void evergreen_delete_compute_state(struct pipe_context *ctx_, void* state)
{
struct r600_context *ctx = (struct r600_context *)ctx_;
COMPUTE_DBG(ctx->screen, "*** evergreen_delete_compute_state\n");
struct r600_pipe_compute *shader = state;
if (!shader)
return;
#ifdef HAVE_OPENCL
#if HAVE_LLVM < 0x0306
for (unsigned i = 0; i < shader->num_kernels; i++) {
struct r600_kernel *kernel = &shader->kernels[i];
LLVMDisposeModule(module);
}
FREE(shader->kernels);
LLVMContextDispose(shader->llvm_ctx);
#else
radeon_shader_binary_clean(&shader->binary);
r600_destroy_shader(&shader->bc);
/* TODO destroy shader->code_bo, shader->const_bo
* we'll need something like r600_buffer_free */
#endif
#endif
FREE(shader);
}
/**
* Free gallivm object's LLVM allocations, but not the gallivm object itself.
*/
static void
free_gallivm_state(struct gallivm_state *gallivm)
{
#if HAVE_LLVM >= 0x207 /* XXX or 0x208? */
/* This leads to crashes w/ some versions of LLVM */
LLVMModuleRef mod;
char *error;
if (gallivm->engine && gallivm->provider)
LLVMRemoveModuleProvider(gallivm->engine, gallivm->provider,
&mod, &error);
#endif
if (gallivm->passmgr) {
LLVMDisposePassManager(gallivm->passmgr);
}
#if 0
/* XXX this seems to crash with all versions of LLVM */
if (gallivm->provider)
LLVMDisposeModuleProvider(gallivm->provider);
#endif
if (HAVE_LLVM >= 0x207 && gallivm->engine) {
/* This will already destroy any associated module */
LLVMDisposeExecutionEngine(gallivm->engine);
} else {
LLVMDisposeModule(gallivm->module);
}
#if !USE_MCJIT
/* Don't free the TargetData, it's owned by the exec engine */
#else
if (gallivm->target) {
LLVMDisposeTargetData(gallivm->target);
}
#endif
/* Never free the LLVM context.
*/
#if 0
if (gallivm->context)
LLVMContextDispose(gallivm->context);
#endif
if (gallivm->builder)
LLVMDisposeBuilder(gallivm->builder);
gallivm->engine = NULL;
gallivm->target = NULL;
gallivm->module = NULL;
gallivm->provider = NULL;
gallivm->passmgr = NULL;
gallivm->context = NULL;
gallivm->builder = NULL;
}
static void codegen_cleanup(compile_t* c)
{
while(c->frame != NULL)
pop_frame(c);
LLVMDisposeBuilder(c->builder);
LLVMDisposeModule(c->module);
LLVMContextDispose(c->context);
LLVMDisposeTargetMachine(c->machine);
reach_free(c->reachable);
}
/**
* Free gallivm object's LLVM allocations, but not the gallivm object itself.
*/
static void
free_gallivm_state(struct gallivm_state *gallivm)
{
#if HAVE_LLVM >= 0x207 /* XXX or 0x208? */
/* This leads to crashes w/ some versions of LLVM */
LLVMModuleRef mod;
char *error;
if (gallivm->engine && gallivm->provider)
LLVMRemoveModuleProvider(gallivm->engine, gallivm->provider,
&mod, &error);
#endif
#if 0
/* XXX this seems to crash with all versions of LLVM */
if (gallivm->provider)
LLVMDisposeModuleProvider(gallivm->provider);
#endif
if (gallivm->passmgr)
LLVMDisposePassManager(gallivm->passmgr);
#if HAVE_LLVM >= 0x207
if (gallivm->module)
LLVMDisposeModule(gallivm->module);
#endif
#if 0
/* Don't free the exec engine, it's a global/singleton */
if (gallivm->engine)
LLVMDisposeExecutionEngine(gallivm->engine);
#endif
#if 0
/* Don't free the TargetData, it's owned by the exec engine */
LLVMDisposeTargetData(gallivm->target);
#endif
if (gallivm->context)
LLVMContextDispose(gallivm->context);
if (gallivm->builder)
LLVMDisposeBuilder(gallivm->builder);
gallivm->engine = NULL;
gallivm->target = NULL;
gallivm->module = NULL;
gallivm->provider = NULL;
gallivm->passmgr = NULL;
gallivm->context = NULL;
gallivm->builder = NULL;
}
static void codegen_cleanup(compile_t* c)
{
while(c->frame != NULL)
pop_frame(c);
LLVMDIBuilderDestroy(c->di);
LLVMDisposeBuilder(c->builder);
LLVMDisposeModule(c->module);
LLVMContextDispose(c->context);
LLVMDisposeTargetMachine(c->machine);
tbaa_metadatas_free(c->tbaa_mds);
reach_free(c->reach);
}
/*
* machine_finish
*
* Context cleanup.
