OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

microsoft/compiler: translate nir to dxil

Browse Source 
Here's the code to emit DXIL code from NIR. It's big and bulky as-is,
and it needs to be split up a bit.

This is the combination of a lot of commits from our development branch,
containing code by several authors.

Co-authored-by: Bill Kristiansen <billkris@microsoft.com>
Co-authored-by: Boris Brezillon <boris.brezillon@collabora.com>
Co-authored-by: Daniel Stone <daniels@collabora.com>
Co-authored-by: Gert Wollny <gert.wollny@collabora.com>
Co-authored-by: Jesse Natalie <jenatali@microsoft.com>
Co-authored-by: Louis-Francis Ratté-Boulianne <lfrb@collabora.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7477>
This commit is contained in:
Erik Faye-Lund
2019-09-04 11:51:29 +02:00
committed by Marge Bot
parent dd5fe2f3de
commit b9c61379ab
9 changed files with 4761 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/compiler/nir/nir_intrinsics.py
View File

@@ -945,6 +945,10 @@ load("global_ir3", [2, 1], indices=[ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN
# Note that this doesn't actually turn into a HW instruction. # Note that this doesn't actually turn into a HW instruction.
intrinsic("bindless_resource_ir3", [1], dest_comp=1, indices=[DESC_SET], flags=[CAN_ELIMINATE, CAN_REORDER]) intrinsic("bindless_resource_ir3", [1], dest_comp=1, indices=[DESC_SET], flags=[CAN_ELIMINATE, CAN_REORDER])
# DXIL specific intrinsics
# src[] = { index, 16-byte-based-offset }
load("ubo_dxil", [1, 1], [], [CAN_ELIMINATE])
# Intrinsics used by the Midgard/Bifrost blend pipeline. These are defined # Intrinsics used by the Midgard/Bifrost blend pipeline. These are defined
# within a blend shader to read/write the raw value from the tile buffer, # within a blend shader to read/write the raw value from the tile buffer,
# without applying any format conversion in the process. If the shader needs # without applying any format conversion in the process. If the shader needs

156
src/microsoft/compiler/dxil_nir.c Normal file
View File

@@ -0,0 +1,156 @@
/*
* Copyright © Microsoft Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#include "dxil_nir.h"
#include "nir_builder.h"
#include "nir_deref.h"
#include "util/u_math.h"
static void
extract_comps_from_vec32(nir_builder *b, nir_ssa_def *vec32,
unsigned dst_bit_size,
nir_ssa_def **dst_comps,
unsigned num_dst_comps)
{
unsigned step = DIV_ROUND_UP(dst_bit_size, 32);
unsigned comps_per32b = 32 / dst_bit_size;
nir_ssa_def *tmp;
for (unsigned i = 0; i < vec32->num_components; i += step) {
switch (dst_bit_size) {
case 64:
tmp = nir_pack_64_2x32_split(b, nir_channel(b, vec32, i),
nir_channel(b, vec32, i + 1));
dst_comps[i / 2] = tmp;
break;
case 32:
dst_comps[i] = nir_channel(b, vec32, i);
break;
case 16:
case 8:
unsigned dst_offs = i * comps_per32b;
tmp = nir_unpack_bits(b, nir_channel(b, vec32, i), dst_bit_size);
for (unsigned j = 0; j < comps_per32b && dst_offs + j < num_dst_comps; j++)
