/**
* Converts int16 half-float to float32
* Note this can be performed in 1 instruction if vcvtph2ps exists (f16c/cvt16)
* [llvm.x86.vcvtph2ps / _mm_cvtph_ps]
*
* @param src value to convert
*
*/
LLVMValueRef
lp_build_half_to_float(struct gallivm_state *gallivm,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef src_type = LLVMTypeOf(src);
unsigned src_length = LLVMGetTypeKind(src_type) == LLVMVectorTypeKind ?
LLVMGetVectorSize(src_type) : 1;
struct lp_type f32_type = lp_type_float_vec(32, 32 * src_length);
struct lp_type i32_type = lp_type_int_vec(32, 32 * src_length);
LLVMTypeRef int_vec_type = lp_build_vec_type(gallivm, i32_type);
LLVMValueRef h;
if (util_cpu_caps.has_f16c && HAVE_LLVM >= 0x0301 &&
(src_length == 4 || src_length == 8)) {
const char *intrinsic = NULL;
if (src_length == 4) {
src = lp_build_pad_vector(gallivm, src, 8);
intrinsic = "llvm.x86.vcvtph2ps.128";
}
else {
intrinsic = "llvm.x86.vcvtph2ps.256";
}
return lp_build_intrinsic_unary(builder, intrinsic,
lp_build_vec_type(gallivm, f32_type), src);
}
/* Convert int16 vector to int32 vector by zero ext (might generate bad code) */
h = LLVMBuildZExt(builder, src, int_vec_type, "");
return lp_build_smallfloat_to_float(gallivm, f32_type, h, 10, 5, 0, true);
}
/**
* Convert srgb int values to linear float values.
* Several possibilities how to do this, e.g.
* - table
* - doing the pow() with int-to-float and float-to-int tricks
* (http://stackoverflow.com/questions/6475373/optimizations-for-pow-with-const-non-integer-exponent)
* - just using standard polynomial approximation
* (3rd order polynomial is required for crappy but just sufficient accuracy)
*
* @param src integer (vector) value(s) to convert
* (chan_bits bit values unpacked to 32 bit already).
*/
LLVMValueRef
lp_build_srgb_to_linear(struct gallivm_state *gallivm,
struct lp_type src_type,
unsigned chan_bits,
LLVMValueRef src)
{
struct lp_type f32_type = lp_type_float_vec(32, src_type.length * 32);
struct lp_build_context f32_bld;
LLVMValueRef srcf, part_lin, part_pow, is_linear, lin_const, lin_thresh;
double coeffs[4] = {0.0023f,
0.0030f / 255.0f,
0.6935f / (255.0f * 255.0f),
0.3012f / (255.0f * 255.0f * 255.0f)
};
assert(src_type.width == 32);
/* Technically this would work with more bits too but would be inaccurate. */
assert(chan_bits <= 8);
lp_build_context_init(&f32_bld, gallivm, f32_type);
/*
* using polynomial: (src * (src * (src * 0.3012 + 0.6935) + 0.0030) + 0.0023)
* ( poly = 0.3012*x^3 + 0.6935*x^2 + 0.0030*x + 0.0023)
* (found with octave polyfit and some magic as I couldn't get the error
* function right). Using the above mentioned error function, the values stay
* within +-0.35, except for the lowest values - hence tweaking linear segment
* to cover the first 16 instead of the first 11 values (the error stays
* just about acceptable there too).
* Hence: lin = src > 15 ? poly : src / 12.6
* This function really only makes sense for vectors, should use LUT otherwise.
* All in all (including float conversion) 11 instructions (with sse4.1),
* 6 constants (polynomial could be done with 1 instruction less at the cost
* of slightly worse dependency chain, fma should also help).
*/
/* doing the 1/255 mul as part of the approximation */
srcf = lp_build_int_to_float(&f32_bld, src);
if (chan_bits != 8) {
/* could adjust all the constants instead */
LLVMValueRef rescale_const = lp_build_const_vec(gallivm, f32_type,
255.0f / ((1 << chan_bits) - 1));
srcf = lp_build_mul(&f32_bld, srcf, rescale_const);
}
lin_const = lp_build_const_vec(gallivm, f32_type, 1.0f / (12.6f * 255.0f));
part_lin = lp_build_mul(&f32_bld, srcf, lin_const);
part_pow = lp_build_polynomial(&f32_bld, srcf, coeffs, 4);
lin_thresh = lp_build_const_vec(gallivm, f32_type, 15.0f);
is_linear = lp_build_compare(gallivm, f32_type, PIPE_FUNC_LEQUAL, srcf, lin_thresh);
return lp_build_select(&f32_bld, is_linear, part_lin, part_pow);
}
/**
* Converts int16 half-float to float32
* Note this can be performed in 1 instruction if vcvtph2ps exists (sse5 i think?)
