-/**
- * Byte swap on element. It will construct a call to intrinsic llvm.bswap
- * based on the type.
- *
- * @param res element to byte swap.
- * @param type int16_t, int32_t, int64_t, float or double
- * @param
- */
-LLVMValueRef
-lp_build_bswap(struct gallivm_state *gallivm,
- LLVMValueRef res,
- struct lp_type type)
-{
- LLVMTypeRef int_type = LLVMIntTypeInContext(gallivm->context,
- type.width);
- const char *intrinsic = NULL;
- if (type.width == 8)
- return res;
- if (type.width == 16)
- intrinsic = "llvm.bswap.i16";
- else if (type.width == 32)
- intrinsic = "llvm.bswap.i32";
- else if (type.width == 64)
- intrinsic = "llvm.bswap.i64";
-
- assert (intrinsic != NULL);
-
- /* In case of a floating-point type cast to a int of same size and then
- * cast back to fp type.
- */
- if (type.floating)
- res = LLVMBuildBitCast(gallivm->builder, res, int_type, "");
- res = lp_build_intrinsic_unary(gallivm->builder, intrinsic, int_type, res);
- if (type.floating)
- res = LLVMBuildBitCast(gallivm->builder, res,
- lp_build_elem_type(gallivm, type), "");
- return res;
-}
-
-
-/**
- * Byte swap every element in the vector.
- *
- * @param packed <vector> to convert
- * @param src_type <vector> type of int16_t, int32_t, int64_t, float or
- * double
- * @param dst_type <vector> type to return
- */
-LLVMValueRef
-lp_build_bswap_vec(struct gallivm_state *gallivm,
- LLVMValueRef packed,
- struct lp_type src_type_vec,
- struct lp_type dst_type_vec)
-{
- LLVMBuilderRef builder = gallivm->builder;
- LLVMTypeRef dst_type = lp_build_elem_type(gallivm, dst_type_vec);
- LLVMValueRef res;
-
- if (src_type_vec.length == 1) {
- res = lp_build_bswap(gallivm, packed, src_type_vec);
- res = LLVMBuildBitCast(gallivm->builder, res, dst_type, "");
- } else {
- unsigned i;
- res = LLVMGetUndef(lp_build_vec_type(gallivm, dst_type_vec));
- for (i = 0; i < src_type_vec.length; ++i) {
- LLVMValueRef index = lp_build_const_int32(gallivm, i);
- LLVMValueRef elem = LLVMBuildExtractElement(builder, packed, index, "");
- elem = lp_build_bswap(gallivm, elem, src_type_vec);
- elem = LLVMBuildBitCast(gallivm->builder, elem, dst_type, "");
- res = LLVMBuildInsertElement(gallivm->builder, res, elem, index, "");
- }
- }
- return res;
-}
-
-
/**
* Converts int16 half-float to float32
* Note this can be performed in 1 instruction if vcvtph2ps exists (f16c/cvt16)
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 &&
+ (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 */
- LLVMValueRef h = LLVMBuildZExt(builder, src, int_vec_type, "");
+ /* 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);
}
/**
* Converts float32 to int16 half-float
- * Note this can be performed in 1 instruction if vcvtps2ph exists (sse5 i think?)
+ * Note this can be performed in 1 instruction if vcvtps2ph exists (f16c/cvt16)
* [llvm.x86.vcvtps2ph / _mm_cvtps_ph]
*
* @param src value to convert
*
- * ref http://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/
- * ref https://gist.github.com/2156668
- *
- * XXX: This is an approximation. It is faster but certain NaNs are converted to
- * infinity, and rounding is not correct.
+ * Convert float32 to half floats, preserving Infs and NaNs,
+ * with rounding towards zero (trunc).
