void
llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
{
struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
struct lp_fragment_shader *shader = fs;
struct lp_fragment_shader_variant *variant;
assert(fs != llvmpipe->fs);
variant = shader->variants;
while(variant) {
struct lp_fragment_shader_variant *next = variant->next;
if(variant->function) {
if(variant->jit_function)
LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
LLVMDeleteFunction(variant->function);
}
FREE(variant);
variant = next;
}
FREE((void *) shader->base.tokens);
FREE(shader);
}
static void
remove_shader_variant(struct llvmpipe_context *lp,
struct lp_fragment_shader_variant *variant)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
unsigned i;
if (gallivm_debug & GALLIVM_DEBUG_IR) {
debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
variant->shader->no, variant->no, variant->shader->variants_created,
variant->shader->variants_cached, lp->nr_fs_variants);
}
for (i = 0; i < Elements(variant->function); i++) {
if (variant->function[i]) {
if (variant->jit_function[i])
LLVMFreeMachineCodeForFunction(screen->engine,
variant->function[i]);
LLVMDeleteFunction(variant->function[i]);
}
}
remove_from_list(&variant->list_item_local);
variant->shader->variants_cached--;
remove_from_list(&variant->list_item_global);
lp->nr_fs_variants--;
FREE(variant);
}
SWIGEXPORT void JNICALL Java_org_jllvm_bindings_ExecutionEngineJNI_LLVMFreeMachineCodeForFunction(JNIEnv *jenv, jclass jcls, jlong jarg1, jlong jarg2) {
LLVMExecutionEngineRef arg1 = (LLVMExecutionEngineRef) 0 ;
LLVMValueRef arg2 = (LLVMValueRef) 0 ;
(void)jenv;
(void)jcls;
arg1 = *(LLVMExecutionEngineRef *)&jarg1;
arg2 = *(LLVMValueRef *)&jarg2;
LLVMFreeMachineCodeForFunction(arg1,arg2);
}
/**
* Free the function (and its machine code).
*/
void
gallivm_free_function(struct gallivm_state *gallivm,
LLVMValueRef func,
const void *code)
{
#if !USE_MCJIT
if (code) {
LLVMFreeMachineCodeForFunction(gallivm->engine, func);
}
LLVMDeleteFunction(func);
#endif
}
void JITImpl::reclaimUnreachableFunctions(JITCoreInfo &coreInfo)
{
std::vector<uint32_t> &unreachableFunctions = coreInfo.unreachableFunctions;
std::map<uint32_t,JITFunctionInfo*> &functionMap = coreInfo.functionMap;
for (std::vector<uint32_t>::iterator it = unreachableFunctions.begin(),
e = unreachableFunctions.end(); it != e; ++it) {
std::map<uint32_t,JITFunctionInfo*>::iterator entry = functionMap.find(*it);
if (entry == functionMap.end())
continue;
LLVMValueRef value = entry->second->value;
LLVMFreeMachineCodeForFunction(executionEngine, value);
LLVMReplaceAllUsesWith(value, LLVMGetUndef(LLVMTypeOf(value)));
LLVMDeleteFunction(value);
delete entry->second;
functionMap.erase(entry);
}
unreachableFunctions.clear();
}
void
llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
{
struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
struct lp_fragment_shader *shader = fs;
struct lp_fragment_shader_variant *variant;
assert(fs != llvmpipe->fs);
(void) llvmpipe;
/*
* XXX: we need to flush the context until we have some sort of reference
* counting in fragment shaders as they may still be binned
*/
draw_flush(llvmpipe->draw);
lp_setup_flush(llvmpipe->setup, 0);
variant = shader->variants;
while(variant) {
struct lp_fragment_shader_variant *next = variant->next;
unsigned i;
for (i = 0; i < Elements(variant->function); i++) {
if (variant->function[i]) {
if (variant->jit_function[i])
LLVMFreeMachineCodeForFunction(screen->engine,
variant->function[i]);