* [llvm.x86.vcvtph2ps / _mm_cvtph_ps]
*
* @param src_type <vector> type of int16
* @param src value to convert
*
* ref http://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/
*/
LLVMValueRef
lp_build_half_to_float(struct gallivm_state *gallivm,
struct lp_type src_type,
LLVMValueRef src)
{
struct lp_type f32_type = lp_type_float_vec(32, 32 * src_type.length);
struct lp_type i32_type = lp_type_int_vec(32, 32 * src_type.length);
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_vec_type(gallivm, i32_type);
LLVMTypeRef float_vec_type = lp_build_vec_type(gallivm, f32_type);
/* Constants */
LLVMValueRef i32_13 = lp_build_const_int_vec(gallivm, i32_type, 13);
LLVMValueRef i32_16 = lp_build_const_int_vec(gallivm, i32_type, 16);
LLVMValueRef i32_mask_nosign = lp_build_const_int_vec(gallivm, i32_type, 0x7fff);
LLVMValueRef i32_was_infnan = lp_build_const_int_vec(gallivm, i32_type, 0x7bff);
LLVMValueRef i32_exp_infnan = lp_build_const_int_vec(gallivm, i32_type, 0xff << 23);
LLVMValueRef f32_magic = LLVMBuildBitCast(builder,
lp_build_const_int_vec(gallivm, i32_type, (254 - 15) << 23),
float_vec_type, "");
/* Convert int16 vector to int32 vector by zero ext */
LLVMValueRef h = LLVMBuildZExt(builder, src, int_vec_type, "");
/* Exponent / mantissa bits */
LLVMValueRef expmant = LLVMBuildAnd(builder, i32_mask_nosign, h, "");
LLVMValueRef shifted = LLVMBuildBitCast(builder, LLVMBuildShl(builder, expmant, i32_13, ""), float_vec_type, "");
/* Exponent adjust */
LLVMValueRef scaled = LLVMBuildBitCast(builder, LLVMBuildFMul(builder, shifted, f32_magic, ""), int_vec_type, "");
/* Make sure Inf/NaN survive */
LLVMValueRef b_wasinfnan = lp_build_compare(gallivm, i32_type, PIPE_FUNC_GREATER, expmant, i32_was_infnan);
LLVMValueRef infnanexp = LLVMBuildAnd(builder, b_wasinfnan, i32_exp_infnan, "");
/* Sign bit */
LLVMValueRef justsign = LLVMBuildXor(builder, h, expmant, "");
LLVMValueRef sign = LLVMBuildShl(builder, justsign, i32_16, "");
/* Combine result */
LLVMValueRef sign_inf = LLVMBuildOr(builder, sign, infnanexp, "");
LLVMValueRef final = LLVMBuildOr(builder, scaled, sign_inf, "");
/* Cast from int32 vector to float32 vector */
return LLVMBuildBitCast(builder, final, float_vec_type, "");
}
/*
* Build LLVM function that exercises the unary operator builder.
