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

amd: lower load_barycentric_at_offset in NIR

Browse Source 
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32782>
This commit is contained in:
Marek Olšák
2024-12-25 12:44:59 -05:00
committed by Marge Bot
parent 16ab05fad1
commit 923f59c971
6 changed files with 19 additions and 124 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
src/amd/common/ac_nir.c
View File

@@ -305,6 +305,25 @@ lower_intrinsic_to_arg(nir_builder *b, nir_instr *instr, void *state)
case nir_intrinsic_load_barycentric_model:
replacement = ac_nir_load_arg(b, s->args, s->args->pull_model);
break;
case nir_intrinsic_load_barycentric_at_offset: {
nir_def *baryc = nir_intrinsic_interp_mode(intrin) == INTERP_MODE_NOPERSPECTIVE ?
ac_nir_load_arg(b, s->args, s->args->linear_center) :
ac_nir_load_arg(b, s->args, s->args->persp_center);
nir_def *i = nir_channel(b, baryc, 0);
nir_def *j = nir_channel(b, baryc, 1);
nir_def *offset_x = nir_channel(b, intrin->src[0].ssa, 0);
nir_def *offset_y = nir_channel(b, intrin->src[0].ssa, 1);
nir_def *ddx_i = nir_ddx(b, i);
nir_def *ddx_j = nir_ddx(b, j);
nir_def *ddy_i = nir_ddy(b, i);
nir_def *ddy_j = nir_ddy(b, j);
/* Interpolate standard barycentrics by offset. */
nir_def *offset_i = nir_ffma(b, ddy_i, offset_y, nir_ffma(b, ddx_i, offset_x, i));
nir_def *offset_j = nir_ffma(b, ddy_j, offset_y, nir_ffma(b, ddx_j, offset_x, j));
replacement = nir_vec2(b, offset_i, offset_j);
break;
}
default:
return false;
}

