X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fauxiliary%2Fgallivm%2Flp_bld_sample_soa.c;h=2a4462b2428128211c6fd856274bf66a0c9fb5fd;hb=c6c55ad3e967f3d151c24795a99634b297c13fde;hp=53cc0c5f34541c7c897a7e4509247ef1283fff72;hpb=0072acd447dc6be652e63752e50215c3105322c8;p=mesa.git diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c index 53cc0c5f345..2a4462b2428 100644 --- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c +++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c @@ -35,11 +35,13 @@ #include "pipe/p_defines.h" #include "pipe/p_state.h" +#include "pipe/p_shader_tokens.h" #include "util/u_debug.h" #include "util/u_dump.h" #include "util/u_memory.h" #include "util/u_math.h" #include "util/u_format.h" +#include "util/u_cpu_detect.h" #include "lp_bld_debug.h" #include "lp_bld_type.h" #include "lp_bld_const.h" @@ -56,6 +58,7 @@ #include "lp_bld_sample_aos.h" #include "lp_bld_struct.h" #include "lp_bld_quad.h" +#include "lp_bld_pack.h" /** @@ -69,7 +72,7 @@ */ static void lp_build_sample_texel_soa(struct lp_build_sample_context *bld, - unsigned unit, + unsigned sampler_unit, LLVMValueRef width, LLVMValueRef height, LLVMValueRef depth, @@ -79,11 +82,13 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, LLVMValueRef y_stride, LLVMValueRef z_stride, LLVMValueRef data_ptr, + LLVMValueRef mipoffsets, LLVMValueRef texel_out[4]) { - const struct lp_sampler_static_state *static_state = bld->static_state; + const struct lp_static_sampler_state *static_state = bld->static_sampler_state; const unsigned dims = bld->dims; struct lp_build_context *int_coord_bld = &bld->int_coord_bld; + LLVMBuilderRef builder = bld->gallivm->builder; LLVMValueRef offset; LLVMValueRef i, j; LLVMValueRef use_border = NULL; @@ -95,7 +100,7 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, LLVMValueRef b1, b2; b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero); b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width); - use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2"); + use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2"); } if (dims >= 2 && @@ -106,11 +111,11 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero); b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height); if (use_border) { - use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1"); - use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2"); + use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1"); + use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2"); } else { - use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2"); + use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2"); } } @@ -122,11 +127,11 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero); b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth); if (use_border) { - use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1"); - use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2"); + use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1"); + use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2"); } else { - use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2"); + use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2"); } } @@ -135,6 +140,9 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, bld->format_desc, x, y, z, y_stride, z_stride, &offset, &i, &j); + if (mipoffsets) { + offset = lp_build_add(&bld->int_coord_bld, offset, mipoffsets); + } if (use_border) { /* If we can sample the border color, it means that texcoords may @@ -148,7 +156,7 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, offset = lp_build_andnot(&bld->int_coord_bld, offset, use_border); } - lp_build_fetch_rgba_soa(bld->builder, + lp_build_fetch_rgba_soa(bld->gallivm, bld->format_desc, bld->texel_type, data_ptr, offset, @@ -171,23 +179,66 @@ lp_build_sample_texel_soa(struct lp_build_sample_context *bld, */ if (use_border) { - /* select texel color or border color depending on use_border */ - LLVMValueRef border_color_ptr = + /* select texel color or border color depending on use_border. */ + LLVMValueRef border_color_ptr = bld->dynamic_state->border_color(bld->dynamic_state, - bld->builder, unit); + bld->gallivm, sampler_unit); + const struct util_format_description *format_desc; int chan; + format_desc = util_format_description(bld->static_texture_state->format); + /* + * Only replace channels which are actually present. The others should + * get optimized away eventually by sampler_view swizzle anyway but it's + * easier too as we'd need some extra logic for channels where we can't + * determine the format directly otherwise. + */ for (chan = 0; chan < 4; chan++) { - LLVMValueRef border_chan = - lp_build_array_get(bld->builder, border_color_ptr, - lp_build_const_int32(chan)); - LLVMValueRef border_chan_vec = - lp_build_broadcast_scalar(&bld->float_vec_bld, border_chan); - texel_out[chan] = lp_build_select(&bld->texel_bld, use_border, - border_chan_vec, texel_out[chan]); + unsigned chan_s; + /* reverse-map channel... */ + for (chan_s = 0; chan_s < 4; chan_s++) { + if (chan_s == format_desc->swizzle[chan]) { + break; + } + } + if (chan_s <= 3) { + LLVMValueRef border_chan = + lp_build_array_get(bld->gallivm, border_color_ptr, + lp_build_const_int32(bld->gallivm, chan)); + LLVMValueRef border_chan_vec = + lp_build_broadcast_scalar(&bld->float_vec_bld, border_chan); + + if (!bld->texel_type.floating) { + border_chan_vec = LLVMBuildBitCast(builder, border_chan_vec, + bld->texel_bld.vec_type, ""); + } + else { + /* + * For normalized format need to clamp border color (technically + * probably should also quantize the data). Really sucks doing this + * here but can't avoid at least for now since this is part of + * sampler state and texture format is part of sampler_view state. + */ + unsigned chan_type = format_desc->channel[chan_s].type; + unsigned chan_norm = format_desc->channel[chan_s].normalized; + if (chan_type == UTIL_FORMAT_TYPE_SIGNED && chan_norm) { + LLVMValueRef clamp_min; + clamp_min = lp_build_const_vec(bld->gallivm, bld->texel_type, -1.0F); + border_chan_vec = lp_build_clamp(&bld->texel_bld, border_chan_vec, + clamp_min, + bld->texel_bld.one); + } + else if (chan_type == UTIL_FORMAT_TYPE_UNSIGNED && chan_norm) { + border_chan_vec = lp_build_clamp(&bld->texel_bld, border_chan_vec, + bld->texel_bld.zero, + bld->texel_bld.one); + } + /* not exactly sure about all others but I think should be ok? */ + } + texel_out[chan] = lp_build_select(&bld->texel_bld, use_border, + border_chan_vec, texel_out[chan]); + } } } - - apply_sampler_swizzle(bld, texel_out); } @@ -205,7 +256,7 @@ lp_build_coord_mirror(struct lp_build_sample_context *bld, lp_build_ifloor_fract(coord_bld, coord, &flr, &fract); /* isOdd = flr & 1 */ - isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, ""); + isOdd = LLVMBuildAnd(bld->gallivm->builder, flr, int_coord_bld->one, ""); /* make coord positive or negative depending on isOdd */ coord = lp_build_set_sign(coord_bld, fract, isOdd); @@ -220,6 +271,41 @@ lp_build_coord_mirror(struct lp_build_sample_context *bld, } +/** + * Helper to compute the first coord and the weight for + * linear wrap repeat npot textures + */ +void +lp_build_coord_repeat_npot_linear(struct lp_build_sample_context *bld, + LLVMValueRef coord_f, + LLVMValueRef length_i, + LLVMValueRef length_f, + LLVMValueRef *coord0_i, + LLVMValueRef *weight_f) +{ + struct lp_build_context *coord_bld = &bld->coord_bld; + struct lp_build_context *int_coord_bld = &bld->int_coord_bld; + LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5); + LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length_i, + int_coord_bld->one); + LLVMValueRef mask; + /* wrap with normalized floats is just fract */ + coord_f = lp_build_fract(coord_bld, coord_f); + /* mul by size and subtract 0.5 */ + coord_f = lp_build_mul(coord_bld, coord_f, length_f); + coord_f = lp_build_sub(coord_bld, coord_f, half); + /* + * we avoided the 0.5/length division before the repeat wrap, + * now need to fix up edge cases with selects + */ + /* convert to int, compute lerp weight */ + lp_build_ifloor_fract(coord_bld, coord_f, coord0_i, weight_f); + 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 wrap mode for linear filtering. * \param x0_out returns first integer texcoord @@ -231,6 +317,7 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, LLVMValueRef coord, LLVMValueRef length, LLVMValueRef length_f, + LLVMValueRef offset, boolean is_pot, unsigned wrap_mode, LLVMValueRef *x0_out, @@ -239,42 +326,55 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, { struct lp_build_context *coord_bld = &bld->coord_bld; struct lp_build_context *int_coord_bld = &bld->int_coord_bld; - LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5); + LLVMBuilderRef builder = bld->gallivm->builder; + LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5); LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one); LLVMValueRef coord0, coord1, weight; switch(wrap_mode) { case PIPE_TEX_WRAP_REPEAT: - /* mul by size and subtract 0.5 */ - coord = lp_build_mul(coord_bld, coord, length_f); - coord = lp_build_sub(coord_bld, coord, half); - /* convert to int, compute lerp weight */ - lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight); - /* repeat wrap */ if (is_pot) { + /* mul by size and subtract 0.5 */ + coord = lp_build_mul(coord_bld, coord, length_f); + coord = lp_build_sub(coord_bld, coord, half); + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + coord = lp_build_add(coord_bld, coord, offset); + } + /* convert to int, compute lerp weight */ + lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight); coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one); - coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, ""); - coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, ""); + /* repeat wrap */ + coord0 = LLVMBuildAnd(builder, coord0, length_minus_one, ""); + coord1 = LLVMBuildAnd(builder, coord1, length_minus_one, ""); } else { - /* Add a bias to the texcoord to handle negative coords */ - LLVMValueRef bias = lp_build_mul_imm(int_coord_bld, length, 1024); LLVMValueRef mask; - coord0 = LLVMBuildAdd(bld->builder, coord0, bias, ""); - coord0 = LLVMBuildURem(bld->builder, coord0, length, ""); - mask = lp_build_compare(bld->builder, int_coord_bld->type, + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + offset = lp_build_div(coord_bld, offset, length_f); + coord = lp_build_add(coord_bld, coord, offset); + } + lp_build_coord_repeat_npot_linear(bld, coord, + length, length_f, + &coord0, &weight); + mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type, PIPE_FUNC_NOTEQUAL, coord0, length_minus_one); - coord1 = LLVMBuildAnd(bld->builder, - lp_build_add(int_coord_bld, coord0, int_coord_bld->one), - mask, ""); + coord1 = LLVMBuildAnd(builder, + lp_build_add(int_coord_bld, coord0, int_coord_bld->one), + mask, ""); } break; case PIPE_TEX_WRAP_CLAMP: - if (bld->static_state->normalized_coords) { + if (bld->static_sampler_state->normalized_coords) { /* scale coord to length */ coord = lp_build_mul(coord_bld, coord, length_f); } + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + coord = lp_build_add(coord_bld, coord, offset); + } /* clamp to [0, length] */ coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f); @@ -291,10 +391,15 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, struct lp_build_context abs_coord_bld = bld->coord_bld; abs_coord_bld.type.sign = FALSE; - if (bld->static_state->normalized_coords) { + if (bld->static_sampler_state->normalized_coords) { /* mul by tex size */ coord = lp_build_mul(coord_bld, coord, length_f); } + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + coord = lp_build_add(coord_bld, coord, offset); + } + /* clamp to length max */ coord = lp_build_min(coord_bld, coord, length_f); /* subtract 0.5 */ @@ -310,20 +415,20 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, } case PIPE_TEX_WRAP_CLAMP_TO_BORDER: - { - LLVMValueRef min; - if (bld->static_state->normalized_coords) { - /* scale coord to length */ - coord = lp_build_mul(coord_bld, coord, length_f); - } - /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */ - coord = lp_build_sub(coord_bld, coord, half); - min = lp_build_const_vec(coord_bld->type, -1.0F); - coord = lp_build_clamp(coord_bld, coord, min, length_f); - /* convert to int, compute lerp weight */ - lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight); - coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one); + if (bld->static_sampler_state->normalized_coords) { + /* scale coord to length */ + coord = lp_build_mul(coord_bld, coord, length_f); } + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + coord = lp_build_add(coord_bld, coord, offset); + } + /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */ + /* can skip clamp (though might not work for very large coord values */ + coord = lp_build_sub(coord_bld, coord, half); + /* convert to int, compute lerp weight */ + lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight); + coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one); break; case PIPE_TEX_WRAP_MIRROR_REPEAT: @@ -333,6 +438,10 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, /* scale coord to length */ coord = lp_build_mul(coord_bld, coord, length_f); coord = lp_build_sub(coord_bld, coord, half); + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + coord = lp_build_add(coord_bld, coord, offset); + } /* convert to int, compute lerp weight */ lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight); @@ -345,12 +454,15 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, break; case PIPE_TEX_WRAP_MIRROR_CLAMP: - coord = lp_build_abs(coord_bld, coord); - - if (bld->static_state->normalized_coords) { + if (bld->static_sampler_state->normalized_coords) { /* scale coord to length */ coord = lp_build_mul(coord_bld, coord, length_f); } + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + coord = lp_build_add(coord_bld, coord, offset); + } + coord = lp_build_abs(coord_bld, coord); /* clamp to [0, length] */ coord = lp_build_min(coord_bld, coord, length_f); @@ -364,42 +476,50 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE: { - LLVMValueRef min, max; struct lp_build_context abs_coord_bld = bld->coord_bld; abs_coord_bld.type.sign = FALSE; - coord = lp_build_abs(coord_bld, coord); - if (bld->static_state->normalized_coords) { + if (bld->static_sampler_state->normalized_coords) { /* scale coord to length */ coord = lp_build_mul(coord_bld, coord, length_f); } + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + coord = lp_build_add(coord_bld, coord, offset); + } + coord = lp_build_abs(coord_bld, coord); - /* clamp to [0.5, length - 0.5] */ - min = half; - max = lp_build_sub(coord_bld, length_f, min); - coord = lp_build_clamp(coord_bld, coord, min, max); - + /* clamp to length max */ + coord = lp_build_min(coord_bld, coord, length_f); + /* subtract 0.5 */ coord = lp_build_sub(coord_bld, coord, half); + /* clamp to [0, length - 0.5] */ + coord = lp_build_max(coord_bld, coord, coord_bld->zero); /* convert to int, compute lerp weight */ lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight); coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one); + /* coord1 = min(coord1, length-1) */ + coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one); } break; case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER: { - coord = lp_build_abs(coord_bld, coord); - - if (bld->static_state->normalized_coords) { + if (bld->static_sampler_state->normalized_coords) { /* scale coord to length */ coord = lp_build_mul(coord_bld, coord, length_f); } + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + coord = lp_build_add(coord_bld, coord, offset); + } + coord = lp_build_abs(coord_bld, coord); /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */ - /* skip -0.5 clamp (always positive), do sub first */ + /* skip clamp - always positive, and other side + only potentially matters for very large coords */ coord = lp_build_sub(coord_bld, coord, half); - coord = lp_build_min(coord_bld, coord, length_f); /* convert to int, compute lerp weight */ lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight); @@ -424,6 +544,8 @@ lp_build_sample_wrap_linear(struct lp_build_sample_context *bld, * Build LLVM code for texture wrap mode for nearest filtering. * \param coord the incoming texcoord (nominally in [0,1]) * \param length the texture size along one dimension, as int vector + * \param length_f the texture size along one dimension, as float vector + * \param offset texel offset along one dimension (as int vector) * \param is_pot if TRUE, length is a power of two * \param wrap_mode one of PIPE_TEX_WRAP_x */ @@ -432,31 +554,42 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, LLVMValueRef coord, LLVMValueRef length, LLVMValueRef length_f, + LLVMValueRef offset, boolean is_pot, unsigned wrap_mode) { struct lp_build_context *coord_bld = &bld->coord_bld; struct lp_build_context *int_coord_bld = &bld->int_coord_bld; + LLVMBuilderRef builder = bld->gallivm->builder; LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one); LLVMValueRef icoord; switch(wrap_mode) { case PIPE_TEX_WRAP_REPEAT: - coord = lp_build_mul(coord_bld, coord, length_f); - icoord = lp_build_ifloor(coord_bld, coord); - if (is_pot) - icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, ""); + if (is_pot) { + coord = lp_build_mul(coord_bld, coord, length_f); + icoord = lp_build_ifloor(coord_bld, coord); + if (offset) { + icoord = lp_build_add(int_coord_bld, icoord, offset); + } + icoord = LLVMBuildAnd(builder, icoord, length_minus_one, ""); + } else { - /* Add a bias to the texcoord to handle negative coords */ - LLVMValueRef bias = lp_build_mul_imm(int_coord_bld, length, 1024); - icoord = LLVMBuildAdd(bld->builder, icoord, bias, ""); - icoord = LLVMBuildURem(bld->builder, icoord, length, ""); + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + offset = lp_build_div(coord_bld, offset, length_f); + coord = lp_build_add(coord_bld, coord, offset); + } + /* take fraction, unnormalize */ + coord = lp_build_fract_safe(coord_bld, coord); + coord = lp_build_mul(coord_bld, coord, length_f); + icoord = lp_build_itrunc(coord_bld, coord); } break; case PIPE_TEX_WRAP_CLAMP: case PIPE_TEX_WRAP_CLAMP_TO_EDGE: - if (bld->static_state->normalized_coords) { + if (bld->static_sampler_state->normalized_coords) { /* scale coord to length */ coord = lp_build_mul(coord_bld, coord, length_f); } @@ -464,6 +597,9 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, /* floor */ /* use itrunc instead since we clamp to 0 anyway */ icoord = lp_build_itrunc(coord_bld, coord); + if (offset) { + icoord = lp_build_add(int_coord_bld, icoord, offset); + } /* clamp to [0, length - 1]. */ icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero, @@ -471,30 +607,28 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, break; case PIPE_TEX_WRAP_CLAMP_TO_BORDER: - /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */ - { - LLVMValueRef