*/
static LLVMValueRef
build_unary_test_func(struct gallivm_state *gallivm,
const struct unary_test_t *test)
{
struct lp_type type = lp_type_float_vec(32, lp_native_vector_width);
LLVMContextRef context = gallivm->context;
LLVMModuleRef module = gallivm->module;
LLVMTypeRef vf32t = lp_build_vec_type(gallivm, type);
LLVMTypeRef args[2] = { LLVMPointerType(vf32t, 0), LLVMPointerType(vf32t, 0) };
LLVMValueRef func = LLVMAddFunction(module, test->name,
LLVMFunctionType(LLVMVoidTypeInContext(context),
args, Elements(args), 0));
LLVMValueRef arg0 = LLVMGetParam(func, 0);
LLVMValueRef arg1 = LLVMGetParam(func, 1);
LLVMBuilderRef builder = gallivm->builder;
LLVMBasicBlockRef block = LLVMAppendBasicBlockInContext(context, func, "entry");
LLVMValueRef ret;
struct lp_build_context bld;
lp_build_context_init(&bld, gallivm, type);
LLVMSetFunctionCallConv(func, LLVMCCallConv);
LLVMPositionBuilderAtEnd(builder, block);
arg1 = LLVMBuildLoad(builder, arg1, "");
ret = test->builder(&bld, arg1);
LLVMBuildStore(builder, ret, arg0);
LLVMBuildRetVoid(builder);
gallivm_verify_function(gallivm, func);
return func;
}
/**
* 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;
LLVMTypeRef dst_vec_type = lp_build_vec_type(gallivm, type);
unsigned vec_len = type.width * type.length;
/*
* 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 <= vec_len);
packed = LLVMBuildBitCast(gallivm->builder, packed, dst_vec_type, "");
#ifdef PIPE_ARCH_BIG_ENDIAN
if (type.floating)
packed = lp_build_bswap_vec(gallivm, packed, type,
lp_type_float_vec(type.width, vec_len));
#endif
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) &&
!format_desc->channel[0].pure_integer) {
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);
}
/* If all channels are of same type and we are not using half-floats */
if (format_desc->is_array &&
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB) {
assert(!format_desc->is_mixed);
return lp_build_fetch_rgba_aos_array(gallivm, format_desc, type, base_ptr, offset);
}
/*
* 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.
*/
LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
LLVMTypeRef pi8t = LLVMPointerType(i8t, 0);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef function;
LLVMValueRef tmp_ptr;
LLVMValueRef tmp;
LLVMValueRef res;
unsigned k;
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: falling back to util_format_%s_fetch_rgba_8unorm\n",
__FUNCTION__, format_desc->short_name);
}
/*
* Declare and bind format_desc->fetch_rgba_8unorm().
*/
{
/*
* 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);
/* 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_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.
*/
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;
unsigned k;
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: falling back to util_format_%s_fetch_rgba_float\n",
__FUNCTION__, format_desc->short_name);
}
/*
* Declare and bind format_desc->fetch_rgba_float().
*/
{
/*
* Function to call looks like:
* fetch(float *dst, const uint8_t *src, unsigned i, unsigned j)
*/
LLVMTypeRef ret_type;
LLVMTypeRef arg_types[4];
ret_type = LLVMVoidTypeInContext(gallivm->context);
arg_types[0] = pf32t;
arg_types[1] = pi8t;
arg_types[2] = i32t;
arg_types[3] = i32t;
function = lp_build_const_func_pointer(gallivm,
func_to_pointer((func_pointer) format_desc->fetch_rgba_float),
ret_type,
arg_types, Elements(arg_types),
format_desc->short_name);
}
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(!util_format_is_pure_integer(format_desc->format));
assert(0);
return lp_build_undef(gallivm, type);
}
/**
* Gather elements from scatter positions in memory into a single vector.
* Use for fetching texels from a texture.
* For SSE, typical values are length=4, src_width=32, dst_width=32.
*
* When src_width < dst_width, the return value can be justified in
* one of two ways:
* "integer justification" is used when the caller treats the destination
* as a packed integer bitmask, as described by the channels' "shift" and
* "width" fields;
* "vector justification" is used when the caller casts the destination
* to a vector and needs channel X to be in vector element 0.
*
* @param length length of the offsets
* @param src_width src element width in bits
* @param dst_type result element type (src will be expanded to fit,
* but truncation is not allowed)
* (this may be a vector, must be pot sized)
* @param aligned whether the data is guaranteed to be aligned (to src_width)
* @param base_ptr base pointer, needs to be a i8 pointer type.
* @param offsets vector with offsets
* @param vector_justify select vector rather than integer justification
*/
LLVMValueRef
lp_build_gather(struct gallivm_state *gallivm,
unsigned length,
unsigned src_width,
struct lp_type dst_type,
boolean aligned,
LLVMValueRef base_ptr,
LLVMValueRef offsets,
boolean vector_justify)
{
LLVMValueRef res;
boolean need_expansion = src_width < dst_type.width * dst_type.length;
boolean vec_fetch;
struct lp_type fetch_type, fetch_dst_type;
LLVMTypeRef src_type;
assert(src_width <= dst_type.width * dst_type.length);
/*
* This is quite a mess...