57
src/amd/compiler/aco_instruction_selection.cpp
View File

@@ -7895,52 +7895,6 @@ emit_rotate_by_constant(isel_context* ctx, Temp& dst, Temp src, unsigned cluster
return dst.id() != 0;
}
void
emit_interp_center(isel_context* ctx, Temp dst, Temp bary, Temp pos1, Temp pos2)
{
Builder bld(ctx->program, ctx->block);
Temp p1 = emit_extract_vector(ctx, bary, 0, v1);
Temp p2 = emit_extract_vector(ctx, bary, 1, v1);
Temp ddx_1, ddx_2, ddy_1, ddy_2;
uint32_t dpp_ctrl0 = dpp_quad_perm(0, 0, 0, 0);
uint32_t dpp_ctrl1 = dpp_quad_perm(1, 1, 1, 1);
uint32_t dpp_ctrl2 = dpp_quad_perm(2, 2, 2, 2);
/* Build DD X/Y */
if (ctx->program->gfx_level >= GFX8) {
Temp tl_1 = bld.vop1_dpp(aco_opcode::v_mov_b32, bld.def(v1), p1, dpp_ctrl0);
ddx_1 = bld.vop2_dpp(aco_opcode::v_sub_f32, bld.def(v1), p1, tl_1, dpp_ctrl1);
ddy_1 = bld.vop2_dpp(aco_opcode::v_sub_f32, bld.def(v1), p1, tl_1, dpp_ctrl2);
Temp tl_2 = bld.vop1_dpp(aco_opcode::v_mov_b32, bld.def(v1), p2, dpp_ctrl0);
ddx_2 = bld.vop2_dpp(aco_opcode::v_sub_f32, bld.def(v1), p2, tl_2, dpp_ctrl1);
ddy_2 = bld.vop2_dpp(aco_opcode::v_sub_f32, bld.def(v1), p2, tl_2, dpp_ctrl2);
} else {
Temp tl_1 = bld.ds(aco_opcode::ds_swizzle_b32, bld.def(v1), p1, (1 << 15) | dpp_ctrl0);
ddx_1 = bld.ds(aco_opcode::ds_swizzle_b32, bld.def(v1), p1, (1 << 15) | dpp_ctrl1);
ddx_1 = bld.vop2(aco_opcode::v_sub_f32, bld.def(v1), ddx_1, tl_1);
ddy_1 = bld.ds(aco_opcode::ds_swizzle_b32, bld.def(v1), p1, (1 << 15) | dpp_ctrl2);
ddy_1 = bld.vop2(aco_opcode::v_sub_f32, bld.def(v1), ddy_1, tl_1);
Temp tl_2 = bld.ds(aco_opcode::ds_swizzle_b32, bld.def(v1), p2, (1 << 15) | dpp_ctrl0);
ddx_2 = bld.ds(aco_opcode::ds_swizzle_b32, bld.def(v1), p2, (1 << 15) | dpp_ctrl1);
ddx_2 = bld.vop2(aco_opcode::v_sub_f32, bld.def(v1), ddx_2, tl_2);
ddy_2 = bld.ds(aco_opcode::ds_swizzle_b32, bld.def(v1), p2, (1 << 15) | dpp_ctrl2);
ddy_2 = bld.vop2(aco_opcode::v_sub_f32, bld.def(v1), ddy_2, tl_2);
}
/* res_k = p_k + ddx_k * pos1 + ddy_k * pos2 */
aco_opcode mad =
ctx->program->gfx_level >= GFX10_3 ? aco_opcode::v_fma_f32 : aco_opcode::v_mad_f32;
Temp tmp1 = bld.vop3(mad, bld.def(v1), ddx_1, pos1, p1);
Temp tmp2 = bld.vop3(mad, bld.def(v1), ddx_2, pos1, p2);
tmp1 = bld.vop3(mad, bld.def(v1), ddy_1, pos2, tmp1);
tmp2 = bld.vop3(mad, bld.def(v1), ddy_2, pos2, tmp2);
bld.pseudo(aco_opcode::p_create_vector, Definition(dst), tmp1, tmp2);
set_wqm(ctx, true);
return;
}
Temp merged_wave_info_to_mask(isel_context* ctx, unsigned i);
Temp lanecount_to_mask(isel_context* ctx, Temp count, unsigned bit_offset);
void pops_await_overlapped_waves(isel_context* ctx);
@@ -8042,17 +7996,6 @@ visit_intrinsic(isel_context* ctx, nir_intrinsic_instr* instr)
{
Builder bld(ctx->program, ctx->block);
switch (instr->intrinsic) {
case nir_intrinsic_load_barycentric_at_offset: {
Temp offset = get_ssa_temp(ctx, instr->src[0].ssa);
RegClass rc = RegClass(offset.type(), 1);
Temp pos1 = bld.tmp(rc), pos2 = bld.tmp(rc);
bld.pseudo(aco_opcode::p_split_vector, Definition(pos1), Definition(pos2), offset);
Temp bary = get_arg(ctx, nir_intrinsic_interp_mode(instr) == INTERP_MODE_NOPERSPECTIVE
? ctx->args->linear_center
: ctx->args->persp_center);
emit_interp_center(ctx, get_ssa_temp(ctx, &instr->def), bary, pos1, pos2);
break;
}
case nir_intrinsic_load_tess_coord: visit_load_tess_coord(ctx, instr); break;
case nir_intrinsic_load_interpolated_input: visit_load_interpolated_input(ctx, instr); break;
case nir_intrinsic_store_output: visit_store_output(ctx, instr); break;

1
src/amd/compiler/aco_instruction_selection_setup.cpp
View File

@@ -550,7 +550,6 @@ init_context(isel_context* ctx, nir_shader* shader)
case nir_intrinsic_load_per_vertex_input:
case nir_intrinsic_load_per_vertex_output:
case nir_intrinsic_load_vertex_id_zero_base:
case nir_intrinsic_load_barycentric_at_offset:
case nir_intrinsic_load_interpolated_input:
case nir_intrinsic_load_local_invocation_index:
case nir_intrinsic_load_subgroup_invocation:

