shader: share format conversion for packed normalized image stores

This commit is contained in:
nmzik
2026-10-12 23:38:45 +02:00
parent aec8c88755
commit c11254a996
8 changed files with 149 additions and 95 deletions
@@ -377,13 +377,16 @@ uint32_t InverseSwizzle(uint32_t swizzle, uint32_t component) {
Format::BufferFormatInfo ImageConversionFormat(const EmitterState& state,
const IR::MemoryInfo& mem) {
const auto format = state.program.info.images[mem.resource].conversion_format;
const auto& image = state.program.info.images[mem.resource];
const auto format = image.conversion_format;
if (format == Prospero::BufferFormat::kInvalid) return {};
const auto info = Format::GetFormatInfo(format);
EXIT_IF(Prospero::RemapTextureFormat(format) == format || info.component_count == 0u ||
info.component_count > 4u);
EXIT_IF(info.type != Format::ComponentType::Uscaled &&
(info.type != Format::ComponentType::Uint || !info.packed_bitfield ||
((info.type != Format::ComponentType::Uint &&
!(info.type == Format::ComponentType::Unorm &&
image.resource_class == IR::ImageResourceClass::Storage)) || !info.packed_bitfield ||
info.byte_size != sizeof(uint32_t)));
return info;
}
@@ -537,47 +540,22 @@ uint32_t EmitOneDimensionalGatherLz(ValueEmitContext& ctx, const IR::MemoryInfo&
return result;
}
uint32_t PackImageTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t texel) {
const auto info = ImageConversionFormat(ctx.state, mem);
if (info.format == Prospero::BufferFormat::kInvalid) return texel;
EXIT_IF(info.type != Format::ComponentType::Uint || !info.packed_bitfield);
auto packed = ConstantU32(ctx.state, 0u);
for (uint32_t component = 0; component < info.component_count; component++) {
const auto value = ctx.state.builder.AllocateId();
ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), value, texel,
component);
const auto maximum =
ConstantU32(ctx.state, info.component_bits[component] == 32u
? UINT32_MAX
: (1u << info.component_bits[component]) - 1u);
const auto within =
Binary(ctx.state, spv::OpULessThan, TypeBool(ctx.state), value, maximum);
const auto clamped = ctx.state.builder.AllocateId();
ctx.state.builder.AddFunction(spv::OpSelect, TypeU32(ctx.state), clamped, within, value,
maximum);
const auto shifted =
info.component_bit_offset[component] == 0u
? clamped
: Binary(ctx.state, spv::OpShiftLeftLogical, TypeU32(ctx.state), clamped,
ConstantU32(ctx.state, info.component_bit_offset[component]));
packed = Binary(ctx.state, spv::OpBitwiseOr, TypeU32(ctx.state), packed, shifted);
}
const auto zero = ConstantU32(ctx.state, 0u);
const auto result = ctx.state.builder.AllocateId();
ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(ctx.state, 4), result,
packed, zero, zero, zero);
return result;
}
uint32_t StoreTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t data, bool integer) {
const auto swizzle = ctx.state.program.info.images[mem.resource].shader_swizzle;
uint32_t values[4] {};
const auto info = ImageConversionFormat(ctx.state, mem);
const bool packed = info.format != Prospero::BufferFormat::kInvalid;
EXIT_IF(packed && (!info.packed_bitfield ||
(info.type != Format::ComponentType::Uint &&
info.type != Format::ComponentType::Unorm)));
const auto zero = ConstantU32(ctx.state, 0);
uint32_t values[4] {zero, zero, zero, zero};
auto packed_value = zero;
const auto dmask = mem.dmask != 0u ? mem.dmask : 1u;
for (uint32_t component = 0; component < 4u; component++) {
for (uint32_t component = 0; component < (packed ? info.component_count : 4u); component++) {
const auto source = InverseSwizzle(swizzle, component);
uint32_t raw = ConstantU32(ctx.state, 0);
if (source < 4u && ((dmask >> source) & 1u) != 0u) {
const bool selected = source < 4u && ((dmask >> source) & 1u) != 0u;
uint32_t raw = zero;
if (selected) {
const auto packed_index = DmaskComponentIndex(dmask, source);
raw = ctx.state.builder.AllocateId();
ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), raw, data,
@@ -587,21 +565,39 @@ uint32_t StoreTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t d
raw = Binary(ctx.state, spv::OpShiftRightLogical, TypeU32(ctx.state), raw,
ConstantU32(ctx.state, 16u));
}
raw = Binary(ctx.state, spv::OpBitwiseAnd, TypeU32(ctx.state), raw,
ConstantU32(ctx.state, 0xffffu));
if (!packed) {
raw = Binary(ctx.state, spv::OpBitwiseAnd, TypeU32(ctx.state), raw,
ConstantU32(ctx.state, 0xffffu));
}
}
}
if (packed) {
if (!selected) continue;
auto value = EncodeFormattedStoreComponent(ctx.state, info, component, raw, mem.data_bits);
if (info.type == Format::ComponentType::Uint) {
value = EmitUMin32(ctx.state, value,
ConstantU32(ctx.state, (1u << info.component_bits[component]) - 1u));
}
if (info.component_bit_offset[component] != 0u) {
value = Binary(ctx.state, spv::OpShiftLeftLogical, TypeU32(ctx.state), value,
ConstantU32(ctx.state, info.component_bit_offset[component]));
}
packed_value = packed_value == zero ? value
: Binary(ctx.state, spv::OpBitwiseOr, TypeU32(ctx.state), packed_value, value);
continue;
}
values[component] = integer ? raw
: mem.data_bits == 16u
? EmitF16BitsToF32(ctx.state, raw)
: Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), raw);
}
if (packed) values[0] = packed_value;
const auto texel = ctx.state.builder.AllocateId();
ctx.state.builder.AddFunction(spv::OpCompositeConstruct,
integer ? TypeU32Vector(ctx.state, 4)
: TypeF32Vector(ctx.state, 4),
texel, values[0], values[1], values[2], values[3]);
return PackImageTexel(ctx, mem, texel);
return texel;
}
template <typename Emit>