* Figure out if the fetch should be done as:
* a) scalar or vector
* b) float or int
*
* As an example, for a 96bit fetch expanded into 4x32bit, it is better
* to use (3x32bit) vector type (then pad the vector). Otherwise, the
* zext will cause extra instructions.
* However, the same isn't true for 3x16bit (the codegen for that is
* completely worthless on x86 simd, and for 3x8bit is is way worse
* still, don't try that... (To get really good code out of llvm for
* these cases, the only way is to decompose the fetches manually
* into 1x32bit/1x16bit, or 1x16/1x8bit respectively, although the latter
* case requires sse41, otherwise simple scalar zext is way better.
* But probably not important enough, so don't bother.)
* Also, we try to honor the floating bit of destination (but isn't
* possible if caller asks for instance for 2x32bit dst_type with
* 48bit fetch - the idea would be to use 3x16bit fetch, pad and
* cast to 2x32f type, so the fetch is always int and on top of that
* we avoid the vec pad and use scalar zext due the above mentioned
* issue).
* Note this is optimized for x86 sse2 and up backend. Could be tweaked
* for other archs if necessary...
*/
if (((src_width % 32) == 0) && ((src_width % dst_type.width) == 0) &&
(dst_type.length > 1)) {
/* use vector fetch (if dst_type is vector) */
vec_fetch = TRUE;
if (dst_type.floating) {
fetch_type = lp_type_float_vec(dst_type.width, src_width);
} else {
fetch_type = lp_type_int_vec(dst_type.width, src_width);
}
/* intentionally not using lp_build_vec_type here */
src_type = LLVMVectorType(lp_build_elem_type(gallivm, fetch_type),
fetch_type.length);
fetch_dst_type = fetch_type;
fetch_dst_type.length = dst_type.length;
} else {
/* use scalar fetch */
vec_fetch = FALSE;
if (dst_type.floating && ((src_width == 32) || (src_width == 64))) {
fetch_type = lp_type_float(src_width);
} else {
fetch_type = lp_type_int(src_width);
}
src_type = lp_build_vec_type(gallivm, fetch_type);
fetch_dst_type = fetch_type;
fetch_dst_type.width = dst_type.width * dst_type.length;
}
if (length == 1) {
/* Scalar */
res = lp_build_gather_elem_vec(gallivm, length,
src_width, src_type, fetch_dst_type,
aligned, base_ptr, offsets, 0,
vector_justify);
return LLVMBuildBitCast(gallivm->builder, res,
lp_build_vec_type(gallivm, dst_type), "");
/*
* Excluding expansion from these paths because if you need it for
* 32bit/64bit fetches you're doing it wrong (this is gather, not
* conversion) and it would be awkward for floats.
*/
} else if (util_cpu_caps.has_avx2 && !need_expansion &&
src_width == 32 && (length == 4 || length == 8)) {
return lp_build_gather_avx2(gallivm, length, src_width, dst_type,
base_ptr, offsets);
/*
* This looks bad on paper wrt throughtput/latency on Haswell.
* Even on Broadwell it doesn't look stellar.
* Albeit no measurements were done (but tested to work).
* Should definitely enable on Skylake.
* (In general, should be more of a win if the fetch is 256bit wide -
* this is true for the 32bit case above too.)
*/
} else if (0 && util_cpu_caps.has_avx2 && !need_expansion &&
src_width == 64 && (length == 2 || length == 4)) {
return lp_build_gather_avx2(gallivm, length, src_width, dst_type,
base_ptr, offsets);
} else {
/* Vector */
LLVMValueRef elems[LP_MAX_VECTOR_WIDTH / 8];
unsigned i;
boolean vec_zext = FALSE;
struct lp_type res_type, gather_res_type;
LLVMTypeRef res_t, gather_res_t;
res_type = fetch_dst_type;
res_type.length *= length;
gather_res_type = res_type;
if (src_width == 16 && dst_type.width == 32 && dst_type.length == 1) {
/*
* Note that llvm is never able to optimize zext/insert combos
* directly (i.e. zero the simd reg, then place the elements into
* the appropriate place directly). (I think this has to do with
* scalar/vector transition.) And scalar 16->32bit zext simd loads
* aren't possible (instead loading to scalar reg first).
* No idea about other archs...
* We could do this manually, but instead we just use a vector
* zext, which is simple enough (and, in fact, llvm might optimize
* this away).
* (We're not trying that with other bit widths as that might not be
* easier, in particular with 8 bit values at least with only sse2.)