dst_comps[dst_offs + j] = nir_channel(b, tmp, j);
break;
}
}
}
static nir_ssa_def *
ubo_load_select_32b_comps(nir_builder *b, nir_ssa_def *vec32,
nir_ssa_def *offset, unsigned num_bytes)
{
assert(num_bytes == 16 || num_bytes == 12 || num_bytes == 8 ||
num_bytes == 4 || num_bytes == 3 || num_bytes == 2 ||
num_bytes == 1);
assert(vec32->num_components == 4);
/* 16 and 12 byte types are always aligned on 16 bytes. */
if (num_bytes > 8)
return vec32;
nir_ssa_def *comps[4];
nir_ssa_def *cond;
for (unsigned i = 0; i < 4; i++)
comps[i] = nir_channel(b, vec32, i);
/* If we have 8bytes or less to load, select which half the vec4 should
* be used.
*/
cond = nir_ine(b, nir_iand(b, offset, nir_imm_int(b, 0x8)),
nir_imm_int(b, 0));
comps[0] = nir_bcsel(b, cond, comps[2], comps[0]);
comps[1] = nir_bcsel(b, cond, comps[3], comps[1]);
/* Thanks to the CL alignment constraints, if we want 8 bytes we're done. */
if (num_bytes == 8)
return nir_vec(b, comps, 2);
/* 4 bytes or less needed, select which of the 32bit component should be
* used and return it. The sub-32bit split is handled in
* extract_comps_from_vec32().
*/
cond = nir_ine(b, nir_iand(b, offset, nir_imm_int(b, 0x4)),
nir_imm_int(b, 0));
return nir_bcsel(b, cond, comps[1], comps[0]);
}
nir_ssa_def *
build_load_ubo_dxil(nir_builder *b, nir_ssa_def *buffer,
nir_ssa_def *offset, unsigned num_components,
unsigned bit_size)
{
nir_ssa_def *idx = nir_ushr(b, offset, nir_imm_int(b, 4));
nir_ssa_def *comps[NIR_MAX_VEC_COMPONENTS];
unsigned num_bits = num_components * bit_size;
unsigned comp_idx = 0;
/* We need to split loads in 16byte chunks because that's the
* granularity of cBufferLoadLegacy().
*/
for (unsigned i = 0; i < num_bits; i += (16 * 8)) {
/* For each 16byte chunk (or smaller) we generate a 32bit ubo vec
* load.
*/
unsigned subload_num_bits = MIN2(num_bits - i, 16 * 8);
nir_intrinsic_instr *load =
nir_intrinsic_instr_create(b->shader,
nir_intrinsic_load_ubo_dxil);
load->num_components = 4;
load->src[0] = nir_src_for_ssa(buffer);
load->src[1] = nir_src_for_ssa(nir_iadd(b, idx, nir_imm_int(b, i / (16 * 8))));
nir_ssa_dest_init(&load->instr, &load->dest, load->num_components,
32, NULL);
nir_builder_instr_insert(b, &load->instr);
nir_ssa_def *vec32 = &load->dest.ssa;
/* First re-arrange the vec32 to account for intra 16-byte offset. */
vec32 = ubo_load_select_32b_comps(b, vec32, offset, subload_num_bits / 8);
/* If we have 2 bytes or less to load we need to adjust the u32 value so
* we can always extract the LSB.
*/
if (subload_num_bits <= 16) {
nir_ssa_def *shift = nir_imul(b, nir_iand(b, offset,
nir_imm_int(b, 3)),
nir_imm_int(b, 8));
vec32 = nir_ushr(b, vec32, shift);
}
/* And now comes the pack/unpack step to match the original type. */
extract_comps_from_vec32(b, vec32, bit_size, &comps[comp_idx],
subload_num_bits / bit_size);
comp_idx += subload_num_bits / bit_size;
}
assert(comp_idx == num_components);
return nir_vec(b, comps, num_components);
}