*/
LLVMValueRef
lp_build_float_to_half(struct gallivm_state *gallivm,
LLVMTypeRef f32_vec_type = LLVMTypeOf(src);
unsigned length = LLVMGetTypeKind(f32_vec_type) == LLVMVectorTypeKind
? LLVMGetVectorSize(f32_vec_type) : 1;
- struct lp_type f32_type = lp_type_float_vec(32, 32 * length);
- struct lp_type u32_type = lp_type_uint_vec(32, 32 * length);
+ struct lp_type i32_type = lp_type_int_vec(32, 32 * length);
struct lp_type i16_type = lp_type_int_vec(16, 16 * length);
- LLVMTypeRef u32_vec_type = lp_build_vec_type(gallivm, u32_type);
- LLVMTypeRef i16_vec_type = lp_build_vec_type(gallivm, i16_type);
- struct lp_build_context f32_bld;
- struct lp_build_context u32_bld;
LLVMValueRef result;
- lp_build_context_init(&f32_bld, gallivm, f32_type);
- lp_build_context_init(&u32_bld, gallivm, u32_type);
-
- {
- /* Constants */
- LLVMValueRef u32_f32inf = lp_build_const_int_vec(gallivm, u32_type, 0xff << 23);
- LLVMValueRef u32_expinf = lp_build_const_int_vec(gallivm, u32_type, 0xe0 << 23);
- LLVMValueRef f32_f16max = lp_build_const_vec(gallivm, f32_type, 65536.0); // 0x8f << 23
- LLVMValueRef f32_magic = lp_build_const_vec(gallivm, f32_type, 1.92592994e-34); // 0x0f << 23
-
- /* Cast from float32 to int32 */
- LLVMValueRef f = LLVMBuildBitCast(builder, src, u32_vec_type, "");
-
- /* Remove sign */
- LLVMValueRef srcabs = lp_build_abs(&f32_bld, src);
- LLVMValueRef fabs = LLVMBuildBitCast(builder, srcabs, u32_vec_type, "");
-
- /* Magic conversion */
- LLVMValueRef clamped = lp_build_min(&f32_bld, f32_f16max, srcabs);
- LLVMValueRef scaled = LLVMBuildBitCast(builder,
- LLVMBuildFMul(builder,
- clamped,
- f32_magic,
- ""),
- u32_vec_type,
- "");
- /* Make sure Inf/NaN and unormalised survive */
- LLVMValueRef infnancase = LLVMBuildXor(builder, u32_expinf, fabs, "");
- LLVMValueRef b_notnormal = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GEQUAL,
- srcabs,
- LLVMBuildBitCast(builder, u32_f32inf, f32_vec_type, ""));
-
- /* Merge normal / unnormal case */
- LLVMValueRef merged = lp_build_select(&u32_bld, b_notnormal, infnancase, scaled);
- LLVMValueRef shifted = lp_build_shr_imm(&u32_bld, merged, 13);
-
- /* Sign bit */
- LLVMValueRef justsign = LLVMBuildXor(builder, f, fabs, "");
- LLVMValueRef signshifted = lp_build_shr_imm(&u32_bld, justsign, 16);
-
- /* Combine result */
- result = LLVMBuildOr(builder, shifted, signshifted, "");
+ if (util_cpu_caps.has_f16c &&
+ (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);
+ }
}
- result = LLVMBuildTrunc(builder, result, i16_vec_type, "");
+ 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.
bias = (double)(1ULL << (mantissa - dst_width));
res = LLVMBuildFMul(builder, src, lp_build_const_vec(gallivm, src_type, scale), "");
+ /* instead of fadd/and could (with sse2) just use lp_build_iround */
res = LLVMBuildFAdd(builder, res, lp_build_const_vec(gallivm, src_type, bias), "");
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
res = LLVMBuildAnd(builder, res,
else if (dst_width == (mantissa + 1)) {
/*
* The destination width matches exactly what can be represented in
- * floating point (i.e., mantissa + 1 bits). So do a straight
- * multiplication followed by casting. No further rounding is necessary.
+ * floating point (i.e., mantissa + 1 bits). Even so correct rounding
+ * still needs to be applied (only for numbers in [0.5-1.0] would
+ * conversion using truncation after scaling be sufficient).