*/
assert(vec_fetch == FALSE);
gather_res_type.width /= 2;
fetch_dst_type = fetch_type;
src_type = lp_build_vec_type(gallivm, fetch_type);
vec_zext = TRUE;
}
res_t = lp_build_vec_type(gallivm, res_type);
gather_res_t = lp_build_vec_type(gallivm, gather_res_type);
res = LLVMGetUndef(gather_res_t);
for (i = 0; i < length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
elems[i] = lp_build_gather_elem_vec(gallivm, length,
src_width, src_type, fetch_dst_type,
aligned, base_ptr, offsets, i,
vector_justify);
if (!vec_fetch) {
res = LLVMBuildInsertElement(gallivm->builder, res, elems[i], index, "");
}
}
if (vec_zext) {
res = LLVMBuildZExt(gallivm->builder, res, res_t, "");
if (vector_justify) {
#ifdef PIPE_ARCH_BIG_ENDIAN
unsigned sv = dst_type.width - src_width;
res = LLVMBuildShl(gallivm->builder, res,
lp_build_const_int_vec(gallivm, res_type, sv), "");
#endif
}
}
if (vec_fetch) {
/*
* Do bitcast now otherwise llvm might get some funny ideas wrt
* float/int types...
*/
for (i = 0; i < length; i++) {
elems[i] = LLVMBuildBitCast(gallivm->builder, elems[i],
lp_build_vec_type(gallivm, dst_type), "");
}
res = lp_build_concat(gallivm, elems, dst_type, length);
} else {
struct lp_type really_final_type = dst_type;
assert(res_type.length * res_type.width ==
dst_type.length * dst_type.width * length);
really_final_type.length *= length;
res = LLVMBuildBitCast(gallivm->builder, res,
lp_build_vec_type(gallivm, really_final_type), "");
}
}
return res;
}
/**
* Fetch a texels from a texture, returning them in SoA layout.
*
* \param type the desired return type for 'rgba'. The vector length
* is the number of texels to fetch
* \param aligned if the offset is guaranteed to be aligned to element width
*
* \param base_ptr points to the base of the texture mip tree.
* \param offset offset to start of the texture image block. For non-
* compressed formats, this simply is an offset to the texel.
* For compressed formats, it is an offset to the start of the
* compressed data block.
*
* \param i, j the sub-block pixel coordinates. For non-compressed formats
* these will always be (0,0). For compressed formats, i will
* be in [0, block_width-1] and j will be in [0, block_height-1].
* \param cache optional value pointing to a lp_build_format_cache structure
*/
void
lp_build_fetch_rgba_soa(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type type,
boolean aligned,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j,
LLVMValueRef cache,
LLVMValueRef rgba_out[4])
{
LLVMBuilderRef builder = gallivm->builder;
enum pipe_format format = format_desc->format;
struct lp_type fetch_type;
if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
format_desc->block.width == 1 &&
format_desc->block.height == 1 &&
format_desc->block.bits <= type.width &&
(format_desc->channel[0].type != UTIL_FORMAT_TYPE_FLOAT ||
format_desc->channel[0].size == 32 ||
format_desc->channel[0].size == 16))
{
/*
* The packed pixel fits into an element of the destination format. Put
* the packed pixels into a vector and extract each component for all
* vector elements in parallel.
*/
LLVMValueRef packed;
/*
* gather the texels from the texture
* Ex: packed = {XYZW, XYZW, XYZW, XYZW}
*/
assert(format_desc->block.bits <= type.width);
fetch_type = lp_type_uint(type.width);
packed = lp_build_gather(gallivm,
type.length,
format_desc->block.bits,
fetch_type,
aligned,
base_ptr, offset, FALSE);
/*
* convert texels to float rgba
*/
lp_build_unpack_rgba_soa(gallivm,
format_desc,
type,
packed, rgba_out);
return;
}
if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB) &&
format_desc->block.width == 1 &&
format_desc->block.height == 1 &&
format_desc->block.bits > type.width &&
((format_desc->block.bits <= type.width * type.length &&
format_desc->channel[0].size <= type.width) ||
(format_desc->channel[0].size == 64 &&
format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT &&
type.floating)))
{
/*
* Similar to above, but the packed pixel is larger than what fits
* into an element of the destination format. The packed pixels will be
* shuffled into SoA vectors appropriately, and then the extraction will
* be done in parallel as much as possible.
