_buffer *buf) { assert(buf); if(!buf) return; + assert(buf->base.refcount == 0); buf->vtbl->destroy(buf); } @@ -193,11 +197,16 @@ static INLINE void pb_reference(struct pb_buffer **dst, struct pb_buffer *src) { - if (src) + if (src) { + assert(src->base.refcount); src->base.refcount++; + } - if (*dst && --(*dst)->base.refcount == 0) - pb_destroy( *dst ); + if (*dst) { + assert((*dst)->base.refcount); + if(--(*dst)->base.refcount == 0) + pb_destroy( *dst ); + } *dst = src; } diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c index 633ee70a75..e2594ea236 100644 --- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c +++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c @@ -86,8 +86,7 @@ fenced_bufmgr_create_buffer(struct pb_manager *mgr, fenced_buf = fenced_buffer_create(fenced_mgr->fenced_list, buf); if(!fenced_buf) { - assert(buf->base.refcount == 1); - pb_destroy(buf); + pb_reference(&buf, NULL); } return fenced_buf; diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c index fe80ca30ee..a976d3041a 100644 --- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c +++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c @@ -100,7 +100,7 @@ mm_buffer_destroy(struct pb_buffer *buf) assert(buf->base.refcount == 0); pipe_mutex_lock(mm->mutex); - mmFreeMem(mm_buf->block); + u_mmFreeMem(mm_buf->block); FREE(buf); pipe_mutex_unlock(mm->mutex); } @@ -175,14 +175,14 @@ mm_bufmgr_create_buffer(struct pb_manager *mgr, mm_buf->mgr = mm; - mm_buf->block = mmAllocMem(mm->heap, size, mm->align2, 0); + mm_buf->block = u_mmAllocMem(mm->heap, size, mm->align2, 0); if(!mm_buf->block) { debug_printf("warning: heap full\n"); #if 0 mmDumpMemInfo(mm->heap); #endif - mm_buf->block = mmAllocMem(mm->heap, size, mm->align2, 0); + mm_buf->block = u_mmAllocMem(mm->heap, size, mm->align2, 0); if(!mm_buf->block) { FREE(mm_buf); pipe_mutex_unlock(mm->mutex); @@ -213,7 +213,7 @@ mm_bufmgr_destroy(struct pb_manager *mgr) pipe_mutex_lock(mm->mutex); - mmDestroy(mm->heap); + u_mmDestroy(mm->heap); pb_unmap(mm->buffer); pb_reference(&mm->buffer, NULL); @@ -254,7 +254,7 @@ mm_bufmgr_create_from_buffer(struct pb_buffer *buffer, if(!mm->map) goto failure; - mm->heap = mmInit(0, size); + mm->heap = u_mmInit(0, size); if (!mm->heap) goto failure; @@ -262,7 +262,7 @@ mm_bufmgr_create_from_buffer(struct pb_buffer *buffer, failure: if(mm->heap) - mmDestroy(mm->heap); + u_mmDestroy(mm->heap); if(mm->map) pb_unmap(mm->buffer); if(mm) diff --git a/src/gallium/auxiliary/rtasm/Makefile b/src/gallium/auxiliary/rtasm/Makefile index 39b8a4dbd7..252dc5274a 100644 --- a/src/gallium/auxiliary/rtasm/Makefile +++ b/src/gallium/auxiliary/rtasm/Makefile @@ -7,6 +7,7 @@ C_SOURCES = \ rtasm_cpu.c \ rtasm_execmem.c \ rtasm_x86sse.c \ + rtasm_ppc.c \ rtasm_ppc_spe.c include ../../Makefile.template diff --git a/src/gallium/auxiliary/rtasm/SConscript b/src/gallium/auxiliary/rtasm/SConscript index 8ea25922aa..eb48368acc 100644 --- a/src/gallium/auxiliary/rtasm/SConscript +++ b/src/gallium/auxiliary/rtasm/SConscript @@ -6,6 +6,7 @@ rtasm = env.ConvenienceLibrary( 'rtasm_cpu.c', 'rtasm_execmem.c', 'rtasm_x86sse.c', + 'rtasm_ppc.c', 'rtasm_ppc_spe.c', ]) diff --git a/src/gallium/auxiliary/rtasm/rtasm_execmem.c b/src/gallium/auxiliary/rtasm/rtasm_execmem.c index f16191cb61..be7433baf8 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_execmem.c +++ b/src/gallium/auxiliary/rtasm/rtasm_execmem.c @@ -38,12 +38,13 @@ #include "rtasm_execmem.h" -#if defined(__linux__) +#if defined(PIPE_OS_LINUX) + /* * Allocate a large block of memory which can hold code then dole it out * in pieces by means of the generic memory manager code. -*/ + */ #include <unistd.h> #include <sys/mman.h> @@ -62,7 +63,7 @@ static void init_heap(void) { if (!exec_heap) - exec_heap = mmInit( 0, EXEC_HEAP_SIZE ); + exec_heap = u_mmInit( 0, EXEC_HEAP_SIZE ); if (!exec_mem) exec_mem = (unsigned char *) mmap(0, EXEC_HEAP_SIZE, @@ -83,7 +84,7 @@ rtasm_exec_malloc(size_t size) if (exec_heap) { size = (size + 31) & ~31; /* next multiple of 32 bytes */ - block = mmAllocMem( exec_heap, size, 5, 0 ); /* 5 -> 32-byte alignment */ + block = u_mmAllocMem( exec_heap, size, 5, 0 ); /* 5 -> 32-byte alignment */ } if (block) @@ -103,17 +104,17 @@ rtasm_exec_free(void *addr) pipe_mutex_lock(exec_mutex); if (exec_heap) { - struct mem_block *block = mmFindBlock(exec_heap, (unsigned char *)addr - exec_mem); + struct mem_block *block = u_mmFindBlock(exec_heap, (unsigned char *)addr - exec_mem); if (block) - mmFreeMem(block); + u_mmFreeMem(block); } pipe_mutex_unlock(exec_mutex); } -#else +#else /* PIPE_OS_LINUX */ /* * Just use regular memory. @@ -133,4 +134,4 @@ rtasm_exec_free(void *addr) } -#endif +#endif /* PIPE_OS_LINUX */ diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc.c b/src/gallium/auxiliary/rtasm/rtasm_ppc.c new file mode 100644 index 0000000000..6d11263be8 --- /dev/null +++ b/src/gallium/auxiliary/rtasm/rtasm_ppc.c @@ -0,0 +1,959 @@ +/************************************************************************** + * + * 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. + * For reference, see http://www.power.org/resources/reading/PowerISA_V2.05.pdf + * ABI info: http://www.cs.utsa.edu/~whaley/teach/cs6463FHPO/LEC/lec12_ho.pdf + * + * Other PPC refs: + * http://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/852569B20050FF778525699600719DF2 + * http://www.ibm.com/developerworks/eserver/library/es-archguide-v2.html + * http://www.freescale.com/files/product/doc/MPCFPE32B.pdf + * + * \author Brian Paul + */ + + +#include <stdio.h> +#include "util/u_memory.h" +#include "pipe/p_debug.h" +#include "rtasm_execmem.h" +#include "rtasm_ppc.h" + + +void +ppc_init_func(struct ppc_function *p) +{ + uint i; + + p->num_inst = 0; + p->max_inst = 100; /* first guess at buffer size */ + p->store = rtasm_exec_malloc(p->max_inst * PPC_INST_SIZE); + p->reg_used = 0x0; + p->fp_used = 0x0; + p->vec_used = 0x0; + + /* only allow using gp registers 3..12 for now */ + for (i = 0; i < 3; i++) + ppc_reserve_register(p, i); + for (i = 12; i < PPC_NUM_REGS; i++) + ppc_reserve_register(p, i); +} + + +void +ppc_release_func(struct ppc_function *p) +{ + assert(p->num_inst <= p->max_inst); + if (p->store != NULL) { + rtasm_exec_free(p->store); + } + p->store = NULL; +} + + +uint +ppc_num_instructions(const struct ppc_function *p) +{ + return p->num_inst; +} + + +void (*ppc_get_func(struct ppc_function *p))(void) +{ +#if 0 + DUMP_END(); + if (DISASSEM && p->store) + debug_printf("disassemble %p %p\n", p->store, p->csr); + + if (p->store == p->error_overflow) + return (void (*)(void)) NULL; + else +#endif + return (void (*)(void)) p->store; +} + + +void +ppc_dump_func(const struct ppc_function *p) +{ + uint i; + for (i = 0; i < p->num_inst; i++) { + debug_printf("%3u: 0x%08x\n", i, p->store[i]); + } +} + + +/** + * Mark a register as being unavailable. + */ +int +ppc_reserve_register(struct ppc_function *p, int reg) +{ + assert(reg < PPC_NUM_REGS); + p->reg_used |= (1 << reg); + return reg; +} + + +/** + * Allocate a general purpose register. + * \return register index or -1 if none left. + */ +int +ppc_allocate_register(struct ppc_function *p) +{ + unsigned i; + for (i = 0; i < PPC_NUM_REGS; i++) { + const uint64_t mask = 1 << i; + if ((p->reg_used & mask) == 0) { + p->reg_used |= mask; + return i; + } + } + return -1; +} + + +/** + * Mark the given general purpose register as "unallocated". + */ +void +ppc_release_register(struct ppc_function *p, int reg) +{ + assert(reg < PPC_NUM_REGS); + assert(p->reg_used & (1 << reg)); + p->reg_used &= ~(1 << reg); +} + + +/** + * Allocate a floating point register. + * \return register index or -1 if none left. + */ +int +ppc_allocate_fp_register(struct ppc_function *p) +{ + unsigned i; + for (i = 0; i < PPC_NUM_FP_REGS; i++) { + const uint64_t mask = 1 << i; + if ((p->fp_used & mask) == 0) { + p->fp_used |= mask; + return i; + } + } + return -1; +} + + +/** + * Mark the given floating point register as "unallocated". + */ +void +ppc_release_fp_register(struct ppc_function *p, int reg) +{ + assert(reg < PPC_NUM_FP_REGS); + assert(p->fp_used & (1 << reg)); + p->fp_used &= ~(1 << reg); +} + + +/** + * Allocate a vector register. + * \return register index or -1 if none left. + */ +int +ppc_allocate_vec_register(struct ppc_function *p) +{ + 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)); + p->vec_used &= ~(1 << reg); +} + + +/** + * Append instruction to instruction buffer. Grow buffer if out of room. + */ +static void +emit_instruction(struct ppc_function *p, uint32_t inst_bits) +{ + if (!p->store) + return; /* out of memory, drop the instruction */ + + if (p->num_inst == p->max_inst) { + /* allocate larger buffer */ + uint32_t *newbuf; + p->max_inst *= 2; /* 2x larger */ + newbuf = rtasm_exec_malloc(p->max_inst * PPC_INST_SIZE); + if (newbuf) { + memcpy(newbuf, p->store, p->num_inst * PPC_INST_SIZE); + } + rtasm_exec_free(p->store); + p->store = newbuf; + if (!p->store) { + /* out of memory */ + p->num_inst = 0; + return; + } + } + + p->store[p->num_inst++] = inst_bits; +} + + +union vx_inst { + uint32_t bits; + struct { + unsigned op:6; + unsigned vD:5; + unsigned vA:5; + unsigned vB:5; + unsigned op2:11; + } 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; + emit_instruction(p, inst.bits); +}; + + +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; +}; + +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; + emit_instruction(p, inst.bits); +}; + + +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_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; + emit_instruction(p, inst.bits); +}; + + +union i_inst { + uint32_t bits; + struct { + unsigned op:6; + unsigned li:24; + unsigned