#include "util/u_debug.h"
#include "util/u_math.h"
+#include "util/u_half.h"
#include "util/u_cpu_detect.h"
#include "lp_bld_type.h"
#include "lp_bld_const.h"
#include "lp_bld_arit.h"
+#include "lp_bld_bitarit.h"
#include "lp_bld_pack.h"
#include "lp_bld_conv.h"
#include "lp_bld_logic.h"
+#include "lp_bld_intr.h"
+#include "lp_bld_printf.h"
+#include "lp_bld_format.h"
+
/**
* Converts int16 half-float to float32
- * Note this can be performed in 1 instruction if vcvtph2ps exists (sse5 i think?)
+ * Note this can be performed in 1 instruction if vcvtph2ps exists (f16c/cvt16)
* [llvm.x86.vcvtph2ps / _mm_cvtph_ps]
*
* @param src value to convert
*
- * ref http://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/
- * ref https://gist.github.com/2144712
*/
LLVMValueRef
lp_build_half_to_float(struct gallivm_state *gallivm,
LLVMValueRef src)
{
- int src_length = LLVMGetVectorSize(LLVMTypeOf(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);
-
- 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, "");
+ 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 (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
+ * Convert float32 to half floats, preserving Infs and NaNs,
+ * with rounding towards zero (trunc).
+ * XXX: For GL, would prefer rounding towards nearest(-even).
*/
LLVMValueRef
lp_build_float_to_half(struct gallivm_state *gallivm,
LLVMValueRef src)
{
- struct lp_type i32_type = lp_type_int_vec(32, 32 * LLVMGetVectorSize(LLVMTypeOf(src)));
-
LLVMBuilderRef builder = gallivm->builder;
- LLVMTypeRef int_vec_type = lp_build_vec_type(gallivm, i32_type);
-
- struct lp_build_context bld;
-
+ 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;
- lp_build_context_init(&bld, gallivm, i32_type);
+ /*
+ * Note: Newer llvm versions (3.6 or so) support fptrunc to 16 bits
+ * directly, without any (x86 or generic) intrinsics.
+ * Albeit the rounding mode cannot be specified (and is undefined,
+ * though in practice on x86 seems to do nearest-even but it may
+ * be dependent on instruction set support), so is essentially
+ * useless.
+ */
- /* Extra scope because lp_build_min needs a build context, le sigh */
- {
- /* 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_fabs = lp_build_const_int_vec(gallivm, i32_type, 0x7fffffff);
- LLVMValueRef i32_f32infty = lp_build_const_int_vec(gallivm, i32_type, 0xff << 23);
- LLVMValueRef i32_expinf = lp_build_const_int_vec(gallivm, i32_type, 0xe0 << 23);
- LLVMValueRef i32_f16max = lp_build_const_int_vec(gallivm, i32_type, 0x8f << 23);
- LLVMValueRef i32_magic = lp_build_const_int_vec(gallivm, i32_type, 0x0f << 23);
-
- /* Cast from float32 to int32 */
- LLVMValueRef f = LLVMBuildBitCast(builder, src, int_vec_type, "");
-
- /* Remove sign */
- LLVMValueRef fabs = LLVMBuildAnd(builder, i32_mask_fabs, f, "");
-
- /* Magic conversion */
- LLVMValueRef clamped = lp_build_min(&bld, i32_f16max, fabs);
- LLVMValueRef scaled = LLVMBuildMul(builder, clamped, i32_magic, "");
-
- /* Make sure Inf/NaN and unormalised survive */
- LLVMValueRef infnancase = LLVMBuildXor(builder, i32_expinf, fabs, "");
- LLVMValueRef b_notnormal = lp_build_compare(gallivm, i32_type, PIPE_FUNC_GREATER, fabs, i32_f32infty);
-
- /* Merge normal / unnormal case */
- LLVMValueRef merge1 = LLVMBuildAnd(builder, infnancase, b_notnormal, "");
- LLVMValueRef merge2 = LLVMBuildNot(builder, LLVMBuildAnd(builder, b_notnormal, scaled, ""), "");
- LLVMValueRef merged = LLVMBuildOr(builder, merge1, merge2, "");
- LLVMValueRef shifted = LLVMBuildLShr(builder, merged, i32_13, "");
-
- /* Sign bit */
- LLVMValueRef justsign = LLVMBuildXor(builder, f, fabs, "");
- LLVMValueRef signshifted = LLVMBuildLShr(builder, justsign, i32_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);
+ }
+ }
+
+ 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), "");
}
- /* Truncate from 32 bit to 16 bit */
- i32_type.width = 16;
- return LLVMBuildTrunc(builder, result, lp_build_vec_type(gallivm, i32_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.
