#include "util/u_debug.h"
#include "util/u_math.h"
+#include "util/u_cpu_detect.h"
#include "lp_bld_type.h"
#include "lp_bld_const.h"
*
* Although the result values can be scaled to an arbitrary bit width specified
* by dst_width, the actual result type will have the same width.
+ *
+ * Ex: src = { float, float, float, float }
+ * return { i32, i32, i32, i32 } where each value is in [0, 2^dst_width-1].
*/
LLVMValueRef
-lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
+lp_build_clamped_float_to_unsigned_norm(struct gallivm_state *gallivm,
struct lp_type src_type,
unsigned dst_width,
LLVMValueRef src)
{
- LLVMTypeRef int_vec_type = lp_build_int_vec_type(src_type);
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, src_type);
LLVMValueRef res;
unsigned mantissa;
- unsigned n;
- unsigned long long ubound;
- unsigned long long mask;
- double scale;
- double bias;
assert(src_type.floating);
+ assert(dst_width <= src_type.width);
+ src_type.sign = FALSE;
mantissa = lp_mantissa(src_type);
- /* We cannot carry more bits than the mantissa */
- n = MIN2(mantissa, dst_width);
+ if (dst_width <= mantissa) {
+ /*
+ * Apply magic coefficients that will make the desired result to appear
+ * in the lowest significant bits of the mantissa, with correct rounding.
+ *
+ * This only works if the destination width fits in the mantissa.
+ */
- /* This magic coefficients will make the desired result to appear in the
- * lowest significant bits of the mantissa.
- */
- ubound = ((unsigned long long)1 << n);
- mask = ubound - 1;
- scale = (double)mask/ubound;
- bias = (double)((unsigned long long)1 << (mantissa - n));
-
- res = LLVMBuildMul(builder, src, lp_build_const_scalar(src_type, scale), "");
- res = LLVMBuildAdd(builder, res, lp_build_const_scalar(src_type, bias), "");
- res = LLVMBuildBitCast(builder, res, int_vec_type, "");
-
- if(dst_width > n) {
- int shift = dst_width - n;
- res = LLVMBuildShl(builder, res, lp_build_int_const_scalar(src_type, shift), "");
-
- /* TODO: Fill in the empty lower bits for additional precision? */
- /* YES: this fixes progs/trivial/tri-z-eq.c.
- * Otherwise vertex Z=1.0 values get converted to something like
- * 0xfffffb00 and the test for equality with 0xffffffff fails.
+ unsigned long long ubound;
+ unsigned long long mask;
+ double scale;
+ double bias;
+
+ ubound = (1ULL << dst_width);
+ mask = ubound - 1;
+ scale = (double)mask/ubound;
+ bias = (double)(1ULL << (mantissa - dst_width));
+
+ res = LLVMBuildFMul(builder, src, lp_build_const_vec(gallivm, src_type, scale), "");
+ res = LLVMBuildFAdd(builder, res, lp_build_const_vec(gallivm, src_type, bias), "");
+ res = LLVMBuildBitCast(builder, res, int_vec_type, "");
+ res = LLVMBuildAnd(builder, res,
+ lp_build_const_int_vec(gallivm, src_type, mask), "");
+ }
+ 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.
*/
-#if 0
- {
- LLVMValueRef msb;
- msb = LLVMBuildLShr(builder, res, lp_build_int_const_scalar(src_type, dst_width - 1), "");
- msb = LLVMBuildShl(builder, msb, lp_build_int_const_scalar(src_type, shift), "");
- msb = LLVMBuildSub(builder, msb, lp_build_int_const_scalar(src_type, 1), "");
- res = LLVMBuildOr(builder, res, msb, "");
- }
-#elif 0
- while(shift > 0) {
- res = LLVMBuildOr(builder, res, LLVMBuildLShr(builder, res, lp_build_int_const_scalar(src_type, n), ""), "");
- shift -= n;
- n *= 2;
+
+ double scale;
+
+ 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, "");
+ }
+ else {
+ /*
+ * The destination exceeds what can be represented in the floating point.
