[Mesa-dev] [PATCH 3/3] gallivm: optimize repeat linear npot code in the aos int path
Roland Scheidegger
sroland at vmware.com
Fri Feb 14 17:06:39 PST 2014
FWIW I've just cleaned 1/3 and 3/3 up a little and splitted it off into
two patches (I really want to be able to track any changes this might
cause separately), and on x86 sse I actually managed to shave off one
instruction by using lp_build_iround() too :-).
2/3 is more of the same just for the nearest filtering path.
In any case though I haven't actually tested any of it yet but the issue
indeed looks very real to me. I actually need to really run some
internal tests with this (piglit is usually not close to sensitive
enough), the whole texture wrap mode stuff is a bit of a nightmare (as
entirely different paths will be run depending on cpu flags AND texture
format which makes bugs in there difficult to detect). At some point I
wanted to unify the coord wrapping in the aos and soa paths since this
doesn't really depend on if aos or soa filtering is used though there
are indeed some dependencies if you want to get optimal code.
Roland
Am 15.02.2014 01:54, schrieb sroland at vmware.com:
> From: Jeff Muizelaar <jmuizelaar at mozilla.com>
>
> Similar to the other cases, shift some weight/coord calculations to int
> space. This should be slightly faster (on x86 sse it should actually safe one
> instruction, and generally int instructions are cheaper).
> ---
> src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c | 74 +++++++++++++++++----
> 1 file changed, 62 insertions(+), 12 deletions(-)
>
> diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c
> index 03a2ed5..e9f8611 100644
> --- a/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c
> +++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c
> @@ -194,6 +194,62 @@ lp_build_sample_wrap_nearest_float(struct lp_build_sample_context *bld,
>
>
> /**
> + * Helper to compute the first coord and the weight for
> + * linear wrap repeat npot textures
> + */
> +static void
> +lp_build_coord_repeat_npot_linear_int(struct lp_build_sample_context *bld,
> + LLVMValueRef coord_f,
> + LLVMValueRef length_i,
> + LLVMValueRef length_f,
> + LLVMValueRef *coord0_i,
> + LLVMValueRef *weight_i)
> +{
> + struct lp_build_context *coord_bld = &bld->coord_bld;
> + struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
> + struct lp_build_context abs_coord_bld;
> + struct lp_type abs_type;
> + LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length_i,
> + int_coord_bld->one);
> + LLVMValueRef mask, i32_c8, i32_c128, i32_c255;
> +
> + /* wrap with normalized floats is just fract */
> + coord_f = lp_build_fract(coord_bld, coord_f);
> + /* mul by size */
> + coord_f = lp_build_mul(coord_bld, coord_f, length_f);
> + /* convert to int, compute lerp weight */
> + coord_f = lp_build_mul_imm(&bld->coord_bld, coord_f, 256);
> +
> + /* At this point we don't have any negative numbers so use non-signed
> + * build context which might help on some archs.
> + */
> + abs_type = coord_bld->type;
> + abs_type.sign = 0;
> + lp_build_context_init(&abs_coord_bld, bld->gallivm, abs_type);
> + *coord0_i = lp_build_iround(&abs_coord_bld, coord_f);
> +
> + /* subtract 0.5 (add -128) */
> + i32_c128 = lp_build_const_int_vec(bld->gallivm, bld->int_coord_type, -128);
> + *coord0_i = LLVMBuildAdd(bld->gallivm->builder, *coord0_i, i32_c128, "");
> +
> + /* compute fractional part (AND with 0xff) */
> + i32_c255 = lp_build_const_int_vec(bld->gallivm, bld->int_coord_type, 255);
> + *weight_i = LLVMBuildAnd(bld->gallivm->builder, *coord0_i, i32_c255, "");
> +
> + /* compute floor (shift right 8) */
> + i32_c8 = lp_build_const_int_vec(bld->gallivm, bld->int_coord_type, 8);
> + *coord0_i = LLVMBuildAShr(bld->gallivm->builder, *coord0_i, i32_c8, "");
> + /*
> + * we avoided the 0.5/length division before the repeat wrap,
> + * now need to fix up edge cases with selects
> + */
> + mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
> + PIPE_FUNC_LESS, *coord0_i, int_coord_bld->zero);
> + *coord0_i = lp_build_select(int_coord_bld, mask, length_minus_one, *coord0_i);
> +}
> +
> +
> +/**
> * Build LLVM code for texture coord wrapping, for linear filtering,
> * for scaled integer texcoords.
> * \param block_length is the length of the pixel block along the
> @@ -251,24 +307,21 @@ lp_build_sample_wrap_linear_int(struct lp_build_sample_context *bld,
> }
> else {
> LLVMValueRef mask;
> - LLVMValueRef weight;
> LLVMValueRef length_f = lp_build_int_to_float(&bld->coord_bld, length);
> if (offset) {
> offset = lp_build_int_to_float(&bld->coord_bld, offset);
> offset = lp_build_div(&bld->coord_bld, offset, length_f);
> coord_f = lp_build_add(&bld->coord_bld, coord_f, offset);
> }
> - lp_build_coord_repeat_npot_linear(bld, coord_f,
> - length, length_f,
> - &coord0, &weight);
> + lp_build_coord_repeat_npot_linear_int(bld, coord_f,
> + length, length_f,
> + &coord0, weight_i);
> mask = lp_build_compare(bld->gallivm, int_coord_bld->type,
> PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
> coord1 = LLVMBuildAnd(builder,
> lp_build_add(int_coord_bld, coord0,
> int_coord_bld->one),
> mask, "");
> - weight = lp_build_mul_imm(&bld->coord_bld, weight, 256);
> - *weight_i = lp_build_itrunc(&bld->coord_bld, weight);
> }
> break;
>
> @@ -308,18 +361,15 @@ lp_build_sample_wrap_linear_int(struct lp_build_sample_context *bld,
> coord0 = LLVMBuildAnd(builder, coord0, length_minus_one, "");
> }
> else {
> - LLVMValueRef weight;
> LLVMValueRef length_f = lp_build_int_to_float(&bld->coord_bld, length);
> if (offset) {
> offset = lp_build_int_to_float(&bld->coord_bld, offset);
> offset = lp_build_div(&bld->coord_bld, offset, length_f);
> coord_f = lp_build_add(&bld->coord_bld, coord_f, offset);
> }
> - lp_build_coord_repeat_npot_linear(bld, coord_f,
> - length, length_f,
> - &coord0, &weight);
> - weight = lp_build_mul_imm(&bld->coord_bld, weight, 256);
> - *weight_i = lp_build_itrunc(&bld->coord_bld, weight);
> + lp_build_coord_repeat_npot_linear_int(bld, coord_f,
> + length, length_f,
> + &coord0, weight_i);
> }
>
> mask = lp_build_compare(bld->gallivm, int_coord_bld->type,
>
More information about the mesa-dev
mailing list