mirror of
https://git.triforceproject.com/triforce/source
synced 2026-11-20 16:49:57 -04:00
Triforce v0.3.6
This commit is contained in:
@@ -117,6 +117,15 @@ void EmitInst(EmitContext& ctx, IR::Inst* inst) {
|
||||
EmitInstOf(ctx, inst);
|
||||
} catch (const NotImplementedException& e) {
|
||||
throw NotImplementedException("{} ({})", e.what(), IR::NameOf(inst->GetOpcode()));
|
||||
} catch (const std::runtime_error& e) { // (super build: sirit's null Id, named by the IR op that asked for it)
|
||||
std::string args;
|
||||
for (size_t i = 0; i < inst->NumArgs(); ++i) {
|
||||
const IR::Value a{inst->Arg(i)};
|
||||
args += a.IsImmediate() ? " imm" : a.IsEmpty() ? " empty" : a.IsIdentity() ? " identity"
|
||||
: (" " + std::string(IR::NameOf(a.InstRecursive()->GetOpcode())) +
|
||||
(a.InstRecursive()->Definition<Id>().value == 0 ? "(undefined)" : ""));
|
||||
}
|
||||
throw LogicError("{} ({}; args:{})", e.what(), IR::NameOf(inst->GetOpcode()), args);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -45,6 +45,13 @@ constexpr u32 MESH_FALLBACK_PAYLOAD_PUSH_OFFSET = MESH_FALLBACK_PUSH_OFFSET;
|
||||
constexpr u32 MESH_FALLBACK_RECORDS_PUSH_OFFSET = MESH_FALLBACK_PUSH_OFFSET + 8;
|
||||
constexpr u32 MESH_FALLBACK_INDIRECT_INPUT_PUSH_OFFSET = MESH_FALLBACK_PUSH_OFFSET + 16;
|
||||
constexpr u32 MESH_FALLBACK_PUSH_END = MESH_FALLBACK_PUSH_OFFSET + 24;
|
||||
// (hookshot) a scene vertex stage's position capture (RuntimeInfo::capture_position), after them: the u64 address its slots
|
||||
// start at (0: none), the vertex index of slot 0 (a u32, the slot taken from VertexIndex wrapping) and the slots there are
|
||||
// (u32). A vertex stores its position (vec4, 16 bytes) at address + (VertexIndex - first) * 16 when that is below the count
|
||||
constexpr u32 CAPTURE_PUSH_OFFSET = MESH_FALLBACK_PUSH_END;
|
||||
constexpr u32 CAPTURE_FIRST_PUSH_OFFSET = CAPTURE_PUSH_OFFSET + 8;
|
||||
constexpr u32 CAPTURE_COUNT_PUSH_OFFSET = CAPTURE_PUSH_OFFSET + 12;
|
||||
constexpr u32 CAPTURE_PUSH_END = CAPTURE_PUSH_OFFSET + 16;
|
||||
// (hookshot) the fallback's mesh workgroups a draw can run: each a slot, a word of the payload buffer saying whether its task
|
||||
// started it; then each task workgroup's payload. It was 1024, a record block each, and Ocarina's floor grass (433 tasks of
|
||||
// 32 slots) lost 13 of 14 meshlets. At 16384, Hyrule Field's far trees (3cb76827/4f0c0b40: 721 tasks of 32) lost the tasks
|
||||
|
||||
@@ -323,7 +323,16 @@ Id EmitGetAttribute(EmitContext& ctx, IR::Attribute attr, Id vertex) {
|
||||
}
|
||||
const Id pointer{
|
||||
AttrPointer(ctx, generic.pointer_type, vertex, generic.id, ctx.Const(element))};
|
||||
const Id value{ctx.OpLoad(generic.component_type, pointer)};
|
||||
Id value{ctx.OpLoad(generic.component_type, pointer)};
|
||||
if (ctx.stage == Stage::Fragment && ctx.noperspective_as_smooth_inputs[index]) {
|
||||
// Written times Position.w and interpolated perspective-correct: undo with 1/w.
|
||||
// Only float inputs are marked, and DefineInputs forces FragCoord for them.