+ * Even if they would now, rounding mode cannot be specified and
+ * is undefined.
+ */
+ 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;
}
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 {
/*
* important, we also get exact results for 0.0 and 1.0.
*/
- unsigned n = MIN2(src_type.width - 1, dst_width);
+ unsigned n = MIN2(src_type.width - 1u, dst_width);
double scale = (double)(1ULL << n);
unsigned lshift = dst_width - n;
}
+/**
+ * Pick a suitable num_dsts for lp_build_conv to ensure optimal cases are used.
+ *
+ * Returns the number of dsts created from src
+ */
+int lp_build_conv_auto(struct gallivm_state *gallivm,
+ struct lp_type src_type,
+ struct lp_type* dst_type,
+ const LLVMValueRef *src,
+ unsigned num_srcs,
+ LLVMValueRef *dst)
+{
+ unsigned i;
+ int num_dsts = num_srcs;
+
+ if (src_type.floating == dst_type->floating &&
+ src_type.width == dst_type->width &&
+ src_type.length == dst_type->length &&
+ src_type.fixed == dst_type->fixed &&
+ src_type.norm == dst_type->norm &&
+ src_type.sign == dst_type->sign)
+ return num_dsts;
+
+ /* Special case 4x4x32 -> 1x16x8 or 2x8x32 -> 1x16x8
+ */
+ if (src_type.norm == 0 &&
+ src_type.width == 32 &&
+ src_type.fixed == 0 &&
+
+ dst_type->floating == 0 &&
+ dst_type->fixed == 0 &&
+ dst_type->width == 8 &&
+
+ ((src_type.floating == 1 && src_type.sign == 1 && dst_type->norm == 1) ||
+ (src_type.floating == 0 && dst_type->floating == 0 &&
+ src_type.sign == dst_type->sign && dst_type->norm == 0))) {
+
+ /* Special case 4x4x32 --> 1x16x8 */
+ if (src_type.length == 4 &&
+ (util_cpu_caps.has_sse2 || util_cpu_caps.has_altivec))
+ {
+ 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 2x8x32 --> 1x16x8 */
+ if (src_type.length == 8 &&
+ util_cpu_caps.has_avx)
+ {
+ 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;
+ }
+ }
+
+ /* lp_build_resize does not support M:N */
+ if (src_type.width == dst_type->width) {
+ lp_build_conv(gallivm, src_type, *dst_type, src, num_srcs, dst, num_dsts);
+ } else {
+ /*
+ * If dst_width is 16 bits and src_width 32 and the dst vector size
+ * 64bit, try feeding 2 vectors at once so pack intrinsics can be used.
+ * (For AVX, this isn't needed, since we usually get 256bit src and
+ * 128bit dst vectors which works ok. If we do AVX2 pack this should
+ * be extended but need to be able to tell conversion code about pack
+ * ordering first.)