aa:1; + unsigned lk:1; + } inst; +}; + +static INLINE void +emit_i(struct ppc_function *p, uint op, uint li, uint aa, uint lk) +{ + union i_inst inst; + inst.inst.op = op; + inst.inst.li = li; + inst.inst.aa = aa; + inst.inst.lk = lk; + emit_instruction(p, inst.bits); +} + + +union xl_inst { + uint32_t bits; + struct { + unsigned op:6; + unsigned bo:5; + unsigned bi:5; + unsigned unused:3; + unsigned bh:2; + unsigned op2:10; + unsigned lk:1; + } inst; +}; + +static INLINE void +emit_xl(struct ppc_function *p, uint op, uint bo, uint bi, uint bh, + uint op2, uint lk) +{ + union xl_inst inst; + inst.inst.op = op; + inst.inst.bo = bo; + inst.inst.bi = bi; + inst.inst.unused = 0x0; + inst.inst.bh = bh; + inst.inst.op2 = op2; + inst.inst.lk = lk; + emit_instruction(p, inst.bits); +} + +static INLINE void +dump_xl(const char *name, uint inst) +{ + union xl_inst i; + + i.bits = inst; + debug_printf("%s = 0x%08x\n", name, inst); + debug_printf(" op: %d 0x%x\n", i.inst.op, i.inst.op); + debug_printf(" bo: %d 0x%x\n", i.inst.bo, i.inst.bo); + debug_printf(" bi: %d 0x%x\n", i.inst.bi, i.inst.bi); + debug_printf(" unused: %d 0x%x\n", i.inst.unused, i.inst.unused); + debug_printf(" bh: %d 0x%x\n", i.inst.bh, i.inst.bh); + debug_printf(" op2: %d 0x%x\n", i.inst.op2, i.inst.op2); + debug_printf(" lk: %d 0x%x\n", i.inst.lk, i.inst.lk); +} + + +union x_inst { + uint32_t bits; + struct { + unsigned op:6; + unsigned vrs:5; + unsigned ra:5; + unsigned rb:5; + unsigned op2:10; + unsigned unused:1; + } inst; +}; + +static INLINE void +emit_x(struct ppc_function *p, uint op, uint vrs, uint ra, uint rb, uint op2) +{ + union x_inst inst; + inst.inst.op = op; + inst.inst.vrs = vrs; + inst.inst.ra = ra; + inst.inst.rb = rb; + inst.inst.op2 = op2; + inst.inst.unused = 0x0; + emit_instruction(p, inst.bits); +} + + +union d_inst { + uint32_t bits; + struct { + unsigned op:6; + unsigned rt:5; + unsigned ra:5; + unsigned si:16; + } inst; +}; + +static inline void +emit_d(struct ppc_function *p, uint op, uint rt, uint ra, int si) +{ + union d_inst inst; + assert(si >= -32768); + assert(si <= 32767); + inst.inst.op = op; + inst.inst.rt = rt; + inst.inst.ra = ra; + inst.inst.si = (unsigned) (si & 0xffff); + emit_instruction(p, inst.bits); +}; + + +union a_inst { + uint32_t bits; + struct { + unsigned op:6; + unsigned frt:5; + unsigned fra:5; + unsigned frb:5; + unsigned unused:5; + unsigned op2:5; + unsigned rc:1; + } inst; +}; + +static inline void +emit_a(struct ppc_function *p, uint op, uint frt, uint fra, uint frb, uint op2, + uint rc) +{ + union a_inst inst; + inst.inst.op = op; + inst.inst.frt = frt; + inst.inst.fra = fra; + inst.inst.frb = frb; + inst.inst.unused = 0x0; + inst.inst.op2 = op2; + inst.inst.rc = rc; + emit_instruction(p, inst.bits); +}; + + +union xo_inst { + uint32_t bits; + struct { + unsigned op:6; + unsigned rt:5; + unsigned ra:5; + unsigned rb:5; + unsigned oe:1; + unsigned op2:9; + unsigned rc:1; + } inst; +}; + +static INLINE void +emit_xo(struct ppc_function *p, uint op, uint rt, uint ra, uint rb, uint oe, + uint op2, uint rc) +{ + union xo_inst inst; + inst.inst.op = op; + inst.inst.rt = rt; + inst.inst.ra = ra; + inst.inst.rb = rb; + inst.inst.oe = oe; + inst.inst.op2 = op2; + inst.inst.rc = rc; + emit_instruction(p, inst.bits); +} + + + + + +/** + ** 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: vD = vA * vB + vC */ +void +ppc_vmaddfp(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC) +{ + emit_va(p, 46, vD, vA, vC, vB); /* note arg order */ +} + +/** vector float negative mult subtract: vD = vA - vB * vC */ +void +ppc_vnmsubfp(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC) +{ + emit_va(p, 47, vD, vB, vA, vC); /* note arg order */ +} + +/** 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); +} + +/** vector store: store vR at mem[vA+vB] */ +void +ppc_stvx(struct ppc_function *p, uint vR, uint vA, uint vB) +{ + emit_x(p, 31, vR, vA, vB, 231); +} + +/** vector load: vR = mem[vA+vB] */ +void +ppc_lvx(struct ppc_function *p, uint vR, uint vA, uint vB) +{ + emit_x(p, 31, vR, vA, vB, 103); +} + +/** load vector element word: vR = mem_word[ra+rb] */ +void +ppc_lvewx(struct ppc_function *p, uint vr, uint ra, uint rb) +{ + emit_x(p, 31, vr, ra, rb, 71); +} + + + + +/** + ** vector 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); +} + +/** Pseudo-instruction: vector move */ +void +ppc_vmove(struct ppc_function *p, uint vD, uint vA) +{ + ppc_vor(p, vD, vA, vA); +} + +/** Set vector register to {0,0,0,0} */ +void +ppc_vzero(struct ppc_function *p, uint vr) +{ + ppc_vxor(p, vr, vr, vr); +} + + + + +/** + ** 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); +} + +/** vector splat signed immediate word */ +void +ppc_vspltisw(struct ppc_function *p, uint vD, int imm) +{ + assert(imm >= -16); + assert(imm < 15); + emit_vx(p, 908, vD, imm, 0); +} + +/** vector shift left word: vD[word] = vA[word] << (vB[word] & 0x1f) */ +void +ppc_vslw(struct ppc_function *p, uint vD, uint vA, uint vB) +{ + emit_vx(p, 388, vD, vA, vB); +} + + + + +/** + ** integer arithmetic + **/ + +/** rt = ra + imm */ +void +ppc_addi(struct ppc_function *p, uint rt, uint ra, int imm) +{ + emit_d(p, 14, rt, ra, imm); +} + +/** rt = ra + (imm << 16) */ +void +ppc_addis(struct ppc_function *p, uint rt, uint ra, int imm) +{ + emit_d(p, 15, rt, ra, imm); +} + +/** rt = ra + rb */ +void +ppc_add(struct ppc_function *p, uint rt, uint ra, uint rb) +{ + emit_xo(p, 31, rt, ra, rb, 0, 266, 0); +} + +/** rt = ra AND ra */ +void +ppc_and(struct ppc_function *p, uint rt, uint ra, uint rb) +{ + emit_x(p, 31, ra, rt, rb, 28); /* note argument order */ +} + +/** rt = ra AND imm */ +void +ppc_andi(struct ppc_function *p, uint rt, uint ra, int imm) +{ + emit_d(p, 28, ra, rt, imm); /* note argument order */ +} + +/** rt = ra OR ra */ +void +ppc_or(struct ppc_function *p, uint rt, uint ra, uint rb) +{ + emit_x(p, 31, ra, rt, rb, 444); /* note argument order */ +} + +/** rt = ra OR imm */ +void +ppc_ori(struct ppc_function *p, uint rt, uint ra, int imm) +{ + emit_d(p, 24, ra, rt, imm); /* note argument order */ +} + +/** rt = ra XOR ra */ +void +ppc_xor(struct ppc_function *p, uint rt, uint ra, uint rb) +{ + emit_x(p, 31, ra, rt, rb, 316); /* note argument order */ +} + +/** rt = ra XOR imm */ +void +ppc_xori(struct ppc_function *p, uint rt, uint ra, int imm) +{ + emit_d(p, 26, ra, rt, imm); /* note argument order */ +} + +/** pseudo instruction: move: rt = ra */ +void +ppc_mr(struct ppc_function *p, uint rt, uint ra) +{ + ppc_or(p, rt, ra, ra); +} + +/** pseudo instruction: load immediate: rt = imm */ +void +ppc_li(struct ppc_function *p, uint rt, int imm) +{ + ppc_addi(p, rt, 0, imm); +} + +/** rt = imm << 16 */ +void +ppc_lis(struct ppc_function *p, uint rt, int imm) +{ + ppc_addis(p, rt, 0, imm); +} + +/** rt = imm */ +void +ppc_load_int(struct ppc_function *p, uint rt, int imm) +{ + ppc_lis(p, rt, (imm >> 16)); /* rt = imm >> 16 */ + ppc_ori(p, rt, rt, (imm & 0xffff)); /* rt = rt | (imm & 0xffff) */ +} + + + + +/** + ** integer load/store + **/ + +/** store rs at memory[(ra)+d], + * then update ra = (ra)+d + */ +void +ppc_stwu(struct ppc_function *p, uint rs, uint ra, int d) +{ + emit_d(p, 37, rs, ra, d); +} + +/** store rs at memory[(ra)+d] */ +void +ppc_stw(struct ppc_function *p, uint rs, uint ra, int d) +{ + emit_d(p, 36, rs, ra, d); +} + +/** Load rt = mem[(ra)+d]; then zero set high 32 bits to zero. */ +void +ppc_lwz(struct ppc_function *p, uint rt, uint ra, int d) +{ + emit_d(p, 32, rt, ra, d); +} + + + +/** + ** Float (non-vector) arithmetic + **/ + +/** add: frt = fra + frb */ +void +ppc_fadd(struct ppc_function *p, uint frt, uint fra, uint frb) +{ + emit_a(p, 63, frt, fra, frb, 21, 0); +} + +/** sub: frt = fra - frb */ +void +ppc_fsub(struct ppc_function *p, uint frt, uint fra, uint frb) +{ + emit_a(p, 63, frt, fra, frb, 20, 0); +} + +/** convert to int: rt = (int) ra */ +void +ppc_fctiwz(struct ppc_function *p, uint rt, uint fra) +{ + emit_x(p, 63, rt, 0, fra, 15); +} + +/** store frs at mem[(ra)+offset] */ +void +ppc_stfs(struct ppc_function *p, uint frs, uint ra, int offset) +{ + emit_d(p, 52, frs, ra, offset); +} + +/** store frs at mem[(ra)+(rb)] */ +void +ppc_stfiwx(struct ppc_function *p, uint frs, uint ra, uint rb) +{ + emit_x(p, 31, frs, ra, rb, 983); +} + +/** load frt = mem[(ra)+offset] */ +void +ppc_lfs(struct ppc_function *p, uint frt, uint ra, int offset) +{ + emit_d(p, 48, frt, ra, offset); +} + + + + + +/** + ** branch instructions + **/ + +/** BLR: Branch to link register (p. 35) */ +void +ppc_blr(struct ppc_function *p) +{ + emit_i(p, 18, 0, 0, 1); +} + +/** Branch Conditional to Link Register (p. 36) */ +void +ppc_bclr(struct ppc_function *p, uint condOp, uint branchHint, uint condReg) +{ + emit_xl(p, 19, condOp, condReg, branchHint, 16, 0); +} + +/** Pseudo instruction: return from subroutine */ +void +ppc_return(struct ppc_function *p) +{ + ppc_bclr(p, BRANCH_COND_ALWAYS, BRANCH_HINT_SUB_RETURN, 0); +} diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc.h b/src/gallium/auxiliary/rtasm/rtasm_ppc.h new file mode 100644 index 0000000000..afb4704c39 --- /dev/null +++ b/src/gallium/auxiliary/rtasm/rtasm_ppc.h @@ -0,0 +1,329 @@ +/************************************************************************** + * + * 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_REGS 32 +#define PPC_NUM_FP_REGS 32 +#define PPC_NUM_VEC_REGS 32 + +/** Stack pointer register */ +#define PPC_REG_SP 1 + +/** Branch conditions */ +#define BRANCH_COND_ALWAYS 0x14 /* binary 1z1zz (z=ignored) */ + +/** Branch hints */ +#define BRANCH_HINT_SUB_RETURN 0x0 /* binary 00 */ + + +struct ppc_function +{ + uint32_t *store; /**< instruction buffer */ + uint num_inst; + uint