|
||||
ASSERT(generic.load_op == InputGenericLoadOp::None && ValidId(ctx.input_position));
|
||||
const Id frag_w{ctx.OpLoad(
|
||||
ctx.F32[1],
|
||||
ctx.OpAccessChain(ctx.input_f32, ctx.input_position, ctx.Const(3u)))};
|
||||
value = ctx.OpFMul(ctx.F32[1], value, frag_w);
|
||||
}
|
||||
return [&ctx, generic, value]() {
|
||||
switch (generic.load_op) {
|
||||
case InputGenericLoadOp::Bitcast:
|
||||
@@ -502,7 +511,7 @@ void EmitSetFragColor(EmitContext& ctx, u32 index, u32 component, Id value) {
|
||||
const Id pointer{ctx.OpAccessChain(ctx.output_f32, ctx.frag_color.at(index), component_id)};
|
||||
if (frag_nan_guard) { // (amd-fix, mma_layout.h) NaN -> 0, the rest clamped to +-1e30
|
||||
const Id clamped{ctx.OpNMin(ctx.F32[1], ctx.OpNMax(ctx.F32[1], value, ctx.Const(-1e30f)), ctx.Const(1e30f))};
|
||||
value = ctx.OpSelect(ctx.F32[1], ctx.OpIsNan(ctx.U1, value), ctx.Const(0.0f), clamped);
|
||||
value = ctx.OpSelect(ctx.F32[1], IsNan(ctx, value, 32), ctx.Const(0.0f), clamped);
|
||||
}
|
||||
ctx.OpStore(pointer, value);
|
||||
} else if (type == AttributeType::UnsignedInt) {
|
||||
|
||||
@@ -155,7 +155,7 @@ Id EmitConvertF16F32(EmitContext& ctx, Id value) {
|
||||
return ctx.OpFConvert(ctx.F16[1], value);
|
||||
#else
|
||||
const auto result = ctx.OpFConvert(ctx.F16[1], value);
|
||||
const auto isOverflowing = ctx.OpIsNan(ctx.U1, result);
|
||||
const auto isOverflowing = IsNan(ctx, result, 16); // (hookshot: mma_layout.h nan_bits)
|
||||
return ctx.OpSelect(ctx.F16[1], isOverflowing, ctx.Constant(ctx.F16[1], 0), result);
|
||||
#endif
|
||||
}
|
||||
|
||||
@@ -4,6 +4,7 @@
|
||||
// SPDX-License-Identifier: GPL-2.0-or-later
|
||||
|
||||
#include "shader_recompiler/backend/spirv/emit_spirv_instructions.h"
|
||||
#include "shader_recompiler/backend/spirv/mma_layout.h"
|
||||
#include "shader_recompiler/backend/spirv/spirv_emit_context.h"
|
||||
#include "shader_recompiler/frontend/ir/modifiers.h"
|
||||
|
||||
@@ -381,16 +382,40 @@ Id EmitFPUnordGreaterThanEqual64(EmitContext& ctx, Id lhs, Id rhs) {
|
||||
return FPUnordCompare(&EmitContext::OpFUnordGreaterThanEqual, ctx, lhs, rhs);
|
||||
}
|
||||
|
||||
// (hookshot) isnan of a scalar float of that many bits. With nan_bits (mma_layout.h: AMD's older Windows shader compiler) it is
|
||||
// a test of the value's bits, a NaN being an exponent of all ones and a mantissa that is not 0, which no float-mode setting of the
|
||||
// compiler can fold away; else OpIsNan, as before
|
||||
Id IsNan(EmitContext& ctx, Id value, u32 bits) {
|
||||
if (!nan_bits) {
|
||||
return ctx.OpIsNan(ctx.U1, value);
|
||||
}
|
||||
if (bits == 64) {
|
||||
const Id words{ctx.OpBitcast(ctx.U32[2], value)};
|
||||
const Id high{ctx.OpBitwiseAnd(ctx.U32[1], ctx.OpCompositeExtract(ctx.U32[1], words, 1U),
|
||||
ctx.Const(0x7fffffffu))};
|
||||
const Id low{ctx.OpCompositeExtract(ctx.U32[1], words, 0U)};
|
||||
const Id infinity{ctx.Const(0x7ff00000u)};
|
||||
return ctx.OpLogicalOr(ctx.U1, ctx.OpUGreaterThan(ctx.U1, high, infinity),
|
||||
ctx.OpLogicalAnd(ctx.U1, ctx.OpIEqual(ctx.U1, high, infinity),
|
||||
ctx.OpINotEqual(ctx.U1, low, ctx.u32_zero_value)));
|
||||
}
|
||||
if (bits == 16) {
|
||||
value = ctx.OpFConvert(ctx.F32[1], value); // (exact: a NaN stays a NaN, and nothing else becomes one)
|
||||
}
|
||||
const Id magnitude{ctx.OpBitwiseAnd(ctx.U32[1], ctx.OpBitcast(ctx.U32[1], value), ctx.Const(0x7fffffffu))};
|
||||
return ctx.OpUGreaterThan(ctx.U1, magnitude, ctx.Const(0x7f800000u));
|