*/
-
double scale;
+ struct lp_build_context uf32_bld;
+ lp_build_context_init(&uf32_bld, gallivm, src_type);
scale = (double)((1ULL << dst_width) - 1);
res = LLVMBuildFMul(builder, src,
lp_build_const_vec(gallivm, src_type, scale), "");
- res = LLVMBuildFPToSI(builder, res, int_vec_type, "");
+ res = lp_build_iround(&uf32_bld, res);
}
else {
/*
dst_type->width == 8)
{
/* Special case 4x4f --> 1x16ub */
- if (src_type.length == 4 && util_cpu_caps.has_sse2)
+ if (src_type.length == 4 &&
+ util_cpu_caps.has_sse2)
{
- assert((num_srcs % 4) == 0);
-
- num_dsts = num_srcs / 4;
- dst_type->length = 16;
+ num_dsts = (num_srcs + 3) / 4;
+ dst_type->length = num_srcs * 4 >= 16 ? 16 : num_srcs * 4;
lp_build_conv(gallivm, src_type, *dst_type, src, num_srcs, dst, num_dsts);
return num_dsts;
}
/* Special case 2x8f --> 1x16ub */
- if (src_type.length == 8 && util_cpu_caps.has_avx)
+ if (src_type.length == 8 &&
+ util_cpu_caps.has_avx)
{
- assert((num_srcs % 2) == 0);
-
- num_dsts = num_srcs / 2;
- dst_type->length = 16;
+ num_dsts = (num_srcs + 1) / 2;
+ dst_type->length = num_srcs * 8 >= 16 ? 16 : num_srcs * 8;
lp_build_conv(gallivm, src_type, *dst_type, src, num_srcs, dst, num_dsts);
return num_dsts;
num_tmps = num_srcs;
- /* Special case 4x4f --> 1x16ub
+ /* Special case 4x4f --> 1x16ub, 2x4f -> 1x8ub, 1x4f -> 1x4ub
*/
if (src_type.floating == 1 &&
src_type.fixed == 0 &&
dst_type.sign == 0 &&
dst_type.norm == 1 &&
dst_type.width == 8 &&
- dst_type.length == 16 &&
- 4 * num_dsts == num_srcs &&
+ ((dst_type.length == 16 && 4 * num_dsts == num_srcs) ||
+ (num_dsts == 1 && dst_type.length * num_srcs == 16 && num_srcs != 3)) &&
util_cpu_caps.has_sse2)
{
struct lp_build_context bld;
- struct lp_type int16_type = dst_type;
- struct lp_type int32_type = dst_type;
+ struct lp_type int16_type, int32_type;
+ struct lp_type dst_type_ext = dst_type;
LLVMValueRef const_255f;
unsigned i, j;
lp_build_context_init(&bld, gallivm, src_type);
+ dst_type_ext.length = 16;
+ int16_type = int32_type = dst_type_ext;
+
int16_type.width *= 2;
int16_type.length /= 2;
int16_type.sign = 1;
for (i = 0; i < num_dsts; ++i, src += 4) {
LLVMValueRef lo, hi;
- for (j = 0; j < 4; ++j) {
+ for (j = 0; j < dst_type.length / 4; ++j) {
tmp[j] = LLVMBuildFMul(builder, src[j], const_255f, "");
tmp[j] = lp_build_iround(&bld, tmp[j]);
}
+ if (num_srcs == 1) {
+ tmp[1] = tmp[0];
+ }
+
/* relying on clamping behavior of sse2 intrinsics here */
lo = lp_build_pack2(gallivm, int32_type, int16_type, tmp[0], tmp[1]);
- hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
- dst[i] = lp_build_pack2(gallivm, int16_type, dst_type, lo, hi);
+
+ if (num_srcs < 4) {
+ hi = lo;
+ }
+ else {
+ hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