+ * So multiply by the largest power two we get away with, and when
+ * subtract the most significant bit to rescale to normalized values.
+ *
+ * The largest power of two factor we can get away is
+ * (1 << (src_type.width - 1)), because we need to use signed . In theory it
+ * should be (1 << (src_type.width - 2)), but IEEE 754 rules states
+ * INT_MIN should be returned in FPToSI, which is the correct result for
+ * values near 1.0!
+ *
+ * This means we get (src_type.width - 1) correct bits for values near 0.0,
+ * and (mantissa + 1) correct bits for values near 1.0. Equally or more
+ * important, we also get exact results for 0.0 and 1.0.
+ */
+
+ unsigned n = MIN2(src_type.width - 1, dst_width);
+
+ double scale = (double)(1ULL << n);
+ unsigned lshift = dst_width - n;
+ unsigned rshift = n;
+ LLVMValueRef lshifted;
+ LLVMValueRef rshifted;
+
+ res = LLVMBuildFMul(builder, src,
+ lp_build_const_vec(gallivm, src_type, scale), "");
+ res = LLVMBuildFPToSI(builder, res, int_vec_type, "");
+
+ /*
+ * Align the most significant bit to its final place.
+ *
+ * This will cause 1.0 to overflow to 0, but the later adjustment will
+ * get it right.
+ */
+ if (lshift) {
+ lshifted = LLVMBuildShl(builder, res,
+ lp_build_const_int_vec(gallivm, src_type,
+ lshift), "");
+ } else {
+ lshifted = res;
}
-#endif
+
+ /*
+ * Align the most significant bit to the right.
+ */
+ rshifted = LLVMBuildAShr(builder, res,
+ lp_build_const_int_vec(gallivm, src_type, rshift),
+ "");
+
+ /*
+ * Subtract the MSB to the LSB, therefore re-scaling from
+ * (1 << dst_width) to ((1 << dst_width) - 1).
+ */
+
+ res = LLVMBuildSub(builder, lshifted, rshifted, "");
}
- else
- res = LLVMBuildAnd(builder, res, lp_build_int_const_scalar(src_type, mask), "");
return res;
}
/**
* Inverse of lp_build_clamped_float_to_unsigned_norm above.
+ * Ex: src = { i32, i32, i32, i32 } with values in range [0, 2^src_width-1]
+ * return {float, float, float, float} with values in range [0, 1].
*/
LLVMValueRef
-lp_build_unsigned_norm_to_float(LLVMBuilderRef builder,
+lp_build_unsigned_norm_to_float(struct gallivm_state *gallivm,
unsigned src_width,
struct lp_type dst_type,
LLVMValueRef src)
{
- LLVMTypeRef vec_type = lp_build_vec_type(dst_type);
- LLVMTypeRef int_vec_type = lp_build_int_vec_type(dst_type);
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMTypeRef vec_type = lp_build_vec_type(gallivm, dst_type);
+ LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, dst_type);
LLVMValueRef bias_;
LLVMValueRef res;
unsigned mantissa;
double scale;
double bias;
+ assert(dst_type.floating);
+
mantissa = lp_mantissa(dst_type);
- n = MIN2(mantissa, src_width);
+ if (src_width <= (mantissa + 1)) {
+ /*
+ * The source width matches fits 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.
+ */
- ubound = ((unsigned long long)1 << n);
- mask = ubound - 1;
- scale = (double)ubound/mask;
- bias = (double)((unsigned long long)1 << (mantissa - n));
+ scale = 1.0/(double)((1ULL << src_width) - 1);
+ res = LLVMBuildSIToFP(builder, src, vec_type, "");
+ res = LLVMBuildFMul(builder, res,
+ lp_build_const_vec(gallivm, dst_type, scale), "");
+ return res;
+ }
+ else {
+ /*
+ * The source width exceeds what can be represented in floating
+ * point. So truncate the incoming values.