||||
}
|
||||
|
||||
Id EmitFPIsNan16(EmitContext& ctx, Id value) {
|
||||
return ctx.OpIsNan(ctx.U1, value);
|
||||
return IsNan(ctx, value, 16);
|
||||
}
|
||||
|
||||
Id EmitFPIsNan32(EmitContext& ctx, Id value) {
|
||||
return ctx.OpIsNan(ctx.U1, value);
|
||||
return IsNan(ctx, value, 32);
|
||||
}
|
||||
|
||||
Id EmitFPIsNan64(EmitContext& ctx, Id value) {
|
||||
return ctx.OpIsNan(ctx.U1, value);
|
||||
return IsNan(ctx, value, 64);
|
||||
}
|
||||
|
||||
} // namespace Shader::Backend::SPIRV
|
||||
|
||||
@@ -9,6 +9,7 @@
|
||||
#include "common/settings.h"
|
||||
#include "shader_recompiler/backend/spirv/emit_spirv.h"
|
||||
#include "shader_recompiler/backend/spirv/emit_spirv_instructions.h"
|
||||
#include "shader_recompiler/backend/spirv/mma_layout.h"
|
||||
#include "shader_recompiler/backend/spirv/spirv_emit_context.h"
|
||||
#include "shader_recompiler/frontend/ir/modifiers.h"
|
||||
|
||||
@@ -445,9 +446,10 @@ Id ImageGatherSubpixelOffset(EmitContext& ctx, const IR::TextureInstInfo& info,
|
||||
case TextureType::Color2DRect:
|
||||
return calculate_coords(2);
|
||||
case TextureType::ColorArray2D:
|
||||
case TextureType::ColorCube:
|
||||
return calculate_coords(3);
|
||||
default:
|
||||
// (hookshot) a cube's coordinates are a direction: no texel to nudge by, and its size query has two components, not
|
||||
// three (the cube case here made invalid SPIR-V: OpImageQuerySizeLod's result of three components on a cube)
|
||||
return coords;
|
||||
}
|
||||
}
|
||||
@@ -488,6 +490,64 @@ void AddOffsetToCoordinates(EmitContext& ctx, const IR::TextureInstInfo& info, I
|
||||
}
|
||||
coords = ctx.OpIAdd(result_type, coords, offset);
|
||||
}
|
||||
// (hookshot, mma_layout.h fetch_bounds) Whether a texel fetch of image at coords (with its level, or its sample, when it has one)
|
||||
// is inside the image, as U1; coords, lod and sample are made ones that are (a coordinate out of range 0, a level or sample past
|
||||
// the last 0), so that the fetch itself never leaves the image, and the caller answers zero for it. NVIDIA's GPUs answer an
|
||||
// out-of-bounds fetch with zero (measured on an RTX 3090 with robustness off: every texture type, past each edge, past the last
|
||||
// level and the last sample; the alpha of a format without one is 1 there, as in range, and 0 here: the translation does not
|
||||
// know the format). Vulkan leaves such a fetch undefined without robustImageAccess, and on AMD the ground fog's depth reads at
|
||||
// points projected off screen took the other branch: a pale box at the screen's edges (found by Ocarina of Lain's author on
|
||||
// an RX 9070). Id{}: a type left as it was: a texel buffer while robustBufferAccess2 is on (on by default: the Robust buffers
|
||||
// row), which answers its out-of-bounds reads with zero already (fetch_bounds_buffers), and the cube types, which no fetch reads
|
||||
Id FetchInBounds(EmitContext& ctx, const IR::TextureInstInfo& info, Id image, Id& coords, Id& lod, Id& sample) {
|
||||
u32 dims{};
|
||||
switch (info.type) {
|
||||
case TextureType::Buffer:
|
||||
if (!fetch_bounds_buffers) {
|
||||
return Id{};
|
||||
}
|
||||
dims = 1;
|
||||
break;
|
||||
case TextureType::Color1D:
|
||||
dims = 1;
|
||||
break;
|
||||
case TextureType::ColorArray1D:
|
||||
case TextureType::Color2D:
|
||||
case TextureType::Color2DRect:
|
||||
dims = 2;
|
||||
break;
|
||||
case TextureType::ColorArray2D:
|
||||
case TextureType::Color3D:
|
||||
dims = 3;
|
||||