+ */
+ unsigned ratio = 1;
+ if (src_type.width == 2 * dst_type->width &&
+ src_type.length == dst_type->length &&
+ dst_type->floating == 0 && (num_srcs % 2 == 0) &&
+ dst_type->width * dst_type->length == 64) {
+ ratio = 2;
+ num_dsts /= 2;
+ dst_type->length *= 2;
+ }
+ for (i = 0; i < num_dsts; i++) {
+ lp_build_conv(gallivm, src_type, *dst_type, &src[i*ratio], ratio, &dst[i], 1);
+ }
+ }
+
+ return num_dsts;
+}
+
+
/**
* Generic type conversion.
*
num_tmps = num_srcs;
- /* Special case 4x4f --> 1x16ub
+ /*
+ * Special case 4x4x32 --> 1x16x8, 2x4x32 -> 1x8x8, 1x4x32 -> 1x4x8
+ * Only float -> s/unorm8 and (u)int32->(u)int8.
+ * XXX: This should cover all interesting backend cases for 8 bit,
+ * but should use same strategy if dst is 16 bit.
*/
- if (src_type.floating == 1 &&
- src_type.fixed == 0 &&
- src_type.sign == 1 &&
- src_type.norm == 0 &&
+ if (src_type.norm == 0 &&
src_type.width == 32 &&
src_type.length == 4 &&
+ src_type.fixed == 0 &&
dst_type.floating == 0 &&
dst_type.fixed == 0 &&
- dst_type.sign == 0 &&
- dst_type.norm == 1 &&
dst_type.width == 8 &&
- dst_type.length == 16 &&
- 4 * num_dsts == num_srcs &&
+ ((src_type.floating == 1 && src_type.sign == 1 && dst_type.norm == 1) ||
+ (src_type.floating == 0 && dst_type.floating == 0 &&
+ src_type.sign == dst_type.sign && dst_type.norm == 0)) &&
- util_cpu_caps.has_sse2)
+ ((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 || util_cpu_caps.has_altivec))
{
struct lp_build_context bld;
- struct lp_type int16_type = dst_type;
- struct lp_type int32_type = dst_type;
- LLVMValueRef const_255f;
+ struct lp_type int16_type, int32_type;
+ struct lp_type dst_type_ext = dst_type;
+ LLVMValueRef const_scale;
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;
int32_type.length /= 4;
int32_type.sign = 1;
- const_255f = lp_build_const_vec(gallivm, src_type, 255.0f);
+ const_scale = lp_build_const_vec(gallivm, src_type, lp_const_scale(dst_type));
for (i = 0; i < num_dsts; ++i, src += 4) {
LLVMValueRef lo, hi;
- for (j = 0; j < 4; ++j) {
- tmp[j] = LLVMBuildFMul(builder, src[j], const_255f, "");
- tmp[j] = lp_build_iround(&bld, tmp[j]);
+ if (src_type.floating) {
+ for (j = 0; j < dst_type.length / 4; ++j) {
+ /*
+ * XXX This is not actually fully correct. The float to int
+ * conversion will produce 0x80000000 value for everything
+ * out of range and NaNs (on x86, llvm.x86.sse2.cvtps2dq).
+ * Hence, NaNs and negatives will get clamped just fine to zero
+ * (relying on clamping pack behavior) when converting to unorm,
+ * however too large values (both finite and infinite) will also
+ * end up as zero, not 255.
+ * For snorm, for now we'll keep bug compatibility with generic
+ * conversion path (meaning too large values are fine, but
+ * NaNs get converted to -128 (purely by luck, as we don't
+ * specify nan behavior for the max there) instead of 0).
+ */
+ if (dst_type.sign) {
+ tmp[j] = lp_build_min(&bld, bld.one, src[j]);
+
+ }
+ else {
+ if (0) {
+ tmp[j] = lp_build_min_ext(&bld, bld.one, src[j],
+ GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN);
+ }
+ tmp[j] = src[j];
+ }
+ tmp[j] = LLVMBuildFMul(builder, tmp[j], const_scale, "");
+ tmp[j] = lp_build_iround(&bld, tmp[j]);
+ }
+ } else {
+ for (j = 0; j < dst_type.length / 4; ++j) {
+ if (!dst_type.sign) {
+ /*
+ * Pack clamp is always signed->unsigned (or signed->signed).