* Good for 16xn (n > 2) and 32xn (n > 1) formats, care is taken so
* the gathered vectors can be shuffled easily (even with avx).
* 64xn float -> 32xn float is handled too but it's a bit special as
* it does the conversion pre-shuffle.
*/
LLVMValueRef packed[4], dst[4], output[4], shuffles[LP_MAX_VECTOR_WIDTH/32];
struct lp_type fetch_type, gather_type = type;
unsigned num_gather, fetch_width, i, j;
struct lp_build_context bld;
boolean fp64 = format_desc->channel[0].size == 64;
lp_build_context_init(&bld, gallivm, type);
assert(type.width == 32);
assert(format_desc->block.bits > type.width);
/*
* First, figure out fetch order.
*/
fetch_width = util_next_power_of_two(format_desc->block.bits);
num_gather = fetch_width / type.width;
/*
* fp64 are treated like fp32 except we fetch twice wide values
* (as we shuffle after trunc). The shuffles for that work out
* mostly fine (slightly suboptimal for 4-wide, perfect for AVX)
* albeit we miss the potential opportunity for hw gather (as it
* only handles native size).
*/
num_gather = fetch_width / type.width;
gather_type.width *= num_gather;
if (fp64) {
num_gather /= 2;
}
gather_type.length /= num_gather;
for (i = 0; i < num_gather; i++) {
LLVMValueRef offsetr, shuf_vec;
if(num_gather == 4) {
for (j = 0; j < gather_type.length; j++) {
unsigned idx = i + 4*j;
shuffles[j] = lp_build_const_int32(gallivm, idx);
}
shuf_vec = LLVMConstVector(shuffles, gather_type.length);
offsetr = LLVMBuildShuffleVector(builder, offset, offset, shuf_vec, "");
}
else if (num_gather == 2) {
assert(num_gather == 2);
for (j = 0; j < gather_type.length; j++) {
unsigned idx = i*2 + (j%2) + (j/2)*4;
shuffles[j] = lp_build_const_int32(gallivm, idx);
}
shuf_vec = LLVMConstVector(shuffles, gather_type.length);
offsetr = LLVMBuildShuffleVector(builder, offset, offset, shuf_vec, "");
}
else {
assert(num_gather == 1);
offsetr = offset;
}
if (gather_type.length == 1) {
LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
offsetr = LLVMBuildExtractElement(builder, offsetr, zero, "");
}
/*
* Determine whether to use float or int loads. This is mostly
* to outsmart the (stupid) llvm int/float shuffle logic, we
* don't really care much if the data is floats or ints...
* But llvm will refuse to use single float shuffle with int data
* and instead use 3 int shuffles instead, the code looks atrocious.
* (Note bitcasts often won't help, as llvm is too smart to be
* fooled by that.)
* Nobody cares about simd float<->int domain transition penalties,
* which usually don't even exist for shuffles anyway.
* With 4x32bit (and 3x32bit) fetch, we use float vec (the data is
* going into transpose, which is unpacks, so doesn't really matter
* much).
* With 2x32bit or 4x16bit fetch, we use float vec, since those
* go into the weird channel separation shuffle. With floats,
* this is (with 128bit vectors):
* - 2 movq, 2 movhpd, 2 shufps
* With ints it would be:
* - 4 movq, 2 punpcklqdq, 4 pshufd, 2 blendw
* I've seen texture functions increase in code size by 15% just due
* to that (there's lots of such fetches in them...)
* (We could chose a different gather order to improve this somewhat
* for the int path, but it would basically just drop the blends,
* so the float path with this order really is optimal.)
* Albeit it is tricky sometimes llvm doesn't ignore the float->int
* casts so must avoid them until we're done with the float shuffle...
* 3x16bit formats (the same is also true for 3x8) are pretty bad but
* there's nothing we can do about them (we could overallocate by
* those couple bytes and use unaligned but pot sized load).
* Note that this is very much x86 specific. I don't know if this
* affect other archs at all.
*/
if (num_gather > 1) {
/*
* We always want some float type here (with x86)
* due to shuffles being float ones afterwards (albeit for
* the num_gather == 4 case int should work fine too
* (unless there's some problems with avx but not avx2).