40
src/microsoft/compiler/dxil_nir.h Normal file
View File

@@ -0,0 +1,40 @@
/*
* Copyright © Microsoft Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#ifndef DXIL_NIR_H
#define DXIL_NIR_H
#include <stdbool.h>
#include "nir.h"
#include "nir_builder.h"
bool dxil_nir_lower_8bit_conv(nir_shader *shader);
bool dxil_nir_lower_x2b(nir_shader *shader);
bool dxil_nir_lower_inot(nir_shader *shader);
nir_ssa_def *
build_load_ubo_dxil(nir_builder *b, nir_ssa_def *buffer,
nir_ssa_def *offset, unsigned num_components,
unsigned bit_size);
#endif /* DXIL_NIR_H */

133
src/microsoft/compiler/dxil_nir_algebraic.py Normal file
View File

@@ -0,0 +1,133 @@
#
# Copyright (C) 2020 Microsoft Corporation
#
# Copyright (C) 2018 Alyssa Rosenzweig
#
# Copyright (C) 2016 Intel Corporation
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice (including the next
# paragraph) shall be included in all copies or substantial portions of the
# Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
# IN THE SOFTWARE.
import argparse
import sys
import math
a = 'a'
# The nir_lower_bit_size() pass gets rid of all 8bit ALUs but insert new u2u8
# and i2i8 operations to convert the result back to the original type after the
# arithmetic operation is done. Those u2u8 and i2i8 operations, as any other
# 8bit operations, are not supported by DXIL and needs to be discarded. The
# dxil_nir_lower_8bit_conv() pass is here for that.
# Similarly, some hardware doesn't support 16bit values
no_8bit_conv = []
no_16bit_conv = []
def remove_unsupported_casts(arr, bit_size, mask, max_unsigned_float, min_signed_float, max_signed_float):
for outer_op_type in ('u2u', 'i2i', 'u2f', 'i2f'):
for outer_op_sz in (16, 32, 64):
if outer_op_sz == bit_size:
continue
outer_op = outer_op_type + str(int(outer_op_sz))
for inner_op_type in ('u2u', 'i2i'):
inner_op = inner_op_type + str(int(bit_size))
for src_sz in (16, 32, 64):
if (src_sz == bit_size):
continue
# Coming from integral, truncate appropriately
orig_seq = (outer_op, (inner_op, 'a@' + str(int(src_sz))))
if (outer_op[0] == 'u'):
new_seq = ('iand', a, mask)
else:
shift = src_sz - bit_size
new_seq = ('ishr', ('ishl', a, shift), shift)
# Make sure the destination is the right type/size
if outer_op_sz != src_sz or outer_op[2] != inner_op[0]:
new_seq = (outer_op, new_seq)
arr += [(orig_seq, new_seq)]
for inner_op_type in ('f2u', 'f2i'):
inner_op = inner_op_type + str(int(bit_size))
if (outer_op[2] == 'f'):
# From float and to float, just truncate via min/max, and ensure the right float size
for src_sz in (16, 32, 64):
if (src_sz == bit_size):
continue
orig_seq = (outer_op, (inner_op, 'a@' + str(int(src_sz))))
if (outer_op[0] == 'u'):
new_seq = ('fmin', ('fmax', a, 0.0), max_unsigned_float)
else:
new_seq = ('fmin', ('fmax', a, min_signed_float), max_signed_float)
if outer_op_sz != src_sz:
new_seq = ('f2f' + str(int(outer_op_sz)), new_seq)
arr += [(orig_seq, new_seq)]
else:
# From float to integral, convert to integral type first, then truncate
orig_seq = (outer_op, (inner_op, a))
float_conv = ('f2' + inner_op[2] + str(int(outer_op_sz)), a)
if (outer_op[0] == 'u'):
new_seq = ('iand', float_conv, mask)
else:
shift = outer_op_sz - bit_size
new_seq = ('ishr', ('ishl', float_conv, shift), shift)
arr += [(orig_seq, new_seq)]
remove_unsupported_casts(no_8bit_conv, 8, 0xff, 255.0, -128.0, 127.0)
remove_unsupported_casts(no_16bit_conv, 16, 0xffff, 65535.0, -32768.0, 32767.0)
lower_x2b = [
(('b2b32', 'a'), ('b2i32', 'a')),
(('b2b1', 'a'), ('i2b1', 'a')),
(('i2b1', 'a'), ('ine', a, 0)),
(('f2b1', 'a'), ('fneu', a, 0)),
]
no_16bit_conv += [
(('f2f32', ('u2u16', 'a@32')), ('unpack_half_2x16_split_x', 'a')),
(('u2u32', ('f2f16_rtz', 'a@32')), ('pack_half_2x16_split', 'a', 0)),
]
lower_inot = [
(('inot', a), ('ixor', a, -1)),
]
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-p', '--import-path', required=True)
args = parser.parse_args()
sys.path.insert(0, args.import_path)
run()
def run():
import nir_algebraic # pylint: disable=import-error
print('#include "dxil_nir.h"')
print(nir_algebraic.AlgebraicPass("dxil_nir_lower_8bit_conv",
no_8bit_conv).render())
print(nir_algebraic.AlgebraicPass("dxil_nir_lower_16bit_conv",
no_16bit_conv).render())
print(nir_algebraic.AlgebraicPass("dxil_nir_lower_x2b",
lower_x2b).render())
print(nir_algebraic.AlgebraicPass("dxil_nir_lower_inot",
lower_inot).render())
if __name__ == '__main__':
main()