max_inst; + uint32_t reg_used; /** used/free general-purpose registers bitmask */ + uint32_t fp_used; /** used/free floating point registers bitmask */ + uint32_t vec_used; /** used/free vector registers bitmask */ +}; + + + +extern void ppc_init_func(struct ppc_function *p); +extern void ppc_release_func(struct ppc_function *p); +extern uint ppc_num_instructions(const struct ppc_function *p); +extern void (*ppc_get_func( struct ppc_function *p ))( void ); +extern void ppc_dump_func(const struct ppc_function *p); + +extern int ppc_reserve_register(struct ppc_function *p, int reg); +extern int ppc_allocate_register(struct ppc_function *p); +extern void ppc_release_register(struct ppc_function *p, int reg); +extern int ppc_allocate_fp_register(struct ppc_function *p); +extern void ppc_release_fp_register(struct ppc_function *p, int reg); +extern int ppc_allocate_vec_register(struct ppc_function *p); +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: vD = vA * vB + vC */ +extern void +ppc_vmaddfp(struct ppc_function *p, uint vD, uint vA, uint vB, uint vC); + +/** vector float negative mult subtract: vD = vA - vB * vC */ +extern void +ppc_vnmsubfp(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); + + +/** vector store: store vR at mem[vA+vB] */ +extern void +ppc_stvx(struct ppc_function *p, uint vR, uint vA, uint vB); + +/** vector load: vR = mem[vA+vB] */ +extern void +ppc_lvx(struct ppc_function *p, uint vR, uint vA, uint vB); + +/** load vector element word: vR = mem_word[vA+vB] */ +extern void +ppc_lvewx(struct ppc_function *p, uint vR, uint vA, uint vB); + + + +/** + ** vector 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); + +/** Pseudo-instruction: vector move */ +extern void +ppc_vmove(struct ppc_function *p, uint vD, uint vA); + +/** Set vector register to {0,0,0,0} */ +extern void +ppc_vzero(struct ppc_function *p, uint vr); + + + +/** + ** 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); + +/** vector splat signed immediate word */ +extern void +ppc_vspltisw(struct ppc_function *p, uint vD, int imm); + +/** vector shift left word: vD[word] = vA[word] << (vB[word] & 0x1f) */ +extern void +ppc_vslw(struct ppc_function *p, uint vD, uint vA, uint vB); + + + +/** + ** scalar arithmetic + **/ + +extern void +ppc_add(struct ppc_function *p, uint rt, uint ra, uint rb); + +extern void +ppc_addi(struct ppc_function *p, uint rt, uint ra, int imm); + +extern void +ppc_and(struct ppc_function *p, uint rt, uint ra, uint rb); + +extern void +ppc_andi(struct ppc_function *p, uint rt, uint ra, int imm); + +extern void +ppc_or(struct ppc_function *p, uint rt, uint ra, uint rb); + +extern void +ppc_ori(struct ppc_function *p, uint rt, uint ra, int imm); + +extern void +ppc_xor(struct ppc_function *p, uint rt, uint ra, uint rb); + +extern void +ppc_xori(struct ppc_function *p, uint rt, uint ra, int imm); + +extern void +ppc_mr(struct ppc_function *p, uint rt, uint ra); + +extern void +ppc_li(struct ppc_function *p, uint rt, int imm); + +extern void +ppc_lis(struct ppc_function *p, uint rt, int imm); + +extern void +ppc_load_int(struct ppc_function *p, uint rt, int imm); + + + +/** + ** scalar load/store + **/ + +extern void +ppc_stwu(struct ppc_function *p, uint rs, uint ra, int d); + +extern void +ppc_stw(struct ppc_function *p, uint rs, uint ra, int d); + +extern void +ppc_lwz(struct ppc_function *p, uint rs, uint ra, int d); + + + +/** + ** Float (non-vector) arithmetic + **/ + +extern void +ppc_fadd(struct ppc_function *p, uint frt, uint fra, uint frb); + +extern void +ppc_fsub(struct ppc_function *p, uint frt, uint fra, uint frb); + +extern void +ppc_fctiwz(struct ppc_function *p, uint rt, uint ra); + +extern void +ppc_stfs(struct ppc_function *p, uint frs, uint ra, int offset); + +extern void +ppc_stfiwx(struct ppc_function *p, uint frs, uint ra, uint rb); + + + +/** + ** branch instructions + **/ + +extern void +ppc_blr(struct ppc_function *p); + +void +ppc_bclr(struct ppc_function *p, uint condOp, uint branchHint, uint condReg); + +extern void +ppc_return(struct ppc_function *p); + + +#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..f8568f690b 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 @@ -143,21 +147,91 @@ 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_instruction(struct spe_function *p, uint32_t inst_bits) +{ + if (!p->store) + return; /* out of memory, drop the instruction */ + + if (p->num_inst == p->max_inst) { + /* allocate larger buffer */ + uint32_t *newbuf; + p->max_inst *= 2; /* 2x larger */ + newbuf = align_malloc(p->max_inst * SPE_INST_SIZE, 16); + if (newbuf) { + memcpy(newbuf, p->store, p->num_inst * SPE_INST_SIZE); + } + align_free(p->store); + p->store = newbuf; + if (!p->store) { + /* out of memory */ + p->num_inst = 0; + return; + } + } + + p->store[p->num_inst++] = inst_bits; +} + + + 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; inst.inst.rB = rB; inst.inst.rA = rA; inst.inst.rT = rT; - p->store[p->num_inst++] = inst.bits; - assert(p->num_inst <= p->max_inst); + emit_instruction(p, inst.bits); + 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; @@ -165,73 +239,105 @@ static void emit_RRR(struct spe_function *p, unsigned op, unsigned rT, inst.inst.rB = rB; inst.inst.rA = rA; inst.inst.rC = rC; - p->store[p->num_inst++] = inst.bits; - assert(p->num_inst <= p->max_inst); + emit_instruction(p, inst.bits); + 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; inst.inst.i7 = imm; inst.inst.rA = rA; inst.inst.rT = rT; - p->store[p->num_inst++] = inst.bits; - assert(p->num_inst <= p->max_inst); + emit_instruction(p, inst.bits); + 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; inst.inst.i8 = imm; inst.inst.rA = rA; inst.inst.rT = rT; - p->store[p->num_inst++] = inst.bits; - assert(p->num_inst <= p->max_inst); + emit_instruction(p, inst.bits); + 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; inst.inst.i10 = imm; inst.inst.rA = rA; inst.inst.rT = rT; - p->store[p->num_inst++] = inst.bits; - assert(p->num_inst <= p->max_inst); + emit_instruction(p, inst.bits); + if (p->print) { + indent(p); + printf("%s\t%s, %s, 0x%x\n", + rem_prefix(name), reg_name(rT), reg_name(rA), imm); + } +} + + +/** As above, but do range checking on signed immediate value */ +static void emit_RI10s(struct spe_function *p, unsigned op, unsigned rT, + unsigned rA, int imm, const char *name) +{ + assert(imm <= 511); + assert(imm >= -512); + emit_RI10(p, op, rT, rA, imm, name); } 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; inst.inst.i16 = imm; inst.inst.rT = rT; - p->store[p->num_inst++] = inst.bits; - assert(p->num_inst <= p->max_inst); + emit_instruction(p, inst.bits); + 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; inst.inst.i18 = imm; inst.inst.rT = rT; - p->store[p->num_inst++] = inst.bits; - assert(p->num_inst <= p->max_inst); + emit_instruction(p, inst.bits); + if (p->print) { + indent(p); + printf("%s\t%s, 0x%x\n", rem_prefix(name), reg_name(rT), imm); + } } @@ -240,80 +346,101 @@ 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_RI10s(_name, _op) \ +void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \ +{ \ + emit_RI10s(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. + * \param code_size initial size of instruction buffer to allocate, in bytes. + * If zero, use a default. */ void spe_init_func(struct spe_function *p, unsigned code_size) { - p->store = align_malloc(code_size, 16); + unsigned int i; + + if (!code_size) + code_size = 64; + p->num_inst = 0; p->max_inst = code_size / SPE_INST_SIZE; + p->store = align_malloc(code_size, 16); + + 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 +454,23 @@ 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 +484,160 @@ 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]--; + } +} - p->regs[idx] |= (1ULL << bit); + +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: offset 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; + + /* offset must be a multiple of 16 */ + assert(offset % 16 == 0); + /* offset must fit in 10-bit signed int field, after shifting */ + assert((offset >> 4) <= 511); + assert((offset >> 4) >= -512); + + p->print = FALSE; + 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: offset 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; + + /* offset must be a multiple of 16 */ + assert(offset % 16 == 0); + /* offset must fit in 10-bit signed int field, after shifting */ + assert((offset >> 4) <= 511); + assert((offset >> 4) >= -512); + + p->print = FALSE; + 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 +651,51 @@ 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 +764,223 @@ 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 & 0x0003ffff) == 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_xor_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 +991,70 @@ 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..d6a3c02f20 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 */ @@ -94,6 +119,9 @@ extern void spe_release_register(struct spe_function *p, int reg); #define EMIT_RI10(_name, _op) \ extern void _name (struct spe_function *p, unsigned rT, unsigned rA, \ int imm) +#define EMIT_RI10s(_name, _op) \ + extern void _name (struct spe_function *p, unsigned rT, unsigned rA, \ + int imm) #define EMIT_RI16(_name, _op) \ extern void _name (struct spe_function *p, unsigned rT, int imm) #define EMIT_RI18(_name, _op) \ @@ -106,11 +134,9 @@ 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); @@ -140,7 +166,7 @@ EMIT_RI16(spe_fsmbi, 0x065); EMIT_RR (spe_ah, 0x0c8); EMIT_RI10(spe_ahi, 0x01d); EMIT_RR (spe_a, 0x0c0); -EMIT_RI10(spe_ai, 0x01c); +EMIT_RI10s(spe_ai, 0x01c); EMIT_RR (spe_sfh, 0x048); EMIT_RI10(spe_sfhi, 0x00d); EMIT_RR (spe_sf, 0x040); @@ -178,19 +204,19 @@ EMIT_R (spe_xshw, 0x2ae); EMIT_R (spe_xswd, 0x2a6); EMIT_RR (spe_and, 0x0c1); EMIT_RR (spe_andc, 0x2c1); -EMIT_RI10(spe_andbi, 0x016); -EMIT_RI10(spe_andhi, 0x015); -EMIT_RI10(spe_andi, 0x014); +EMIT_RI10s(spe_andbi, 0x016); +EMIT_RI10s(spe_andhi, 0x015); +EMIT_RI10s(spe_andi, 0x014); EMIT_RR (spe_or, 0x041); EMIT_RR (spe_orc, 0x2c9); -EMIT_RI10(spe_orbi, 0x006); -EMIT_RI10(spe_orhi, 0x005); -EMIT_RI10(spe_ori, 0x004); +EMIT_RI10s(spe_orbi, 0x006); +EMIT_RI10s(spe_orhi, 0x005); +EMIT_RI10s(spe_ori, 0x004); EMIT_R (spe_orx, 0x1f0); EMIT_RR (spe_xor, 0x241); -EMIT_RI10(spe_xorbi, 0x026); -EMIT_RI10(spe_xorhi, 0x025); -EMIT_RI10(spe_xori, 0x024); +EMIT_RI10s(spe_xorbi, 0x026); +EMIT_RI10s(spe_xorhi, 0x025); +EMIT_RI10s(spe_xori, 0x024); EMIT_RR (spe_nand, 0x0c9); EMIT_RR (spe_nor, 0x049); EMIT_RR (spe_eqv, 0x249); @@ -268,6 +294,12 @@ 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 +324,33 @@ 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 +360,18 @@ 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 */ @@ -361,6 +425,7 @@ EMIT_R (spe_wrch, 0x10d); #undef EMIT_RI7 #undef EMIT_RI8 #undef EMIT_RI10 +#undef EMIT_RI10s #undef EMIT_RI16 #undef EMIT_RI18 #undef EMIT_I16 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/Makefile b/src/gallium/auxiliary/tgsi/Makefile index c7155a9316..d7df9490cf 100644 --- a/src/gallium/auxiliary/tgsi/Makefile +++ b/src/gallium/auxiliary/tgsi/Makefile @@ -11,6 +11,7 @@ C_SOURCES = \ tgsi_info.c \ tgsi_iterate.c \ tgsi_parse.c \ + tgsi_ppc.c \ tgsi_scan.c \ tgsi_sse2.c \ tgsi_text.c \ diff --git a/src/gallium/auxiliary/tgsi/SConscript b/src/gallium/auxiliary/tgsi/SConscript index 45bf3f6d57..8200cce42f 100644 --- a/src/gallium/auxiliary/tgsi/SConscript +++ b/src/gallium/auxiliary/tgsi/SConscript @@ -12,6 +12,7 @@ tgsi = env.ConvenienceLibrary( 'tgsi_parse.c', 'tgsi_sanity.c', 'tgsi_scan.c', + 'tgsi_ppc.c', 'tgsi_sse2.c', 'tgsi_text.c', 'tgsi_transform.c', 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 41dffc3dba..8d135f8777 100644 --- a/src/gallium/auxiliary/tgsi/tgsi_exec.c +++ b/src/gallium/auxiliary/tgsi/tgsi_exec.c @@ -1701,6 +1701,7 @@ 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_ppc.c b/src/gallium/auxiliary/tgsi/tgsi_ppc.c new file mode 100644 index 0000000000..a92b1902e3 --- /dev/null +++ b/src/gallium/auxiliary/tgsi/tgsi_ppc.c @@ -0,0 +1,1329 @@ +/************************************************************************** + * + * 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. + * + **************************************************************************/ + +/** + * TGSI to PowerPC code generation. + */ + +#include "pipe/p_config.h" + +#if defined(PIPE_ARCH_PPC) + +#include "pipe/p_debug.h" +#include "pipe/p_shader_tokens.h" +#include "util/u_math.h" +#include "util/u_memory.h" +#include "util/u_sse.h" +#include "tgsi/tgsi_parse.h" +#include "tgsi/tgsi_util.h" +#include "tgsi_dump.h" +#include "tgsi_exec.h" +#include "tgsi_ppc.h" +#include "rtasm/rtasm_ppc.h" + + +/** + * Since it's pretty much impossible to form PPC vector immediates, load + * them from memory here: + */ +const float ppc_builtin_constants[] ALIGN16_ATTRIB = { + 1.0f, -128.0f, 128.0, 0.0 +}; + + +#define FOR_EACH_CHANNEL( CHAN )\ + for (CHAN = 0; CHAN < NUM_CHANNELS; CHAN++) + +#define IS_DST0_CHANNEL_ENABLED( INST, CHAN )\ + ((INST).FullDstRegisters[0].DstRegister.WriteMask & (1 << (CHAN))) + +#define IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\ + if (IS_DST0_CHANNEL_ENABLED( INST, CHAN )) + +#define FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\ + FOR_EACH_CHANNEL( CHAN )\ + IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN ) + +#define CHAN_X 0 +#define CHAN_Y 1 +#define CHAN_Z 2 +#define CHAN_W 3 + + +/** + * How many TGSI temps should be implemented with real PPC vector registers + * rather than memory. + */ +#define MAX_PPC_TEMPS 4 + + +struct reg_chan_vec +{ + struct tgsi_full_src_register src; + uint chan; + uint vec; +}; + + +/** + * Context/state used during code gen. + */ +struct gen_context +{ + struct ppc_function *f; + int inputs_reg; /**< GP register pointing to input params */ + int outputs_reg; /**< GP register pointing to output params */ + int temps_reg; /**< GP register pointing to temporary "registers" */ + int immed_reg; /**< GP register pointing to immediates buffer */ + int const_reg; /**< GP register pointing to constants buffer */ + int builtins_reg; /**< GP register pointint to built-in constants */ + + int offset_reg; /**< used to reduce redundant li instructions */ + int offset_value; + + int one_vec; /**< vector register with {1.0, 1.0, 1.0, 1.0} */ + int bit31_vec; /**< vector register with {1<<31, 1<<31, 1<<31, 1<<31} */ + + /** + * Map TGSI temps to PPC vector temps. + * We have 32 PPC vector regs. Use 16 of them for storing 4 TGSI temps. + * XXX currently only do this for TGSI temps [0..MAX_PPC_TEMPS-1]. + */ + int temps_map[MAX_PPC_TEMPS][4]; + + /** + * Cache of src registers. + * This is used to avoid redundant load instructions. + */ + struct { + struct tgsi_full_src_register src; + uint chan; + uint vec; + } regs[12]; /* 3 src regs, 4 channels */ + uint num_regs; +}; + + +/** + * Initialize code generation context. + */ +static void +init_gen_context(struct gen_context *gen, struct ppc_function *func) +{ + uint i; + + memset(gen, 0, sizeof(*gen)); + gen->f = func; + gen->inputs_reg = ppc_reserve_register(func, 3); /* first function param */ + gen->outputs_reg = ppc_reserve_register(func, 4); /* second function param */ + gen->temps_reg = ppc_reserve_register(func, 5); /* ... */ + gen->immed_reg = ppc_reserve_register(func, 6); + gen->const_reg = ppc_reserve_register(func, 7); + gen->builtins_reg = ppc_reserve_register(func, 8); + gen->one_vec = -1; + gen->bit31_vec = -1; + gen->offset_reg = -1; + gen->offset_value = -9999999; + for (i = 0; i < MAX_PPC_TEMPS; i++) { + gen->temps_map[i][0] = ppc_allocate_vec_register(gen->f); + gen->temps_map[i][1] = ppc_allocate_vec_register(gen->f); + gen->temps_map[i][2] = ppc_allocate_vec_register(gen->f); + gen->temps_map[i][3] = ppc_allocate_vec_register(gen->f); + } +} + + +/** + * All PPC vector load/store instructions form an effective address + * by adding the contents of two registers. For example: + * lvx v2,r8,r9 # v2 = memory[r8 + r9] + * stvx v2,r8,r9 # memory[r8 + r9] = v2; + * So our lvx/stvx instructions are typically preceded by an 'li' instruction + * to load r9 (above) with an immediate (an offset). + * This code emits that 'li' instruction, but only if the offset value is + * different than the previous 'li'. + * This optimization seems to save about 10% in the instruction count. + * Note that we need to unconditionally emit an 'li' inside basic blocks + * (such as inside loops). + */ +static int +emit_li_offset(struct gen_context *gen, int offset) +{ + if (gen->offset_reg <= 0) { + /* allocate a GP register for storing load/store offset */ + gen->offset_reg = ppc_allocate_register(gen->f); + } + + /* emit new 'li' if offset is changing */ + if (gen->offset_value < 0 || gen->offset_value != offset) { + gen->offset_value = offset; + ppc_li(gen->f, gen->offset_reg, offset); + } + + return gen->offset_reg; +} + + +/** + * Forces subsequent emit_li_offset() calls to emit an 'li'. + * To be called at the top of basic blocks. + */ +static void +reset_li_offset(struct gen_context *gen) +{ + gen->offset_value = -9999999; +} + + + +/** + * Load the given vector register with {value, value, value, value}. + * The value must be in the ppu_builtin_constants[] array. + * We wouldn't need this if there was a simple way to load PPC vector + * registers with immediate values! + */ +static void +load_constant_vec(struct gen_context *gen, int dst_vec, float value) +{ + uint pos; + for (pos = 0; pos < Elements(ppc_builtin_constants); pos++) { + if (ppc_builtin_constants[pos] == value) { + int offset = pos * 4; + int offset_reg = emit_li_offset(gen, offset); + + /* Load 4-byte word into vector register. + * The vector slot depends on the effective address we load from. + * We know that our builtins start at a 16-byte boundary so we + * know that 'swizzle' tells us which vector slot will have the + * loaded word. The other vector slots will be undefined. + */ + ppc_lvewx(gen->f, dst_vec, gen->builtins_reg, offset_reg); + /* splat word[pos % 4] across the vector reg */ + ppc_vspltw(gen->f, dst_vec, dst_vec, pos % 4); + return; + } + } + assert(0 && "Need to add new constant to ppc_builtin_constants array"); +} + + +/** + * Return index of vector register containing {1.0, 1.0, 1.0, 1.0}. + */ +static int +gen_one_vec(struct gen_context *gen) +{ + if (gen->one_vec < 0) { + gen->one_vec = ppc_allocate_vec_register(gen->f); + load_constant_vec(gen, gen->one_vec, 1.0f); + } + return gen->one_vec; +} + +/** + * Return index of vector register containing {1<<31, 1<<31, 1<<31, 1<<31}. + */ +static int +gen_get_bit31_vec(struct gen_context *gen) +{ + if (gen->bit31_vec < 0) { + gen->bit31_vec = ppc_allocate_vec_register(gen->f); + ppc_vspltisw(gen->f, gen->bit31_vec, -1); + ppc_vslw(gen->f, gen->bit31_vec, gen->bit31_vec, gen->bit31_vec); + } + return gen->bit31_vec; +} + + +/** + * Register fetch. Return PPC vector register with result. + */ +static int +emit_fetch(struct gen_context *gen, + const struct tgsi_full_src_register *reg, + const unsigned chan_index) +{ + uint swizzle = tgsi_util_get_full_src_register_extswizzle(reg, chan_index); + int dst_vec = -1; + + switch (swizzle) { + case TGSI_EXTSWIZZLE_X: + case TGSI_EXTSWIZZLE_Y: + case TGSI_EXTSWIZZLE_Z: + case TGSI_EXTSWIZZLE_W: + switch (reg->SrcRegister.File) { + case TGSI_FILE_INPUT: + { + int offset = (reg->SrcRegister.Index * 4 + swizzle) * 16; + int offset_reg = emit_li_offset(gen, offset); + dst_vec = ppc_allocate_vec_register(gen->f); + ppc_lvx(gen->f, dst_vec, gen->inputs_reg, offset_reg); + } + break; + case TGSI_FILE_TEMPORARY: + if (reg->SrcRegister.Index < MAX_PPC_TEMPS) { + /* use PPC vec register */ + dst_vec = gen->temps_map[reg->SrcRegister.Index][swizzle]; + } + else { + /* use memory-based temp register "file" */ + int offset = (reg->SrcRegister.Index * 4 + swizzle) * 16; + int offset_reg = emit_li_offset(gen, offset); + dst_vec = ppc_allocate_vec_register(gen->f); + ppc_lvx(gen->f, dst_vec, gen->temps_reg, offset_reg); + } + break; + case TGSI_FILE_IMMEDIATE: + { + int offset = (reg->SrcRegister.Index * 4 + swizzle) * 4; + int offset_reg = emit_li_offset(gen, offset); + dst_vec = ppc_allocate_vec_register(gen->f); + /* Load 4-byte word into vector register. + * The vector slot depends on the effective address we load from. + * We know that our immediates start at a 16-byte boundary so we + * know that 'swizzle' tells us which vector slot will have the + * loaded word. The other vector slots will be undefined. + */ + ppc_lvewx(gen->f, dst_vec, gen->immed_reg, offset_reg); + /* splat word[swizzle] across the vector reg */ + ppc_vspltw(gen->f, dst_vec, dst_vec, swizzle); + } + break; + case TGSI_FILE_CONSTANT: + { + int offset = (reg->SrcRegister.Index * 4 + swizzle) * 4; + int offset_reg = emit_li_offset(gen, offset); + dst_vec = ppc_allocate_vec_register(gen->f); + /* Load 4-byte word into vector register. + * The vector slot depends on the effective address we load from. + * We know that our constants start at a 16-byte boundary so we + * know that 'swizzle' tells us which vector slot will have the + * loaded word. The other vector slots will be undefined. + */ + ppc_lvewx(gen->f, dst_vec, gen->const_reg, offset_reg); + /* splat word[swizzle] across the vector reg */ + ppc_vspltw(gen->f, dst_vec, dst_vec, swizzle); + } + break; + default: + assert( 0 ); + } + break; + case TGSI_EXTSWIZZLE_ZERO: + ppc_vzero(gen->f, dst_vec); + break; + case TGSI_EXTSWIZZLE_ONE: + { + int one_vec = gen_one_vec(gen); + dst_vec = ppc_allocate_vec_register(gen->f); + ppc_vmove(gen->f, dst_vec, one_vec); + } + break; + default: + assert( 0 ); + } + + assert(dst_vec >= 0); + + { + uint sign_op = tgsi_util_get_full_src_register_sign_mode(reg, chan_index); + if (sign_op != TGSI_UTIL_SIGN_KEEP) { + int bit31_vec = gen_get_bit31_vec(gen); + + switch (sign_op) { + case TGSI_UTIL_SIGN_CLEAR: + /* vec = vec & ~bit31 */ + ppc_vandc(gen->f, dst_vec, dst_vec, bit31_vec); + break; + case TGSI_UTIL_SIGN_SET: + /* vec = vec | bit31 */ + ppc_vor(gen->f, dst_vec, dst_vec, bit31_vec); + break; + case TGSI_UTIL_SIGN_TOGGLE: + /* vec = vec ^ bit31 */ + ppc_vxor(gen->f, dst_vec, dst_vec, bit31_vec); + break; + default: + assert(0); + } + } + } + + return dst_vec; +} + + + +/** + * Test if two TGSI src registers refer to the same memory location. + * We use this to avoid redundant register loads. + */ +static boolean +equal_src_locs(const struct tgsi_full_src_register *a, uint chan_a, + const struct tgsi_full_src_register *b, uint chan_b) +{ + int swz_a, swz_b; + int sign_a, sign_b; + if (a->SrcRegister.File != b->SrcRegister.File) + return FALSE; + if (a->SrcRegister.Index != b->SrcRegister.Index) + return FALSE; + swz_a = tgsi_util_get_full_src_register_extswizzle(a, chan_a); + swz_b = tgsi_util_get_full_src_register_extswizzle(b, chan_b); + if (swz_a != swz_b) + return FALSE; + sign_a = tgsi_util_get_full_src_register_sign_mode(a, chan_a); + sign_b = tgsi_util_get_full_src_register_sign_mode(b, chan_b); + if (sign_a != sign_b) + return FALSE; + return TRUE; +} + + +/** + * Given a TGSI src register and channel index, return the PPC vector + * register containing the value. We use a cache to prevent re-loading + * the same register multiple times. + * \return index of PPC vector register with the desired src operand + */ +static int +get_src_vec(struct gen_context *gen, + struct tgsi_full_instruction *inst, int src_reg, uint chan) +{ + const const struct tgsi_full_src_register *src = + &inst->FullSrcRegisters[src_reg]; + int vec; + uint i; + + /* check the cache */ + for (i = 0; i < gen->num_regs; i++) { + if (equal_src_locs(&gen->regs[i].src, gen->regs[i].chan, src, chan)) { + /* cache hit */ + assert(gen->regs[i].vec >= 0); + return gen->regs[i].vec; + } + } + + /* cache miss: allocate new vec reg and emit fetch/load code */ + vec = emit_fetch(gen, src, chan); + gen->regs[gen->num_regs].src = *src; + gen->regs[gen->num_regs].chan = chan; + gen->regs[gen->num_regs].vec = vec; + gen->num_regs++; + + assert(gen->num_regs <= Elements(gen->regs)); + + assert(vec >= 0); + + return vec; +} + + +/** + * Clear the src operand cache. To be called at the end of each emit function. + */ +static void +release_src_vecs(struct gen_context *gen) +{ + uint i; + for (i = 0; i < gen->num_regs; i++) { + const const struct tgsi_full_src_register src = gen->regs[i].src; + if (!(src.SrcRegister.File == TGSI_FILE_TEMPORARY && + src.SrcRegister.Index < MAX_PPC_TEMPS)) { + ppc_release_vec_register(gen->f, gen->regs[i].vec); + } + } + gen->num_regs = 0; +} + + + +static int +get_dst_vec(struct gen_context *gen, + const struct tgsi_full_instruction *inst, + unsigned chan_index) +{ + const struct tgsi_full_dst_register *reg = &inst->FullDstRegisters[0]; + + if (reg->DstRegister.File == TGSI_FILE_TEMPORARY && + reg->DstRegister.Index < MAX_PPC_TEMPS) { + int vec = gen->temps_map[reg->DstRegister.Index][chan_index]; + return vec; + } + else { + return ppc_allocate_vec_register(gen->f); + } +} + + +/** + * Register store. Store 'src_vec' at location indicated by 'reg'. + * \param free_vec Should the src_vec be released when done? + */ +static void +emit_store(struct gen_context *gen, + int src_vec, + const struct tgsi_full_instruction *inst, + unsigned chan_index, + boolean free_vec) +{ + const struct tgsi_full_dst_register *reg = &inst->FullDstRegisters[0]; + + switch (reg->DstRegister.File) { + case TGSI_FILE_OUTPUT: + { + int offset = (reg->DstRegister.Index * 4 + chan_index) * 16; + int offset_reg = emit_li_offset(gen, offset); + ppc_stvx(gen->f, src_vec, gen->outputs_reg, offset_reg); + } + break; + case TGSI_FILE_TEMPORARY: + if (reg->DstRegister.Index < MAX_PPC_TEMPS) { + if (!free_vec) { + int dst_vec = gen->temps_map[reg->DstRegister.Index][chan_index]; + if (dst_vec != src_vec) + ppc_vmove(gen->f, dst_vec, src_vec); + } + free_vec = FALSE; + } + else { + int offset = (reg->DstRegister.Index * 4 + chan_index) * 16; + int offset_reg = emit_li_offset(gen, offset); + ppc_stvx(gen->f, src_vec, gen->temps_reg, offset_reg); + } + break; +#if 0 + case TGSI_FILE_ADDRESS: + emit_addrs( + func, + xmm, + reg->DstRegister.Index, + chan_index ); + break; +#endif + default: + assert( 0 ); + } + +#if 0 + switch( inst->Instruction.Saturate ) { + case TGSI_SAT_NONE: + break; + + case TGSI_SAT_ZERO_ONE: + /* assert( 0 ); */ + break; + + case TGSI_SAT_MINUS_PLUS_ONE: + assert( 0 ); + break; + } +#endif + + if (free_vec) + ppc_release_vec_register(gen->f, src_vec); +} + + +static void +emit_scalar_unaryop(struct gen_context *gen, struct tgsi_full_instruction *inst) +{ + int v0, v1; + uint chan_index; + + v0 = get_src_vec(gen, inst, 0, CHAN_X); + v1 = ppc_allocate_vec_register(gen->f); + + switch (inst->Instruction.Opcode) { + case TGSI_OPCODE_RSQ: + /* v1 = 1.0 / sqrt(v0) */ + ppc_vrsqrtefp(gen->f, v1, v0); + break; + case TGSI_OPCODE_RCP: + /* v1 = 1.0 / v0 */ + ppc_vrefp(gen->f, v1, v0); + break; + default: + assert(0); + } + + FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) { + emit_store(gen, v1, inst, chan_index, FALSE); + } + + release_src_vecs(gen); + ppc_release_vec_register(gen->f, v1); +} + + +static void +emit_unaryop(struct gen_context *gen, struct tgsi_full_instruction *inst) +{ + uint chan_index; + FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan_index) { + int v0 = get_src_vec(gen, inst, 0, chan_index); /* v0 = srcreg[0] */ + int v1 = get_dst_vec(gen, inst, chan_index); + switch (inst->Instruction.Opcode) { + case TGSI_OPCODE_ABS: + /* turn off the most significant bit of each vector float word */ + { + int bit31_vec = gen_get_bit31_vec(gen); + ppc_vandc(gen->f, v1, v0, bit31_vec); /* v1 = v0 & ~bit31 */ + } + break; + case TGSI_OPCODE_FLOOR: + ppc_vrfim(gen->f, v1, v0); /* v1 = floor(v0) */ + break; + case TGSI_OPCODE_FRAC: + ppc_vrfim(gen->f, v1, v0); /* tmp = floor(v0) */ + ppc_vsubfp(gen->f, v1, v0, v1); /* v1 = v0 - v1 */ + break; + case TGSI_OPCODE_EXPBASE2: + ppc_vexptefp(gen->f, v1, v0); /* v1 = 2^v0 */ + break; + case TGSI_OPCODE_LOGBASE2: + /* XXX this may be broken! */ + ppc_vlogefp(gen->f, v1, v0); /* v1 = log2(v0) */ + break; + case TGSI_OPCODE_MOV: + case TGSI_OPCODE_SWZ: + if (v0 != v1) + ppc_vmove(gen->f, v1, v0); + break; + default: + assert(0); + } + emit_store(gen, v1, inst, chan_index, TRUE); /* store v0 */ + } + + release_src_vecs(gen); +} + + +static void +emit_binop(struct gen_context *gen, struct tgsi_full_instruction *inst) +{ + int zero_vec = -1; + uint chan; + + if (inst->Instruction.Opcode == TGSI_OPCODE_MUL) { + zero_vec = ppc_allocate_vec_register(gen->f); + ppc_vzero(gen->f, zero_vec); + } + + FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan) { + /* fetch src operands */ + int v0 = get_src_vec(gen, inst, 0, chan); + int v1 = get_src_vec(gen, inst, 1, chan); + int v2 = get_dst_vec(gen, inst, chan); + + /* emit binop */ + switch (inst->Instruction.Opcode) { + case TGSI_OPCODE_ADD: + ppc_vaddfp(gen->f, v2, v0, v1); + break; + case TGSI_OPCODE_SUB: + ppc_vsubfp(gen->f, v2, v0, v1); + break; + case TGSI_OPCODE_MUL: + ppc_vmaddfp(gen->f, v2, v0, v1, zero_vec); + break; + case TGSI_OPCODE_MIN: + ppc_vminfp(gen->f, v2, v0, v1); + break; + case TGSI_OPCODE_MAX: + ppc_vmaxfp(gen->f, v2, v0, v1); + break; + default: + assert(0); + } + + /* store v2 */ + emit_store(gen, v2, inst, chan, TRUE); + } + + if (inst->Instruction.Opcode == TGSI_OPCODE_MUL) + ppc_release_vec_register(gen->f, zero_vec); + + release_src_vecs(gen); +} + + +static void +emit_triop(struct gen_context *gen, struct tgsi_full_instruction *inst) +{ + uint chan; + + FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan) { + /* fetch src operands */ + int v0 = get_src_vec(gen, inst, 0, chan); + int v1 = get_src_vec(gen, inst, 1, chan); + int v2 = get_src_vec(gen, inst, 2, chan); + int v3 = get_dst_vec(gen, inst, chan); + + /* emit ALU */ + switch (inst->Instruction.Opcode) { + case TGSI_OPCODE_MAD: + ppc_vmaddfp(gen->f, v3, v0, v1, v2); /* v3 = v0 * v1 + v2 */ + break; + case TGSI_OPCODE_LRP: + ppc_vsubfp(gen->f, v3, v1, v2); /* v3 = v1 - v2 */ + ppc_vmaddfp(gen->f, v3, v0, v3, v2); /* v3 = v0 * v3 + v2 */ + break; + default: + assert(0); + } + + /* store v3 */ + emit_store(gen, v3, inst, chan, TRUE); + } + + release_src_vecs(gen); +} + + +/** + * Vector comparisons, resulting in 1.0 or 0.0 values. + */ +static void +emit_inequality(struct gen_context *gen, struct tgsi_full_instruction *inst) +{ + uint chan; + int one_vec = gen_one_vec(gen); + + FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan) { + /* fetch src operands */ + int v0 = get_src_vec(gen, inst, 0, chan); + int v1 = get_src_vec(gen, inst, 1, chan); + int v2 = get_dst_vec(gen, inst, chan); + boolean complement = FALSE; + + switch (inst->Instruction.Opcode) { + case TGSI_OPCODE_SNE: + complement = TRUE; + /* fall-through */ + case TGSI_OPCODE_SEQ: + ppc_vcmpeqfpx(gen->f, v2, v0, v1); /* v2 = v0 == v1 ? ~0 : 0 */ + break; + + case TGSI_OPCODE_SGE: + complement = TRUE; + /* fall-through */ + case TGSI_OPCODE_SLT: + ppc_vcmpgtfpx(gen->f, v2, v1, v0); /* v2 = v1 > v0 ? ~0 : 0 */ + break; + + case TGSI_OPCODE_SLE: + complement = TRUE; + /* fall-through */ + case TGSI_OPCODE_SGT: + ppc_vcmpgtfpx(gen->f, v2, v0, v1); /* v2 = v0 > v1 ? ~0 : 0 */ + break; + default: + assert(0); + } + + /* v2 is now {0,0,0,0} or {~0,~0,~0,~0} */ + + if (complement) + ppc_vandc(gen->f, v2, one_vec, v2); /* v2 = one_vec & ~v2 */ + else + ppc_vand(gen->f, v2, one_vec, v2); /* v2 = one_vec & v2 */ + + /* store v2 */ + emit_store(gen, v2, inst, chan, TRUE); + } + + release_src_vecs(gen); +} + + +static void +emit_dotprod(struct gen_context *gen, struct tgsi_full_instruction *inst) +{ + int v0, v1, v2; + uint chan_index; + + v2 = ppc_allocate_vec_register(gen->f); + + ppc_vxor(gen->f, v2, v2, v2); /* v2 = {0, 0, 0, 0} */ + + v0 = get_src_vec(gen, inst, 0, CHAN_X); /* v0 = src0.XXXX */ + v1 = get_src_vec(gen, inst, 1, CHAN_X); /* v1 = src1.XXXX */ + ppc_vmaddfp(gen->f, v2, v0, v1, v2); /* v2 = v0 * v1 + v2 */ + + v0 = get_src_vec(gen, inst, 0, CHAN_Y); /* v0 = src0.YYYY */ + v1 = get_src_vec(gen, inst, 1, CHAN_Y); /* v1 = src1.YYYY */ + ppc_vmaddfp(gen->f, v2, v0, v1, v2); /* v2 = v0 * v1 + v2 */ + + v0 = get_src_vec(gen, inst, 0, CHAN_Z); /* v0 = src0.ZZZZ */ + v1 = get_src_vec(gen, inst, 1, CHAN_Z); /* v1 = src1.ZZZZ */ + ppc_vmaddfp(gen->f, v2, v0, v1, v2); /* v2 = v0 * v1 + v2 */ + + if (inst->Instruction.Opcode == TGSI_OPCODE_DP4) { + v0 = get_src_vec(gen, inst, 0, CHAN_W); /* v0 = src0.WWWW */ + v1 = get_src_vec(gen, inst, 1, CHAN_W); /* v1 = src1.WWWW */ + ppc_vmaddfp(gen->f, v2, v0, v1, v2); /* v2 = v0 * v1 + v2 */ + } + else if (inst->Instruction.Opcode == TGSI_OPCODE_DPH) { + v1 = get_src_vec(gen, inst, 1, CHAN_W); /* v1 = src1.WWWW */ + ppc_vaddfp(gen->f, v2, v2, v1); /* v2 = v2 + v1 */ + } + + FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan_index) { + emit_store(gen, v2, inst, chan_index, FALSE); /* store v2, free v2 later */ + } + + release_src_vecs(gen); + + ppc_release_vec_register(gen->f, v2); +} + + +/** Approximation for vr = pow(va, vb) */ +static void +ppc_vec_pow(struct ppc_function *f, int vr, int va, int vb) +{ + /* pow(a,b) ~= exp2(log2(a) * b) */ + int t_vec = ppc_allocate_vec_register(f); + int zero_vec = ppc_allocate_vec_register(f); + + ppc_vzero(f, zero_vec); + + ppc_vlogefp(f, t_vec, va); /* t = log2(va) */ + ppc_vmaddfp(f, t_vec, t_vec, vb, zero_vec); /* t = t * vb */ + ppc_vexptefp(f, vr, t_vec); /* vr = 2^t */ + + ppc_release_vec_register(f, t_vec); + ppc_release_vec_register(f, zero_vec); +} + + +static void +emit_lit(struct gen_context *gen, struct tgsi_full_instruction *inst) +{ + int one_vec = gen_one_vec(gen); + + /* Compute X */ + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X)) { + emit_store(gen, one_vec, inst, CHAN_X, FALSE); + } + + /* Compute Y, Z */ + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y) || + IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) { + int x_vec; + int zero_vec = ppc_allocate_vec_register(gen->f); + + x_vec = get_src_vec(gen, inst, 0, CHAN_X); /* x_vec = src[0].x */ + + ppc_vzero(gen->f, zero_vec); /* zero = {0,0,0,0} */ + ppc_vmaxfp(gen->f, x_vec, x_vec, zero_vec); /* x_vec = max(x_vec, 0) */ + + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) { + emit_store(gen, x_vec, inst, CHAN_Y, FALSE); + } + + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) { + int y_vec, w_vec; + int z_vec = ppc_allocate_vec_register(gen->f); + int pow_vec = ppc_allocate_vec_register(gen->f); + int pos_vec = ppc_allocate_vec_register(gen->f); + int p128_vec = ppc_allocate_vec_register(gen->f); + int n128_vec = ppc_allocate_vec_register(gen->f); + + y_vec = get_src_vec(gen, inst, 0, CHAN_Y); /* y_vec = src[0].y */ + ppc_vmaxfp(gen->f, y_vec, y_vec, zero_vec); /* y_vec = max(y_vec, 0) */ + + w_vec = get_src_vec(gen, inst, 0, CHAN_W); /* w_vec = src[0].w */ + + /* clamp W to [-128, 128] */ + load_constant_vec(gen, p128_vec, 128.0f); + load_constant_vec(gen, n128_vec, -128.0f); + ppc_vmaxfp(gen->f, w_vec, w_vec, n128_vec); /* w = max(w, -128) */ + ppc_vminfp(gen->f, w_vec, w_vec, p128_vec); /* w = min(w, 128) */ + + /* if temp.x > 0 + * z = pow(tmp.y, tmp.w) + * else + * z = 0.0 + */ + ppc_vec_pow(gen->f, pow_vec, y_vec, w_vec); /* pow = pow(y, w) */ + ppc_vcmpgtfpx(gen->f, pos_vec, x_vec, zero_vec); /* pos = x > 0 */ + ppc_vand(gen->f, z_vec, pow_vec, pos_vec); /* z = pow & pos */ + + emit_store(gen, z_vec, inst, CHAN_Z, FALSE); + + ppc_release_vec_register(gen->f, z_vec); + ppc_release_vec_register(gen->f, pow_vec); + ppc_release_vec_register(gen->f, pos_vec); + ppc_release_vec_register(gen->f, p128_vec); + ppc_release_vec_register(gen->f, n128_vec); + } + + ppc_release_vec_register(gen->f, zero_vec); + } + + /* Compute W */ + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W)) { + emit_store(gen, one_vec, inst, CHAN_W, FALSE); + } + + release_src_vecs(gen); +} + + +static void +emit_exp(struct gen_context *gen, struct tgsi_full_instruction *inst) +{ + const int one_vec = gen_one_vec(gen); + int src_vec; + + /* get src arg */ + src_vec = get_src_vec(gen, inst, 0, CHAN_X); + + /* Compute X = 2^floor(src) */ + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X)) { + int dst_vec = get_dst_vec(gen, inst, CHAN_X); + int tmp_vec = ppc_allocate_vec_register(gen->f); + ppc_vrfim(gen->f, tmp_vec, src_vec); /* tmp = floor(src); */ + ppc_vexptefp(gen->f, dst_vec, tmp_vec); /* dst = 2 ^ tmp */ + emit_store(gen, dst_vec, inst, CHAN_X, TRUE); + ppc_release_vec_register(gen->f, tmp_vec); + } + + /* Compute Y = src - floor(src) */ + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) { + int dst_vec = get_dst_vec(gen, inst, CHAN_Y); + int tmp_vec = ppc_allocate_vec_register(gen->f); + ppc_vrfim(gen->f, tmp_vec, src_vec); /* tmp = floor(src); */ + ppc_vsubfp(gen->f, dst_vec, src_vec, tmp_vec); /* dst = src - tmp */ + emit_store(gen, dst_vec, inst, CHAN_Y, TRUE); + ppc_release_vec_register(gen->f, tmp_vec); + } + + /* Compute Z = RoughApprox2ToX(src) */ + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) { + int dst_vec = get_dst_vec(gen, inst, CHAN_Z); + ppc_vexptefp(gen->f, dst_vec, src_vec); /* dst = 2 ^ src */ + emit_store(gen, dst_vec, inst, CHAN_Z, TRUE); + } + + /* Compute W = 1.0 */ + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W)) { + emit_store(gen, one_vec, inst, CHAN_W, FALSE); + } + + release_src_vecs(gen); +} + + +static void +emit_log(struct gen_context *gen, struct tgsi_full_instruction *inst) +{ + const int bit31_vec = gen_get_bit31_vec(gen); + const int one_vec = gen_one_vec(gen); + int src_vec, abs_vec; + + /* get src arg */ + src_vec = get_src_vec(gen, inst, 0, CHAN_X); + + /* compute abs(src) */ + abs_vec = ppc_allocate_vec_register(gen->f); + ppc_vandc(gen->f, abs_vec, src_vec, bit31_vec); /* abs = src & ~bit31 */ + + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) && + IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) { + + /* compute tmp = floor(log2(abs)) */ + int tmp_vec = ppc_allocate_vec_register(gen->f); + ppc_vlogefp(gen->f, tmp_vec, abs_vec); /* tmp = log2(abs) */ + ppc_vrfim(gen->f, tmp_vec, tmp_vec); /* tmp = floor(tmp); */ + + /* Compute X = tmp */ + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X)) { + emit_store(gen, tmp_vec, inst, CHAN_X, FALSE); + } + + /* Compute Y = abs / 2^tmp */ + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) { + const int zero_vec = ppc_allocate_vec_register(gen->f); + ppc_vzero(gen->f, zero_vec); + ppc_vexptefp(gen->f, tmp_vec, tmp_vec); /* tmp = 2 ^ tmp */ + ppc_vrefp(gen->f, tmp_vec, tmp_vec); /* tmp = 1 / tmp */ + /* tmp = abs * tmp + zero */ + ppc_vmaddfp(gen->f, tmp_vec, abs_vec, tmp_vec, zero_vec); + emit_store(gen, tmp_vec, inst, CHAN_Y, FALSE); + ppc_release_vec_register(gen->f, zero_vec); + } + + ppc_release_vec_register(gen->f, tmp_vec); + } + + /* Compute