summaryrefslogtreecommitdiff
path: root/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium/auxiliary/gallivm/lp_bld_format_soa.c')
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_format_soa.c91
1 files changed, 77 insertions, 14 deletions
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c
index e1b94adc85..9f405921b0 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c
@@ -36,7 +36,7 @@
#include "lp_bld_const.h"
#include "lp_bld_conv.h"
#include "lp_bld_swizzle.h"
-#include "lp_bld_sample.h" /* for lp_build_gather */
+#include "lp_bld_gather.h"
#include "lp_bld_format.h"
@@ -251,6 +251,41 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
}
+void
+lp_build_rgba8_to_f32_soa(LLVMBuilderRef builder,
+ struct lp_type dst_type,
+ LLVMValueRef packed,
+ LLVMValueRef *rgba)
+{
+ LLVMValueRef mask = lp_build_const_int_vec(dst_type, 0xff);
+ unsigned chan;
+
+ packed = LLVMBuildBitCast(builder, packed,
+ lp_build_int_vec_type(dst_type), "");
+
+ /* Decode the input vector components */
+ for (chan = 0; chan < 4; ++chan) {
+ unsigned start = chan*8;
+ unsigned stop = start + 8;
+ LLVMValueRef input;
+
+ input = packed;
+
+ if (start)
+ input = LLVMBuildLShr(builder, input,
+ lp_build_const_int_vec(dst_type, start), "");
+
+ if (stop < 32)
+ input = LLVMBuildAnd(builder, input, mask, "");
+
+ input = lp_build_unsigned_norm_to_float(builder, 8, dst_type, input);
+
+ rgba[chan] = input;
+ }
+}
+
+
+
/**
* Fetch a texels from a texture, returning them in SoA layout.
*
@@ -311,20 +346,49 @@ lp_build_fetch_rgba_soa(LLVMBuilderRef builder,
format_desc,
type,
packed, rgba_out);
+ return;
}
- else {
- /*
- * 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.
- */
+ /*
+ * Try calling lp_build_fetch_rgba_aos for all pixels.
+ */
+
+ if (util_format_fits_8unorm(format_desc) &&
+ type.floating && type.width == 32 && type.length == 4) {
+ struct lp_type tmp_type;
+ LLVMValueRef tmp;
+
+ memset(&tmp_type, 0, sizeof tmp_type);
+ tmp_type.width = 8;
+ tmp_type.length = type.length * 4;
+ tmp_type.norm = TRUE;
+
+ tmp = lp_build_fetch_rgba_aos(builder, format_desc, tmp_type,
+ base_ptr, offset, i, j);
+
+ lp_build_rgba8_to_f32_soa(builder,
+ type,
+ tmp,
+ rgba_out);
+
+ 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.
+ */
+
+ {
unsigned k, chan;
+ struct lp_type tmp_type;
- assert(type.floating);
+ tmp_type = type;
+ tmp_type.length = 4;
for (chan = 0; chan < 4; ++chan) {
rgba_out[chan] = lp_build_undef(type);
@@ -334,18 +398,17 @@ lp_build_fetch_rgba_soa(LLVMBuilderRef builder,
for(k = 0; k < type.length; ++k) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), k, 0);
LLVMValueRef offset_elem;
- LLVMValueRef ptr;
LLVMValueRef i_elem, j_elem;
LLVMValueRef tmp;
offset_elem = LLVMBuildExtractElement(builder, offset, index, "");
- ptr = LLVMBuildGEP(builder, base_ptr, &offset_elem, 1, "");
i_elem = LLVMBuildExtractElement(builder, i, index, "");
j_elem = LLVMBuildExtractElement(builder, j, index, "");
/* Get a single float[4]={R,G,B,A} pixel */
- tmp = lp_build_fetch_rgba_aos(builder, format_desc, ptr,
+ tmp = lp_build_fetch_rgba_aos(builder, format_desc, tmp_type,
+ base_ptr, offset_elem,
i_elem, j_elem);
/*