min, max; - - if (bld->static_state->normalized_coords) { - /* scale coord to length */ - coord = lp_build_mul(coord_bld, coord, length_f); - } - - icoord = lp_build_ifloor(coord_bld, coord); - - /* clamp to [-1, length] */ - min = lp_build_negate(int_coord_bld, int_coord_bld->one); - max = length; - icoord = lp_build_clamp(int_coord_bld, icoord, min, max); + if (bld->static_sampler_state->normalized_coords) { + /* scale coord to length */ + coord = lp_build_mul(coord_bld, coord, length_f); + } + /* no clamp necessary, border masking will handle this */ + icoord = lp_build_ifloor(coord_bld, coord); + if (offset) { + icoord = lp_build_add(int_coord_bld, icoord, offset); } break; case PIPE_TEX_WRAP_MIRROR_REPEAT: + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + offset = lp_build_div(coord_bld, offset, length_f); + coord = lp_build_add(coord_bld, coord, offset); + } /* compute mirror function */ coord = lp_build_coord_mirror(bld, coord); /* scale coord to length */ - assert(bld->static_state->normalized_coords); + assert(bld->static_sampler_state->normalized_coords); coord = lp_build_mul(coord_bld, coord, length_f); /* itrunc == ifloor here */ @@ -506,12 +640,15 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, case PIPE_TEX_WRAP_MIRROR_CLAMP: case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE: - coord = lp_build_abs(coord_bld, coord); - - if (bld->static_state->normalized_coords) { + if (bld->static_sampler_state->normalized_coords) { /* scale coord to length */ coord = lp_build_mul(coord_bld, coord, length_f); } + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + coord = lp_build_add(coord_bld, coord, offset); + } + coord = lp_build_abs(coord_bld, coord); /* itrunc == ifloor here */ icoord = lp_build_itrunc(coord_bld, coord); @@ -521,18 +658,18 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, break; case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER: - coord = lp_build_abs(coord_bld, coord); - - if (bld->static_state->normalized_coords) { + if (bld->static_sampler_state->normalized_coords) { /* scale coord to length */ coord = lp_build_mul(coord_bld, coord, length_f); } + if (offset) { + offset = lp_build_int_to_float(coord_bld, offset); + coord = lp_build_add(coord_bld, coord, offset); + } + coord = lp_build_abs(coord_bld, coord); /* itrunc == ifloor here */ icoord = lp_build_itrunc(coord_bld, coord); - - /* clamp to [0, length] */ - icoord = lp_build_min(int_coord_bld, icoord, length); break; default: @@ -550,14 +687,16 @@ lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld, */ static void lp_build_sample_image_nearest(struct lp_build_sample_context *bld, - unsigned unit, + unsigned sampler_unit, LLVMValueRef size, LLVMValueRef row_stride_vec, LLVMValueRef img_stride_vec, LLVMValueRef data_ptr, + LLVMValueRef mipoffsets, LLVMValueRef s, LLVMValueRef t, LLVMValueRef r, + const LLVMValueRef *offsets, LLVMValueRef colors_out[4]) { const unsigned dims = bld->dims; @@ -568,10 +707,10 @@ lp_build_sample_image_nearest(struct lp_build_sample_context *bld, LLVMValueRef flt_width_vec; LLVMValueRef flt_height_vec; LLVMValueRef flt_depth_vec; - LLVMValueRef x, y, z; + LLVMValueRef x, y = NULL, z = NULL; lp_build_extract_image_sizes(bld, - bld->int_size_type, + &bld->int_size_bld, bld->int_coord_type, size, &width_vec, &height_vec, &depth_vec); @@ -579,7 +718,7 @@ lp_build_sample_image_nearest(struct lp_build_sample_context *bld, flt_size = lp_build_int_to_float(&bld->float_size_bld, size); lp_build_extract_image_sizes(bld, - bld->float_size_type, + &bld->float_size_bld, bld->coord_type, flt_size, &flt_width_vec, &flt_height_vec, &flt_depth_vec); @@ -587,42 +726,39 @@ lp_build_sample_image_nearest(struct lp_build_sample_context *bld, /* * Compute integer texcoords. */ - x = lp_build_sample_wrap_nearest(bld, s, width_vec, flt_width_vec, - bld->static_state->pot_width, - bld->static_state->wrap_s); + x = lp_build_sample_wrap_nearest(bld, s, width_vec, flt_width_vec, offsets[0], + bld->static_texture_state->pot_width, + bld->static_sampler_state->wrap_s); lp_build_name(x, "tex.x.wrapped"); if (dims >= 2) { - y = lp_build_sample_wrap_nearest(bld, t, height_vec, flt_height_vec, - bld->static_state->pot_height, - bld->static_state->wrap_t); + y = lp_build_sample_wrap_nearest(bld, t, height_vec, flt_height_vec, offsets[1], + bld->static_texture_state->pot_height, + bld->static_sampler_state->wrap_t); lp_build_name(y, "tex.y.wrapped"); if (dims == 3) { - z = lp_build_sample_wrap_nearest(bld, r, depth_vec, flt_depth_vec, - bld->static_state->pot_depth, - bld->static_state->wrap_r); + z = lp_build_sample_wrap_nearest(bld, r, depth_vec, flt_depth_vec, offsets[2], + bld->static_texture_state->pot_depth, + bld->static_sampler_state->wrap_r); lp_build_name(z, "tex.z.wrapped"); } - else if (bld->static_state->target == PIPE_TEXTURE_CUBE) { - z = r; - } - else { - z = NULL; - } } - else { - y = z = NULL; + if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE || + bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY || + bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) { + z = r; + lp_build_name(z, "tex.z.layer"); } /* * Get texture colors. */ - lp_build_sample_texel_soa(bld, unit, + lp_build_sample_texel_soa(bld, sampler_unit, width_vec, height_vec, depth_vec, x, y, z, row_stride_vec, img_stride_vec, - data_ptr, colors_out); + data_ptr, mipoffsets, colors_out); } @@ -632,14 +768,16 @@ lp_build_sample_image_nearest(struct lp_build_sample_context *bld, */ static void lp_build_sample_image_linear(struct lp_build_sample_context *bld, - unsigned unit, + unsigned sampler_unit, LLVMValueRef size, LLVMValueRef row_stride_vec, LLVMValueRef img_stride_vec, LLVMValueRef data_ptr, + LLVMValueRef mipoffsets, LLVMValueRef s, LLVMValueRef t, LLVMValueRef r, + const LLVMValueRef *offsets, LLVMValueRef colors_out[4]) { const unsigned dims = bld->dims; @@ -650,13 +788,13 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, LLVMValueRef flt_width_vec; LLVMValueRef flt_height_vec; LLVMValueRef flt_depth_vec; - LLVMValueRef x0, y0, z0, x1, y1, z1; - LLVMValueRef s_fpart, t_fpart, r_fpart; + LLVMValueRef x0, y0 = NULL, z0 = NULL, x1, y1 = NULL, z1 = NULL; + LLVMValueRef s_fpart, t_fpart = NULL, r_fpart = NULL; LLVMValueRef neighbors[2][2][4]; int chan; lp_build_extract_image_sizes(bld, - bld->int_size_type, + &bld->int_size_bld, bld->int_coord_type, size, &width_vec, &height_vec, &depth_vec); @@ -664,7 +802,7 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, flt_size = lp_build_int_to_float(&bld->float_size_bld, size); lp_build_extract_image_sizes(bld, - bld->float_size_type, + &bld->float_size_bld, bld->coord_type, flt_size, &flt_width_vec, &flt_height_vec, &flt_depth_vec); @@ -672,64 +810,61 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, /* * Compute integer texcoords. */ - lp_build_sample_wrap_linear(bld, s, width_vec, flt_width_vec, - bld->static_state->pot_width, - bld->static_state->wrap_s, + lp_build_sample_wrap_linear(bld, s, width_vec, flt_width_vec, offsets[0], + bld->static_texture_state->pot_width, + bld->static_sampler_state->wrap_s, &x0, &x1, &s_fpart); lp_build_name(x0, "tex.x0.wrapped"); lp_build_name(x1, "tex.x1.wrapped"); if (dims >= 2) { - lp_build_sample_wrap_linear(bld, t, height_vec, flt_height_vec, - bld->static_state->pot_height, - bld->static_state->wrap_t, + lp_build_sample_wrap_linear(bld, t, height_vec, flt_height_vec, offsets[1], + bld->static_texture_state->pot_height, + bld->static_sampler_state->wrap_t, &y0, &y1, &t_fpart); lp_build_name(y0, "tex.y0.wrapped"); lp_build_name(y1, "tex.y1.wrapped"); if (dims == 3) { - lp_build_sample_wrap_linear(bld, r, depth_vec, flt_depth_vec, - bld->static_state->pot_depth, - bld->static_state->wrap_r, + lp_build_sample_wrap_linear(bld, r, depth_vec, flt_depth_vec, offsets[2], + bld->static_texture_state->pot_depth, + bld->static_sampler_state->wrap_r, &z0, &z1, &r_fpart); lp_build_name(z0, "tex.z0.wrapped"); lp_build_name(z1, "tex.z1.wrapped"); } - else if (bld->static_state->target == PIPE_TEXTURE_CUBE) { - z0 = z1 = r; /* cube face */ - r_fpart = NULL; - } - else { - z0 = z1 = NULL; - r_fpart = NULL; - } } - else { - y0 = y1 = t_fpart = NULL; - z0 = z1 = r_fpart = NULL; + if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE || + bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY || + bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) { + z0 = z1 = r; /* cube face or array layer */ + lp_build_name(z0, "tex.z0.layer"); + lp_build_name(z1, "tex.z1.layer"); } + /* * Get texture colors. */ /* get x0/x1 texels */ - lp_build_sample_texel_soa(bld, unit, + lp_build_sample_texel_soa(bld, sampler_unit, width_vec, height_vec, depth_vec, x0, y0, z0, row_stride_vec, img_stride_vec, - data_ptr, neighbors[0][0]); - lp_build_sample_texel_soa(bld, unit, + data_ptr, mipoffsets, neighbors[0][0]); + lp_build_sample_texel_soa(bld, sampler_unit, width_vec, height_vec, depth_vec, x1, y0, z0, row_stride_vec, img_stride_vec, - data_ptr, neighbors[0][1]); + data_ptr, mipoffsets, neighbors[0][1]); if (dims == 1) { /* Interpolate two samples from 1D image to produce one color */ for (chan = 0; chan < 4; chan++) { colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart, neighbors[0][0][chan], - neighbors[0][1][chan]); + neighbors[0][1][chan], + 0); } } else { @@ -737,16 +872,16 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, LLVMValueRef colors0[4]; /* get x0/x1 texels at y1 */ - lp_build_sample_texel_soa(bld, unit, + lp_build_sample_texel_soa(bld, sampler_unit, width_vec, height_vec, depth_vec, x0, y1, z0, row_stride_vec, img_stride_vec, - data_ptr, neighbors[1][0]); - lp_build_sample_texel_soa(bld, unit, + data_ptr, mipoffsets, neighbors[1][0]); + lp_build_sample_texel_soa(bld, sampler_unit, width_vec, height_vec, depth_vec, x1, y1, z0, row_stride_vec, img_stride_vec, - data_ptr, neighbors[1][1]); + data_ptr, mipoffsets, neighbors[1][1]); /* Bilinear interpolate the four samples from the 2D image / 3D slice */ for (chan = 0; chan < 4; chan++) { @@ -755,7 +890,8 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, neighbors[0][0][chan], neighbors[0][1][chan], neighbors[1][0][chan], - neighbors[1][1][chan]); + neighbors[1][1][chan], + 0); } if (dims == 3) { @@ -763,26 +899,26 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, LLVMValueRef colors1[4]; /* get x0/x1/y0/y1 texels at z1 */ - lp_build_sample_texel_soa(bld, unit, + lp_build_sample_texel_soa(bld, sampler_unit, width_vec, height_vec, depth_vec, x0, y0, z1, row_stride_vec, img_stride_vec, - data_ptr, neighbors1[0][0]); - lp_build_sample_texel_soa(bld, unit, + data_ptr, mipoffsets, neighbors1[0][0]); + lp_build_sample_texel_soa(bld, sampler_unit, width_vec, height_vec, depth_vec, x1, y0, z1, row_stride_vec, img_stride_vec, - data_ptr, neighbors1[0][1]); - lp_build_sample_texel_soa(bld, unit, + data_ptr, mipoffsets, neighbors1[0][1]); + lp_build_sample_texel_soa(bld, sampler_unit, width_vec, height_vec, depth_vec, x0, y1, z1, row_stride_vec, img_stride_vec, - data_ptr, neighbors1[1][0]); - lp_build_sample_texel_soa(bld, unit, + data_ptr, mipoffsets, neighbors1[1][0]); + lp_build_sample_texel_soa(bld, sampler_unit, width_vec, height_vec, depth_vec, x1, y1, z1, row_stride_vec, img_stride_vec, - data_ptr, neighbors1[1][1]); + data_ptr, mipoffsets, neighbors1[1][1]); /* Bilinear interpolate the four samples from the second Z slice */ for (chan = 0; chan < 4; chan++) { @@ -791,14 +927,16 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, neighbors1[0][0][chan], neighbors1[0][1][chan], neighbors1[1][0][chan], - neighbors1[1][1][chan]); + neighbors1[1][1][chan], + 0); } /* Linearly interpolate the two samples from the two 3D slices */ for (chan = 0; chan < 4; chan++) { colors_out[chan] = lp_build_lerp(&bld->texel_bld, r_fpart, - colors0[chan], colors1[chan]); + colors0[chan], colors1[chan], + 0); } } else { @@ -819,26 +957,29 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld, */ static void lp_build_sample_mipmap(struct lp_build_sample_context *bld, - unsigned unit, + unsigned sampler_unit, unsigned img_filter, unsigned mip_filter, LLVMValueRef s, LLVMValueRef t, LLVMValueRef r, + const LLVMValueRef *offsets, LLVMValueRef ilevel0, LLVMValueRef ilevel1, LLVMValueRef lod_fpart, LLVMValueRef *colors_out) { - LLVMBuilderRef builder = bld->builder; - LLVMValueRef size0; - LLVMValueRef size1; - LLVMValueRef row_stride0_vec; - LLVMValueRef row_stride1_vec; - LLVMValueRef img_stride0_vec; - LLVMValueRef img_stride1_vec; - LLVMValueRef data_ptr0; - LLVMValueRef data_ptr1; + LLVMBuilderRef builder = bld->gallivm->builder; + LLVMValueRef size0 = NULL; + LLVMValueRef size1 = NULL; + LLVMValueRef row_stride0_vec = NULL; + LLVMValueRef row_stride1_vec = NULL; + LLVMValueRef img_stride0_vec = NULL; + LLVMValueRef img_stride1_vec = NULL; + LLVMValueRef data_ptr0 = NULL; + LLVMValueRef data_ptr1 = NULL; + LLVMValueRef mipoff0 = NULL; + LLVMValueRef mipoff1 = NULL; LLVMValueRef colors0[4], colors1[4]; unsigned chan; @@ -846,20 +987,27 @@ lp_build_sample_mipmap(struct lp_build_sample_context *bld, lp_build_mipmap_level_sizes(bld, ilevel0, &size0, &row_stride0_vec, &img_stride0_vec); - data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0); + if (bld->num_lods == 1) { + data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0); + } + else { + /* This path should work for num_lods 1 too but slightly less efficient */ + data_ptr0 = bld->base_ptr; + mipoff0 = lp_build_get_mip_offsets(bld, ilevel0); + } if (img_filter == PIPE_TEX_FILTER_NEAREST) { - lp_build_sample_image_nearest(bld, unit, + lp_build_sample_image_nearest(bld, sampler_unit, size0, row_stride0_vec, img_stride0_vec, - data_ptr0, s, t, r, + data_ptr0, mipoff0, s, t, r, offsets, colors0); } else { assert(img_filter == PIPE_TEX_FILTER_LINEAR); - lp_build_sample_image_linear(bld, unit, + lp_build_sample_image_linear(bld, sampler_unit, size0, row_stride0_vec, img_stride0_vec, - data_ptr0, s, t, r, + data_ptr0, mipoff0, s, t, r, offsets, colors0); } @@ -873,40 +1021,71 @@ lp_build_sample_mipmap(struct lp_build_sample_context *bld, LLVMValueRef need_lerp; /* need_lerp = lod_fpart > 0 */ - need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT, - lod_fpart, - bld->float_bld.zero, - "need_lerp"); + if (bld->num_lods == 1) { + need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT, + lod_fpart, bld->levelf_bld.zero, + "need_lerp"); + } + else { + /* + * We'll do mip filtering if any of the quads (or individual + * pixel in case of per-pixel lod) need it. + * It might be better to split the vectors here and only fetch/filter + * quads which need it. + */ + /* + * We unfortunately need to clamp lod_fpart here since we can get + * negative values which would screw up filtering if not all + * lod_fpart values have same sign. + */ + lod_fpart = lp_build_max(&bld->levelf_bld, lod_fpart, + bld->levelf_bld.zero); + need_lerp = lp_build_compare(bld->gallivm, bld->levelf_bld.type, + PIPE_FUNC_GREATER, + lod_fpart, bld->levelf_bld.zero); + need_lerp = lp_build_any_true_range(&bld->leveli_bld, bld->num_lods, need_lerp); + } - lp_build_if(&if_ctx, builder, need_lerp); + lp_build_if(&if_ctx, bld->gallivm, need_lerp); { /* sample the second mipmap level */ lp_build_mipmap_level_sizes(bld, ilevel1, &size1, &row_stride1_vec, &img_stride1_vec); - data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1); + if (bld->num_lods == 1) { + data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1); + } + else { + data_ptr1 = bld->base_ptr; + mipoff1 = lp_build_get_mip_offsets(bld, ilevel1); + } if (img_filter == PIPE_TEX_FILTER_NEAREST) { - lp_build_sample_image_nearest(bld, unit, + lp_build_sample_image_nearest(bld, sampler_unit, size1, row_stride1_vec, img_stride1_vec, - data_ptr1, s, t, r, + data_ptr1, mipoff1, s, t, r, offsets, colors1); } else { - lp_build_sample_image_linear(bld, unit, + lp_build_sample_image_linear(bld, sampler_unit, size1, row_stride1_vec, img_stride1_vec, - data_ptr1, s, t, r, + data_ptr1, mipoff1, s, t, r, offsets, colors1); } /* interpolate samples from the two mipmap levels */ - lod_fpart = lp_build_broadcast_scalar(&bld->texel_bld, lod_fpart); + if (bld->num_lods != bld->coord_type.length) + lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm, + bld->levelf_bld.type, + bld->texel_bld.type, + lod_fpart); for (chan = 0; chan < 4; chan++) { colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart, - colors0[chan], colors1[chan]); + colors0[chan], colors1[chan], + 0); LLVMBuildStore(builder, colors0[chan], colors_out[chan]); } } @@ -915,36 +1094,63 @@ lp_build_sample_mipmap(struct lp_build_sample_context *bld, } +/** + * Build (per-coord) layer value. + * Either clamp layer to valid values or fill in optional out_of_bounds + * value and just return value unclamped. + */ +static LLVMValueRef +lp_build_layer_coord(struct lp_build_sample_context *bld, + unsigned texture_unit, + LLVMValueRef layer, + LLVMValueRef *out_of_bounds) +{ + LLVMValueRef num_layers; + struct lp_build_context *int_coord_bld = &bld->int_coord_bld; + + num_layers = bld->dynamic_state->depth(bld->dynamic_state, + bld->gallivm, texture_unit); + + if (out_of_bounds) { + LLVMValueRef out1, out; + num_layers = lp_build_broadcast_scalar(int_coord_bld, num_layers); + out = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, layer, int_coord_bld->zero); + out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, layer, num_layers); + *out_of_bounds = lp_build_or(int_coord_bld, out, out1); + return layer; + } + else { + LLVMValueRef maxlayer; + maxlayer = lp_build_sub(&bld->int_bld, num_layers, bld->int_bld.one); + maxlayer = lp_build_broadcast_scalar(int_coord_bld, maxlayer); + return lp_build_clamp(int_coord_bld, layer, int_coord_bld->zero, maxlayer); + } +} + /** - * General texture sampling codegen. - * This function handles texture sampling for all texture targets (1D, - * 2D, 3D, cube) and all filtering modes. + * Calculate cube face, lod, mip levels. */ static void -lp_build_sample_general(struct lp_build_sample_context *bld, - unsigned unit, - LLVMValueRef s, - LLVMValueRef t, - LLVMValueRef r, - const LLVMValueRef *ddx, - const LLVMValueRef *ddy, - LLVMValueRef lod_bias, /* optional */ - LLVMValueRef explicit_lod, /* optional */ - LLVMValueRef *colors_out) +lp_build_sample_common(struct lp_build_sample_context *bld, + unsigned texture_index, + unsigned sampler_index, + LLVMValueRef *s, + LLVMValueRef *t, + LLVMValueRef *r, + const struct lp_derivatives *derivs, /* optional */ + LLVMValueRef lod_bias, /* optional */ + LLVMValueRef explicit_lod, /* optional */ + LLVMValueRef *lod_ipart, + LLVMValueRef *lod_fpart, + LLVMValueRef *ilevel0, + LLVMValueRef *ilevel1) { - struct lp_build_context *int_bld = &bld->int_bld; - LLVMBuilderRef builder = bld->builder; - const unsigned mip_filter = bld->static_state->min_mip_filter; - const unsigned min_filter = bld->static_state->min_img_filter; - const unsigned mag_filter = bld->static_state->mag_img_filter; - LLVMValueRef lod_ipart = NULL, lod_fpart = NULL; - LLVMValueRef ilevel0, ilevel1 = NULL; - LLVMValueRef face_ddx[4], face_ddy[4]; - LLVMValueRef texels[4]; - LLVMTypeRef i32t = LLVMInt32Type(); - LLVMValueRef i32t_zero = LLVMConstInt(i32t, 0, 0); - unsigned chan; + const unsigned mip_filter = bld->static_sampler_state->min_mip_filter; + const unsigned min_filter = bld->static_sampler_state->min_img_filter; + const unsigned mag_filter = bld->static_sampler_state->mag_img_filter; + const unsigned target = bld->static_texture_state->target; + LLVMValueRef first_level, cube_rho = NULL; /* printf("%s mip %d min %d mag %d\n", __FUNCTION__, @@ -952,27 +1158,30 @@ lp_build_sample_general(struct lp_build_sample_context *bld, */ /* - * Choose cube face, recompute texcoords and derivatives for the chosen face. + * Choose cube face, recompute texcoords for the chosen face and + * compute rho here too (as it requires transform of derivatives). */ - if (bld->static_state->target == PIPE_TEXTURE_CUBE) { + if (target == PIPE_TEXTURE_CUBE) { LLVMValueRef face, face_s, face_t; - lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t); - s = face_s; /* vec */ - t = face_t; /* vec */ + boolean need_derivs; + need_derivs = ((min_filter != mag_filter || + mip_filter != PIPE_TEX_MIPFILTER_NONE) && + !bld->static_sampler_state->min_max_lod_equal && + !explicit_lod); + lp_build_cube_lookup(bld, *s, *t, *r, derivs, &face, &face_s, &face_t, + &cube_rho, need_derivs); + *s = face_s; /* vec */ + *t = face_t; /* vec */ /* use 'r' to indicate cube face */ - r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */ - - /* recompute ddx, ddy using the new (s,t) face texcoords */ - face_ddx[0] = lp_build_scalar_ddx(&bld->coord_bld, s); - face_ddx[1] = lp_build_scalar_ddx(&bld->coord_bld, t); - face_ddx[2] = NULL; - face_ddx[3] = NULL; - face_ddy[0] = lp_build_scalar_ddy(&bld->coord_bld, s); - face_ddy[1] = lp_build_scalar_ddy(&bld->coord_bld, t); - face_ddy[2] = NULL; - face_ddy[3] = NULL; - ddx = face_ddx; - ddy = face_ddy; + *r = face; /* vec */ + } + else if (target == PIPE_TEXTURE_1D_ARRAY) { + *r = lp_build_iround(&bld->coord_bld, *t); + *r = lp_build_layer_coord(bld, texture_index, *r, NULL); + } + else if (target == PIPE_TEXTURE_2D_ARRAY) { + *r = lp_build_iround(&bld->coord_bld, *r); + *r = lp_build_layer_coord(bld, texture_index, *r, NULL); } /* @@ -983,12 +1192,13 @@ lp_build_sample_general(struct lp_build_sample_context *bld, /* Need to compute lod either to choose mipmap levels or to * distinguish between minification/magnification with one mipmap level. */ - lp_build_lod_selector(bld, unit, ddx, ddy, - lod_bias, explicit_lod, + lp_build_lod_selector(bld, texture_index, sampler_index, + *s, *t, *r, cube_rho, + derivs, lod_bias, explicit_lod, mip_filter, - &lod_ipart, &lod_fpart); + lod_ipart, lod_fpart); } else { - lod_ipart = i32t_zero; + *lod_ipart = bld->leveli_bld.zero; } /* @@ -1000,45 +1210,75 @@ lp_build_sample_general(struct lp_build_sample_context *bld, /* fall-through */ case PIPE_TEX_MIPFILTER_NONE: /* always use mip level 0 */ - if (bld->static_state->target == PIPE_TEXTURE_CUBE) { + if (HAVE_LLVM == 0x0207 && target == PIPE_TEXTURE_CUBE) { /* XXX this is a work-around for an apparent bug in LLVM 2.7. * We should be able to set ilevel0 = const(0) but that causes * bad x86 code to be emitted. */ - assert(lod_ipart); - lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0); + assert(*lod_ipart); + lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0, NULL); } else { - ilevel0 = i32t_zero; + first_level = bld->dynamic_state->first_level(bld->dynamic_state, + bld->gallivm, texture_index); + first_level = lp_build_broadcast_scalar(&bld->leveli_bld, first_level); + *ilevel0 = first_level; } break; case PIPE_TEX_MIPFILTER_NEAREST: - assert(lod_ipart); - lp_build_nearest_mip_level(bld, unit, lod_ipart, &ilevel0); + assert(*lod_ipart); + lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0, NULL); break; case PIPE_TEX_MIPFILTER_LINEAR: - assert(lod_ipart); - assert(lod_fpart); - lp_build_linear_mip_levels(bld, unit, - lod_ipart, &lod_fpart, - &ilevel0, &ilevel1); + assert(*lod_ipart); + assert(*lod_fpart); + lp_build_linear_mip_levels(bld, texture_index, + *lod_ipart, lod_fpart, + ilevel0, ilevel1); break; } +} + +/** + * General texture sampling codegen. + * This function handles texture sampling for all texture targets (1D, + * 2D, 3D, cube) and all filtering modes. + */ +static void +lp_build_sample_general(struct lp_build_sample_context *bld, + unsigned sampler_unit, + LLVMValueRef s, + LLVMValueRef t, + LLVMValueRef r, + const LLVMValueRef *offsets, + LLVMValueRef lod_ipart, + LLVMValueRef lod_fpart, + LLVMValueRef ilevel0, + LLVMValueRef ilevel1, + LLVMValueRef *colors_out) +{ + struct lp_build_context *int_bld = &bld->int_bld; + LLVMBuilderRef builder = bld->gallivm->builder; + const unsigned mip_filter = bld->static_sampler_state->min_mip_filter; + const unsigned min_filter = bld->static_sampler_state->min_img_filter; + const unsigned mag_filter = bld->static_sampler_state->mag_img_filter; + LLVMValueRef texels[4]; + unsigned chan; /* * Get/interpolate texture colors. */ for (chan = 0; chan < 4; ++chan) { - texels[chan] = lp_build_alloca(builder, bld->texel_bld.vec_type, ""); - lp_build_name(texels[chan], "sampler%u_texel_%c_var", unit, "xyzw"[chan]); + texels[chan] = lp_build_alloca(bld->gallivm, bld->texel_bld.vec_type, ""); + lp_build_name(texels[chan], "sampler%u_texel_%c_var", sampler_unit, "xyzw"[chan]); } if (min_filter == mag_filter) { - /* no need to distinquish between minification and magnification */ - lp_build_sample_mipmap(bld, unit, + /* no need to distinguish between minification and magnification */ + lp_build_sample_mipmap(bld, sampler_unit, min_filter, mip_filter, - s, t, r, + s, t, r, offsets, ilevel0, ilevel1, lod_fpart, texels); } @@ -1049,26 +1289,39 @@ lp_build_sample_general(struct lp_build_sample_context *bld, struct lp_build_if_state if_ctx; LLVMValueRef minify; + /* + * XXX this should to all lods into account, if some are min + * some max probably could hack up the coords/weights in the linear + * path with selects to work for nearest. + * If that's just two quads sitting next to each other it seems + * quite ok to do the same filtering method on both though, at + * least unless we have explicit lod (and who uses different + * min/mag filter with that?) + */ + if (bld->num_lods > 1) + lod_ipart = LLVMBuildExtractElement(builder, lod_ipart, + lp_build_const_int32(bld->gallivm, 0), ""); + /* minify = lod >= 0.0 */ minify = LLVMBuildICmp(builder, LLVMIntSGE, lod_ipart, int_bld->zero, ""); - lp_build_if(&if_ctx, builder, minify); + lp_build_if(&if_ctx, bld->gallivm, minify); { /* Use the minification filter */ - lp_build_sample_mipmap(bld, unit, + lp_build_sample_mipmap(bld, sampler_unit, min_filter, mip_filter, - s, t, r, + s, t, r, offsets, ilevel0, ilevel1, lod_fpart, texels); } lp_build_else(&if_ctx); { /* Use the magnification filter */ - lp_build_sample_mipmap(bld, unit, + lp_build_sample_mipmap(bld, sampler_unit, mag_filter, PIPE_TEX_MIPFILTER_NONE, - s, t, r, - i32t_zero, NULL, NULL, + s, t, r, offsets, + ilevel0, NULL, NULL, texels); } lp_build_endif(&if_ctx); @@ -1076,44 +1329,221 @@ lp_build_sample_general(struct lp_build_sample_context *bld, for (chan = 0; chan < 4; ++chan) { colors_out[chan] = LLVMBuildLoad(builder, texels[chan], ""); - lp_build_name(colors_out[chan], "sampler%u_texel_%c", unit, "xyzw"[chan]); + lp_build_name(colors_out[chan], "sampler%u_texel_%c", sampler_unit, "xyzw"[chan]); + } +} + + +/** + * Texel fetch function. + * In contrast to general sampling there is no filtering, no coord minification, + * lod (if any) is always explicit uint, coords are uints (in terms of texel units) + * directly to be applied to the selected mip level (after adding texel offsets). + * This function handles texel fetch for all targets where texel fetch is supported + * (no cube maps, but 1d, 2d, 3d are supported, arrays and buffers should be too). + */ +static void +lp_build_fetch_texel(struct lp_build_sample_context *bld, + unsigned texture_unit, + const LLVMValueRef *coords, + LLVMValueRef explicit_lod, + const LLVMValueRef *offsets, + LLVMValueRef *colors_out) +{ + struct lp_build_context *perquadi_bld = &bld->leveli_bld; + struct lp_build_context *int_coord_bld = &bld->int_coord_bld; + unsigned dims = bld->dims, chan; + unsigned target = bld->static_texture_state->target; + boolean out_of_bound_ret_zero = TRUE; + LLVMValueRef size, ilevel; + LLVMValueRef row_stride_vec = NULL, img_stride_vec = NULL; + LLVMValueRef x = coords[0], y = coords[1], z = coords[2]; + LLVMValueRef width, height, depth, i, j; + LLVMValueRef offset, out_of_bounds, out1; + + out_of_bounds = int_coord_bld->zero; + + if (explicit_lod && bld->static_texture_state->target != PIPE_BUFFER) { + if (bld->num_lods != int_coord_bld->type.length) { + ilevel = lp_build_pack_aos_scalars(bld->gallivm, int_coord_bld->type, + perquadi_bld->type, explicit_lod, 0); + } + else { + ilevel = explicit_lod; + } + lp_build_nearest_mip_level(bld, texture_unit, ilevel, &ilevel, + out_of_bound_ret_zero ? &out_of_bounds : NULL); + } + else { + assert(bld->num_lods == 1); + if (bld->static_texture_state->target != PIPE_BUFFER) { + ilevel = bld->dynamic_state->first_level(bld->dynamic_state, + bld->gallivm, texture_unit); + } + else { + ilevel = lp_build_const_int32(bld->gallivm, 