break;
|
||||
default:
|
||||
return Id{};
|
||||
}
|
||||
const Id size_type{ctx.U32[dims]};
|
||||
Id level_ok{};
|
||||
Id size{};
|
||||
if (Sirit::ValidId(lod)) {
|
||||
const Id levels{ctx.OpImageQueryLevels(ctx.U32[1], image)};
|
||||
level_ok = ctx.OpULessThan(ctx.U1, lod, levels);
|
||||
lod = ctx.OpSelect(ctx.U32[1], level_ok, lod, ctx.u32_zero_value);
|
||||
size = ctx.OpImageQuerySizeLod(size_type, image, lod);
|
||||
} else {
|
||||
size = ctx.OpImageQuerySize(size_type, image); // (a texel buffer, or a multisampled image)
|
||||
}
|
||||
// (unsigned: a negative coordinate is past the end too)
|
||||
const Id inside{ctx.OpULessThan(dims == 1 ? ctx.U1 : ctx.TypeVector(ctx.U1, dims), coords, size)};
|
||||
coords = ctx.OpSelect(size_type, inside, coords, ctx.ConstantNull(size_type));
|
||||
Id in_bounds{dims == 1 ? inside : ctx.OpAll(ctx.U1, inside)};
|
||||
if (Sirit::ValidId(level_ok)) {
|
||||
in_bounds = ctx.OpLogicalAnd(ctx.U1, in_bounds, level_ok);
|
||||
}
|
||||
if (Sirit::ValidId(sample)) {
|
||||
const Id sample_ok{ctx.OpULessThan(ctx.U1, sample, ctx.OpImageQuerySamples(ctx.U32[1], image))};
|
||||
sample = ctx.OpSelect(ctx.U32[1], sample_ok, sample, ctx.u32_zero_value);
|
||||
in_bounds = ctx.OpLogicalAnd(ctx.U1, in_bounds, sample_ok);
|
||||
}
|
||||
return in_bounds;
|
||||
}
|
||||
} // Anonymous namespace
|
||||
|
||||
Id EmitBindlessImageSampleImplicitLod(EmitContext&) {
|
||||
@@ -724,6 +784,17 @@ Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id c
|
||||
// This image is multisampled, lod must be implicit
|
||||
lod = Id{};
|
||||
}
|
||||
if (fetch_bounds) { // (hookshot: FetchInBounds)
|
||||
const Id image{TextureImage(ctx, info, index)};
|
||||
const Id in_bounds{FetchInBounds(ctx, info, image, coords, lod, ms)};
|
||||
const ImageOperands operands(lod, ms);
|
||||
Id color{Emit(&EmitContext::OpImageSparseFetch, &EmitContext::OpImageFetch, ctx, inst,
|
||||
result_type, image, coords, operands.MaskOptional(), operands.Span())};
|
||||
if (Sirit::ValidId(in_bounds)) {
|
||||
color = ctx.OpSelect(result_type, in_bounds, color, ctx.ConstantNull(result_type));
|
||||
}
|
||||
return is_integer ? ctx.OpBitcast(ctx.F32[4], color) : color;
|
||||
}
|
||||
const ImageOperands operands(lod, ms);
|
||||
const Id color{Emit(&EmitContext::OpImageSparseFetch, &EmitContext::OpImageFetch, ctx, inst,
|
||||
result_type, TextureImage(ctx, info, index), coords,
|
||||
|
||||
@@ -644,4 +644,7 @@ Id EmitDPdyFine(EmitContext& ctx, Id op_a);
|
||||
Id EmitDPdxCoarse(EmitContext& ctx, Id op_a);
|
||||
Id EmitDPdyCoarse(EmitContext& ctx, Id op_a);
|
||||
|
||||
// (hookshot) Helpers the emitters share, not IR instructions
|
||||
Id IsNan(EmitContext& ctx, Id value, u32 bits); // (OpIsNan, or a test of the bits: mma_layout.h nan_bits)
|
||||
|
||||
} // namespace Shader::Backend::SPIRV
|
||||
|
||||
@@ -4,6 +4,9 @@
|
||||
// SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project
|
||||
// SPDX-License-Identifier: GPL-2.0-or-later
|
||||
|
||||
#include <utility>
|
||||
#include <vector>
|
||||
|
||||
#include "shader_recompiler/backend/spirv/emit_spirv.h"
|
||||
#include "shader_recompiler/backend/spirv/emit_spirv_instructions.h"
|
||||
#include "shader_recompiler/backend/spirv/spirv_emit_context.h"
|
||||
@@ -20,6 +23,74 @@ void ConvertDepthMode(EmitContext& ctx) {
|
||||
ctx.OpStore(ctx.output_position, vector);
|
||||
}
|
||||
|
||||
/// Output variable and the unscaled value it held before ScaleNoPerspectiveOutputs.