555
src/microsoft/compiler/dxil_nir_lower_int_samplers.c Normal file
View File

@@ -0,0 +1,555 @@
/*
* Copyright © Microsoft Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* on the rights to use, copy, modify, merge, publish, distribute, sub
* license, and/or sell copies of the Software, and to permit persons to whom
* the Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
* THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
* USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
#include "dxil_nir_lower_int_samplers.h"
#include "nir_builder.h"
#include "nir_builtin_builder.h"
bool
lower_sample_to_txf_for_integer_tex_filter(const nir_instr *instr,
UNUSED const void *_options)
{
if (instr->type != nir_instr_type_tex)
return false;
nir_tex_instr *tex = nir_instr_as_tex(instr);
if (tex->op != nir_texop_tex &&
tex->op != nir_texop_txb &&
tex->op != nir_texop_txl &&
tex->op != nir_texop_txd)
return false;
return (tex->dest_type & (nir_type_int | nir_type_uint));
}
nir_ssa_def *
dx_get_texture_lod(nir_builder *b, nir_tex_instr *tex)
{
nir_tex_instr *tql;
unsigned num_srcs = 0;
for (unsigned i = 0; i < tex->num_srcs; i++) {
if (tex->src[i].src_type == nir_tex_src_coord ||
tex->src[i].src_type == nir_tex_src_texture_deref ||
tex->src[i].src_type == nir_tex_src_sampler_deref ||
tex->src[i].src_type == nir_tex_src_texture_offset ||
tex->src[i].src_type == nir_tex_src_sampler_offset ||
tex->src[i].src_type == nir_tex_src_texture_handle ||
tex->src[i].src_type == nir_tex_src_sampler_handle)
num_srcs++;
}
tql = nir_tex_instr_create(b->shader, num_srcs);
tql->op = nir_texop_lod;
unsigned coord_components = tex->coord_components;
if (tex->is_array)
--coord_components;
tql->coord_components = coord_components;
tql->sampler_dim = tex->sampler_dim;
tql->is_shadow = tex->is_shadow;
tql->is_new_style_shadow = tex->is_new_style_shadow;
tql->texture_index = tex->texture_index;
tql->sampler_index = tex->sampler_index;
tql->dest_type = nir_type_float;
/* The coordinate needs special handling because we might have
* to strip the array index. Don't clutter the code with an additional
* check for is_array though, in the worst case we create an additional
* move the the optimization will remove later again. */
int coord_index = nir_tex_instr_src_index(tex, nir_tex_src_coord);
nir_ssa_def *ssa_src = nir_channels(b, tex->src[coord_index].src.ssa,
(1 << coord_components) - 1);
nir_src src = nir_src_for_ssa(ssa_src);
nir_src_copy(&tql->src[0].src, &src, tql);
tql->src[0].src_type = nir_tex_src_coord;
unsigned idx = 1;
for (unsigned i = 0; i < tex->num_srcs; i++) {
if (tex->src[i].src_type == nir_tex_src_texture_deref ||
tex->src[i].src_type == nir_tex_src_sampler_deref ||
tex->src[i].src_type == nir_tex_src_texture_offset ||
tex->src[i].src_type == nir_tex_src_sampler_offset ||
tex->src[i].src_type == nir_tex_src_texture_handle ||
tex->src[i].src_type == nir_tex_src_sampler_handle) {
nir_src_copy(&tql->src[idx].src, &tex->src[i].src, tql);
tql->src[idx].src_type = tex->src[i].src_type;
idx++;
}
}
nir_ssa_dest_init(&tql->instr, &tql->dest, 2, 32, NULL);
nir_builder_instr_insert(b, &tql->instr);
/* DirectX LOD only has a value in x channel */
return nir_channel(b, &tql->dest.ssa, 0);
}
typedef struct {
nir_ssa_def *coords;
nir_ssa_def *use_border_color;
} wrap_result_t;
typedef struct {
nir_ssa_def *lod;
nir_ssa_def *size;
int ncoord_comp;
wrap_result_t wrap[3];
} wrap_lower_param_t;
static void
wrap_clamp_to_edge(nir_builder *b, wrap_result_t *wrap_params, nir_ssa_def *size)
{
/* clamp(coord, 0, size - 1) */
wrap_params->coords = nir_fmin(b, nir_fsub(b, size, nir_imm_float(b, 1.0f)),
nir_fmax(b, wrap_params->coords, nir_imm_float(b, 0.0f)));
}
static void
wrap_repeat(nir_builder *b, wrap_result_t *wrap_params, nir_ssa_def *size)
{
/* mod(coord, size)
* This instruction must be exact, otherwise certain sizes result in
* incorrect sampling */
wrap_params->coords = nir_fmod(b, wrap_params->coords, size);
nir_instr_as_alu(wrap_params->coords->parent_instr)->exact = true;
}
static nir_ssa_def *
mirror(nir_builder *b, nir_ssa_def *coord)
{
/* coord if >= 0, otherwise -(1 + coord) */
return nir_bcsel(b, nir_fge(b, coord, nir_imm_float(b, 0.0f)), coord,
nir_fneg(b, nir_fadd(b, nir_imm_float(b, 1.0f), coord)));
}
static void
wrap_mirror_repeat(nir_builder *b, wrap_result_t *wrap_params, nir_ssa_def *size)