LLVMDeleteFunction(variant->function[i]);
}
}
FREE(variant);
variant = next;
}
FREE((void *) shader->base.tokens);
FREE(shader);
}
/* llvalue -> ExecutionEngine.t -> unit */
CAMLprim value llvm_ee_free_machine_code(LLVMValueRef F,
LLVMExecutionEngineRef EE) {
LLVMFreeMachineCodeForFunction(EE, F);
return Val_unit;
}
ALIGN_STACK
static boolean
test_one(unsigned verbose,
FILE *fp,
const struct pipe_blend_state *blend,
enum vector_mode mode,
struct lp_type type)
{
LLVMModuleRef module = NULL;
LLVMValueRef func = NULL;
LLVMExecutionEngineRef engine = NULL;
LLVMModuleProviderRef provider = NULL;
LLVMPassManagerRef pass = NULL;
char *error = NULL;
blend_test_ptr_t blend_test_ptr;
boolean success;
const unsigned n = LP_TEST_NUM_SAMPLES;
int64_t cycles[LP_TEST_NUM_SAMPLES];
double cycles_avg = 0.0;
unsigned i, j;
if(verbose >= 1)
dump_blend_type(stdout, blend, mode, type);
module = LLVMModuleCreateWithName("test");
func = add_blend_test(module, blend, mode, type);
if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) {
LLVMDumpModule(module);
abort();
}
LLVMDisposeMessage(error);
provider = LLVMCreateModuleProviderForExistingModule(module);
if (LLVMCreateJITCompiler(&engine, provider, 1, &error)) {
if(verbose < 1)
dump_blend_type(stderr, blend, mode, type);
fprintf(stderr, "%s\n", error);
LLVMDisposeMessage(error);
abort();
}
#if 0
pass = LLVMCreatePassManager();
LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass);
/* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
* but there are more on SVN. */
LLVMAddConstantPropagationPass(pass);
LLVMAddInstructionCombiningPass(pass);
LLVMAddPromoteMemoryToRegisterPass(pass);
LLVMAddGVNPass(pass);
LLVMAddCFGSimplificationPass(pass);
LLVMRunPassManager(pass, module);
#else
(void)pass;
#endif
if(verbose >= 2)
LLVMDumpModule(module);
blend_test_ptr = (blend_test_ptr_t)LLVMGetPointerToGlobal(engine, func);
if(verbose >= 2)
lp_disassemble(blend_test_ptr);
success = TRUE;
for(i = 0; i < n && success; ++i) {
if(mode == AoS) {
ALIGN16_ATTRIB uint8_t src[LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t dst[LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t con[LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t res[LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t ref[LP_NATIVE_VECTOR_WIDTH/8];
int64_t start_counter = 0;
int64_t end_counter = 0;
random_vec(type, src);
random_vec(type, dst);
random_vec(type, con);
{
double fsrc[LP_MAX_VECTOR_LENGTH];
double fdst[LP_MAX_VECTOR_LENGTH];
double fcon[LP_MAX_VECTOR_LENGTH];
double fref[LP_MAX_VECTOR_LENGTH];
read_vec(type, src, fsrc);
read_vec(type, dst, fdst);
read_vec(type, con, fcon);
for(j = 0; j < type.length; j += 4)
compute_blend_ref(blend, fsrc + j, fdst + j, fcon + j, fref + j);
write_vec(type, ref, fref);
}
start_counter = rdtsc();
blend_test_ptr(src, dst, con, res);
end_counter = rdtsc();
cycles[i] = end_counter - start_counter;
if(!compare_vec(type, res, ref)) {
success = FALSE;
if(verbose < 1)
dump_blend_type(stderr, blend, mode, type);
fprintf(stderr, "MISMATCH\n");
fprintf(stderr, " Src: ");
dump_vec(stderr, type, src);
fprintf(stderr, "\n");
fprintf(stderr, " Dst: ");
dump_vec(stderr, type, dst);
fprintf(stderr, "\n");
fprintf(stderr, " Con: ");
dump_vec(stderr, type, con);
fprintf(stderr, "\n");
fprintf(stderr, " Res: ");
dump_vec(stderr, type, res);
fprintf(stderr, "\n");
fprintf(stderr, " Ref: ");
dump_vec(stderr, type, ref);
fprintf(stderr, "\n");
}
}
if(mode == SoA) {
const unsigned stride = type.length*type.width/8;