+ }
+ dst[i] = lp_build_pack2(gallivm, int16_type, dst_type_ext, lo, hi);
+ }
+ if (num_srcs < 4) {
+ dst[0] = lp_build_extract_range(gallivm, dst[0], 0, dst_type.length);
}
return;
}
- /* Special case 2x8f --> 1x16ub
+ /* Special case 2x8f --> 1x16ub, 1x8f ->1x8ub
*/
else if (src_type.floating == 1 &&
src_type.fixed == 0 &&
dst_type.sign == 0 &&
dst_type.norm == 1 &&
dst_type.width == 8 &&
- dst_type.length == 16 &&
- 2 * num_dsts == num_srcs &&
+ ((dst_type.length == 16 && 2 * num_dsts == num_srcs) ||
+ (num_dsts == 1 && dst_type.length * num_srcs == 8)) &&
util_cpu_caps.has_avx) {
struct lp_build_context bld;
- struct lp_type int16_type = dst_type;
- struct lp_type int32_type = dst_type;
+ struct lp_type int16_type, int32_type;
+ struct lp_type dst_type_ext = dst_type;
LLVMValueRef const_255f;
unsigned i;
lp_build_context_init(&bld, gallivm, src_type);
+ dst_type_ext.length = 16;
+ int16_type = int32_type = dst_type_ext;
+
int16_type.width *= 2;
int16_type.length /= 2;
int16_type.sign = 1;
LLVMValueRef lo, hi, a, b;
a = LLVMBuildFMul(builder, src[0], const_255f, "");
- b = LLVMBuildFMul(builder, src[1], const_255f, "");
-
a = lp_build_iround(&bld, a);
- b = lp_build_iround(&bld, b);
-
tmp[0] = lp_build_extract_range(gallivm, a, 0, 4);
tmp[1] = lp_build_extract_range(gallivm, a, 4, 4);
- tmp[2] = lp_build_extract_range(gallivm, b, 0, 4);
- tmp[3] = lp_build_extract_range(gallivm, b, 4, 4);
-
/* relying on clamping behavior of sse2 intrinsics here */
lo = lp_build_pack2(gallivm, int32_type, int16_type, tmp[0], tmp[1]);
- hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
- dst[i] = lp_build_pack2(gallivm, int16_type, dst_type, lo, hi);
+
+ if (num_srcs == 1) {
+ hi = lo;
+ }
+ else {
+ b = LLVMBuildFMul(builder, src[1], const_255f, "");
+ b = lp_build_iround(&bld, b);
+ tmp[2] = lp_build_extract_range(gallivm, b, 0, 4);
+ tmp[3] = lp_build_extract_range(gallivm, b, 4, 4);
+ hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
+
+ }
+ dst[i] = lp_build_pack2(gallivm, int16_type, dst_type_ext, lo, hi);
}
+
+ if (num_srcs == 1) {
+ dst[0] = lp_build_extract_range(gallivm, dst[0], 0, dst_type.length);
+ }
+
return;
}
}
else {
double dst_scale = lp_const_scale(dst_type);
- LLVMTypeRef tmp_vec_type;
if (dst_scale != 1.0) {
LLVMValueRef scale = lp_build_const_vec(gallivm, tmp_type, dst_scale);
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
- /* Use an equally sized integer for intermediate computations */
- tmp_type.floating = FALSE;
- tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
- for(i = 0; i < num_tmps; ++i) {
+ /*
+ * these functions will use fptosi in some form which won't work
+ * with 32bit uint dst. Causes lp_test_conv failures though.