|
||||
using ScaledOutputs = std::vector<std::pair<Id, Id>>;
|
||||
|
||||
/// Write screen-linear generic outputs times Position.w (see RuntimeInfo::noperspective_outputs).
|
||||
/// Returns the unscaled values so a geometry shader can restore them after EmitVertex.
|
||||
ScaledOutputs ScaleNoPerspectiveOutputs(EmitContext& ctx) {
|
||||
ScaledOutputs saved;
|
||||
if (ctx.runtime_info.noperspective_outputs.none() || !Sirit::ValidId(ctx.output_position)) {
|
||||
return saved;
|
||||
}
|
||||
const Id position{ctx.OpLoad(ctx.F32[4], ctx.output_position)};
|
||||
const Id w{ctx.OpCompositeExtract(ctx.F32[1], position, 3u)};
|
||||
for (size_t index = 0; index < ctx.output_generics.size(); ++index) {
|
||||
if (!ctx.runtime_info.noperspective_outputs.test(index)) {
|
||||
continue;
|
||||
}
|
||||
Id last{};
|
||||
for (const GenericElementInfo& element : ctx.output_generics[index]) {
|
||||
if (!Sirit::ValidId(element.id) || element.id.value == last.value ||
|
||||
element.num_components == 0) {
|
||||
continue;
|
||||
}
|
||||
last = element.id;
|
||||
const Id type{ctx.F32[element.num_components]};
|
||||
const Id value{ctx.OpLoad(type, element.id)};
|
||||
const Id scaled{element.num_components == 1 ? ctx.OpFMul(type, value, w)
|
||||
: ctx.OpVectorTimesScalar(type, value, w)};
|
||||
ctx.OpStore(element.id, scaled);
|
||||
saved.emplace_back(element.id, value);
|
||||
}
|
||||
}
|
||||
return saved;
|
||||
}
|
||||
|
||||
/// A geometry shader keeps its outputs between EmitVertex calls: restore the unscaled values.
|
||||
void RestoreNoPerspectiveOutputs(EmitContext& ctx, const ScaledOutputs& saved) {
|
||||
for (const auto& [id, value] : saved) {
|
||||
ctx.OpStore(id, value);
|
||||
}
|
||||
}
|
||||
|
||||
/// (hookshot) RuntimeInfo::capture_position: the position as the stage leaves it, stored at the pushed address's slot for
|
||||
/// this vertex (CAPTURE_PUSH_OFFSET), when the address is not 0 and the slot below the pushed count. What is drawn is not
|
||||
/// changed: the store is the stage's last act, and reads only what it wrote
|
||||
void CapturePosition(EmitContext& ctx) {
|
||||
const auto push = [&](Id type, u32 member) {
|
||||
const Id pointer{ctx.TypePointer(spv::StorageClass::PushConstant, type)};
|
||||
return ctx.OpLoad(type, ctx.OpAccessChain(pointer, ctx.rescaling_push_constants, ctx.Const(member)));
|
||||
};
|
||||
const Id address{push(ctx.U64, ctx.capture_address_member_index)};
|
||||
const Id first{push(ctx.U32[1], ctx.capture_first_member_index)};
|
||||
const Id count{push(ctx.U32[1], ctx.capture_count_member_index)};
|
||||
const Id slot{ctx.OpISub(ctx.U32[1], ctx.OpLoad(ctx.U32[1], ctx.capture_vertex_index), first)};
|
||||
const Id zero{ctx.Constant(ctx.U64, u64{0})};
|
||||
const Id wanted{ctx.OpLogicalAnd(ctx.U1, ctx.OpINotEqual(ctx.U1, address, zero), ctx.OpULessThan(ctx.U1, slot, count))};
|
||||
const Id store_label{ctx.OpLabel()};
|
||||
const Id end_label{ctx.OpLabel()};
|
||||
ctx.OpSelectionMerge(end_label, spv::SelectionControlMask::MaskNone);
|
||||
ctx.OpBranchConditional(wanted, store_label, end_label);
|
||||
ctx.AddLabel(store_label);
|
||||
const Id offset{ctx.OpIMul(ctx.U64, ctx.OpUConvert(ctx.U64, slot), ctx.Constant(ctx.U64, u64{16}))};
|
||||
const Id at{ctx.OpIAdd(ctx.U64, address, offset)};
|
||||
const Id position{ctx.OpLoad(ctx.F32[4], ctx.output_position)};
|
||||
ctx.OpStore(ctx.OpConvertUToPtr(ctx.capture_pointer, at), position, spv::MemoryAccessMask::Aligned, 16U);