Z = RoughApproxLog2(abs) */ + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) { + int dst_vec = get_dst_vec(gen, inst, CHAN_Z); + ppc_vlogefp(gen->f, dst_vec, abs_vec); /* dst = log2(abs) */ + emit_store(gen, dst_vec, inst, CHAN_Z, TRUE); + } + + /* Compute W = 1.0 */ + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W)) { + emit_store(gen, one_vec, inst, CHAN_W, FALSE); + } + + ppc_release_vec_register(gen->f, abs_vec); + release_src_vecs(gen); +} + + +static void +emit_pow(struct gen_context *gen, struct tgsi_full_instruction *inst) +{ + int s0_vec = get_src_vec(gen, inst, 0, CHAN_X); + int s1_vec = get_src_vec(gen, inst, 1, CHAN_X); + int pow_vec = ppc_allocate_vec_register(gen->f); + int chan; + + ppc_vec_pow(gen->f, pow_vec, s0_vec, s1_vec); + + FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan) { + emit_store(gen, pow_vec, inst, chan, FALSE); + } + + ppc_release_vec_register(gen->f, pow_vec); + + release_src_vecs(gen); +} + + +static void +emit_xpd(struct gen_context *gen, struct tgsi_full_instruction *inst) +{ + int x0_vec, y0_vec, z0_vec; + int x1_vec, y1_vec, z1_vec; + int zero_vec, tmp_vec; + int tmp2_vec; + + zero_vec = ppc_allocate_vec_register(gen->f); + ppc_vzero(gen->f, zero_vec); + + tmp_vec = ppc_allocate_vec_register(gen->f); + tmp2_vec = ppc_allocate_vec_register(gen->f); + + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y) || + IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) { + x0_vec = get_src_vec(gen, inst, 0, CHAN_X); + x1_vec = get_src_vec(gen, inst, 1, CHAN_X); + } + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) || + IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) { + y0_vec = get_src_vec(gen, inst, 0, CHAN_Y); + y1_vec = get_src_vec(gen, inst, 1, CHAN_Y); + } + if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) || + IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) { + z0_vec = get_src_vec(gen, inst, 0, CHAN_Z); + z1_vec = get_src_vec(gen, inst, 1, CHAN_Z); + } + + IF_IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) { + /* tmp = y0 * z1 */ + ppc_vmaddfp(gen->f, tmp_vec, y0_vec, z1_vec, zero_vec); + /* tmp = tmp - z0 * y1*/ + ppc_vnmsubfp(gen->f, tmp_vec, tmp_vec, z0_vec, y1_vec); + emit_store(gen, tmp_vec, inst, CHAN_X, FALSE); + } + IF_IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y) { + /* tmp = z0 * x1 */ + ppc_vmaddfp(gen->f, tmp_vec, z0_vec, x1_vec, zero_vec); + /* tmp = tmp - x0 * z1 */ + ppc_vnmsubfp(gen->f, tmp_vec, tmp_vec, x0_vec, z1_vec); + emit_store(gen, tmp_vec, inst, CHAN_Y, FALSE); + } + IF_IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z) { + /* tmp = x0 * y1 */ + ppc_vmaddfp(gen->f, tmp_vec, x0_vec, y1_vec, zero_vec); + /* tmp = tmp - y0 * x1 */ + ppc_vnmsubfp(gen->f, tmp_vec, tmp_vec, y0_vec, x1_vec); + emit_store(gen, tmp_vec, inst, CHAN_Z, FALSE); + } + /* W is undefined */ + + ppc_release_vec_register(gen->f, tmp_vec); + ppc_release_vec_register(gen->f, zero_vec); + release_src_vecs(gen); +} + +static int +emit_instruction(struct gen_context *gen, + struct tgsi_full_instruction *inst) +{ + switch (inst->Instruction.Opcode) { + case TGSI_OPCODE_MOV: + case TGSI_OPCODE_SWZ: + case TGSI_OPCODE_ABS: + case TGSI_OPCODE_FLOOR: + case TGSI_OPCODE_FRAC: + case TGSI_OPCODE_EXPBASE2: + case TGSI_OPCODE_LOGBASE2: + emit_unaryop(gen, inst); + break; + case TGSI_OPCODE_RSQ: + case TGSI_OPCODE_RCP: + emit_scalar_unaryop(gen, inst); + break; + case TGSI_OPCODE_ADD: + case TGSI_OPCODE_SUB: + case TGSI_OPCODE_MUL: + case TGSI_OPCODE_MIN: + case TGSI_OPCODE_MAX: + emit_binop(gen, inst); + break; + case TGSI_OPCODE_SEQ: + case TGSI_OPCODE_SNE: + case TGSI_OPCODE_SLT: + case TGSI_OPCODE_SGT: + case TGSI_OPCODE_SLE: + case TGSI_OPCODE_SGE: + emit_inequality(gen, inst); + break; + case TGSI_OPCODE_MAD: + case TGSI_OPCODE_LRP: + emit_triop(gen, inst); + break; + case TGSI_OPCODE_DP3: + case TGSI_OPCODE_DP4: + case TGSI_OPCODE_DPH: + emit_dotprod(gen, inst); + break; + case TGSI_OPCODE_LIT: + emit_lit(gen, inst); + break; + case TGSI_OPCODE_LOG: + emit_log(gen, inst); + break; + case TGSI_OPCODE_EXP: + emit_exp(gen, inst); + break; + case TGSI_OPCODE_POW: + emit_pow(gen, inst); + break; + case TGSI_OPCODE_XPD: + emit_xpd(gen, inst); + break; + case TGSI_OPCODE_END: + /* normal end */ + return 1; + default: + return 0; + } + return 1; +} + + +static void +emit_declaration( + struct ppc_function *func, + struct tgsi_full_declaration *decl ) +{ + if( decl->Declaration.File == TGSI_FILE_INPUT ) { +#if 0 + unsigned first, last, mask; + unsigned i, j; + + first = decl->DeclarationRange.First; + last = decl->DeclarationRange.Last; + mask = decl->Declaration.UsageMask; + + for( i = first; i <= last; i++ ) { + for( j = 0; j < NUM_CHANNELS; j++ ) { + if( mask & (1 << j) ) { + switch( decl->Declaration.Interpolate ) { + case TGSI_INTERPOLATE_CONSTANT: + emit_coef_a0( func, 0, i, j ); + emit_inputs( func, 0, i, j ); + break; + + case TGSI_INTERPOLATE_LINEAR: + emit_tempf( func, 0, 0, TGSI_SWIZZLE_X ); + emit_coef_dadx( func, 1, i, j ); + emit_tempf( func, 2, 0, TGSI_SWIZZLE_Y ); + emit_coef_dady( func, 3, i, j ); + emit_mul( func, 0, 1 ); /* x * dadx */ + emit_coef_a0( func, 4, i, j ); + emit_mul( func, 2, 3 ); /* y * dady */ + emit_add( func, 0, 4 ); /* x * dadx + a0 */ + emit_add( func, 0, 2 ); /* x * dadx + y * dady + a0 */ + emit_inputs( func, 0, i, j ); + break; + + case TGSI_INTERPOLATE_PERSPECTIVE: + emit_tempf( func, 0, 0, TGSI_SWIZZLE_X ); + emit_coef_dadx( func, 1, i, j ); + emit_tempf( func, 2, 0, TGSI_SWIZZLE_Y ); + emit_coef_dady( func, 3, i, j ); + emit_mul( func, 0, 1 ); /* x * dadx */ + emit_tempf( func, 4, 0, TGSI_SWIZZLE_W ); + emit_coef_a0( func, 5, i, j ); + emit_rcp( func, 4, 4 ); /* 1.0 / w */ + emit_mul( func, 2, 3 ); /* y * dady */ + emit_add( func, 0, 5 ); /* x * dadx + a0 */ + emit_add( func, 0, 2 ); /* x * dadx + y * dady + a0 */ + emit_mul( func, 0, 4 ); /* (x * dadx + y * dady + a0) / w */ + emit_inputs( func, 0, i, j ); + break; + + default: + assert( 0 ); + break; + } + } + } + } +#endif + } +} + + + +static void +emit_prologue(struct ppc_function *func) +{ + /* XXX set up stack frame */ +} + + +static void +emit_epilogue(struct ppc_function *func) +{ + ppc_return(func); + /* XXX restore prev stack frame */ + debug_printf("PPC: Emitted %u instructions\n", func->num_inst); +} + + + +/** + * Translate a TGSI vertex/fragment shader to PPC code. + * + * \param tokens the TGSI input shader + * \param func the output PPC code/function + * \param immediates buffer to place immediates, later passed to PPC func + * \return TRUE for success, FALSE if translation failed + */ +boolean +tgsi_emit_ppc(const struct tgsi_token *tokens, + struct ppc_function *func, + float (*immediates)[4], + boolean do_swizzles ) +{ + static int use_ppc_asm = -1; + struct tgsi_parse_context parse; + /*boolean instruction_phase = FALSE;*/ + unsigned ok = 1; + uint num_immediates = 0; + struct gen_context gen; + + if (use_ppc_asm < 0) { + /* If GALLIUM_NOPPC is set, don't use PPC codegen */ + use_ppc_asm = !debug_get_bool_option("GALLIUM_NOPPC", FALSE); + } + if (!use_ppc_asm) + return FALSE; + + if (0) { + debug_printf("\n********* TGSI->PPC ********\n"); + tgsi_dump(tokens, 0); + } + + util_init_math(); + + init_gen_context(&gen, func); + + emit_prologue(func); + + tgsi_parse_init( &parse, tokens ); + + while (!tgsi_parse_end_of_tokens(&parse) && ok) { + tgsi_parse_token(&parse); + + switch (parse.FullToken.Token.Type) { + case TGSI_TOKEN_TYPE_DECLARATION: + if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_FRAGMENT) { + emit_declaration(func, &parse.FullToken.FullDeclaration ); + } + break; + + case TGSI_TOKEN_TYPE_INSTRUCTION: + ok = emit_instruction(&gen, &parse.FullToken.FullInstruction); + + if (!ok) { + debug_printf("failed to translate tgsi opcode %d to PPC (%s)\n", + parse.FullToken.FullInstruction.Instruction.Opcode, + parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX ? + "vertex shader" : "fragment shader"); + } + break; + + case TGSI_TOKEN_TYPE_IMMEDIATE: + /* splat each immediate component into a float[4] vector for SoA */ + { + const uint size = parse.FullToken.FullImmediate.Immediate.Size - 1; + uint i; + assert(size <= 4); + assert(num_immediates < TGSI_EXEC_NUM_IMMEDIATES); + for (i = 0; i < size; i++) { + immediates[num_immediates][i] = + parse.FullToken.FullImmediate.u.ImmediateFloat32[i].Float; + } + num_immediates++; + } + break; + + default: + ok = 0; + assert( 0 ); + } + } + + emit_epilogue(func); + + tgsi_parse_free( &parse ); + + if (ppc_num_instructions(func) == 0) { + /* ran out of memory for instructions */ + ok = FALSE; + } + + if (!ok) + debug_printf("TGSI->PPC translation failed\n"); + + return ok; +} + +#endif /* PIPE_ARCH_PPC */ diff --git a/src/gallium/auxiliary/tgsi/tgsi_ppc.h b/src/gallium/auxiliary/tgsi/tgsi_ppc.h new file mode 100644 index 0000000000..829ec075e7 --- /dev/null +++ b/src/gallium/auxiliary/tgsi/tgsi_ppc.h @@ -0,0 +1,51 @@ +/************************************************************************** + * + * 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 TGSI_PPC_H +#define TGSI_PPC_H + +#if defined __cplusplus +extern "C" { +#endif + +struct tgsi_token; +struct ppc_function; + +extern const float ppc_builtin_constants[]; + + +boolean +tgsi_emit_ppc(const struct tgsi_token *tokens, + struct ppc_function *function, + float (*immediates)[4], + boolean do_swizzles); + +#if defined __cplusplus +} +#endif + +#endif /* TGSI_PPC_H */ diff --git a/src/gallium/auxiliary/tgsi/tgsi_sse2.c b/src/gallium/auxiliary/tgsi/tgsi_sse2.c index 4d59106dbf..3df0c5db3f 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: @@ -509,10 +512,31 @@ 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) ); @@ -522,12 +546,49 @@ 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( @@ -539,39 +600,6 @@ 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, @@ -582,6 +610,7 @@ 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)() ) @@ -593,10 +622,111 @@ 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. */ @@ -639,38 +769,35 @@ 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 ); } @@ -710,10 +837,12 @@ 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 ); } @@ -731,31 +860,35 @@ 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 ); } @@ -797,14 +930,14 @@ 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] ); @@ -816,11 +949,13 @@ 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 ); @@ -913,10 +1048,12 @@ 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 ); } @@ -1336,7 +1473,7 @@ 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, @@ -1382,11 +1519,11 @@ 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) */ @@ -1398,7 +1535,7 @@ 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 ); } } @@ -1416,21 +1553,21 @@ 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 ); @@ -1620,7 +1757,7 @@ 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; @@ -1633,7 +1770,7 @@ 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; @@ -1645,7 +1782,7 @@ 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 ); } @@ -1654,7 +1791,7 @@ 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 ); } @@ -1664,7 +1801,7 @@ 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 ); } @@ -1755,7 +1892,7 @@ 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 ); } @@ -1814,7 +1951,7 @@ 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 ); } @@ -1908,12 +2045,12 @@ 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..a1a51d7ef2 100644 --- a/src/gallium/auxiliary/util/p_debug.c +++ b/src/gallium/auxiliary/util/p_debug.c @@ -36,6 +36,13 @@ #include <windows.h> #include <winddi.h> +#elif defined(PIPE_SUBSYSTEM_WINDOWS_CE) + +#include <stdio.h> +#include <stdlib.h> +#include <windows.h> +#include <types.h> + #elif defined(PIPE_SUBSYSTEM_WINDOWS_USER) #ifndef WIN32_LEAN_AND_MEAN @@ -98,7 +105,35 @@ void _debug_vprintf(const char *format, va_list ap) OutputDebugStringA(buf); buf[0] = '\0'; } -#elif defined(PIPE_SUBSYSTEM_WINDOWS_CE) || defined(PIPE_SUBSYSTEM_WINDOWS_MINIPORT) +#elif defined(PIPE_SUBSYSTEM_WINDOWS_CE) + wchar_t *wide_format; + long wide_str_len; + char buf[512]; + int ret; +#if (_WIN32_WCE < 600) + ret = vsprintf(buf, format, ap); + if(ret < 0){ + sprintf(buf, "Cant handle debug print!"); + ret = 25; + } +#else + ret = vsprintf_s(buf, 512, format, ap); + if(ret < 0){ + sprintf_s(buf, 512, "Cant handle debug print!"); + ret = 25; + } +#endif + buf[ret] = '\0'; + /* Format is ascii - needs to be converted to wchar_t for printing */ + wide_str_len = MultiByteToWideChar(CP_ACP, 0, (const char *) buf, -1, NULL, 0); + wide_format = (wchar_t *) malloc((wide_str_len+1) * sizeof(wchar_t)); + if (wide_format) { + MultiByteToWideChar(CP_ACP, 0, (const char *) buf, -1, + wide_format, wide_str_len); + NKDbgPrintfW(wide_format, wide_format); + free(wide_format); + } +#elif defined(PIPE_SUBSYSTEM_WINDOWS_MINIPORT) /* TODO */ #else /* !PIPE_SUBSYSTEM_WINDOWS */ vfprintf(stderr, format, ap); @@ -306,6 +341,13 @@ 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 +357,12 @@ 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; } @@ -625,6 +672,7 @@ void debug_dump_surface_bmp(const char *filename, struct pipe_surface *surface) { +#ifndef PIPE_SUBSYSTEM_WINDOWS_MINIPORT struct util_stream *stream; unsigned surface_usage; struct bmp_file_header bmfh; @@ -691,6 +739,7 @@ error2: FREE(rgba); error1: ; +#endif } #endif 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..d2eaa2e7f7 100644 --- a/src/gallium/auxiliary/util/u_math.h +++ b/src/gallium/auxiliary/util/u_math.h @@ -68,7 +68,7 @@ __inline double ceil(double val) return ceil_val; } -#ifndef PIPE_SUBSYSTEM_WINDOWS_CE +#ifndef PIPE_SUBSYSTEM_WINDOWS_CE_OGL __inline double floor(double val) { double floor_val; @@ -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_mm.c b/src/gallium/auxiliary/util/u_mm.c index 01dd67c810..45ce257b5e 100644 --- a/src/gallium/auxiliary/util/u_mm.c +++ b/src/gallium/auxiliary/util/u_mm.c @@ -31,7 +31,7 @@ void -mmDumpMemInfo(const struct mem_block *heap) +u_mmDumpMemInfo(const struct mem_block *heap) { debug_printf("Memory heap %p:\n", (void *)heap); if (heap == 0) { @@ -58,7 +58,7 @@ mmDumpMemInfo(const struct mem_block *heap) } struct mem_block * -mmInit(int ofs, int size) +u_mmInit(int ofs, int size) { struct mem_block *heap, *block; @@ -165,7 +165,7 @@ SliceBlock(struct mem_block *p, struct mem_block * -mmAllocMem(struct mem_block *heap, int size, int align2, int startSearch) +u_mmAllocMem(struct mem_block *heap, int size, int align2, int startSearch) { struct mem_block *p; const int mask = (1 << align2)-1; @@ -202,7 +202,7 @@ mmAllocMem(struct mem_block *heap, int size, int align2, int startSearch) struct mem_block * -mmFindBlock(struct mem_block *heap, int start) +u_mmFindBlock(struct mem_block *heap, int start) { struct mem_block *p; @@ -241,7 +241,7 @@ Join2Blocks(struct mem_block *p) } int -mmFreeMem(struct mem_block *b) +u_mmFreeMem(struct mem_block *b) { if (!b) return 0; @@ -270,7 +270,7 @@ mmFreeMem(struct mem_block *b) void -mmDestroy(struct mem_block *heap) +u_mmDestroy(struct mem_block *heap) { struct mem_block *p; diff --git a/src/gallium/auxiliary/util/u_mm.h b/src/gallium/auxiliary/util/u_mm.h index b226b101cb..ce20e48763 100644 --- a/src/gallium/auxiliary/util/u_mm.h +++ b/src/gallium/auxiliary/util/u_mm.h @@ -49,7 +49,7 @@ struct mem_block { * input: total size in bytes * return: a heap pointer if OK, NULL if error */ -extern struct mem_block *mmInit(int ofs, int size); +extern struct mem_block *u_mmInit(int ofs, int size); /** * Allocate 'size' bytes with 2^align2 bytes alignment, @@ -61,7 +61,7 @@ extern struct mem_block *mmInit(int ofs, int size); * startSearch = linear offset from start of heap to begin search * return: pointer to the allocated block, 0 if error */ -extern struct mem_block *mmAllocMem(struct mem_block *heap, int size, int align2, +extern struct mem_block *u_mmAllocMem(struct mem_block *heap, int size, int align2, int startSearch); /** @@ -69,23 +69,23 @@ extern struct mem_block *mmAllocMem(struct mem_block *heap, int size, int align2 * input: pointer to a block * return: 0 if OK, -1 if error */ -extern int mmFreeMem(struct mem_block *b); +extern int u_mmFreeMem(struct mem_block *b); /** * Free block starts at offset * input: pointer to a heap, start offset * return: pointer to a block */ -extern struct mem_block *mmFindBlock(struct mem_block *heap, int start); +extern struct mem_block *u_mmFindBlock(struct mem_block *heap, int start); /** * destroy MM */ -extern void mmDestroy(struct mem_block *mmInit); +extern void u_mmDestroy(struct mem_block *mmInit); /** * For debuging purpose. */ -extern void mmDumpMemInfo(const struct mem_block *mmInit); +extern void u_mmDumpMemInfo(const struct mem_block *mmInit); #endif diff --git a/src/gallium/auxiliary/util/u_rect.c b/src/gallium/auxiliary/util/u_rect.c index f5619ef791..30f32413d7 100644 --- a/src/gallium/auxiliary/util/u_rect.c +++ b/src/gallium/auxiliary/util/u_rect.c @@ -222,7 +222,8 @@ util_surface_copy(struct pipe_context *pipe, w, h, src_map, do_flip ? -(int) src->stride : src->stride, - src_x, src_y); + src_x, + do_flip ? w - src_y : src_y); } pipe->screen->surface_unmap(pipe->screen, src); 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_ */ diff --git a/src/gallium/auxiliary/util/u_tile.c b/src/gallium/auxiliary/util/u_tile.c index 853c503f4f..32f6b072a0 100644 --- a/src/gallium/auxiliary/util/u_tile.c +++ b/src/gallium/auxiliary/util/u_tile.c @@ -460,7 +460,7 @@ l8_put_tile_rgba(ubyte *dst, for (j = 0; j < w; j++, pRow += 4) { unsigned r; r = float_to_ubyte(pRow[0]); - *dst++ = r; + *dst++ = (ubyte) r; } p += src_stride; } @@ -504,7 +504,7 @@ a8_put_tile_rgba(ubyte *dst, for (j = 0; j < w; j++, pRow += 4) { unsigned a; a = float_to_ubyte(pRow[3]); - *dst++ = a; + *dst++ = (ubyte) a; } p += src_stride; } @@ -634,7 +634,7 @@ i8_put_tile_rgba(ubyte *dst, for (j = 0; j < w; j++, pRow += 4) { unsigned r; r = float_to_ubyte(pRow[0]); - *dst++ = r; + *dst++ = (ubyte) r; } p += src_stride; } @@ -769,6 +769,32 @@ z24s8_get_tile_rgba(const unsigned *src, } +/*** PIPE_FORMAT_Z32_FLOAT ***/ + +/** + * Return each Z value as four floats in [0,1]. + */ +static void +z32f_get_tile_rgba(const float *src, + unsigned w, unsigned h, + float *p, + unsigned dst_stride) +{ + unsigned i, j; + + for (i = 0; i < h; i++) { + float *pRow = p; + for (j = 0; j < w; j++, pRow += 4) { + pRow[0] = + pRow[1] = + pRow[2] = + pRow[3] = *src++; + } + p += dst_stride; + } +} + + /*** PIPE_FORMAT_YCBCR / PIPE_FORMAT_YCBCR_REV ***/ /** @@ -913,6 +939,9 @@ pipe_tile_raw_to_rgba(enum pipe_format format, case PIPE_FORMAT_Z24S8_UNORM: z24s8_get_tile_rgba((unsigned *) src, w, h, dst, dst_stride); break; + case PIPE_FORMAT_Z32_FLOAT: + z32f_get_tile_rgba((float *) src, w, h, dst, dst_stride); + break; case PIPE_FORMAT_YCBCR: ycbcr_get_tile_rgba((ushort *) src, w, h, dst, dst_stride, FALSE); break; diff --git a/src/gallium/auxiliary/util/u_time.c b/src/gallium/auxiliary/util/u_time.c index bf7d1d1c8d..57b80e5604 100644 --- a/src/gallium/auxiliary/util/u_time.c +++ b/src/gallium/auxiliary/util/u_time.c @@ -200,7 +200,7 @@ util_time_timeout(const struct util_time *start, } -#if defined(PIPE_SUBSYSYEM_WINDOWS_DISPLAY) +#if defined(PIPE_SUBSYSTEM_WINDOWS_DISPLAY) void util_time_sleep(unsigned usecs) { LONGLONG start, curr, end; |