*/
if (format_desc->channel[0].size == 64) {
fetch_type = lp_type_float_vec(64, gather_type.width);
} else {
fetch_type = lp_type_int_vec(32, gather_type.width);
}
}
else {
/* type doesn't matter much */
if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT &&
(format_desc->channel[0].size == 32 ||
format_desc->channel[0].size == 64)) {
fetch_type = lp_type_float(gather_type.width);
} else {
fetch_type = lp_type_uint(gather_type.width);
}
}
/* Now finally gather the values */
packed[i] = lp_build_gather(gallivm, gather_type.length,
format_desc->block.bits,
fetch_type, aligned,
base_ptr, offsetr, FALSE);
if (fp64) {
struct lp_type conv_type = type;
conv_type.width *= 2;
packed[i] = LLVMBuildBitCast(builder, packed[i],
lp_build_vec_type(gallivm, conv_type), "");
packed[i] = LLVMBuildFPTrunc(builder, packed[i], bld.vec_type, "");
}
}
/* shuffle the gathered values to SoA */
if (num_gather == 2) {
for (i = 0; i < num_gather; i++) {
for (j = 0; j < type.length; j++) {
unsigned idx = (j%2)*2 + (j/4)*4 + i;
if ((j/2)%2)
idx += type.length;
shuffles[j] = lp_build_const_int32(gallivm, idx);
}
dst[i] = LLVMBuildShuffleVector(builder, packed[0], packed[1],
LLVMConstVector(shuffles, type.length), "");
}
}
else if (num_gather == 4) {
lp_build_transpose_aos(gallivm, lp_int_type(type), packed, dst);
}
else {
assert(num_gather == 1);
dst[0] = packed[0];
}
/*
* And finally unpack exactly as above, except that
* chan shift is adjusted and the right vector selected.
*/
if (!fp64) {
for (i = 0; i < num_gather; i++) {
dst[i] = LLVMBuildBitCast(builder, dst[i], bld.int_vec_type, "");
}
for (i = 0; i < format_desc->nr_channels; i++) {
struct util_format_channel_description chan_desc = format_desc->channel[i];
unsigned blockbits = type.width;
unsigned vec_nr = chan_desc.shift / type.width;
chan_desc.shift %= type.width;
output[i] = lp_build_extract_soa_chan(&bld,
blockbits,
FALSE,
chan_desc,
dst[vec_nr]);
}
}
else {
for (i = 0; i < format_desc->nr_channels; i++) {
output[i] = dst[i];
}
}
lp_build_format_swizzle_soa(format_desc, &bld, output, rgba_out);
return;
}
if (format == PIPE_FORMAT_R11G11B10_FLOAT ||
format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
/*
* similar conceptually to above but requiring special
* AoS packed -> SoA float conversion code.
*/
LLVMValueRef packed;
struct lp_type fetch_type = lp_type_uint(type.width);
assert(type.floating);
assert(type.width == 32);
packed = lp_build_gather(gallivm, type.length,
format_desc->block.bits,
fetch_type, aligned,
base_ptr, offset, FALSE);
if (format == PIPE_FORMAT_R11G11B10_FLOAT) {
lp_build_r11g11b10_to_float(gallivm, packed, rgba_out);
}
else {
lp_build_rgb9e5_to_float(gallivm, packed, rgba_out);
}
return;
}
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS &&
format_desc->block.bits == 64) {
/*
* special case the format is 64 bits but we only require
* 32bit (or 8bit) from each block.
*/
LLVMValueRef packed;
struct lp_type fetch_type = lp_type_uint(type.width);
if (format == PIPE_FORMAT_X32_S8X24_UINT) {
/*
* for stencil simply fix up offsets - could in fact change
* base_ptr instead even outside the shader.
*/
unsigned mask = (1 << 8) - 1;
LLVMValueRef s_offset = lp_build_const_int_vec(gallivm, type, 4);
offset = LLVMBuildAdd(builder, offset, s_offset, "");
packed = lp_build_gather(gallivm, type.length, 32, fetch_type,
aligned, base_ptr, offset, FALSE);
packed = LLVMBuildAnd(builder, packed,
lp_build_const_int_vec(gallivm, type, mask), "");
}
else {
assert (format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT);
packed = lp_build_gather(gallivm, type.length, 32, fetch_type,
aligned, base_ptr, offset, TRUE);
packed = LLVMBuildBitCast(builder, packed,
lp_build_vec_type(gallivm, type), "");
}
/* for consistency with lp_build_unpack_rgba_soa() return sss1 or zzz1 */
rgba_out[0] = rgba_out[1] = rgba_out[2] = packed;
rgba_out[3] = lp_build_const_vec(gallivm, type, 1.0f);
return;
}
/*
* Try calling lp_build_fetch_rgba_aos for all pixels.