+ */
- res = src;
+ n = MIN2(mantissa, src_width);
- if(src_width > mantissa) {
- int shift = src_width - mantissa;
- res = LLVMBuildLShr(builder, res, lp_build_int_const_scalar(dst_type, shift), "");
- }
+ ubound = ((unsigned long long)1 << n);
+ mask = ubound - 1;
+ scale = (double)ubound/mask;
+ bias = (double)((unsigned long long)1 << (mantissa - n));
- bias_ = lp_build_const_scalar(dst_type, bias);
+ res = src;
- res = LLVMBuildOr(builder,
- res,
- LLVMBuildBitCast(builder, bias_, int_vec_type, ""), "");
+ if (src_width > mantissa) {
+ int shift = src_width - mantissa;
+ res = LLVMBuildLShr(builder, res,
+ lp_build_const_int_vec(gallivm, dst_type, shift), "");
+ }
+
+ bias_ = lp_build_const_vec(gallivm, dst_type, bias);
- res = LLVMBuildBitCast(builder, res, vec_type, "");
+ res = LLVMBuildOr(builder,
+ res,
+ LLVMBuildBitCast(builder, bias_, int_vec_type, ""), "");
- res = LLVMBuildSub(builder, res, bias_, "");
- res = LLVMBuildMul(builder, res, lp_build_const_scalar(dst_type, scale), "");
+ res = LLVMBuildBitCast(builder, res, vec_type, "");
+
+ res = LLVMBuildFSub(builder, res, bias_, "");
+ res = LLVMBuildFMul(builder, res, lp_build_const_vec(gallivm, dst_type, scale), "");
+ }
return res;
}
* to the lp_type union.
*/
void
-lp_build_conv(LLVMBuilderRef builder,
+lp_build_conv(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
const LLVMValueRef *src, unsigned num_srcs,
LLVMValueRef *dst, unsigned num_dsts)
{
+ LLVMBuilderRef builder = gallivm->builder;
struct lp_type tmp_type;
LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
unsigned num_tmps;
unsigned i;
- /* Register width must remain constant */
- assert(src_type.width * src_type.length == dst_type.width * dst_type.length);
-
/* We must not loose or gain channels. Only precision */
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
assert(src_type.length <= LP_MAX_VECTOR_LENGTH);
assert(dst_type.length <= LP_MAX_VECTOR_LENGTH);
+ assert(num_srcs <= LP_MAX_VECTOR_LENGTH);
+ assert(num_dsts <= LP_MAX_VECTOR_LENGTH);
tmp_type = src_type;
- for(i = 0; i < num_srcs; ++i)
+ for(i = 0; i < num_srcs; ++i) {
+ assert(lp_check_value(src_type, src[i]));
tmp[i] = src[i];
+ }
num_tmps = num_srcs;
+
+ /* Special case 4x4f --> 1x16ub
+ */
+ if (src_type.floating == 1 &&
+ src_type.fixed == 0 &&
+ src_type.sign == 1 &&
+ src_type.norm == 0 &&
+ src_type.width == 32 &&
+ src_type.length == 4 &&
+
+ dst_type.floating == 0 &&
+ dst_type.fixed == 0 &&
+ dst_type.sign == 0 &&
+ dst_type.norm == 1 &&
+ dst_type.width == 8 &&
+ dst_type.length == 16 &&
+
+ util_cpu_caps.has_sse2)
+ {
+ int i;
+
+ for (i = 0; i < num_dsts; i++, src += 4) {
+ struct lp_type int16_type = dst_type;
+ struct lp_type int32_type = dst_type;
+ LLVMValueRef lo, hi;
+ LLVMValueRef src_int0;
+ LLVMValueRef src_int1;
+ LLVMValueRef src_int2;
+ LLVMValueRef src_int3;
+ LLVMTypeRef int16_vec_type;
+ LLVMTypeRef int32_vec_type;
+ LLVMTypeRef src_vec_type;
+ LLVMTypeRef dst_vec_type;
+ LLVMValueRef const_255f;
+ LLVMValueRef a, b, c, d;
+
+ int16_type.width *= 2;
+ int16_type.length /= 2;
+ int16_type.sign = 1;