ALIGN16_ATTRIB uint8_t src[4*LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t dst[4*LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t con[4*LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t res[4*LP_NATIVE_VECTOR_WIDTH/8];
ALIGN16_ATTRIB uint8_t ref[4*LP_NATIVE_VECTOR_WIDTH/8];
int64_t start_counter = 0;
int64_t end_counter = 0;
boolean mismatch;
for(j = 0; j < 4; ++j) {
random_vec(type, src + j*stride);
random_vec(type, dst + j*stride);
random_vec(type, con + j*stride);
}
{
double fsrc[4];
double fdst[4];
double fcon[4];
double fref[4];
unsigned k;
for(k = 0; k < type.length; ++k) {
for(j = 0; j < 4; ++j) {
fsrc[j] = read_elem(type, src + j*stride, k);
fdst[j] = read_elem(type, dst + j*stride, k);
fcon[j] = read_elem(type, con + j*stride, k);
}
compute_blend_ref(blend, fsrc, fdst, fcon, fref);
for(j = 0; j < 4; ++j)
write_elem(type, ref + j*stride, k, fref[j]);
}
}
start_counter = rdtsc();
blend_test_ptr(src, dst, con, res);
end_counter = rdtsc();
cycles[i] = end_counter - start_counter;
mismatch = FALSE;
for (j = 0; j < 4; ++j)
if(!compare_vec(type, res + j*stride, ref + j*stride))
mismatch = TRUE;
if (mismatch) {
success = FALSE;
if(verbose < 1)
dump_blend_type(stderr, blend, mode, type);
fprintf(stderr, "MISMATCH\n");
for(j = 0; j < 4; ++j) {
char channel = "RGBA"[j];
fprintf(stderr, " Src%c: ", channel);
dump_vec(stderr, type, src + j*stride);
fprintf(stderr, "\n");
fprintf(stderr, " Dst%c: ", channel);
dump_vec(stderr, type, dst + j*stride);
fprintf(stderr, "\n");
fprintf(stderr, " Con%c: ", channel);
dump_vec(stderr, type, con + j*stride);
fprintf(stderr, "\n");
fprintf(stderr, " Res%c: ", channel);
dump_vec(stderr, type, res + j*stride);
fprintf(stderr, "\n");
fprintf(stderr, " Ref%c: ", channel);
dump_vec(stderr, type, ref + j*stride);
fprintf(stderr, "\n");
}
}
}
}
/*
* Unfortunately the output of cycle counter is not very reliable as it comes
* -- sometimes we get outliers (due IRQs perhaps?) which are
* better removed to avoid random or biased data.
*/
{
double sum = 0.0, sum2 = 0.0;
double avg, std;
unsigned m;
for(i = 0; i < n; ++i) {
sum += cycles[i];
sum2 += cycles[i]*cycles[i];
}
avg = sum/n;
std = sqrtf((sum2 - n*avg*avg)/n);
m = 0;
sum = 0.0;
for(i = 0; i < n; ++i) {
if(fabs(cycles[i] - avg) <= 4.0*std) {
sum += cycles[i];
++m;
}
}
cycles_avg = sum/m;
}
if(fp)
write_tsv_row(fp, blend, mode, type, cycles_avg, success);
if (!success) {
if(verbose < 2)
LLVMDumpModule(module);
LLVMWriteBitcodeToFile(module, "blend.bc");
fprintf(stderr, "blend.bc written\n");
fprintf(stderr, "Invoke as \"llc -o - blend.bc\"\n");
abort();
}
LLVMFreeMachineCodeForFunction(engine, func);
LLVMDisposeExecutionEngine(engine);
if(pass)
LLVMDisposePassManager(pass);
return success;
}
PIPE_ALIGN_STACK
static boolean
test_round(unsigned verbose, FILE *fp)
{
LLVMModuleRef module = NULL;
LLVMValueRef test_round = NULL, test_trunc, test_floor, test_ceil;
LLVMExecutionEngineRef engine = lp_build_engine;
LLVMPassManagerRef pass = NULL;
char *error = NULL;
test_round_t round_func, trunc_func, floor_func, ceil_func;
float unpacked[4];
unsigned packed;
boolean success = TRUE;
int i;
module = LLVMModuleCreateWithName("test");
test_round = add_test(module, "round", lp_build_round);
test_trunc = add_test(module, "trunc", lp_build_trunc);
test_floor = add_test(module, "floor", lp_build_floor);
test_ceil = add_test(module, "ceil", lp_build_ceil);
if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) {