0); + } + } + lp_build_mipmap_level_sizes(bld, ilevel, + &size, + &row_stride_vec, &img_stride_vec); + lp_build_extract_image_sizes(bld, &bld->int_size_bld, int_coord_bld->type, + size, &width, &height, &depth); + + if (target == PIPE_TEXTURE_1D_ARRAY || + target == PIPE_TEXTURE_2D_ARRAY) { + if (target == PIPE_TEXTURE_1D_ARRAY) { + z = y; + } + if (out_of_bound_ret_zero) { + z = lp_build_layer_coord(bld, texture_unit, z, &out1); + out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1); + } + else { + z = lp_build_layer_coord(bld, texture_unit, z, NULL); + } + } + + /* This is a lot like border sampling */ + if (offsets[0]) { + /* + * coords are really unsigned, offsets are signed, but I don't think + * exceeding 31 bits is possible + */ + x = lp_build_add(int_coord_bld, x, offsets[0]); + } + out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero); + out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1); + out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width); + out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1); + + if (dims >= 2) { + if (offsets[1]) { + y = lp_build_add(int_coord_bld, y, offsets[1]); + } + out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero); + out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1); + out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height); + out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1); + + if (dims >= 3) { + if (offsets[2]) { + z = lp_build_add(int_coord_bld, z, offsets[2]); + } + out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero); + out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1); + out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth); + out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1); + } + } + + lp_build_sample_offset(int_coord_bld, + bld->format_desc, + x, y, z, row_stride_vec, img_stride_vec, + &offset, &i, &j); + + if (bld->static_texture_state->target != PIPE_BUFFER) { + offset = lp_build_add(int_coord_bld, offset, + lp_build_get_mip_offsets(bld, ilevel)); + } + + offset = lp_build_andnot(int_coord_bld, offset, out_of_bounds); + + lp_build_fetch_rgba_soa(bld->gallivm, + bld->format_desc, + bld->texel_type, + bld->base_ptr, offset, + i, j, + colors_out); + + if (out_of_bound_ret_zero) { + /* + * Only needed for ARB_robust_buffer_access_behavior and d3d10. + * Could use min/max above instead of out-of-bounds comparisons + * if we don't care about the result returned for out-of-bounds. + */ + for (chan = 0; chan < 4; chan++) { + colors_out[chan] = lp_build_select(&bld->texel_bld, out_of_bounds, + bld->texel_bld.zero, colors_out[chan]); + } } } /** * Do shadow test/comparison. - * \param p the texcoord Z (aka R, aka P) component + * \param coords incoming texcoords * \param texel the texel to compare against (use the X channel) + * Ideally this should really be done per-sample. */ static void lp_build_sample_compare(struct lp_build_sample_context *bld, - LLVMValueRef p, + const LLVMValueRef *coords, LLVMValueRef texel[4]) { struct lp_build_context *texel_bld = &bld->texel_bld; - LLVMValueRef res; + LLVMBuilderRef builder = bld->gallivm->builder; + LLVMValueRef res, p; const unsigned chan = 0; + unsigned chan_type; + const struct util_format_description *format_desc; - if (bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE) + if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) return; + if (bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY || + bld->static_texture_state->target == PIPE_TEXTURE_CUBE) { + p = coords[3]; + } + else { + p = coords[2]; + } + /* debug code */ if (0) { - LLVMValueRef indx = lp_build_const_int32(0); - LLVMValueRef coord = LLVMBuildExtractElement(bld->builder, p, indx, ""); - LLVMValueRef tex = LLVMBuildExtractElement(bld->builder, - texel[chan], indx, ""); - lp_build_printf(bld->builder, "shadow compare coord %f to texture %f\n", + LLVMValueRef indx = lp_build_const_int32(bld->gallivm, 0); + LLVMValueRef coord = LLVMBuildExtractElement(builder, p, indx, ""); + LLVMValueRef tex = LLVMBuildExtractElement(builder, texel[chan], indx, ""); + lp_build_printf(bld->gallivm, "shadow compare coord %f to texture %f\n", coord, tex); } + /* Clamp p coords to [0,1] for fixed function depth texture format */ + format_desc = util_format_description(bld->static_texture_state->format); + /* not entirely sure we couldn't end up with non-valid swizzle here */ + chan_type = format_desc->swizzle[0] <= UTIL_FORMAT_SWIZZLE_W ? + format_desc->channel[format_desc->swizzle[0]].type : + UTIL_FORMAT_TYPE_FLOAT; + if (chan_type != UTIL_FORMAT_TYPE_FLOAT) { + p = lp_build_clamp(&bld->coord_bld, p, + bld->coord_bld.zero, bld->coord_bld.one); + } + + /* + * technically this is not entirely correct for unorm depth as the ref value + * should be converted to the depth format (quantization!) and comparison + * then done in texture format. + */ + /* result = (p FUNC texel) ? 1 : 0 */ - res = lp_build_cmp(texel_bld, bld->static_state->compare_func, - p, texel[chan]); + /* + * honor d3d10 floating point rules here, which state that comparisons + * are ordered except NOT_EQUAL which is unordered. + */ + if (bld->static_sampler_state->compare_func != PIPE_FUNC_NOTEQUAL) { + res = lp_build_cmp_ordered(texel_bld, bld->static_sampler_state->compare_func, + p, texel[chan]); + } + else { + res = lp_build_cmp(texel_bld, bld->static_sampler_state->compare_func, + p, texel[chan]); + } res = lp_build_select(texel_bld, res, texel_bld->one, texel_bld->zero); - /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */ + /* + * returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE. + * This should be ok because sampler swizzle is applied on top of it. + */ texel[0] = texel[1] = texel[2] = res; @@ -1126,10 +1556,12 @@ lp_build_sample_compare(struct lp_build_sample_context *bld, * For debugging. */ void -lp_build_sample_nop(struct lp_type type, +lp_build_sample_nop(struct gallivm_state *gallivm, + struct lp_type type, + const LLVMValueRef *coords, LLVMValueRef texel_out[4]) { - LLVMValueRef one = lp_build_one(type); + LLVMValueRef one = lp_build_one(gallivm, type); unsigned chan; for (chan = 0; chan < 4; chan++) { @@ -1143,34 +1575,40 @@ lp_build_sample_nop(struct lp_type type, * 'texel' will return a vector of four LLVMValueRefs corresponding to * R, G, B, A. * \param type vector float type to use for coords, etc. - * \param ddx partial derivatives of (s,t,r,q) with respect to x - * \param ddy partial derivatives of (s,t,r,q) with respect to y + * \param is_fetch if this is a texel fetch instruction. + * \param derivs partial derivatives of (s,t,r,q) with respect to x and y */ void -lp_build_sample_soa(LLVMBuilderRef builder, - const struct lp_sampler_static_state *static_state, +lp_build_sample_soa(struct gallivm_state *gallivm, + const struct lp_static_texture_state *static_texture_state, + const struct lp_static_sampler_state *static_sampler_state, struct lp_sampler_dynamic_state *dynamic_state, struct lp_type type, - unsigned unit, - unsigned num_coords, + boolean is_fetch, + unsigned texture_index, + unsigned sampler_index, const LLVMValueRef *coords, - const LLVMValueRef ddx[4], - const LLVMValueRef ddy[4], + const LLVMValueRef *offsets, + const struct lp_derivatives *derivs, /* optional */ LLVMValueRef lod_bias, /* optional */ LLVMValueRef explicit_lod, /* optional */ + boolean scalar_lod, LLVMValueRef texel_out[4]) { - unsigned dims = texture_dims(static_state->target); + unsigned dims = texture_dims(static_texture_state->target); + unsigned num_quads = type.length / 4; + unsigned mip_filter; struct lp_build_sample_context bld; - LLVMTypeRef i32t = LLVMInt32Type(); - + struct lp_static_sampler_state derived_sampler_state = *static_sampler_state; + LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context); + LLVMBuilderRef builder = gallivm->builder; + LLVMValueRef tex_width; LLVMValueRef s; LLVMValueRef t; LLVMValueRef r; - struct lp_type float_vec_type; if (0) { - enum pipe_format fmt = static_state->format; + enum pipe_format fmt = static_texture_state->format; debug_printf("Sample from %s\n", util_format_name(fmt)); } @@ -1178,40 +1616,120 @@ lp_build_sample_soa(LLVMBuilderRef builder, /* Setup our build context */ memset(&bld, 0, sizeof bld); - bld.builder = builder; - bld.static_state = static_state; + bld.gallivm = gallivm; + bld.static_sampler_state = &derived_sampler_state; + bld.static_texture_state = static_texture_state; bld.dynamic_state = dynamic_state; - bld.format_desc = util_format_description(static_state->format); + bld.format_desc = util_format_description(static_texture_state->format); bld.dims = dims; + bld.vector_width = lp_type_width(type); + bld.float_type = lp_type_float(32); bld.int_type = lp_type_int(32); bld.coord_type = type; bld.int_coord_type = lp_int_type(type); - bld.float_size_type = lp_type_float(32); - bld.float_size_type.length = dims > 1 ? 4 : 1; - bld.int_size_type = lp_int_type(bld.float_size_type); + bld.float_size_in_type = lp_type_float(32); + bld.float_size_in_type.length = dims > 1 ? 