+ * Hence need min.
+ */
+ LLVMValueRef const_max;
+ const_max = lp_build_const_int_vec(gallivm, src_type, 255);
+ tmp[j] = lp_build_min(&bld, src[j], const_max);
+ } else {
+ tmp[j] = src[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 2x8x32 --> 1x16x8, 1x8x32 ->1x8x8
*/
- else if (src_type.floating == 1 &&
- src_type.fixed == 0 &&
- src_type.sign == 1 &&
- src_type.norm == 0 &&
- src_type.width == 32 &&
- src_type.length == 8 &&
-
- dst_type.floating == 0 &&
- dst_type.fixed == 0 &&
- dst_type.sign == 0 &&
- dst_type.norm == 1 &&
- dst_type.width == 8 &&
- dst_type.length == 16 &&
-
- 2 * num_dsts == num_srcs &&
+ else if (src_type.norm == 0 &&
+ src_type.width == 32 &&
+ src_type.length == 8 &&
+ src_type.fixed == 0 &&
+
+ dst_type.floating == 0 &&
+ dst_type.fixed == 0 &&
+ dst_type.width == 8 &&
+
+ ((src_type.floating == 1 && src_type.sign == 1 && dst_type.norm == 1) ||
+ (src_type.floating == 0 && dst_type.floating == 0 &&
+ src_type.sign == dst_type.sign && dst_type.norm == 0)) &&
+
+ ((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;
- LLVMValueRef const_255f;
+ struct lp_type int16_type, int32_type;
+ struct lp_type dst_type_ext = dst_type;
+ LLVMValueRef const_scale;
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;
int32_type.length /= 4;
int32_type.sign = 1;
- const_255f = lp_build_const_vec(gallivm, src_type, 255.0f);
+ const_scale = lp_build_const_vec(gallivm, src_type, lp_const_scale(dst_type));
for (i = 0; i < num_dsts; ++i, src += 2) {
- 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);
+ unsigned j;
+ for (j = 0; j < (num_srcs == 1 ? 1 : 2); j++) {
+ LLVMValueRef lo, hi, a;
+
+ a = src[j];
+ if (src_type.floating) {
+ if (dst_type.sign) {
+ a = lp_build_min(&bld, bld.one, a);
+
+ }
+ else {
+ if (0) {
+ a = lp_build_min_ext(&bld, bld.one, a,
+ GALLIVM_NAN_RETURN_NAN_FIRST_NONNAN);
+ }
+ }
+ a = LLVMBuildFMul(builder, a, const_scale, "");
+ a = lp_build_iround(&bld, a);
+ } else {
+ if (!dst_type.sign) {
+ LLVMValueRef const_max;
+ const_max = lp_build_const_int_vec(gallivm, src_type, 255);
+ a = lp_build_min(&bld, a, const_max);
+ }
+ }
+ lo = lp_build_extract_range(gallivm, a, 0, 4);
+ hi = lp_build_extract_range(gallivm, a, 4, 4);
+ /* relying on clamping behavior of sse2 intrinsics here */
+ tmp[j] = lp_build_pack2(gallivm, int32_type, int16_type, lo, hi);
+ }
- 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);
+ if (num_srcs == 1) {
+ tmp[1] = tmp[0];
+ }
+ dst[i] = lp_build_pack2(gallivm, int16_type, dst_type_ext, tmp[0], tmp[1]);
+ }
- /* 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) {
+ 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);
}
}
new_type.width = dst_type.width;
new_type.length = dst_type.length;
+ /*
+ * Note that resize when using packs can sometimes get min/max
+ * clamping for free. Should be able to exploit this...
+ */
lp_build_resize(gallivm, tmp_type, new_type, tmp, num_srcs, tmp, num_dsts);
tmp_type = new_type;
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 */
projects / mesa.git / blobdiff