+
+ int32_type.width *= 4;
+ int32_type.length /= 4;
+ int32_type.sign = 1;
+
+ src_vec_type = lp_build_vec_type(gallivm, src_type);
+ dst_vec_type = lp_build_vec_type(gallivm, dst_type);
+ int16_vec_type = lp_build_vec_type(gallivm, int16_type);
+ int32_vec_type = lp_build_vec_type(gallivm, int32_type);
+
+ const_255f = lp_build_const_vec(gallivm, src_type, 255.0f);
+
+ a = LLVMBuildFMul(builder, src[0], const_255f, "");
+ b = LLVMBuildFMul(builder, src[1], const_255f, "");
+ c = LLVMBuildFMul(builder, src[2], const_255f, "");
+ d = LLVMBuildFMul(builder, src[3], const_255f, "");
+
+ {
+ struct lp_build_context bld;
+
+ bld.gallivm = gallivm;
+ bld.type = src_type;
+ bld.vec_type = src_vec_type;
+ bld.int_elem_type = lp_build_elem_type(gallivm, int32_type);
+ bld.int_vec_type = int32_vec_type;
+ bld.undef = lp_build_undef(gallivm, src_type);
+ bld.zero = lp_build_zero(gallivm, src_type);
+ bld.one = lp_build_one(gallivm, src_type);
+
+ src_int0 = lp_build_iround(&bld, a);
+ src_int1 = lp_build_iround(&bld, b);
+ src_int2 = lp_build_iround(&bld, c);
+ src_int3 = lp_build_iround(&bld, d);
+ }
+ /* relying on clamping behavior of sse2 intrinsics here */
+ lo = lp_build_pack2(gallivm, int32_type, int16_type, src_int0, src_int1);
+ hi = lp_build_pack2(gallivm, int32_type, int16_type, src_int2, src_int3);
+ dst[i] = lp_build_pack2(gallivm, int16_type, dst_type, lo, hi);
+ }
+ return;
+ }
+
/*
* Clamp if necessary
*/
double dst_max = lp_const_max(dst_type);
LLVMValueRef thres;
- lp_build_context_init(&bld, builder, tmp_type);
+ lp_build_context_init(&bld, gallivm, tmp_type);
if(src_min < dst_min) {
if(dst_min == 0.0)
thres = bld.zero;
else
- thres = lp_build_const_scalar(src_type, dst_min);
+ thres = lp_build_const_vec(gallivm, src_type, dst_min);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_max(&bld, tmp[i], thres);
}
if(dst_max == 1.0)
thres = bld.one;
else
- thres = lp_build_const_scalar(src_type, dst_max);
+ thres = lp_build_const_vec(gallivm, src_type, dst_max);
for(i = 0; i < num_tmps; ++i)
tmp[i] = lp_build_min(&bld, tmp[i], thres);
}
else if(tmp_type.floating) {
if(!dst_type.fixed && !dst_type.sign && dst_type.norm) {
for(i = 0; i < num_tmps; ++i) {
- tmp[i] = lp_build_clamped_float_to_unsigned_norm(builder,
+ tmp[i] = lp_build_clamped_float_to_unsigned_norm(gallivm,
tmp_type,
dst_type.width,
tmp[i]);
LLVMTypeRef tmp_vec_type;
if (dst_scale != 1.0) {
- LLVMValueRef scale = lp_build_const_scalar(tmp_type, dst_scale);
+ LLVMValueRef scale = lp_build_const_vec(gallivm, tmp_type, dst_scale);
for(i = 0; i < num_tmps; ++i)
- tmp[i] = LLVMBuildMul(builder, tmp[i], 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(tmp_type);
+ tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
/* FIXME: compensate different offsets too */
if(src_shift > dst_shift) {
- LLVMValueRef shift = lp_build_int_const_scalar(tmp_type, 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, "");
/*
* Truncate or expand bit width
+ *
+ * No data conversion should happen here, although the sign bits are
+ * crucial to avoid bad clamping.
