/**
* Generates LLVM IR to call debug_printf.
*/
static LLVMValueRef
lp_build_print_args(struct gallivm_state* gallivm,
int argcount,
LLVMValueRef* args)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMContextRef context = gallivm->context;
LLVMValueRef func_printf;
LLVMTypeRef printf_type;
int i;
assert(args);
assert(argcount > 0);
assert(LLVMTypeOf(args[0]) == LLVMPointerType(LLVMInt8TypeInContext(context), 0));
/* Cast any float arguments to doubles as printf expects */
for (i = 1; i < argcount; i++) {
LLVMTypeRef type = LLVMTypeOf(args[i]);
if (LLVMGetTypeKind(type) == LLVMFloatTypeKind)
args[i] = LLVMBuildFPExt(builder, args[i], LLVMDoubleTypeInContext(context), "");
}
printf_type = LLVMFunctionType(LLVMInt32TypeInContext(context), NULL, 0, 1);
func_printf = lp_build_const_int_pointer(gallivm, func_to_pointer((func_pointer)debug_printf));
func_printf = LLVMBuildBitCast(builder, func_printf, LLVMPointerType(printf_type, 0), "debug_printf");
return LLVMBuildCall(builder, func_printf, args, argcount, "");
}
/**
* Fetch a pixel into a 4 float AoS.
*
* \param format_desc describes format of the image we're fetching from
* \param ptr address of the pixel block (or the texel if uncompressed)
* \param i, j the sub-block pixel coordinates. For non-compressed formats
* these will always be (0, 0).
* \return a 4 element vector with the pixel's RGBA values.
*/
LLVMValueRef
lp_build_fetch_rgba_aos(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type type,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j)
{
LLVMBuilderRef builder = gallivm->builder;
unsigned num_pixels = type.length / 4;
struct lp_build_context bld;
assert(type.length <= LP_MAX_VECTOR_LENGTH);
assert(type.length % 4 == 0);
lp_build_context_init(&bld, gallivm, type);
/*
* Trivial case
*
* The format matches the type (apart of a swizzle) so no need for
* scaling or converting.
*/
if (format_matches_type(format_desc, type) &&
format_desc->block.bits <= type.width * 4 &&
util_is_power_of_two(format_desc->block.bits)) {
LLVMValueRef packed;
/*
* The format matches the type (apart of a swizzle) so no need for
* scaling or converting.
*/
packed = lp_build_gather(gallivm, type.length/4,
format_desc->block.bits, type.width*4,
base_ptr, offset);
assert(format_desc->block.bits <= type.width * type.length);
packed = LLVMBuildBitCast(gallivm->builder, packed,
lp_build_vec_type(gallivm, type), "");
return lp_build_format_swizzle_aos(format_desc, &bld, packed);
}
/*
* Bit arithmetic
*/
if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
format_desc->block.width == 1 &&
format_desc->block.height == 1 &&
util_is_power_of_two(format_desc->block.bits) &&
format_desc->block.bits <= 32 &&
format_desc->is_bitmask &&
!format_desc->is_mixed &&
(format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED ||
format_desc->channel[1].type == UTIL_FORMAT_TYPE_UNSIGNED)) {
LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
LLVMValueRef res;
unsigned k;
/*
* Unpack a pixel at a time into a <4 x float> RGBA vector
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef packed;
packed = lp_build_gather_elem(gallivm, num_pixels,
format_desc->block.bits, 32,
base_ptr, offset, k);
tmps[k] = lp_build_unpack_arith_rgba_aos(gallivm,
format_desc,
packed);
}
/*
* Type conversion.
*
* TODO: We could avoid floating conversion for integer to
* integer conversions.
*/
if (gallivm_debug & GALLIVM_DEBUG_PERF && !type.floating) {
debug_printf("%s: unpacking %s with floating point\n",
__FUNCTION__, format_desc->short_name);
}
lp_build_conv(gallivm,
lp_float32_vec4_type(),
type,
tmps, num_pixels, &res, 1);
return lp_build_format_swizzle_aos(format_desc, &bld, res);
}
/*
* YUV / subsampled formats
*/
if (format_desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
struct lp_type tmp_type;
LLVMValueRef tmp;
memset(&tmp_type, 0, sizeof tmp_type);
tmp_type.width = 8;
tmp_type.length = num_pixels * 4;
tmp_type.norm = TRUE;
tmp = lp_build_fetch_subsampled_rgba_aos(gallivm,
format_desc,
num_pixels,
base_ptr,
offset,
i, j);
lp_build_conv(gallivm,
tmp_type, type,
&tmp, 1, &tmp, 1);
return tmp;
}
/*
* Fallback to util_format_description::fetch_rgba_8unorm().