printf("LLVMVerifyModule: %s\n", error);
LLVMDumpModule(module);
abort();
}
LLVMDisposeMessage(error);
#if 0
pass = LLVMCreatePassManager();
LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass);
/* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
* but there are more on SVN. */
LLVMAddConstantPropagationPass(pass);
LLVMAddInstructionCombiningPass(pass);
LLVMAddPromoteMemoryToRegisterPass(pass);
LLVMAddGVNPass(pass);
LLVMAddCFGSimplificationPass(pass);
LLVMRunPassManager(pass, module);
#else
(void)pass;
#endif
round_func = (test_round_t) pointer_to_func(LLVMGetPointerToGlobal(engine, test_round));
trunc_func = (test_round_t) pointer_to_func(LLVMGetPointerToGlobal(engine, test_trunc));
floor_func = (test_round_t) pointer_to_func(LLVMGetPointerToGlobal(engine, test_floor));
ceil_func = (test_round_t) pointer_to_func(LLVMGetPointerToGlobal(engine, test_ceil));
memset(unpacked, 0, sizeof unpacked);
packed = 0;
if (0)
LLVMDumpModule(module);
for (i = 0; i < 3; i++) {
v4sf xvals[3] = {
{-10.0, -1, 0, 12.0},
{-1.5, -0.25, 1.25, 2.5},
{-0.99, -0.01, 0.01, 0.99}
};
v4sf x = xvals[i];
v4sf y, ref;
float *xp = (float *) &x;
float *refp = (float *) &ref;
printf("\n");
printv("x ", x);
refp[0] = round(xp[0]);
refp[1] = round(xp[1]);
refp[2] = round(xp[2]);
refp[3] = round(xp[3]);
y = round_func(x);
printv("C round(x) ", ref);
printv("LLVM round(x)", y);
compare(ref, y);
refp[0] = trunc(xp[0]);
refp[1] = trunc(xp[1]);
refp[2] = trunc(xp[2]);
refp[3] = trunc(xp[3]);
y = trunc_func(x);
printv("C trunc(x) ", ref);
printv("LLVM trunc(x)", y);
compare(ref, y);
refp[0] = floor(xp[0]);
refp[1] = floor(xp[1]);
refp[2] = floor(xp[2]);
refp[3] = floor(xp[3]);
y = floor_func(x);
printv("C floor(x) ", ref);
printv("LLVM floor(x)", y);
compare(ref, y);
refp[0] = ceil(xp[0]);
refp[1] = ceil(xp[1]);
refp[2] = ceil(xp[2]);
refp[3] = ceil(xp[3]);
y = ceil_func(x);
printv("C ceil(x) ", ref);
printv("LLVM ceil(x) ", y);
compare(ref, y);
}
LLVMFreeMachineCodeForFunction(engine, test_round);
LLVMFreeMachineCodeForFunction(engine, test_trunc);
LLVMFreeMachineCodeForFunction(engine, test_floor);
LLVMFreeMachineCodeForFunction(engine, test_ceil);
LLVMDisposeExecutionEngine(engine);
if(pass)
LLVMDisposePassManager(pass);
return success;
}
PIPE_ALIGN_STACK
static boolean
test_one(unsigned verbose,
FILE *fp,
struct lp_type src_type,
struct lp_type dst_type)
{
LLVMModuleRef module = NULL;
LLVMValueRef func = NULL;
LLVMExecutionEngineRef engine = NULL;
LLVMModuleProviderRef provider = NULL;
LLVMPassManagerRef pass = NULL;
char *error = NULL;
conv_test_ptr_t conv_test_ptr;
boolean success;
const unsigned n = LP_TEST_NUM_SAMPLES;
int64_t cycles[LP_TEST_NUM_SAMPLES];
double cycles_avg = 0.0;
unsigned num_srcs;
unsigned num_dsts;
double eps;
unsigned i, j;
if(verbose >= 1)
dump_conv_types(stdout, src_type, dst_type);
if(src_type.length > dst_type.length) {
num_srcs = 1;
num_dsts = src_type.length/dst_type.length;
}
else {
num_dsts = 1;
num_srcs = dst_type.length/src_type.length;
}
assert(src_type.width * src_type.length == dst_type.width * dst_type.length);
/* We must not loose or gain channels. Only precision */
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
eps = MAX2(lp_const_eps(src_type), lp_const_eps(dst_type));
module = LLVMModuleCreateWithName("test");
func = add_conv_test(module, src_type, num_srcs, dst_type, num_dsts);
if(LLVMVerifyModule(module, LLVMPrintMessageAction, &error)) {
LLVMDumpModule(module);
abort();
}
LLVMDisposeMessage(error);