|
||||
ctx.OpBranch(end_label);
|
||||
ctx.AddLabel(end_label);
|
||||
}
|
||||
|
||||
void SetFixedPipelinePointSize(EmitContext& ctx) {
|
||||
if (ctx.runtime_info.fixed_state_point_size) {
|
||||
const float point_size{*ctx.runtime_info.fixed_state_point_size};
|
||||
@@ -151,6 +222,12 @@ void EmitEpilogue(EmitContext& ctx) {
|
||||
!ctx.profile.support_native_ndc) {
|
||||
ConvertDepthMode(ctx);
|
||||
}
|
||||
if (ctx.stage == Stage::VertexB || ctx.stage == Stage::TessellationEval) {
|
||||
ScaleNoPerspectiveOutputs(ctx);
|
||||
}
|
||||
if (ctx.stage == Stage::VertexB && ctx.wants_capture) { // (after the depth's conversion: the position as drawn)
|
||||
CapturePosition(ctx);
|
||||
}
|
||||
if (ctx.stage == Stage::Fragment) {
|
||||
AlphaTest(ctx);
|
||||
}
|
||||
@@ -160,6 +237,7 @@ void EmitEmitVertex(EmitContext& ctx, const IR::Value& stream) {
|
||||
if (ctx.runtime_info.convert_depth_mode && !ctx.profile.support_native_ndc) {
|
||||
ConvertDepthMode(ctx);
|
||||
}
|
||||
const ScaledOutputs scaled{ScaleNoPerspectiveOutputs(ctx)};
|
||||
if (!ctx.profile.support_geometry_streams) {
|
||||
throw NotImplementedException("Geometry streams");
|
||||
} else if (stream.IsImmediate()) {
|
||||
@@ -170,6 +248,7 @@ void EmitEmitVertex(EmitContext& ctx, const IR::Value& stream) {
|
||||
}
|
||||
// Restore fixed pipeline point size after emitting the vertex
|
||||
SetFixedPipelinePointSize(ctx);
|
||||
RestoreNoPerspectiveOutputs(ctx, scaled);
|
||||
}
|
||||
|
||||
void EmitEndPrimitive(EmitContext& ctx, const IR::Value& stream) {
|
||||
|
||||
@@ -25,5 +25,15 @@ inline bool frag_nan_guard = false;
|
||||
// the colour pass is then rejected in spots: holes showing the bright sky, blown up by bloom (the glowing flowers).
|
||||
// On for every card but NVIDIA (vk_coopmat.cpp); RX_VK_NO_INVARIANT=1 off. From NX2navi patch 0007.
|
||||
inline bool position_invariant = false;
|
||||
// (hookshot) An out-of-bounds texel fetch answers zero, as NVIDIA's GPUs answer it (measured on an RTX 3090), rather than
|
||||
// whatever the host's does (EmitImageFetch). Vulkan leaves it undefined without robustImageAccess; on AMD the ground fog's
|
||||
// off-screen depth reads drew a pale box at the screen's edges. Every card but NVIDIA (gpu/vk/vk_vendor.h); RX_VK_FETCH_BOUNDS
|
||||
inline bool fetch_bounds = false;
|
||||
// (hookshot) and texel buffers' fetches too: only when robustBufferAccess2 is off (the Debug tab's Robust buffers Off, RX_VK_ROBUST=0,
|
||||
// or a driver without it), which otherwise answers their out-of-bounds reads with zero already (vk_coopmat.cpp)
|
||||
inline bool fetch_bounds_buffers = false;
|
||||
// (hookshot) isnan tested on the value's bits (IsNan in emit_spirv_floating_point.cpp), not OpIsNan, for AMD's older Windows
|
||||
// shader compiler (RX 5000 and RX 6000: gpu/vk/vk_vendor.h); RX_VK_NAN_BITS=0 or 1
|
||||
inline bool nan_bits = false;
|
||||
|
||||
} // namespace Shader::Backend::SPIRV
|
||||
|
||||
@@ -524,6 +524,12 @@ EmitContext::EmitContext(const Profile& profile_, const RuntimeInfo& runtime_inf
|
||||
if (runtime_info.mesh_fallback || runtime_info.mesh_fallback_vertex) {
|
||||
wants_mesh_fallback = true; // (before any DefineRescalingInputPushConstant, which is made once)
|
||||
}
|
||||
// (hookshot) a scene vertex stage's position capture: its members go in the one push block, so decided before it is made.