*/
if (format_desc->fetch_rgba_8unorm &&
!type.floating && type.width == 8 && !type.sign && type.norm) {
/*
* Fallback to calling util_format_description::fetch_rgba_8unorm.
*
* This is definitely not the most efficient way of fetching pixels, as
* we miss the opportunity to do vectorization, but this it is a
* convenient for formats or scenarios for which there was no opportunity
* or incentive to optimize.
*/
LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
char name[256];
LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
LLVMTypeRef pi8t = LLVMPointerType(i8t, 0);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef function;
LLVMValueRef tmp_ptr;
LLVMValueRef tmp;
LLVMValueRef res;
LLVMValueRef callee;
unsigned k;
util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_8unorm",
format_desc->short_name);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: falling back to %s\n", __FUNCTION__, name);
}
/*
* Declare and bind format_desc->fetch_rgba_8unorm().
*/
function = LLVMGetNamedFunction(module, name);
if (!function) {
/*
* Function to call looks like:
* fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j)
*/
LLVMTypeRef ret_type;
LLVMTypeRef arg_types[4];
LLVMTypeRef function_type;
ret_type = LLVMVoidTypeInContext(gallivm->context);
arg_types[0] = pi8t;
arg_types[1] = pi8t;
arg_types[2] = i32t;
arg_types[3] = i32t;
function_type = LLVMFunctionType(ret_type, arg_types,
Elements(arg_types), 0);
function = LLVMAddFunction(module, name, function_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
LLVMSetLinkage(function, LLVMExternalLinkage);
assert(LLVMIsDeclaration(function));
}
/* make const pointer for the C fetch_rgba_float function */
callee = lp_build_const_int_pointer(gallivm,
func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm));
/* cast the callee pointer to the function's type */
function = LLVMBuildBitCast(builder, callee,
LLVMTypeOf(function), "cast callee");
tmp_ptr = lp_build_alloca(gallivm, i32t, "");
res = LLVMGetUndef(LLVMVectorType(i32t, num_pixels));
/*
* Invoke format_desc->fetch_rgba_8unorm() for each pixel and insert the result
* in the SoA vectors.
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef index = lp_build_const_int32(gallivm, k);
LLVMValueRef args[4];
args[0] = LLVMBuildBitCast(builder, tmp_ptr, pi8t, "");
args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
base_ptr, offset, k);
if (num_pixels == 1) {
args[2] = i;
args[3] = j;
}
else {
args[2] = LLVMBuildExtractElement(builder, i, index, "");
args[3] = LLVMBuildExtractElement(builder, j, index, "");
}
LLVMBuildCall(builder, function, args, Elements(args), "");
tmp = LLVMBuildLoad(builder, tmp_ptr, "");
if (num_pixels == 1) {
res = tmp;
}
else {
res = LLVMBuildInsertElement(builder, res, tmp, index, "");
}
}
/* Bitcast from <n x i32> to <4n x i8> */
res = LLVMBuildBitCast(builder, res, bld.vec_type, "");
return res;
}
/*
* Fallback to util_format_description::fetch_rgba_float().
*/
if (format_desc->fetch_rgba_float) {
/*
* Fallback to calling util_format_description::fetch_rgba_float.
*
* This is definitely not the most efficient way of fetching pixels, as
* we miss the opportunity to do vectorization, but this it is a
* convenient for formats or scenarios for which there was no opportunity
* or incentive to optimize.
*/
LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)));
char name[256];
LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context);
LLVMTypeRef f32x4t = LLVMVectorType(f32t, 4);
LLVMTypeRef pf32t = LLVMPointerType(f32t, 0);
LLVMTypeRef pi8t = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef function;
LLVMValueRef tmp_ptr;
LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
LLVMValueRef res;
LLVMValueRef callee;
unsigned k;
util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_float",
format_desc->short_name);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: falling back to %s\n", __FUNCTION__, name);
}
/*
* Declare and bind format_desc->fetch_rgba_float().