{
/* (size − 1) − mirror(mod(coord, 2 * size) − size) */
nir_ssa_def *coord_mod2size = nir_fmod(b, wrap_params->coords, nir_fmul(b, nir_imm_float(b, 2.0f), size));
nir_instr_as_alu(coord_mod2size->parent_instr)->exact = true;
nir_ssa_def *a = nir_fsub(b, coord_mod2size, size);
wrap_params->coords = nir_fsub(b, nir_fsub(b, size, nir_imm_float(b, 1.0f)), mirror(b, a));
}
static void
wrap_mirror_clamp_to_edge(nir_builder *b, wrap_result_t *wrap_params, nir_ssa_def *size)
{
/* clamp(mirror(coord), 0, size - 1) */
wrap_params->coords = nir_fmin(b, nir_fsub(b, size, nir_imm_float(b, 1.0f)),
nir_fmax(b, mirror(b, wrap_params->coords), nir_imm_float(b, 0.0f)));
}
static void
wrap_clamp(nir_builder *b, wrap_result_t *wrap_params, nir_ssa_def *size)
{
nir_ssa_def *is_low = nir_flt(b, wrap_params->coords, nir_imm_float(b, 0.0));
nir_ssa_def *is_high = nir_fge(b, wrap_params->coords, size);
wrap_params->use_border_color = nir_ior(b, is_low, is_high);
}
static void
wrap_mirror_clamp(nir_builder *b, wrap_result_t *wrap_params, nir_ssa_def *size)
{
/* We have to take care of the boundaries */
nir_ssa_def *is_low = nir_flt(b, wrap_params->coords, nir_fmul(b, size, nir_imm_float(b, -1.0)));
nir_ssa_def *is_high = nir_flt(b, nir_fmul(b, size, nir_imm_float(b, 2.0)), wrap_params->coords);
wrap_params->use_border_color = nir_ior(b, is_low, is_high);
/* Within the boundaries this acts like mirror_repeat */
wrap_mirror_repeat(b, wrap_params, size);
}
static wrap_result_t
wrap_coords(nir_builder *b, nir_ssa_def *coords, enum pipe_tex_wrap wrap,
nir_ssa_def *size)
{
wrap_result_t result = {coords, nir_imm_false(b)};
switch (wrap) {
case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
wrap_clamp_to_edge(b, &result, size);
break;
case PIPE_TEX_WRAP_REPEAT:
wrap_repeat(b, &result, size);
break;
case PIPE_TEX_WRAP_MIRROR_REPEAT:
wrap_mirror_repeat(b, &result, size);
break;
case PIPE_TEX_WRAP_MIRROR_CLAMP:
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
wrap_mirror_clamp_to_edge(b, &result, size);
break;
case PIPE_TEX_WRAP_CLAMP:
case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
wrap_clamp(b, &result, size);
break;
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
wrap_mirror_clamp(b, &result, size);
break;
}
return result;
}
static nir_ssa_def *
load_bordercolor(nir_builder *b, nir_tex_instr *tex, dxil_wrap_sampler_state *active_state,
const dxil_texture_swizzle_state *tex_swizzle)
{
nir_const_value const_value[4] = {{0}};
int ndest_comp = nir_dest_num_components(tex->dest);
unsigned swizzle[4] = {
tex_swizzle->swizzle_r,
tex_swizzle->swizzle_g,
tex_swizzle->swizzle_b,
tex_swizzle->swizzle_a
};
for (int i = 0; i < ndest_comp; ++i) {
switch (swizzle[i]) {
case PIPE_SWIZZLE_0:
const_value[i].f32 = 0;
break;
case PIPE_SWIZZLE_1:
const_value[i].i32 = 1;
break;
case PIPE_SWIZZLE_X:
case PIPE_SWIZZLE_Y:
case PIPE_SWIZZLE_Z:
case PIPE_SWIZZLE_W:
const_value[i].f32 = active_state->border_color[swizzle[i]];
break;
default:
unreachable("Unexpected swizzle value");
}
}
return nir_build_imm(b, ndest_comp, 32, const_value);
}
nir_tex_instr *
create_txf_from_tex(nir_builder *b, nir_tex_instr *tex)
{
nir_tex_instr *txf;
unsigned num_srcs = 0;
for (unsigned i = 0; i < tex->num_srcs; i++) {
if (tex->src[i].src_type == nir_tex_src_texture_deref ||
tex->src[i].src_type == nir_tex_src_texture_offset ||
tex->src[i].src_type == nir_tex_src_texture_handle)
num_srcs++;
}
txf = nir_tex_instr_create(b->shader, num_srcs);
txf->op = nir_texop_txf;
txf->sampler_dim = tex->sampler_dim;
txf->is_array = tex->is_array;
txf->is_shadow = tex->is_shadow;
txf->is_new_style_shadow = tex->is_new_style_shadow;
txf->texture_index = tex->texture_index;
txf->sampler_index = tex->sampler_index;
txf->dest_type = tex->dest_type;
unsigned idx = 0;
for (unsigned i = 0; i < tex->num_srcs; i++) {
if (tex->src[i].src_type == nir_tex_src_texture_deref ||
tex->src[i].src_type == nir_tex_src_texture_offset ||
tex->src[i].src_type == nir_tex_src_texture_handle) {
nir_src_copy(&txf->src[idx].src, &tex->src[i].src, txf);
txf->src[idx].src_type = tex->src[i].src_type;
idx++;
}
}
nir_ssa_dest_init(&txf->instr, &txf->dest,
nir_tex_instr_dest_size(txf), 32, NULL);
nir_builder_instr_insert(b, &txf->instr);
return txf;
}
static nir_ssa_def *
load_texel(nir_builder *b, nir_tex_instr *tex, wrap_lower_param_t *params)