|
||||
// Not with the render area's block (a second push block; full-screen passes, never the scene's geometry)
|
||||
if (runtime_info.capture_position && stage == Stage::VertexB && !runtime_info.mesh_fallback_vertex &&
|
||||
profile.unified_descriptor_binding && profile.support_int64 && !program.info.uses_render_area) {
|
||||
wants_capture = true;
|
||||
}
|
||||
DefineCommonTypes(program.info);
|
||||
DefineCommonConstants();
|
||||
DefineInterfaces(program);
|
||||
@@ -576,6 +582,21 @@ EmitContext::EmitContext(const Profile& profile_, const RuntimeInfo& runtime_inf
|
||||
mesh_info = program.mesh;
|
||||
DefineRescalingInputPushConstant();
|
||||
}
|
||||
if (wants_capture) { // (the store itself: EmitEpilogue, emit_spirv_special.cpp)
|
||||
if (!Sirit::ValidId(U64)) {
|
||||
AddCapability(spv::Capability::Int64);
|
||||
U64 = Name(TypeInt(64, false), "u64");
|
||||
}
|
||||
AddCapability(spv::Capability::PhysicalStorageBufferAddresses);
|
||||
AddExtension("SPV_KHR_physical_storage_buffer");
|
||||
SetMemoryModel(spv::AddressingModel::PhysicalStorageBuffer64, spv::MemoryModel::GLSL450);
|
||||
// (Vulkan's VertexIndex, base vertex included: the redraw draws the same indices with the same base)
|
||||
capture_vertex_index = Sirit::ValidId(vertex_index)
|
||||
? vertex_index
|
||||
: DefineInput(*this, U32[1], true, spv::BuiltIn::VertexIndex);
|
||||
capture_pointer = TypePointer(spv::StorageClass::PhysicalStorageBuffer, F32[4]);
|
||||
program.info.captures_position = true;
|
||||
}
|
||||
DefineMesh(program);
|
||||
DefineLocalMemory(program);
|
||||
DefineSharedMemory(program);
|
||||
@@ -592,6 +613,13 @@ EmitContext::EmitContext(const Profile& profile_, const RuntimeInfo& runtime_inf
|
||||
DefineGlobalMemoryFunctions(program.info);
|
||||
DefineRescalingInput(program.info);
|
||||
DefineRenderArea(program.info);
|
||||
// (the capture's push block, when nothing above made it. Not made before them: the block is made once, with the members
|
||||
// wanted by then, and DefineGlobalMemoryFunctions asks for the page table's (wants_guest_pages) only as it runs. Made
|
||||
// earlier, the block had no such member, the stage read the table's address from member 0, and the GPU was lost at the
|
||||
// first frame in play)
|
||||
if (wants_capture) {
|
||||
DefineRescalingInputPushConstant();
|
||||
}
|
||||
}
|
||||
|
||||
EmitContext::~EmitContext() = default;
|
||||
@@ -1603,6 +1631,14 @@ void EmitContext::DefineRescalingInputPushConstant() {
|
||||
mesh_fallback_indirect_input_member_index = member_index++;
|
||||
}
|
||||
}
|
||||
if (wants_capture) {
|
||||
members.push_back(U64);
|
||||
capture_address_member_index = member_index++;
|
||||
members.push_back(U32[1]);
|
||||
capture_first_member_index = member_index++;
|
||||
members.push_back(U32[1]);
|
||||
capture_count_member_index = member_index++;
|
||||
}
|
||||
const Id push_constant_struct{TypeStruct(std::span(members.data(), members.size()))};
|
||||
Decorate(push_constant_struct, spv::Decoration::Block);
|
||||
Name(push_constant_struct, "ResolutionInfo");
|
||||
@@ -1641,6 +1677,14 @@ void EmitContext::DefineRescalingInputPushConstant() {
|
||||
MESH_FALLBACK_INDIRECT_INPUT_PUSH_OFFSET);
|
||||
}
|
||||
}
|
||||
if (wants_capture) {
|
||||
MemberDecorate(push_constant_struct, capture_address_member_index, spv::Decoration::Offset, CAPTURE_PUSH_OFFSET);
|
||||
MemberDecorate(push_constant_struct, capture_first_member_index, spv::Decoration::Offset, CAPTURE_FIRST_PUSH_OFFSET);
|
||||
MemberDecorate(push_constant_struct, capture_count_member_index, spv::Decoration::Offset, CAPTURE_COUNT_PUSH_OFFSET);