provider = LLVMCreateModuleProviderForExistingModule(module);
if (LLVMCreateJITCompiler(&engine, provider, 1, &error)) {
if(verbose < 1)
dump_conv_types(stderr, src_type, dst_type);
fprintf(stderr, "%s\n", error);
LLVMDisposeMessage(error);
abort();
}
#if 0
pass = LLVMCreatePassManager();
LLVMAddTargetData(LLVMGetExecutionEngineTargetData(engine), pass);
/* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
* but there are more on SVN. */
LLVMAddConstantPropagationPass(pass);
LLVMAddInstructionCombiningPass(pass);
LLVMAddPromoteMemoryToRegisterPass(pass);
LLVMAddGVNPass(pass);
LLVMAddCFGSimplificationPass(pass);
LLVMRunPassManager(pass, module);
#else
(void)pass;
#endif
if(verbose >= 2)
LLVMDumpModule(module);
conv_test_ptr = (conv_test_ptr_t)LLVMGetPointerToGlobal(engine, func);
if(verbose >= 2)
lp_disassemble(conv_test_ptr);
success = TRUE;
for(i = 0; i < n && success; ++i) {
unsigned src_stride = src_type.length*src_type.width/8;
unsigned dst_stride = dst_type.length*dst_type.width/8;
PIPE_ALIGN_VAR(16) uint8_t src[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH];
PIPE_ALIGN_VAR(16) uint8_t dst[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH];
double fref[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH];
uint8_t ref[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH];
int64_t start_counter = 0;
int64_t end_counter = 0;
for(j = 0; j < num_srcs; ++j) {
random_vec(src_type, src + j*src_stride);
read_vec(src_type, src + j*src_stride, fref + j*src_type.length);
}
for(j = 0; j < num_dsts; ++j) {
write_vec(dst_type, ref + j*dst_stride, fref + j*dst_type.length);
}
start_counter = rdtsc();
conv_test_ptr(src, dst);
end_counter = rdtsc();
cycles[i] = end_counter - start_counter;
for(j = 0; j < num_dsts; ++j) {
if(!compare_vec_with_eps(dst_type, dst + j*dst_stride, ref + j*dst_stride, eps))
success = FALSE;
}
if (!success) {
if(verbose < 1)
dump_conv_types(stderr, src_type, dst_type);
fprintf(stderr, "MISMATCH\n");
for(j = 0; j < num_srcs; ++j) {
fprintf(stderr, " Src%u: ", j);
dump_vec(stderr, src_type, src + j*src_stride);
fprintf(stderr, "\n");
}
#if 1
fprintf(stderr, " Ref: ");
for(j = 0; j < src_type.length*num_srcs; ++j)
fprintf(stderr, " %f", fref[j]);
fprintf(stderr, "\n");
#endif
for(j = 0; j < num_dsts; ++j) {
fprintf(stderr, " Dst%u: ", j);
dump_vec(stderr, dst_type, dst + j*dst_stride);
fprintf(stderr, "\n");
fprintf(stderr, " Ref%u: ", j);
dump_vec(stderr, dst_type, ref + j*dst_stride);
fprintf(stderr, "\n");
}
}
}
/*
* Unfortunately the output of cycle counter is not very reliable as it comes
* -- sometimes we get outliers (due IRQs perhaps?) which are
* better removed to avoid random or biased data.
*/
{
double sum = 0.0, sum2 = 0.0;
double avg, std;
unsigned m;
for(i = 0; i < n; ++i) {
sum += cycles[i];
sum2 += cycles[i]*cycles[i];
}
avg = sum/n;
std = sqrtf((sum2 - n*avg*avg)/n);
m = 0;
sum = 0.0;
for(i = 0; i < n; ++i) {
if(fabs(cycles[i] - avg) <= 4.0*std) {
sum += cycles[i];
++m;
}
}
cycles_avg = sum/m;
}
if(fp)
write_tsv_row(fp, src_type, dst_type, cycles_avg, success);
if (!success) {
static boolean firsttime = TRUE;
if(firsttime) {
if(verbose < 2)
LLVMDumpModule(module);
LLVMWriteBitcodeToFile(module, "conv.bc");
fprintf(stderr, "conv.bc written\n");
fprintf(stderr, "Invoke as \"llc -o - conv.bc\"\n");
firsttime = FALSE;
/* abort(); */
}
}
LLVMFreeMachineCodeForFunction(engine, func);
LLVMDisposeExecutionEngine(engine);
if(pass)
LLVMDisposePassManager(pass);
return success;
}