{
nir_ssa_def *texcoord = NULL;
/* Put coordinates back together */
switch (tex->coord_components) {
case 1:
texcoord = params->wrap[0].coords;
break;
case 2:
texcoord = nir_vec2(b, params->wrap[0].coords, params->wrap[1].coords);
break;
case 3:
texcoord = nir_vec3(b, params->wrap[0].coords, params->wrap[1].coords, params->wrap[2].coords);
break;
default:
;
}
texcoord = nir_f2i32(b, texcoord);
nir_tex_instr *load = create_txf_from_tex(b, tex);
nir_tex_instr_add_src(load, nir_tex_src_lod, nir_src_for_ssa(params->lod));
nir_tex_instr_add_src(load, nir_tex_src_coord, nir_src_for_ssa(texcoord));
b->cursor = nir_after_instr(&load->instr);
return &load->dest.ssa;
}
typedef struct {
dxil_wrap_sampler_state *aws;
float max_bias;
nir_ssa_def *size;
int ncoord_comp;
} lod_params;
static nir_ssa_def *
evalute_active_lod(nir_builder *b, nir_tex_instr *tex, lod_params *params)
{
static nir_ssa_def *lod = NULL;
/* Later we use min_lod for clamping the LOD to a legal value */
float min_lod = MAX2(params->aws->min_lod, 0.0f);
/* Evaluate the LOD to be used for the texel fetch */
if (unlikely(tex->op == nir_texop_txl)) {
int lod_index = nir_tex_instr_src_index(tex, nir_tex_src_lod);
/* if we have an explicite LOD, take it */
lod = tex->src[lod_index].src.ssa;
} else if (unlikely(tex->op == nir_texop_txd)) {
int ddx_index = nir_tex_instr_src_index(tex, nir_tex_src_ddx);
int ddy_index = nir_tex_instr_src_index(tex, nir_tex_src_ddy);
assert(ddx_index >= 0 && ddy_index >= 0);
nir_ssa_def *grad = nir_fmax(b,
tex->src[ddx_index].src.ssa,
tex->src[ddy_index].src.ssa);
nir_ssa_def *r = nir_fmul(b, grad, nir_i2f32(b, params->size));
nir_ssa_def *rho = nir_channel(b, r, 0);
for (int i = 1; i < params->ncoord_comp; ++i)
rho = nir_fmax(b, rho, nir_channel(b, r, i));
lod = nir_flog2(b, rho);
} else if (b->shader->info.stage == MESA_SHADER_FRAGMENT){
lod = dx_get_texture_lod(b, tex);
} else {
/* Only fragment shaders provide the gradient information to evaluate a LOD,
* so force 0 otherwise */
lod = nir_imm_float(b, 0.0);
}
/* Evaluate bias according to OpenGL (4.6 (Compatibility Profile) October 22, 2019),
* sec. 8.14.1, eq. (8.9)
*
* lod' = lambda + CLAMP(bias_texobj + bias_texunit + bias_shader)
*
* bias_texobj is the value of TEXTURE_LOD_BIAS for the bound texture object. ...
* bias_textunt is the value of TEXTURE_LOD_BIAS for the current texture unit, ...
* bias shader is the value of the optional bias parameter in the texture
* lookup functions available to fragment shaders. ... The sum of these values
* is clamped to the range [−bias_max, bias_max] where bias_max is the value
* of the implementation defined constant MAX_TEXTURE_LOD_BIAS.
* In core contexts the value bias_texunit is dropped from above equation.
*
* Gallium provides the value lod_bias as the sum of bias_texobj and bias_texunit
* in compatibility contexts and as bias_texobj in core contexts, hence the
* implementation here is the same in both cases.
*/
nir_ssa_def *lod_bias = nir_imm_float(b, params->aws->lod_bias);
if (unlikely(tex->op == nir_texop_txb)) {
int bias_index = nir_tex_instr_src_index(tex, nir_tex_src_bias);
lod_bias = nir_fadd(b, lod_bias, tex->src[bias_index].src.ssa);
}
lod = nir_fadd(b, lod, nir_fclamp(b, lod_bias,
nir_imm_float(b, -params->max_bias),
nir_imm_float(b, params->max_bias)));
/* Clamp lod according to ibid. eq. (8.10) */
lod = nir_fmax(b, lod, nir_imm_float(b, min_lod));
/* If the max lod is > max_bias = log2(max_texture_size), the lod will be clamped
* by the number of levels, no need to clamp it againt the max_lod first. */
if (params->aws->max_lod <= params->max_bias)
lod = nir_fmin(b, lod, nir_imm_float(b, params->aws->max_lod));
/* Pick nearest LOD */
lod = nir_f2i32(b, nir_fround_even(b, lod));
/* cap actual lod by number of available levels */
return nir_imin(b, lod, nir_imm_int(b, params->aws->last_level));
}
typedef struct {
dxil_wrap_sampler_state *wrap_states;
dxil_texture_swizzle_state *tex_swizzles;
float max_bias;
} sampler_states;
static nir_ssa_def *
lower_sample_to_txf_for_integer_tex_impl(nir_builder *b, nir_instr *instr,
void *options)
{
sampler_states *states = (sampler_states *)options;
wrap_lower_param_t params = {0};
nir_tex_instr *tex = nir_instr_as_tex(instr);
dxil_wrap_sampler_state *active_wrap_state = &states->wrap_states[tex->sampler_index];