*/
void
machine_finish (machinedef_t *mach)
{
machine_ctx_t m;
if (mach == 0) return;
m = machine_context(mach);
if (m == 0) return;
if (m->target_machine != 0) LLVMDisposeTargetMachine(m->target_machine);
if (m->llvmctx != 0) LLVMContextDispose(m->llvmctx);
if (m->outfile != 0) free(m->outfile);
free(m);
} /* machine_finish */
void evergreen_delete_compute_state(struct pipe_context *ctx, void* state)
{
struct r600_pipe_compute *shader = (struct r600_pipe_compute *)state;
if (!shader)
return;
FREE(shader->kernels);
#ifdef HAVE_OPENCL
if (shader->llvm_ctx){
LLVMContextDispose(shader->llvm_ctx);
}
#endif
FREE(shader);
}
PIPE_ALIGN_STACK
static boolean
test_one(unsigned verbose,
FILE *fp,
struct lp_type src_type,
struct lp_type dst_type)
{
LLVMContextRef context;
struct gallivm_state *gallivm;
LLVMValueRef func = 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 ((src_type.width >= dst_type.width && src_type.length > dst_type.length) ||
(src_type.width <= dst_type.width && src_type.length < dst_type.length)) {
return TRUE;
}
/* Known failures
* - fixed point 32 -> float 32
* - float 32 -> signed normalised integer 32
*/
if ((src_type.floating && !dst_type.floating && dst_type.sign && dst_type.norm && src_type.width == dst_type.width) ||
(!src_type.floating && dst_type.floating && src_type.fixed && src_type.width == dst_type.width)) {
return TRUE;
}
/* Known failures
* - fixed point 32 -> float 32
* - float 32 -> signed normalised integer 32
*/
if ((src_type.floating && !dst_type.floating && dst_type.sign && dst_type.norm && src_type.width == dst_type.width) ||
(!src_type.floating && dst_type.floating && src_type.fixed && src_type.width == dst_type.width)) {
return TRUE;
}
if(verbose >= 1)
dump_conv_types(stderr, src_type, dst_type);
if (src_type.length > dst_type.length) {
num_srcs = 1;
num_dsts = src_type.length/dst_type.length;
}
else if (src_type.length < dst_type.length) {
num_dsts = 1;
num_srcs = dst_type.length/src_type.length;
}
else {
num_dsts = 1;
num_srcs = 1;
}
/* 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));
if (dst_type.norm && dst_type.sign && src_type.sign && !src_type.floating) {
/*
* This is quite inaccurate due to shift being used.
* I don't think it's possible to hit such conversions with
* llvmpipe though.
*/
eps *= 2;
}
context = LLVMContextCreate();
gallivm = gallivm_create("test_module", context);
func = add_conv_test(gallivm, src_type, num_srcs, dst_type, num_dsts);
gallivm_compile_module(gallivm);
conv_test_ptr = (conv_test_ptr_t)gallivm_jit_function(gallivm, func);
gallivm_free_ir(gallivm);
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(LP_MIN_VECTOR_ALIGN) uint8_t src[LP_MAX_VECTOR_LENGTH*LP_MAX_VECTOR_LENGTH];
PIPE_ALIGN_VAR(LP_MIN_VECTOR_ALIGN) 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 || verbose >= 3) {
if(verbose < 1)
dump_conv_types(stderr, src_type, dst_type);
if (success) {
fprintf(stderr, "PASS\n");
}
else {
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);
gallivm_destroy(gallivm);
LLVMContextDispose(context);
return success;
}
PIPE_ALIGN_STACK
static boolean
test_one(unsigned verbose,
FILE *fp,
const struct pipe_blend_state *blend,
struct lp_type type)
{
LLVMContextRef context;
struct gallivm_state *gallivm;
LLVMValueRef func = 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;
const unsigned stride = lp_type_width(type)/8;
if(verbose >= 1)
dump_blend_type(stdout, blend, type);
context = LLVMContextCreate();
gallivm = gallivm_create("test_module", context);
func = add_blend_test(gallivm, blend, type);
gallivm_compile_module(gallivm);
blend_test_ptr = (blend_test_ptr_t)gallivm_jit_function(gallivm, func);
gallivm_free_ir(gallivm);
success = TRUE;
{
uint8_t *src, *src1, *dst, *con, *res, *ref;
src = align_malloc(stride, stride);
src1 = align_malloc(stride, stride);
dst = align_malloc(stride, stride);
con = align_malloc(stride, stride);
res = align_malloc(stride, stride);
ref = align_malloc(stride, stride);
for(i = 0; i < n && success; ++i) {
int64_t start_counter = 0;
int64_t end_counter = 0;
random_vec(type, src);
random_vec(type, src1);
random_vec(type, dst);
random_vec(type, con);
{
double fsrc[LP_MAX_VECTOR_LENGTH];
double fsrc1[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, src1, fsrc1);
read_vec(type, dst, fdst);
read_vec(type, con, fcon);
for(j = 0; j < type.length; j += 4)
compute_blend_ref(blend, fsrc + j, fsrc1 + j, fdst + j, fcon + j, fref + j);
write_vec(type, ref, fref);
}
start_counter = rdtsc();
blend_test_ptr(src, src1, 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, type);
fprintf(stderr, "MISMATCH\n");
fprintf(stderr, " Src: ");
dump_vec(stderr, type, src);
fprintf(stderr, "\n");
fprintf(stderr, " Src1: ");
dump_vec(stderr, type, src1);
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");
}
}
align_free(src);
align_free(src1);
align_free(dst);
align_free(con);
align_free(res);
align_free(ref);
}
/*
* 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, type, cycles_avg, success);
gallivm_destroy(gallivm);
LLVMContextDispose(context);
return success;
}