4 : 1; + bld.int_size_in_type = lp_int_type(bld.float_size_in_type); bld.texel_type = type; - float_vec_type = lp_type_float_vec(32); + /* always using the first channel hopefully should be safe, + * if not things WILL break in other places anyway. + */ + if (bld.format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB && + bld.format_desc->channel[0].pure_integer) { + if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) { + bld.texel_type = lp_type_int_vec(type.width, type.width * type.length); + } + else if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) { + bld.texel_type = lp_type_uint_vec(type.width, type.width * type.length); + } + } + else if (util_format_has_stencil(bld.format_desc) && + !util_format_has_depth(bld.format_desc)) { + /* for stencil only formats, sample stencil (uint) */ + bld.texel_type = lp_type_int_vec(type.width, type.width * type.length); + } + + if (!static_texture_state->level_zero_only) { + derived_sampler_state.min_mip_filter = static_sampler_state->min_mip_filter; + } else { + derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE; + } + mip_filter = derived_sampler_state.min_mip_filter; + + if (0) { + debug_printf(" .min_mip_filter = %u\n", derived_sampler_state.min_mip_filter); + } - lp_build_context_init(&bld.float_bld, builder, bld.float_type); - lp_build_context_init(&bld.float_vec_bld, builder, float_vec_type); - lp_build_context_init(&bld.int_bld, builder, bld.int_type); - lp_build_context_init(&bld.coord_bld, builder, bld.coord_type); - lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type); - lp_build_context_init(&bld.int_size_bld, builder, bld.int_size_type); - lp_build_context_init(&bld.float_size_bld, builder, bld.float_size_type); - lp_build_context_init(&bld.texel_bld, builder, bld.texel_type); + /* + * This is all a bit complicated different paths are chosen for performance + * reasons. + * Essentially, there can be 1 lod per element, 1 lod per quad or 1 lod for + * everything (the last two options are equivalent for 4-wide case). + * If there's per-quad lod but we split to 4-wide so we can use AoS, per-quad + * lod is calculated then the lod value extracted afterwards so making this + * case basically the same as far as lod handling is concerned for the + * further sample/filter code as the 1 lod for everything case. + * Different lod handling mostly shows up when building mipmap sizes + * (lp_build_mipmap_level_sizes() and friends) and also in filtering + * (getting the fractional part of the lod to the right texels). + */ + + /* + * There are other situations where at least the multiple int lods could be + * avoided like min and max lod being equal. + */ + if (explicit_lod && !scalar_lod && + ((is_fetch && bld.static_texture_state->target != PIPE_BUFFER) || + (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE))) + bld.num_lods = type.length; + /* TODO: for true scalar_lod should only use 1 lod value */ + else if ((is_fetch && explicit_lod && bld.static_texture_state->target != PIPE_BUFFER ) || + (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) { + bld.num_lods = num_quads; + } + else { + bld.num_lods = 1; + } + + bld.levelf_type = type; + /* we want native vector size to be able to use our intrinsics */ + if (bld.num_lods != type.length) { + bld.levelf_type.length = type.length > 4 ? ((type.length + 15) / 16) * 4 : 1; + } + bld.leveli_type = lp_int_type(bld.levelf_type); + bld.float_size_type = bld.float_size_in_type; + /* Note: size vectors may not be native. They contain minified w/h/d/_ values, + * with per-element lod that is w0/h0/d0/_/w1/h1/d1_/... so up to 8x4f32 */ + if (bld.num_lods > 1) { + bld.float_size_type.length = bld.num_lods == type.length ? + bld.num_lods * bld.float_size_in_type.length : + type.length; + } + bld.int_size_type = lp_int_type(bld.float_size_type); + + lp_build_context_init(&bld.float_bld, gallivm, bld.float_type); + lp_build_context_init(&bld.float_vec_bld, gallivm, type); + lp_build_context_init(&bld.int_bld, gallivm, bld.int_type); + lp_build_context_init(&bld.coord_bld, gallivm, bld.coord_type); + lp_build_context_init(&bld.int_coord_bld, gallivm, bld.int_coord_type); + lp_build_context_init(&bld.int_size_in_bld, gallivm, bld.int_size_in_type); + lp_build_context_init(&bld.float_size_in_bld, gallivm, bld.float_size_in_type); + lp_build_context_init(&bld.int_size_bld, gallivm, bld.int_size_type); + lp_build_context_init(&bld.float_size_bld, gallivm, bld.float_size_type); + lp_build_context_init(&bld.texel_bld, gallivm, bld.texel_type); + lp_build_context_init(&bld.levelf_bld, gallivm, bld.levelf_type); + lp_build_context_init(&bld.leveli_bld, gallivm, bld.leveli_type); /* Get the dynamic state */ - bld.width = dynamic_state->width(dynamic_state, builder, unit); - bld.height = dynamic_state->height(dynamic_state, builder, unit); - bld.depth = dynamic_state->depth(dynamic_state, builder, unit); - bld.row_stride_array = dynamic_state->row_stride(dynamic_state, builder, unit); - bld.img_stride_array = dynamic_state->img_stride(dynamic_state, builder, unit); - bld.data_array = dynamic_state->data_ptr(dynamic_state, builder, unit); - /* Note that data_array is an array[level] of pointers to texture images */ + tex_width = dynamic_state->width(dynamic_state, gallivm, texture_index); + bld.row_stride_array = dynamic_state->row_stride(dynamic_state, gallivm, texture_index); + bld.img_stride_array = dynamic_state->img_stride(dynamic_state, gallivm, texture_index); + bld.base_ptr = dynamic_state->base_ptr(dynamic_state, gallivm, texture_index); + bld.mip_offsets = dynamic_state->mip_offsets(dynamic_state, gallivm, texture_index); + /* Note that mip_offsets is an array[level] of offsets to texture images */ s = coords[0]; t = coords[1]; @@ -1219,51 +1737,399 @@ lp_build_sample_soa(LLVMBuilderRef builder, /* width, height, depth as single int vector */ if (dims <= 1) { - bld.int_size = bld.width; + bld.int_size = tex_width; } else { - bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_bld.undef, - bld.width, LLVMConstInt(i32t, 0, 0), ""); + bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef, + tex_width, LLVMConstInt(i32t, 0, 0), ""); if (dims >= 2) { + LLVMValueRef tex_height = + dynamic_state->height(dynamic_state, gallivm, texture_index); bld.int_size = LLVMBuildInsertElement(builder, bld.int_size, - bld.height, LLVMConstInt(i32t, 1, 0), ""); + tex_height, LLVMConstInt(i32t, 1, 0), ""); if (dims >= 3) { + LLVMValueRef tex_depth = + dynamic_state->depth(dynamic_state, gallivm, texture_index); bld.int_size = LLVMBuildInsertElement(builder, bld.int_size, - bld.depth, LLVMConstInt(i32t, 2, 0), ""); + tex_depth, LLVMConstInt(i32t, 2, 0), ""); } } } if (0) { /* For debug: no-op texture sampling */ - lp_build_sample_nop(bld.texel_type, texel_out); - } - else if (util_format_fits_8unorm(bld.format_desc) && - lp_is_simple_wrap_mode(static_state->wrap_s) && - lp_is_simple_wrap_mode(static_state->wrap_t)) { - /* do sampling/filtering with fixed pt arithmetic */ - lp_build_sample_aos(&bld, unit, s, t, r, ddx, ddy, - lod_bias, explicit_lod, + lp_build_sample_nop(gallivm, + bld.texel_type, + coords, texel_out); } + else if (is_fetch) { + lp_build_fetch_texel(&bld, texture_index, coords, + explicit_lod, offsets, + texel_out); + } + else { + LLVMValueRef lod_ipart = NULL, lod_fpart = NULL; + LLVMValueRef ilevel0 = NULL, ilevel1 = NULL; + boolean use_aos = util_format_fits_8unorm(bld.format_desc) && + lp_is_simple_wrap_mode(static_sampler_state->wrap_s) && + lp_is_simple_wrap_mode(static_sampler_state->wrap_t); + if ((gallivm_debug & GALLIVM_DEBUG_PERF) && - util_format_fits_8unorm(bld.format_desc)) { + !use_aos && util_format_fits_8unorm(bld.format_desc)) { debug_printf("%s: using floating point linear filtering for %s\n", __FUNCTION__, bld.format_desc->short_name); debug_printf(" min_img %d mag_img %d mip %d wraps %d wrapt %d\n", - static_state->min_img_filter, - static_state->mag_img_filter, - static_state->min_mip_filter, - static_state->wrap_s, - static_state->wrap_t); + static_sampler_state->min_img_filter, + static_sampler_state->mag_img_filter, + static_sampler_state->min_mip_filter, + static_sampler_state->wrap_s, + static_sampler_state->wrap_t); } - lp_build_sample_general(&bld, unit, s, t, r, ddx, ddy, - lod_bias, explicit_lod, - texel_out); + lp_build_sample_common(&bld, texture_index, sampler_index, + &s, &t, &r, + derivs, lod_bias, explicit_lod, + &lod_ipart, &lod_fpart, + &ilevel0, &ilevel1); + + /* + * we only try 8-wide sampling with soa as it appears to + * be a loss with aos with AVX (but it should work). + * (It should be faster if we'd support avx2) + */ + if (num_quads == 1 || !use_aos) { + + if (num_quads > 1) { + if (mip_filter == PIPE_TEX_MIPFILTER_NONE) { + LLVMValueRef index0 = lp_build_const_int32(gallivm, 0); + /* + * These parameters are the same for all quads, + * could probably simplify. + */ + lod_ipart = LLVMBuildExtractElement(builder, lod_ipart, index0, ""); + ilevel0 = LLVMBuildExtractElement(builder, ilevel0, index0, ""); + } + } + if (use_aos) { + /* do sampling/filtering with fixed pt arithmetic */ + lp_build_sample_aos(&bld, sampler_index, + s, t, r, offsets, + lod_ipart, lod_fpart, + ilevel0, ilevel1, + texel_out); + } + + else { + lp_build_sample_general(&bld, sampler_index, + s, t, r, offsets, + lod_ipart, lod_fpart, + ilevel0, ilevel1, + texel_out); + } + } + else { + unsigned j; + struct lp_build_sample_context bld4; + struct lp_type type4 = type; + unsigned i; + LLVMValueRef texelout4[4]; + LLVMValueRef texelouttmp[4][LP_MAX_VECTOR_LENGTH/16]; + + type4.length = 4; + + /* Setup our build context */ + memset(&bld4, 0, sizeof bld4); + bld4.gallivm = bld.gallivm; + bld4.static_texture_state = bld.static_texture_state; + bld4.static_sampler_state = bld.static_sampler_state; + bld4.dynamic_state = bld.dynamic_state; + bld4.format_desc = bld.format_desc; + bld4.dims = bld.dims; + bld4.row_stride_array = bld.row_stride_array; + bld4.img_stride_array = bld.img_stride_array; + bld4.base_ptr = bld.base_ptr; + bld4.mip_offsets = bld.mip_offsets; + bld4.int_size = bld.int_size; + + bld4.vector_width = lp_type_width(type4); + + bld4.float_type = lp_type_float(32); + bld4.int_type = lp_type_int(32); + bld4.coord_type = type4; + bld4.int_coord_type = lp_int_type(type4); + bld4.float_size_in_type = lp_type_float(32); + bld4.float_size_in_type.length = dims > 1 ? 