b->cursor = nir_before_instr(instr);
int coord_index = nir_tex_instr_src_index(tex, nir_tex_src_coord);
nir_ssa_def *old_coord = tex->src[coord_index].src.ssa;
params.ncoord_comp = tex->coord_components;
if (tex->is_array)
params.ncoord_comp -= 1;
/* This helper to get the texture size always uses LOD 0, and DirectX doesn't support
* giving another LOD when querying the texture size */
nir_ssa_def *size0 = nir_get_texture_size(b, tex);
params.lod = nir_imm_int(b, 0);
if (active_wrap_state->last_level > 0) {
lod_params p = {
.aws = active_wrap_state,
.max_bias = states->max_bias,
.size = size0,
.ncoord_comp = params.ncoord_comp
};
params.lod = evalute_active_lod(b, tex, &p);
/* Evaluate actual level size*/
params.size = nir_i2f32(b, nir_imax(b, nir_ishr(b, size0, params.lod),
nir_imm_int(b, 1)));
} else {
params.size = nir_i2f32(b, size0);
}
nir_ssa_def *new_coord = old_coord;
if (!active_wrap_state->is_nonnormalized_coords) {
/* Evaluate the integer lookup coordinates for the requested LOD, don't touch the
* array index */
if (!tex->is_array) {
new_coord = nir_fmul(b, params.size, old_coord);
} else {
nir_ssa_def *array_index = nir_channel(b, old_coord, params.ncoord_comp);
int mask = (1 << params.ncoord_comp) - 1;
nir_ssa_def *coord = nir_fmul(b, nir_channels(b, params.size, mask),
nir_channels(b, old_coord, mask));
switch (params.ncoord_comp) {
case 1:
new_coord = nir_vec2(b, coord, array_index);
break;
case 2:
new_coord = nir_vec3(b, nir_channel(b, coord, 0),
nir_channel(b, coord, 1),
array_index);
break;
default:
unreachable("unsupported number of non-array coordinates");
}
}
}
nir_ssa_def *coord_help[3];
for (int i = 0; i < params.ncoord_comp; ++i)
coord_help[i] = nir_ffloor(b, nir_channel(b, new_coord, i));
// Note: array index needs to be rounded to nearest before clamp rather than floored
if (tex->is_array)
coord_help[params.ncoord_comp] = nir_fround_even(b, nir_channel(b, new_coord, params.ncoord_comp));
/* Correct the texture coordinates for the offsets. */
int offset_index = nir_tex_instr_src_index(tex, nir_tex_src_offset);
if (offset_index >= 0) {
nir_ssa_def *offset = tex->src[offset_index].src.ssa;
for (int i = 0; i < params.ncoord_comp; ++i)
coord_help[i] = nir_fadd(b, coord_help[i], nir_i2f32(b, nir_channel(b, offset, i)));
}
nir_ssa_def *use_border_color = nir_imm_false(b);
if (!active_wrap_state->skip_boundary_conditions) {
for (int i = 0; i < params.ncoord_comp; ++i) {
params.wrap[i] = wrap_coords(b, coord_help[i], active_wrap_state->wrap[i], nir_channel(b, params.size, i));
use_border_color = nir_ior(b, use_border_color, params.wrap[i].use_border_color);
}
if (tex->is_array)
params.wrap[params.ncoord_comp] =
wrap_coords(b, coord_help[params.ncoord_comp],
PIPE_TEX_WRAP_CLAMP_TO_EDGE,
nir_i2f32(b, nir_channel(b, size0, params.ncoord_comp)));
} else {
/* When we emulate a cube map by using a texture array, the coordinates are always
* in range, and we don't have to take care of boundary conditions */
for (unsigned i = 0; i < 3; ++i) {
params.wrap[i].coords = coord_help[i];
params.wrap[i].use_border_color = nir_imm_false(b);
}
}
const dxil_texture_swizzle_state one2one = {
PIPE_SWIZZLE_X, PIPE_SWIZZLE_Y, PIPE_SWIZZLE_Z, PIPE_SWIZZLE_W
};
nir_if *border_if = nir_push_if(b, use_border_color);
const dxil_texture_swizzle_state *swizzle = states->tex_swizzles ?
&states->tex_swizzles[tex->sampler_index]:
&one2one;
nir_ssa_def *border_color = load_bordercolor(b, tex, active_wrap_state, swizzle);
nir_if *border_else = nir_push_else(b, border_if);
nir_ssa_def *sampler_color = load_texel(b, tex, &params);
nir_pop_if(b, border_else);
return nir_if_phi(b, border_color, sampler_color);
}
/* Sampling from integer textures is not allowed in DirectX, so we have
* to use texel fetches. For this we have to scale the coordiantes
* to be integer based, and evaluate the LOD the texel fetch has to be
* applied on, and take care of the boundary conditions .
*/
bool
dxil_lower_sample_to_txf_for_integer_tex(nir_shader *s,
dxil_wrap_sampler_state *wrap_states,
dxil_texture_swizzle_state *tex_swizzles,
float max_bias)
{
sampler_states states = {wrap_states, tex_swizzles, max_bias};
bool result =
nir_shader_lower_instructions(s,
lower_sample_to_txf_for_integer_tex_filter,
lower_sample_to_txf_for_integer_tex_impl,
&states);
return result;
}