+ */
+ if (0)
+ assert(dst_type.sign || dst_type.width < 32);
+
+ if (dst_type.sign && dst_type.norm && !dst_type.fixed) {
+ struct lp_build_context bld;
+
+ lp_build_context_init(&bld, gallivm, tmp_type);
+ for(i = 0; i < num_tmps; ++i) {
+ tmp[i] = lp_build_iround(&bld, tmp[i]);
+ }
+ tmp_type.floating = FALSE;
+ }
+ else {
+ LLVMTypeRef tmp_vec_type;
+
+ tmp_type.floating = FALSE;
+ tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
+ for(i = 0; i < num_tmps; ++i) {
#if 0
- if(dst_type.sign)
- tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
- else
- tmp[i] = LLVMBuildFPToUI(builder, tmp[i], tmp_vec_type, "");
+ if(dst_type.sign)
+ tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
+ else
+ tmp[i] = LLVMBuildFPToUI(builder, tmp[i], tmp_vec_type, "");
#else
- /* FIXME: there is no SSE counterpart for LLVMBuildFPToUI */
- tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
+ /* FIXME: there is no SSE counterpart for LLVMBuildFPToUI */
+ tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
#endif
+ }
}
}
}
unsigned dst_shift = lp_const_shift(dst_type);
unsigned src_offset = lp_const_offset(src_type);
unsigned dst_offset = lp_const_offset(dst_type);
+ struct lp_build_context bld;
+ lp_build_context_init(&bld, gallivm, tmp_type);
/* Compensate for different offsets */
- if (dst_offset > src_offset && src_type.width > dst_type.width) {
+ /* sscaled -> unorm and similar would cause negative shift count, skip */
+ if (dst_offset > src_offset && src_type.width > dst_type.width && src_shift > 0) {
for (i = 0; i < num_tmps; ++i) {
LLVMValueRef shifted;
- LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, src_shift - 1);
- if(src_type.sign)
- shifted = LLVMBuildAShr(builder, tmp[i], shift, "");
- else
- shifted = LLVMBuildLShr(builder, tmp[i], shift, "");
+ shifted = lp_build_shr_imm(&bld, tmp[i], src_shift - 1);
tmp[i] = LLVMBuildSub(builder, tmp[i], shifted, "");
}
}
if(src_shift > dst_shift) {
- LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type,
- src_shift - dst_shift);
for(i = 0; i < num_tmps; ++i)
- if(src_type.sign)
- tmp[i] = LLVMBuildAShr(builder, tmp[i], shift, "");
- else
- tmp[i] = LLVMBuildLShr(builder, tmp[i], shift, "");
+ tmp[i] = lp_build_shr_imm(&bld, tmp[i], src_shift - dst_shift);
}
}
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
+
+ /* the formula above will produce value below -1.0 for most negative
+ * value but everything seems happy with that hence disable for now */
+ if (0 && !src_type.fixed && src_type.norm && src_type.sign) {
+ struct lp_build_context bld;
+
+ lp_build_context_init(&bld, gallivm, dst_type);
+ for(i = 0; i < num_tmps; ++i) {
+ tmp[i] = lp_build_max(&bld, tmp[i],
+ lp_build_const_vec(gallivm, dst_type, -1.0f));
+ }
+ }
}
}
else {
unsigned dst_shift = lp_const_shift(dst_type);
unsigned src_offset = lp_const_offset(src_type);
unsigned dst_offset = lp_const_offset(dst_type);
+ struct lp_build_context bld;
+ lp_build_context_init(&bld, gallivm, tmp_type);
if (src_shift < dst_shift) {
LLVMValueRef pre_shift[LP_MAX_VECTOR_LENGTH];
- LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, dst_shift - src_shift);
- for (i = 0; i < num_tmps; ++i) {
- pre_shift[i] = tmp[i];
- tmp[i] = LLVMBuildShl(builder, tmp[i], shift, "");
+ if (dst_shift - src_shift < dst_type.width) {
+ for (i = 0; i < num_tmps; ++i) {
+ pre_shift[i] = tmp[i];
+ tmp[i] = lp_build_shl_imm(&bld, tmp[i], dst_shift - src_shift);
+ }
+ }
+ else {
+ /*
+ * This happens for things like sscaled -> unorm conversions. Shift
+ * counts equal to bit width cause undefined results, so hack around it.
+ */
+ for (i = 0; i < num_tmps; ++i) {
+ pre_shift[i] = tmp[i];
+ tmp[i] = lp_build_zero(gallivm, dst_type);
+ }
}
/* Compensate for different offsets */