boolean compare_vec_with_eps(struct lp_type type, const void *res, const void *ref, double eps) { unsigned i; for (i = 0; i < type.length; ++i) { double res_elem = read_elem(type, res, i); double ref_elem = read_elem(type, ref, i); double delta = fabs(res_elem - ref_elem); if(delta >= 2.0*eps) return FALSE; } return TRUE; }
static awk_bool_t read_array(int fd, awk_array_t array) { uint32_t i; uint32_t count; awk_element_t new_elem; if (read(fd, & count, sizeof(count)) != sizeof(count)) { return awk_false; } count = ntohl(count); for (i = 0; i < count; i++) { if (read_elem(fd, & new_elem)) { /* add to array */ if (! set_array_element_by_elem(array, & new_elem)) { fprintf(stderr, _("read_array: set_array_element failed\n")); return awk_false; } } else break; } if (i != count) return awk_false; return awk_true; }
void read_vec(struct lp_type type, const void *src, double *dst) { unsigned i; for (i = 0; i < type.length; ++i) dst[i] = read_elem(type, src, i); }
static int read_array(int fd, NODE *array) { int i; uint32_t count; uint32_t array_sz; int index; NODE *new_elem; if (read(fd, & count, sizeof(count)) != sizeof(count)) { return -1; } array->table_size = ntohl(count); if (read(fd, & array_sz, sizeof(array_sz)) != sizeof(array_sz)) { return -1; } array->array_size = ntohl(array_sz); /* malloc var_array */ array->var_array = (NODE **) malloc(array->array_size * sizeof(NODE *)); memset(array->var_array, '\0', array->array_size * sizeof(NODE *)); for (i = 0; i < array->table_size; i++) { if ((new_elem = read_elem(fd, & index, array)) != NULL) { new_elem->ahnext = array->var_array[index]; array->var_array[index] = new_elem; } else break; } if (i != array->table_size) return -1; return 0; }
boolean compare_vec_with_eps(struct lp_type type, const void *res, const void *ref, double eps) { unsigned i; eps *= type.floating ? 8.0 : 2.0; for (i = 0; i < type.length; ++i) { double res_elem = read_elem(type, res, i); double ref_elem = read_elem(type, ref, i); double delta = res_elem - ref_elem; if (ref_elem < -1.0 || ref_elem > 1.0) { delta /= ref_elem; } delta = fabs(delta); if (delta >= eps) { return FALSE; } } return TRUE; }
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_one(unsigned verbose, FILE *fp, const struct pipe_blend_state *blend, enum vector_mode mode, struct lp_type type) { 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, mode, type); gallivm = gallivm_create(); func = add_blend_test(gallivm, blend, mode, type); gallivm_compile_module(gallivm); blend_test_ptr = (blend_test_ptr_t)gallivm_jit_function(gallivm, func); success = TRUE; if(mode == AoS) { uint8_t *src, *dst, *con, *res, *ref; src = 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, 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"); } } align_free(src); align_free(dst); align_free(con); align_free(res); align_free(ref); } else if(mode == SoA) { uint8_t *src, *dst, *con, *res, *ref; src = align_malloc(4*stride, stride); dst = align_malloc(4*stride, stride); con = align_malloc(4*stride, stride); res = align_malloc(4*stride, stride); ref = align_malloc(4*stride, stride); for(i = 0; i < n && success; ++i) { 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"); fprintf(stderr, "\n"); } } } align_free(src); 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, mode, type, cycles_avg, success); gallivm_free_function(gallivm, func, blend_test_ptr); gallivm_destroy(gallivm); return success; }