64
src/microsoft/compiler/dxil_nir_lower_int_samplers.h Normal file
View File

@@ -0,0 +1,64 @@
/*
* Copyright © Microsoft Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* on the rights to use, copy, modify, merge, publish, distribute, sub
* license, and/or sell copies of the Software, and to permit persons to whom
* the Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
* THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
* USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
#ifndef DXIL_NIR_LOWER_INT_SAMPLERS_H
#define DXIL_NIR_LOWER_INT_SAMPLERS_H
#include "pipe/p_state.h"
#include "nir.h"
#ifdef __cplusplus
extern "C" {
#endif
typedef struct {
unsigned swizzle_r:3;
unsigned swizzle_g:3;
unsigned swizzle_b:3;
unsigned swizzle_a:3;
} dxil_texture_swizzle_state;
typedef struct {
float border_color[4];
float lod_bias;
float min_lod, max_lod;
int last_level;
uint8_t wrap[3];
uint8_t is_int_sampler:1;
uint8_t is_nonnormalized_coords:1;
uint8_t is_linear_filtering:1;
uint8_t skip_boundary_conditions:1;
uint8_t unused:4;
} dxil_wrap_sampler_state;
bool
dxil_lower_sample_to_txf_for_integer_tex(nir_shader *s,
dxil_wrap_sampler_state *wrap_states,
dxil_texture_swizzle_state *tex_swizzles,
float max_bias);
#ifdef __cplusplus
}
#endif
#endif // DXIL_NIR_LOWER_INT_SAMPLERS_H

17
src/microsoft/compiler/meson.build
View File

@@ -26,12 +26,27 @@ files_libdxil_compiler = files(
'dxil_enums.c', 'dxil_enums.c',
'dxil_function.c', 'dxil_function.c',
'dxil_module.c', 'dxil_module.c',
'dxil_nir.c',
'dxil_nir_lower_int_samplers.c',
'dxil_signature.c', 'dxil_signature.c',
'nir_to_dxil.c',
)
dxil_nir_algebraic_c = custom_target(
'dxil_nir_algebraic.c',
input : 'dxil_nir_algebraic.py',
output : 'dxil_nir_algebraic.c',
command : [
prog_python, '@INPUT@',
'-p', join_paths(meson.source_root(), 'src/compiler/nir/'),
],
capture : true,
depend_files : nir_algebraic_py,
) )
libdxil_compiler = static_library( libdxil_compiler = static_library(
'dxil_compiler', 'dxil_compiler',
[files_libdxil_compiler, sha1_h], [files_libdxil_compiler, dxil_nir_algebraic_c, sha1_h],
include_directories : [inc_include, inc_src, inc_mapi, inc_mesa, inc_compiler, inc_gallium], include_directories : [inc_include, inc_src, inc_mapi, inc_mesa, inc_compiler, inc_gallium],
dependencies: [idep_nir_headers], dependencies: [idep_nir_headers],
gnu_symbol_visibility : 'hidden', gnu_symbol_visibility : 'hidden',

3725
src/microsoft/compiler/nir_to_dxil.c Normal file
View File

File diff suppressed because it is too large Load Diff

68
src/microsoft/compiler/nir_to_dxil.h Normal file
View File

@@ -0,0 +1,68 @@
/*
* Copyright © Microsoft Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#ifndef NIR_TO_DXIL_H
#define NIR_TO_DXIL_H
#include <stdbool.h>
#include "nir.h"
#ifdef __cplusplus
extern "C" {
#endif
struct blob;
const char *
dxil_vs_attr_index_to_name(unsigned index);
enum dxil_sysvalue_type {
DXIL_NO_SYSVALUE = 0,
DXIL_SYSVALUE,
DXIL_GENERATED_SYSVALUE
};
enum dxil_sysvalue_type
nir_var_to_dxil_sysvalue_type(nir_variable *var, uint64_t other_stage_mask);
struct nir_to_dxil_options {
bool interpolate_at_vertex;
bool lower_int16;
bool disable_math_refactoring;
unsigned ubo_binding_offset;
unsigned provoking_vertex;
};
bool
nir_to_dxil(struct nir_shader *s, const struct nir_to_dxil_options *opts,
struct blob *blob);
const nir_shader_compiler_options*
dxil_get_nir_compiler_options(void);
#ifdef __cplusplus
}
#endif
#endif