* Should only really hit subsampled, compressed
* (for s3tc srgb too, for rgtc the unorm ones only) by now.
* (This is invalid for plain 8unorm formats because we're lazy with
* the swizzle since some results would arrive swizzled, some not.)
*/
if ((format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN) &&
(util_format_fits_8unorm(format_desc) ||
format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) &&
type.floating && type.width == 32 &&
(type.length == 1 || (type.length % 4 == 0))) {
struct lp_type tmp_type;
struct lp_build_context bld;
LLVMValueRef packed, rgba[4];
const struct util_format_description *flinear_desc;
const struct util_format_description *frgba8_desc;
unsigned chan;
lp_build_context_init(&bld, gallivm, type);
/*
* Make sure the conversion in aos really only does convert to rgba8
* and not anything more (so use linear format, adjust type).
*/
flinear_desc = util_format_description(util_format_linear(format));
memset(&tmp_type, 0, sizeof tmp_type);
tmp_type.width = 8;
tmp_type.length = type.length * 4;
tmp_type.norm = TRUE;
packed = lp_build_fetch_rgba_aos(gallivm, flinear_desc, tmp_type,
aligned, base_ptr, offset, i, j, cache);
packed = LLVMBuildBitCast(builder, packed, bld.int_vec_type, "");
/*
* The values are now packed so they match ordinary (srgb) RGBA8 format,
* hence need to use matching format for unpack.
*/
frgba8_desc = util_format_description(PIPE_FORMAT_R8G8B8A8_UNORM);
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
assert(format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC);
frgba8_desc = util_format_description(PIPE_FORMAT_R8G8B8A8_SRGB);
}
lp_build_unpack_rgba_soa(gallivm,
frgba8_desc,
type,
packed, rgba);
/*
* We converted 4 channels. Make sure llvm can drop unneeded ones
* (luckily the rgba order is fixed, only LA needs special case).
*/
for (chan = 0; chan < 4; chan++) {
enum pipe_swizzle swizzle = format_desc->swizzle[chan];
if (chan == 3 && util_format_is_luminance_alpha(format)) {
swizzle = PIPE_SWIZZLE_W;
}
rgba_out[chan] = lp_build_swizzle_soa_channel(&bld, rgba, swizzle);
}
return;
}
/*
* Fallback to calling lp_build_fetch_rgba_aos for each pixel.
*
* This is not the most efficient way of fetching pixels, as we
* miss some opportunities to do vectorization, but this is
* convenient for formats or scenarios for which there was no
* opportunity or incentive to optimize.
*
* We do NOT want to end up here, this typically is quite terrible,
* in particular if the formats have less than 4 channels.
*
* Right now, this should only be hit for:
* - RGTC snorm formats
* (those miss fast fetch functions hence they are terrible anyway)
*/
{
unsigned k;
struct lp_type tmp_type;
LLVMValueRef aos_fetch[LP_MAX_VECTOR_WIDTH / 32];
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: AoS fetch fallback for %s\n",
__FUNCTION__, format_desc->short_name);
}
tmp_type = type;
tmp_type.length = 4;
/*
* Note that vector transpose can be worse compared to insert/extract
* for aos->soa conversion (for formats with 1 or 2 channels). However,
* we should try to avoid getting here for just about all formats, so
* don't bother.
*/
/* loop over number of pixels */
for(k = 0; k < type.length; ++k) {
LLVMValueRef index = lp_build_const_int32(gallivm, k);
LLVMValueRef offset_elem;
LLVMValueRef i_elem, j_elem;
offset_elem = LLVMBuildExtractElement(builder, offset,
index, "");
i_elem = LLVMBuildExtractElement(builder, i, index, "");
j_elem = LLVMBuildExtractElement(builder, j, index, "");
/* Get a single float[4]={R,G,B,A} pixel */
aos_fetch[k] = lp_build_fetch_rgba_aos(gallivm, format_desc, tmp_type,
aligned, base_ptr, offset_elem,
i_elem, j_elem, cache);
}
convert_to_soa(gallivm, aos_fetch, rgba_out, type);
}
}