18
src/amd/llvm/ac_llvm_build.c
View File

@@ -3483,24 +3483,6 @@ LLVMValueRef ac_build_canonicalize(struct ac_llvm_context *ctx, LLVMValueRef src
return ac_build_intrinsic(ctx, intr, type, params, 1, 0);
}
/*
* this takes an I,J coordinate pair,
* and works out the X and Y derivatives.
* it returns DDX(I), DDX(J), DDY(I), DDY(J).
*/
LLVMValueRef ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij)
{
LLVMValueRef result[4], a;
unsigned i;
for (i = 0; i < 2; i++) {
a = LLVMBuildExtractElement(ctx->builder, interp_ij, LLVMConstInt(ctx->i32, i, false), "");
result[i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 1, a);
result[2 + i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 2, a);
}
return ac_build_gather_values(ctx, result, 4);
}
LLVMValueRef ac_build_load_helper_invocation(struct ac_llvm_context *ctx)
{
LLVMValueRef result = ac_build_intrinsic(ctx, "llvm.amdgcn.live.mask", ctx->i1, NULL, 0, 0);

2
src/amd/llvm/ac_llvm_build.h
View File

@@ -459,8 +459,6 @@ LLVMValueRef ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
LLVMValueRef ac_build_canonicalize(struct ac_llvm_context *ctx, LLVMValueRef src0,
unsigned bitsize);
LLVMValueRef ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
LLVMValueRef ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMTypeRef fn_type, LLVMValueRef func,

46
src/amd/llvm/ac_nir_to_llvm.c
View File

@@ -2657,47 +2657,6 @@ static LLVMValueRef visit_var_atomic(struct ac_nir_context *ctx, const nir_intri
return result;
}
static LLVMValueRef barycentric_offset(struct ac_nir_context *ctx, unsigned mode,
LLVMValueRef offset)
{
LLVMValueRef interp_param = mode == INTERP_MODE_NOPERSPECTIVE ?
ac_get_arg(&ctx->ac, ctx->args->linear_center) :
ac_get_arg(&ctx->ac, ctx->args->persp_center);
LLVMValueRef src_c0 =
ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, offset, ctx->ac.i32_0, ""));
LLVMValueRef src_c1 =
ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, offset, ctx->ac.i32_1, ""));
LLVMValueRef ij_out[2];
LLVMValueRef ddxy_out = ac_build_ddxy_interp(&ctx->ac, interp_param);
/*
* take the I then J parameters, and the DDX/Y for it, and
* calculate the IJ inputs for the interpolator.
* temp1 = ddx * offset/sample.x + I;
* interp_param.I = ddy * offset/sample.y + temp1;
* temp1 = ddx * offset/sample.x + J;
* interp_param.J = ddy * offset/sample.y + temp1;
*/
for (unsigned i = 0; i < 2; i++) {
LLVMValueRef ix_ll = LLVMConstInt(ctx->ac.i32, i, false);
LLVMValueRef iy_ll = LLVMConstInt(ctx->ac.i32, i + 2, false);
LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->ac.builder, ddxy_out, ix_ll, "");
LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder, ddxy_out, iy_ll, "");
LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder, interp_param, ix_ll, "");
LLVMValueRef temp1, temp2;
interp_el = LLVMBuildBitCast(ctx->ac.builder, interp_el, ctx->ac.f32, "");
temp1 = ac_build_fmad(&ctx->ac, ddx_el, src_c0, interp_el);
temp2 = ac_build_fmad(&ctx->ac, ddy_el, src_c1, temp1);
ij_out[i] = LLVMBuildBitCast(ctx->ac.builder, temp2, ctx->ac.i32, "");
}
interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
}
static LLVMValueRef load_interpolated_input(struct ac_nir_context *ctx, LLVMValueRef interp_param,
unsigned index, unsigned comp_start,
unsigned num_components, unsigned bitsize,
@@ -3010,11 +2969,6 @@ static bool visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *ins
result = visit_var_atomic(ctx, instr, ptr, 1);
break;
}
case nir_intrinsic_load_barycentric_at_offset: {
LLVMValueRef offset = ac_to_float(&ctx->ac, get_src(ctx, instr->src[0]));
result = barycentric_offset(ctx, nir_intrinsic_interp_mode(instr), offset);
break;
}
case nir_intrinsic_load_interpolated_input: {
/* We assume any indirect loads have been lowered away */
ASSERTED nir_const_value *offset = nir_src_as_const_value(instr->src[1]);