/* XXX: SSE3 has PSHUFB which should be better than bitmasks, but forcing
* using shuffles here actually causes worst results. More investigation is
* needed. */
- if (n <= 4) {
+ if (type.width >= 16) {
/*
* Shuffle.
*/
* YY00 YY00 .... YY00
* YYYY YYYY .... YYYY <= output
*/
- struct lp_type type4 = type;
+ struct lp_type type4;
const char shifts[4][2] = {
{ 1, 2},
{-1, 2},
a = LLVMBuildAnd(bld->builder, a, lp_build_const_mask_aos(type, cond), "");
+ /*
+ * Build a type where each element is an integer that cover the four
+ * channels.
+ */
+
+ type4 = type;
+ type4.floating = FALSE;
type4.width *= 4;
type4.length /= 4;
LLVMValueRef
-lp_build_swizzle1_aos(struct lp_build_context *bld,
- LLVMValueRef a,
- const unsigned char swizzle[4])
+lp_build_swizzle_aos(struct lp_build_context *bld,
+ LLVMValueRef a,
+ const unsigned char swizzles[4])
{
- const unsigned n = bld->type.length;
+ const struct lp_type type = bld->type;
+ const unsigned n = type.length;
unsigned i, j;
- if(a == bld->undef || a == bld->zero || a == bld->one)
+ if (swizzles[0] == PIPE_SWIZZLE_RED &&
+ swizzles[1] == PIPE_SWIZZLE_GREEN &&
+ swizzles[2] == PIPE_SWIZZLE_BLUE &&
+ swizzles[3] == PIPE_SWIZZLE_ALPHA) {
return a;
+ }
- if(swizzle[0] == swizzle[1] && swizzle[1] == swizzle[2] && swizzle[2] == swizzle[3])
- return lp_build_broadcast_aos(bld, a, swizzle[0]);
+ if (swizzles[0] == swizzles[1] &&
+ swizzles[1] == swizzles[2] &&
+ swizzles[2] == swizzles[3]) {
+ switch (swizzles[0]) {
+ case PIPE_SWIZZLE_RED:
+ case PIPE_SWIZZLE_GREEN:
+ case PIPE_SWIZZLE_BLUE:
+ case PIPE_SWIZZLE_ALPHA:
+ return lp_build_broadcast_aos(bld, a, swizzles[0]);
+ case PIPE_SWIZZLE_ZERO:
+ return bld->zero;
+ case PIPE_SWIZZLE_ONE:
+ return bld->one;
+ default:
+ assert(0);
+ return bld->undef;
+ }
+ }
- {
+ if (type.width >= 16) {
/*
* Shuffle.
*/
- LLVMTypeRef elem_type = LLVMInt32Type();
+ LLVMValueRef undef = LLVMGetUndef(lp_build_elem_type(type));
+ LLVMTypeRef i32t = LLVMInt32Type();
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
+ LLVMValueRef aux[LP_MAX_VECTOR_LENGTH];
+
+ memset(aux, 0, sizeof aux);
+
+ for(j = 0; j < n; j += 4) {
+ for(i = 0; i < 4; ++i) {
+ unsigned shuffle;
+ switch (swizzles[i]) {
+ default:
+ assert(0);
+ /* fall through */
+ case PIPE_SWIZZLE_RED:
+ case PIPE_SWIZZLE_GREEN:
+ case PIPE_SWIZZLE_BLUE:
+ case PIPE_SWIZZLE_ALPHA:
+ shuffle = j + swizzles[i];
+ break;
+ case PIPE_SWIZZLE_ZERO:
+ shuffle = type.length + 0;
+ if (!aux[0]) {
+ aux[0] = lp_build_const_elem(type, 0.0);
+ }
+ break;
+ case PIPE_SWIZZLE_ONE:
+ shuffle = type.length + 1;
+ if (!aux[1]) {
+ aux[1] = lp_build_const_elem(type, 1.0);
+ }
+ break;
+ }
+ shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
+ }
+ }
- for(j = 0; j < n; j += 4)
- for(i = 0; i < 4; ++i)
- shuffles[j + i] = LLVMConstInt(elem_type, j + swizzle[i], 0);
+ for (i = 0; i < n; ++i) {
+ if (!aux[i]) {
+ aux[i] = undef;
+ }
+ }
- return LLVMBuildShuffleVector(bld->builder, a, bld->undef, LLVMConstVector(shuffles, n), "");
+ return LLVMBuildShuffleVector(bld->builder, a,
+ LLVMConstVector(aux, n),
+ LLVMConstVector(shuffles, n), "");
+ } else {
+ /*
+ * Bit mask and shifts.
+ *
+ * For example, this will convert BGRA to RGBA by doing
+ *
+ * rgba = (bgra & 0x00ff0000) >> 16
+ * | (bgra & 0xff00ff00)
+ * | (bgra & 0x000000ff) << 16
+ *
+ * This is necessary not only for faster cause, but because X86 backend
+ * will refuse shuffles of <4 x i8> vectors
+ */
+ LLVMValueRef res;
+ struct lp_type type4;
+ boolean cond[4];
+ unsigned chan;
+ int shift;
+
+ /*
+ * Start with a mixture of 1 and 0.
+ */
+ for (chan = 0; chan < 4; ++chan) {
+ cond[chan] = swizzles[chan] == PIPE_SWIZZLE_ONE ? TRUE : FALSE;
+ }
+ res = lp_build_select_aos(bld, bld->one, bld->zero, cond);
+
+ /*
+ * Build a type where each element is an integer that cover the four
+ * channels.
+ */
+ type4 = type;
+ type4.floating = FALSE;
+ type4.width *= 4;
+ type4.length /= 4;
+
+ a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(type4), "");
+ res = LLVMBuildBitCast(bld->builder, res, lp_build_vec_type(type4), "");
+
+ /*
+ * Mask and shift the channels, trying to group as many channels in the
+ * same shift as possible
+ */
+ for (shift = -3; shift <= 3; ++shift) {
+ unsigned long long mask = 0;
+
+ assert(type4.width <= sizeof(mask)*8);
+
+ for (chan = 0; chan < 4; ++chan) {
+ /* FIXME: big endian */
+ if (swizzles[chan] < 4 &&
+ chan - swizzles[chan] == shift) {
+ mask |= ((1ULL << type.width) - 1) << (swizzles[chan] * type.width);
+ }
+ }
+
+ if (mask) {
+ LLVMValueRef masked;
+ LLVMValueRef shifted;
+
+ if (0)
+ debug_printf("shift = %i, mask = 0x%08llx\n", shift, mask);
+
+ masked = LLVMBuildAnd(bld->builder, a,
+ lp_build_const_int_vec(type4, mask), "");
+ if (shift > 0) {
+ shifted = LLVMBuildShl(bld->builder, masked,
+ lp_build_const_int_vec(type4, shift*type.width), "");
+ } else if (shift < 0) {
+ shifted = LLVMBuildLShr(bld->builder, masked,
+ lp_build_const_int_vec(type4, -shift*type.width), "");
+ } else {
+ shifted = masked;
+ }
+
+ res = LLVMBuildOr(bld->builder, res, shifted, "");
+ }
+ }
+
+ return LLVMBuildBitCast(bld->builder, res, lp_build_vec_type(type), "");
}
}
projects / mesa.git / commitdiff