*/
- assert(!tmp_type.floating || tmp_type.width == dst_type.width);
+ {
+ struct lp_type new_type;
- if(tmp_type.width > dst_type.width) {
- assert(num_dsts == 1);
- tmp[0] = lp_build_pack(builder, tmp_type, dst_type, TRUE, tmp, num_tmps);
- tmp_type.width = dst_type.width;
- tmp_type.length = dst_type.length;
- num_tmps = 1;
- }
+ new_type = tmp_type;
+ new_type.sign = dst_type.sign;
+ new_type.width = dst_type.width;
+ new_type.length = dst_type.length;
- if(tmp_type.width < dst_type.width) {
- assert(num_tmps == 1);
- lp_build_unpack(builder, tmp_type, dst_type, tmp[0], tmp, num_dsts);
- tmp_type.width = dst_type.width;
- tmp_type.length = dst_type.length;
+ lp_build_resize(gallivm, tmp_type, new_type, tmp, num_srcs, tmp, num_dsts);
+
+ tmp_type = new_type;
num_tmps = num_dsts;
}
- assert(tmp_type.width == dst_type.width);
- assert(tmp_type.length == dst_type.length);
- assert(num_tmps == num_dsts);
-
/*
* Scale to the widest range
*/
else if(!src_type.floating && dst_type.floating) {
if(!src_type.fixed && !src_type.sign && src_type.norm) {
for(i = 0; i < num_tmps; ++i) {
- tmp[i] = lp_build_unsigned_norm_to_float(builder,
+ tmp[i] = lp_build_unsigned_norm_to_float(gallivm,
src_type.width,
dst_type,
tmp[i]);
/* Use an equally sized integer for intermediate computations */
tmp_type.floating = TRUE;
tmp_type.sign = TRUE;
- tmp_vec_type = lp_build_vec_type(tmp_type);
+ tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
for(i = 0; i < num_tmps; ++i) {
#if 0
if(dst_type.sign)
}
if (src_scale != 1.0) {
- LLVMValueRef scale = lp_build_const_scalar(tmp_type, 1.0/src_scale);
+ LLVMValueRef scale = lp_build_const_vec(gallivm, tmp_type, 1.0/src_scale);
for(i = 0; i < num_tmps; ++i)
- tmp[i] = LLVMBuildMul(builder, tmp[i], scale, "");
+ tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
}
}
}
/* FIXME: compensate different offsets too */
if(src_shift < dst_shift) {
- LLVMValueRef shift = lp_build_int_const_scalar(tmp_type, dst_shift - src_shift);
+ LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, dst_shift - src_shift);
for(i = 0; i < num_tmps; ++i)
tmp[i] = LLVMBuildShl(builder, tmp[i], shift, "");
}
}
- for(i = 0; i < num_dsts; ++i)
+ for(i = 0; i < num_dsts; ++i) {
dst[i] = tmp[i];
+ assert(lp_check_value(dst_type, dst[i]));
+ }
}
* This is basically a very trimmed down version of lp_build_conv.
*/
void
-lp_build_conv_mask(LLVMBuilderRef builder,
+lp_build_conv_mask(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
const LLVMValueRef *src, unsigned num_srcs,
if(src_type.width > dst_type.width) {
assert(num_dsts == 1);
- dst[0] = lp_build_pack(builder, src_type, dst_type, TRUE, src, num_srcs);
+ dst[0] = lp_build_pack(gallivm, src_type, dst_type, TRUE, src, num_srcs);
}
else if(src_type.width < dst_type.width) {
assert(num_srcs == 1);
- lp_build_unpack(builder, src_type, dst_type, src[0], dst, num_dsts);
+ lp_build_unpack(gallivm, src_type, dst_type, src[0], dst, num_dsts);
}
else {
assert(num_srcs == num_dsts);