*/
function = LLVMGetNamedFunction(module, name);
if (!function) {
/*
* Function to call looks like:
* fetch(float *dst, const uint8_t *src, unsigned i, unsigned j)
*/
LLVMTypeRef ret_type;
LLVMTypeRef arg_types[4];
LLVMTypeRef function_type;
ret_type = LLVMVoidTypeInContext(gallivm->context);
arg_types[0] = pf32t;
arg_types[1] = pi8t;
arg_types[2] = i32t;
arg_types[3] = i32t;
function_type = LLVMFunctionType(ret_type, arg_types,
Elements(arg_types), 0);
function = LLVMAddFunction(module, name, function_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
LLVMSetLinkage(function, LLVMExternalLinkage);
assert(LLVMIsDeclaration(function));
}
/* Note: we're using this casting here instead of LLVMAddGlobalMapping()
* to work around a bug in LLVM 2.6.
*/
/* make const pointer for the C fetch_rgba_float function */
callee = lp_build_const_int_pointer(gallivm,
func_to_pointer((func_pointer) format_desc->fetch_rgba_float));
/* cast the callee pointer to the function's type */
function = LLVMBuildBitCast(builder, callee,
LLVMTypeOf(function), "cast callee");
tmp_ptr = lp_build_alloca(gallivm, f32x4t, "");
/*
* Invoke format_desc->fetch_rgba_float() for each pixel and insert the result
* in the SoA vectors.
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef args[4];
args[0] = LLVMBuildBitCast(builder, tmp_ptr, pf32t, "");
args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
base_ptr, offset, k);
if (num_pixels == 1) {
args[2] = i;
args[3] = j;
}
else {
LLVMValueRef index = lp_build_const_int32(gallivm, k);
args[2] = LLVMBuildExtractElement(builder, i, index, "");
args[3] = LLVMBuildExtractElement(builder, j, index, "");
}
LLVMBuildCall(builder, function, args, Elements(args), "");
tmps[k] = LLVMBuildLoad(builder, tmp_ptr, "");
}
lp_build_conv(gallivm,
lp_float32_vec4_type(),
type,
tmps, num_pixels, &res, 1);
return res;
}
assert(0);
return lp_build_undef(gallivm, type);
}
static void
update_cached_block(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
LLVMValueRef ptr_addr,
LLVMValueRef hash_index,
LLVMValueRef cache)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
LLVMTypeRef pi8t = LLVMPointerType(i8t, 0);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef i32x4 = LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4);
LLVMValueRef function;
LLVMValueRef tag_value, tmp_ptr;
LLVMValueRef col[4];
unsigned i, j;
/*
* Use format_desc->fetch_rgba_8unorm() for each pixel in the block.
* This doesn't actually make any sense whatsoever, someone would need
* to write a function doing this for all pixels in a block (either as
* an external c function or with generated code). Don't ask.
*/
{
/*
* Function to call looks like:
* fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j)
*/
LLVMTypeRef ret_type;
LLVMTypeRef arg_types[4];
LLVMTypeRef function_type;
assert(format_desc->fetch_rgba_8unorm);
ret_type = LLVMVoidTypeInContext(gallivm->context);
arg_types[0] = pi8t;
arg_types[1] = pi8t;
arg_types[2] = i32t;
arg_types[3] = i32t;
function_type = LLVMFunctionType(ret_type, arg_types,
ARRAY_SIZE(arg_types), 0);
/* make const pointer for the C fetch_rgba_8unorm function */
function = lp_build_const_int_pointer(gallivm,
func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm));
/* cast the callee pointer to the function's type */
function = LLVMBuildBitCast(builder, function,
LLVMPointerType(function_type, 0),
"cast callee");
}
tmp_ptr = lp_build_array_alloca(gallivm, i32x4,
lp_build_const_int32(gallivm, 16),
"tmp_decode_store");
tmp_ptr = LLVMBuildBitCast(builder, tmp_ptr, pi8t, "");
/*
* Invoke format_desc->fetch_rgba_8unorm() for each pixel.
* This is going to be really really slow.
* Note: the block store format is actually
* x0y0x0y1x0y2x0y3 x1y0x1y1x1y2x1y3 ...
*/
for (i = 0; i < 4; ++i) {
for (j = 0; j < 4; ++j) {
LLVMValueRef args[4];
LLVMValueRef dst_offset = lp_build_const_int32(gallivm, (i * 4 + j) * 4);
/*
* Note we actually supply a pointer to the start of the block,
* not the start of the texture.