|
||||
MemberName(push_constant_struct, capture_address_member_index, "capture_address");
|
||||
MemberName(push_constant_struct, capture_first_member_index, "capture_first");
|
||||
MemberName(push_constant_struct, capture_count_member_index, "capture_count");
|
||||
}
|
||||
const Id pointer_type{TypePointer(spv::StorageClass::PushConstant, push_constant_struct)};
|
||||
rescaling_push_constants = AddGlobalVariable(pointer_type, spv::StorageClass::PushConstant);
|
||||
Name(rescaling_push_constants, "rescaling_push_constants");
|
||||
@@ -2017,6 +2061,12 @@ void EmitContext::DefineImages(const Info& info, u32& binding, u32& scaling_inde
|
||||
void EmitContext::DefineInputs(const IR::Program& program) {
|
||||
const Info& info{program.info};
|
||||
const VaryingState loads{info.loads.mask | info.passthrough.mask};
|
||||
const bool noperspective_as_smooth{
|
||||
stage == Stage::Fragment && runtime_info.scaled_noperspective_inputs &&
|
||||
std::ranges::any_of(info.interpolation,
|
||||
[](Interpolation mode) {
|
||||
return mode == Interpolation::NoPerspective;
|
||||
})};
|
||||
|
||||
if (info.uses_workgroup_id) {
|
||||
workgroup_id = DefineInput(*this, U32[3], false, spv::BuiltIn::WorkgroupId);
|
||||
@@ -2084,7 +2134,8 @@ void EmitContext::DefineInputs(const IR::Program& program) {
|
||||
layer = DefineInput(*this, U32[1], false, spv::BuiltIn::Layer);
|
||||
Decorate(layer, spv::Decoration::Flat);
|
||||
}
|
||||
if (loads.AnyComponent(IR::Attribute::PositionX)) {
|
||||
if (loads.AnyComponent(IR::Attribute::PositionX) ||
|
||||
info.passthrough.AnyComponent(IR::Attribute::PositionX) || noperspective_as_smooth) {
|
||||
const bool is_fragment{stage == Stage::Fragment};
|
||||
if (!is_fragment && profile.has_broken_spirv_position_input) {
|
||||
need_input_position_indirect = true;
|
||||
@@ -2189,7 +2240,12 @@ void EmitContext::DefineInputs(const IR::Program& program) {
|
||||
case Interpolation::Smooth:
|
||||
break;
|
||||
case Interpolation::NoPerspective:
|
||||
Decorate(id, spv::Decoration::NoPerspective);
|
||||
// The previous stage may write it times Position.w (noperspective_outputs).
|
||||
if (!runtime_info.scaled_noperspective_inputs) {
|
||||
Decorate(id, spv::Decoration::NoPerspective);
|
||||
} else {
|
||||
noperspective_as_smooth_inputs[index] = true;
|
||||
}
|
||||
break;
|
||||
case Interpolation::Flat:
|
||||
Decorate(id, spv::Decoration::Flat);
|
||||
|
||||
@@ -369,6 +369,8 @@ public:
|
||||
bool need_input_position_indirect{};
|
||||
Id input_position{};
|
||||
std::array<InputGenericInfo, 32> input_generics{};
|
||||
/// Screen-linear fragment inputs read through perspective interpolation, times FragCoord.w.
|
||||
std::array<bool, 32> noperspective_as_smooth_inputs{};
|
||||
|
||||
Id output_point_size{};
|
||||
Id output_position{};
|
||||
@@ -429,6 +431,13 @@ public:
|
||||
u32 mesh_fallback_payload_member_index{};
|
||||
u32 mesh_fallback_records_member_index{};
|
||||
u32 mesh_fallback_indirect_input_member_index{};
|
||||
/// (hookshot) RuntimeInfo::capture_position: the push block's capture members, the vertex index and a vec4 pointer type
|
||||
bool wants_capture{};
|
||||
u32 capture_address_member_index{};
|
||||
u32 capture_first_member_index{};
|
||||
u32 capture_count_member_index{};
|
||||
Id capture_vertex_index{};
|
||||
Id capture_pointer{};
|
||||
/// A pointer to the u32 at a PhysicalStorageBuffer device address
|
||||
Id MeshFallbackWord(Id address);
|
||||
|
||||
|
||||
Reference in New Issue
Block a user