summaryrefslogtreecommitdiffstats
path: root/src/gallium
diff options
context:
space:
mode:
authorRoland Scheidegger <[email protected]>2010-09-24 15:02:24 +0200
committerJosé Fonseca <[email protected]>2010-09-25 12:19:31 +0100
commit46d05d4ef99857e50d978247917f3e16574418f4 (patch)
treefdb7bd877ce5cac758448196873375a51d6934ef /src/gallium
parent26dc60d0a32f3e5b8084fda5991b762a721662e8 (diff)
gallivm: don't use URem/UDiv when calculating offsets for blocks
While it's true that llvm can and will indeed replace this with bit arithmetic (since block height/width is POT), it does so (llvm 2.7) by element and hence extracts/shifts/reinserts each element individually. This costs about 16 instructions (and extract is not really fast) vs. 1...
Diffstat (limited to 'src/gallium')
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_sample.c12
1 files changed, 11 insertions, 1 deletions
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample.c b/src/gallium/auxiliary/gallivm/lp_bld_sample.c
index 19e380a8dce..44f44ff1aa4 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample.c
@@ -655,11 +655,21 @@ lp_build_sample_partial_offset(struct lp_build_context *bld,
* Pixel blocks have power of two dimensions. LLVM should convert the
* rem/div to bit arithmetic.
* TODO: Verify this.
+ * It does indeed BUT it does transform it to scalar (and back) when doing so
+ * (using roughly extract, shift/and, mov, unpack) (llvm 2.7).
+ * The generated code looks seriously unfunny and is quite expensive.
*/
-
+#if 0
LLVMValueRef block_width = lp_build_const_int_vec(bld->type, block_length);
subcoord = LLVMBuildURem(bld->builder, coord, block_width, "");
coord = LLVMBuildUDiv(bld->builder, coord, block_width, "");
+#else
+ unsigned logbase2 = util_unsigned_logbase2(block_length);
+ LLVMValueRef block_shift = lp_build_const_int_vec(bld->type, logbase2);
+ LLVMValueRef block_mask = lp_build_const_int_vec(bld->type, block_length - 1);
+ subcoord = LLVMBuildAnd(bld->builder, coord, block_mask, "");
+ coord = LLVMBuildLShr(bld->builder, coord, block_shift, "");
+#endif
}
offset = lp_build_mul(bld, coord, stride);