4 : 1; + bld4.int_size_in_type = lp_int_type(bld4.float_size_in_type); + bld4.texel_type = bld.texel_type; + bld4.texel_type.length = 4; + bld4.levelf_type = type4; + /* we want native vector size to be able to use our intrinsics */ + bld4.levelf_type.length = 1; + bld4.leveli_type = lp_int_type(bld4.levelf_type); + + if (explicit_lod && !scalar_lod && + ((is_fetch && bld.static_texture_state->target != PIPE_BUFFER) || + (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE))) + bld4.num_lods = type4.length; + else + bld4.num_lods = 1; + + bld4.levelf_type = type4; + /* we want native vector size to be able to use our intrinsics */ + if (bld4.num_lods != type4.length) { + bld4.levelf_type.length = 1; + } + bld4.leveli_type = lp_int_type(bld4.levelf_type); + bld4.float_size_type = bld4.float_size_in_type; + if (bld4.num_lods > 1) { + bld4.float_size_type.length = bld4.num_lods == type4.length ? + bld4.num_lods * bld4.float_size_in_type.length : + type4.length; + } + bld4.int_size_type = lp_int_type(bld4.float_size_type); + + lp_build_context_init(&bld4.float_bld, gallivm, bld4.float_type); + lp_build_context_init(&bld4.float_vec_bld, gallivm, type4); + lp_build_context_init(&bld4.int_bld, gallivm, bld4.int_type); + lp_build_context_init(&bld4.coord_bld, gallivm, bld4.coord_type); + lp_build_context_init(&bld4.int_coord_bld, gallivm, bld4.int_coord_type); + lp_build_context_init(&bld4.int_size_in_bld, gallivm, bld4.int_size_in_type); + lp_build_context_init(&bld4.float_size_in_bld, gallivm, bld4.float_size_in_type); + lp_build_context_init(&bld4.int_size_bld, gallivm, bld4.int_size_type); + lp_build_context_init(&bld4.float_size_bld, gallivm, bld4.float_size_type); + lp_build_context_init(&bld4.texel_bld, gallivm, bld4.texel_type); + lp_build_context_init(&bld4.levelf_bld, gallivm, bld4.levelf_type); + lp_build_context_init(&bld4.leveli_bld, gallivm, bld4.leveli_type); + + for (i = 0; i < num_quads; i++) { + LLVMValueRef s4, t4, r4; + LLVMValueRef lod_ipart4, lod_fpart4 = NULL; + LLVMValueRef ilevel04, ilevel14 = NULL; + LLVMValueRef offsets4[4] = { NULL }; + unsigned num_lods = bld4.num_lods; + + s4 = lp_build_extract_range(gallivm, s, 4*i, 4); + t4 = lp_build_extract_range(gallivm, t, 4*i, 4); + r4 = lp_build_extract_range(gallivm, r, 4*i, 4); + + if (offsets[0]) { + offsets4[0] = lp_build_extract_range(gallivm, offsets[0], 4*i, 4); + if (dims > 1) { + offsets4[1] = lp_build_extract_range(gallivm, offsets[1], 4*i, 4); + if (dims > 2) { + offsets4[2] = lp_build_extract_range(gallivm, offsets[2], 4*i, 4); + } + } + } + lod_ipart4 = lp_build_extract_range(gallivm, lod_ipart, num_lods * i, num_lods); + ilevel04 = lp_build_extract_range(gallivm, ilevel0, num_lods * i, num_lods); + if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) { + ilevel14 = lp_build_extract_range(gallivm, ilevel1, num_lods * i, num_lods); + lod_fpart4 = lp_build_extract_range(gallivm, lod_fpart, num_lods * i, num_lods); + } + + if (use_aos) { + /* do sampling/filtering with fixed pt arithmetic */ + lp_build_sample_aos(&bld4, sampler_index, + s4, t4, r4, offsets4, + lod_ipart4, lod_fpart4, + ilevel04, ilevel14, + texelout4); + } + + else { + lp_build_sample_general(&bld4, sampler_index, + s4, t4, r4, offsets4, + lod_ipart4, lod_fpart4, + ilevel04, ilevel14, + texelout4); + } + for (j = 0; j < 4; j++) { + texelouttmp[j][i] = texelout4[j]; + } + } + + for (j = 0; j < 4; j++) { + texel_out[j] = lp_build_concat(gallivm, texelouttmp[j], type4, num_quads); + } + } + + lp_build_sample_compare(&bld, coords, texel_out); + } + + if (static_texture_state->target != PIPE_BUFFER) { + apply_sampler_swizzle(&bld, texel_out); + } + + /* + * texel type can be a (32bit) int/uint (for pure int formats only), + * however we are expected to always return floats (storage is untyped). + */ + if (!bld.texel_type.floating) { + unsigned chan; + for (chan = 0; chan < 4; chan++) { + texel_out[chan] = LLVMBuildBitCast(builder, texel_out[chan], + lp_build_vec_type(gallivm, type), ""); + } } +} - lp_build_sample_compare(&bld, r, texel_out); +void +lp_build_size_query_soa(struct gallivm_state *gallivm, + const struct lp_static_texture_state *static_state, + struct lp_sampler_dynamic_state *dynamic_state, + struct lp_type int_type, + unsigned texture_unit, + unsigned target, + boolean is_sviewinfo, + boolean scalar_lod, + LLVMValueRef explicit_lod, + LLVMValueRef *sizes_out) +{ + LLVMValueRef lod, level, size; + LLVMValueRef first_level = NULL; + int dims, i; + boolean has_array; + unsigned num_lods = 1; + struct lp_build_context bld_int_vec4; + + /* + * Do some sanity verification about bound texture and shader dcl target. + * Not entirely sure what's possible but assume array/non-array + * always compatible (probably not ok for OpenGL but d3d10 has no + * distinction of arrays at the resource level). + * Everything else looks bogus (though not entirely sure about rect/2d). + * Currently disabled because it causes assertion failures if there's + * nothing bound (or rather a dummy texture, not that this case would + * return the right values). + */ + if (0 && static_state->target != target) { + if (static_state->target == PIPE_TEXTURE_1D) + assert(target == PIPE_TEXTURE_1D_ARRAY); + else if (static_state->target == PIPE_TEXTURE_1D_ARRAY) + assert(target == PIPE_TEXTURE_1D); + else if (static_state->target == PIPE_TEXTURE_2D) + assert(target == PIPE_TEXTURE_2D_ARRAY); + else if (static_state->target == PIPE_TEXTURE_2D_ARRAY) + assert(target == PIPE_TEXTURE_2D); + else if (static_state->target == PIPE_TEXTURE_CUBE) + assert(target == PIPE_TEXTURE_CUBE_ARRAY); + else if (static_state->target == PIPE_TEXTURE_CUBE_ARRAY) + assert(target == PIPE_TEXTURE_CUBE); + else + assert(0); + } + + dims = texture_dims(target); + + switch (target) { + case PIPE_TEXTURE_1D_ARRAY: + case PIPE_TEXTURE_2D_ARRAY: + has_array = TRUE; + break; + default: + has_array = FALSE; + break; + } + + assert(!int_type.floating); + + lp_build_context_init(&bld_int_vec4, gallivm, lp_type_int_vec(32, 128)); + + if (explicit_lod) { + /* FIXME: this needs to honor per-element lod */ + lod = LLVMBuildExtractElement(gallivm->builder, explicit_lod, lp_build_const_int32(gallivm, 0), ""); + first_level = dynamic_state->first_level(dynamic_state, gallivm, texture_unit); + level = LLVMBuildAdd(gallivm->builder, lod, first_level, "level"); + lod = lp_build_broadcast_scalar(&bld_int_vec4, level); + } else { + lod = bld_int_vec4.zero; + } + + size = bld_int_vec4.undef; + + size = LLVMBuildInsertElement(gallivm->builder, size, + dynamic_state->width(dynamic_state, gallivm, texture_unit), + lp_build_const_int32(gallivm, 0), ""); + + if (dims >= 2) { + size = LLVMBuildInsertElement(gallivm->builder, size, + dynamic_state->height(dynamic_state, gallivm, texture_unit), + lp_build_const_int32(gallivm, 1), ""); + } + + if (dims >= 3) { + size = LLVMBuildInsertElement(gallivm->builder, size, + dynamic_state->depth(dynamic_state, gallivm, texture_unit), + lp_build_const_int32(gallivm, 2), ""); + } + + size = lp_build_minify(&bld_int_vec4, size, lod); + + if (has_array) + size = LLVMBuildInsertElement(gallivm->builder, size, + dynamic_state->depth(dynamic_state, gallivm, texture_unit), + lp_build_const_int32(gallivm, dims), ""); + + /* + * d3d10 requires zero for x/y/z values (but not w, i.e. mip levels) + * if level is out of bounds (note this can't cover unbound texture + * here, which also requires returning zero). + */ + if (explicit_lod && is_sviewinfo) { + LLVMValueRef last_level, out, out1; + struct lp_build_context leveli_bld; + + /* everything is scalar for now */ + lp_build_context_init(&leveli_bld, gallivm, lp_type_int_vec(32, 32)); + last_level = dynamic_state->last_level(dynamic_state, gallivm, texture_unit); + + out = lp_build_cmp(&leveli_bld, PIPE_FUNC_LESS, level, first_level); + out1 = lp_build_cmp(&leveli_bld, PIPE_FUNC_GREATER, level, last_level); + out = lp_build_or(&leveli_bld, out, out1); + if (num_lods == 1) { + out = lp_build_broadcast_scalar(&bld_int_vec4, out); + } + else { + /* TODO */ + assert(0); + } + size = lp_build_andnot(&bld_int_vec4, size, out); + } + for (i = 0; i < dims + (has_array ? 1 : 0); i++) { + sizes_out[i] = lp_build_extract_broadcast(gallivm, bld_int_vec4.type, int_type, + size, + lp_build_const_int32(gallivm, i)); + } + if (is_sviewinfo) { + for (; i < 4; i++) { + sizes_out[i] = lp_build_const_vec(gallivm, int_type, 0.0); + } + } + + /* + * if there's no explicit_lod (buffers, rects) queries requiring nr of + * mips would be illegal. + */ + if (is_sviewinfo && explicit_lod) { + struct lp_build_context bld_int_scalar; + LLVMValueRef num_levels; + lp_build_context_init(&bld_int_scalar, gallivm, lp_type_int(32)); + + if (static_state->level_zero_only) { + num_levels = bld_int_scalar.one; + } + else { + LLVMValueRef last_level; + + last_level = dynamic_state->last_level(dynamic_state, gallivm, texture_unit); + num_levels = lp_build_sub(&bld_int_scalar, last_level, first_level); + num_levels = lp_build_add(&bld_int_scalar, num_levels, bld_int_scalar.one); + } + sizes_out[3] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, int_type), + num_levels); + } }