*/
args[0] = LLVMBuildGEP(gallivm->builder, tmp_ptr, &dst_offset, 1, "");
args[1] = ptr_addr;
args[2] = LLVMConstInt(i32t, i, 0);
args[3] = LLVMConstInt(i32t, j, 0);
LLVMBuildCall(builder, function, args, ARRAY_SIZE(args), "");
}
}
/* Finally store the block - pointless mem copy + update tag. */
tmp_ptr = LLVMBuildBitCast(builder, tmp_ptr, LLVMPointerType(i32x4, 0), "");
for (i = 0; i < 4; ++i) {
LLVMValueRef tmp_offset = lp_build_const_int32(gallivm, i);
LLVMValueRef ptr = LLVMBuildGEP(gallivm->builder, tmp_ptr, &tmp_offset, 1, "");
col[i] = LLVMBuildLoad(builder, ptr, "");
}
tag_value = LLVMBuildPtrToInt(gallivm->builder, ptr_addr,
LLVMInt64TypeInContext(gallivm->context), "");
store_cached_block(gallivm, col, tag_value, hash_index, cache);
}
/**
* Converts float32 to int16 half-float
* Note this can be performed in 1 instruction if vcvtps2ph exists (f16c/cvt16)
* [llvm.x86.vcvtps2ph / _mm_cvtps_ph]
*
* @param src value to convert
*
* Convert float32 to half floats, preserving Infs and NaNs,
* with rounding towards zero (trunc).
*/
LLVMValueRef
lp_build_float_to_half(struct gallivm_state *gallivm,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef f32_vec_type = LLVMTypeOf(src);
unsigned length = LLVMGetTypeKind(f32_vec_type) == LLVMVectorTypeKind
? LLVMGetVectorSize(f32_vec_type) : 1;
struct lp_type i32_type = lp_type_int_vec(32, 32 * length);
struct lp_type i16_type = lp_type_int_vec(16, 16 * length);
LLVMValueRef result;
if (util_cpu_caps.has_f16c && HAVE_LLVM >= 0x0301 &&
(length == 4 || length == 8)) {
struct lp_type i168_type = lp_type_int_vec(16, 16 * 8);
unsigned mode = 3; /* same as LP_BUILD_ROUND_TRUNCATE */
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
const char *intrinsic = NULL;
if (length == 4) {
intrinsic = "llvm.x86.vcvtps2ph.128";
}
else {
intrinsic = "llvm.x86.vcvtps2ph.256";
}
result = lp_build_intrinsic_binary(builder, intrinsic,
lp_build_vec_type(gallivm, i168_type),
src, LLVMConstInt(i32t, mode, 0));
if (length == 4) {
result = lp_build_extract_range(gallivm, result, 0, 4);
}
}
else {
result = lp_build_float_to_smallfloat(gallivm, i32_type, src, 10, 5, 0, true);
/* Convert int32 vector to int16 vector by trunc (might generate bad code) */
result = LLVMBuildTrunc(builder, result, lp_build_vec_type(gallivm, i16_type), "");
}
/*
* Debugging code.
*/
if (0) {
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef i16t = LLVMInt16TypeInContext(gallivm->context);
LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context);
LLVMValueRef ref_result = LLVMGetUndef(LLVMVectorType(i16t, length));
unsigned i;
LLVMTypeRef func_type = LLVMFunctionType(i16t, &f32t, 1, 0);
LLVMValueRef func = lp_build_const_int_pointer(gallivm, func_to_pointer((func_pointer)util_float_to_half));
func = LLVMBuildBitCast(builder, func, LLVMPointerType(func_type, 0), "util_float_to_half");
for (i = 0; i < length; ++i) {
LLVMValueRef index = LLVMConstInt(i32t, i, 0);
LLVMValueRef f32 = LLVMBuildExtractElement(builder, src, index, "");
#if 0
/* XXX: not really supported by backends */
LLVMValueRef f16 = lp_build_intrinsic_unary(builder, "llvm.convert.to.fp16", i16t, f32);
#else
LLVMValueRef f16 = LLVMBuildCall(builder, func, &f32, 1, "");
#endif
ref_result = LLVMBuildInsertElement(builder, ref_result, f16, index, "");
}
lp_build_print_value(gallivm, "src = ", src);
lp_build_print_value(gallivm, "llvm = ", result);
lp_build_print_value(gallivm, "util = ", ref_result);
lp_build_printf(gallivm, "\n");
}
return result;
}
