shader_recompiler: handle fetch shader address offsets (#538)
* shader_recompiler: handle fetch shader address offsets parse index & offset sgpr from fetch shader and propagate them to vkBindVertexBuffers * shader_recompiler: fix fetch_shader when offset is not present * video_core: propagate index/offset SGPRs to vkCmdDraw instead of offsetting the buffer address * video_core: add vertex_offset to non-indexed draw calls renamed fetch offset fields
This commit is contained in:
parent
208575d392
commit
9e4fc17e6c
|
@ -3,6 +3,7 @@
|
|||
|
||||
#include <algorithm>
|
||||
#include <boost/container/static_vector.hpp>
|
||||
#include "common/assert.h"
|
||||
#include "shader_recompiler/frontend/decode.h"
|
||||
#include "shader_recompiler/frontend/fetch_shader.h"
|
||||
|
||||
|
@ -33,8 +34,8 @@ namespace Shader::Gcn {
|
|||
* We take the reverse way, extract the original input semantics from these instructions.
|
||||
**/
|
||||
|
||||
std::vector<VertexAttribute> ParseFetchShader(const u32* code, u32* out_size) {
|
||||
std::vector<VertexAttribute> attributes;
|
||||
FetchShaderData ParseFetchShader(const u32* code, u32* out_size) {
|
||||
FetchShaderData data{};
|
||||
GcnCodeSlice code_slice(code, code + std::numeric_limits<u32>::max());
|
||||
GcnDecodeContext decoder;
|
||||
|
||||
|
@ -59,6 +60,21 @@ std::vector<VertexAttribute> ParseFetchShader(const u32* code, u32* out_size) {
|
|||
continue;
|
||||
}
|
||||
|
||||
if (inst.opcode == Opcode::V_ADD_I32) {
|
||||
const auto vgpr = inst.dst[0].code;
|
||||
const auto sgpr = s8(inst.src[0].code);
|
||||
switch (vgpr) {
|
||||
case 0: // V0 is always the vertex offset
|
||||
data.vertex_offset_sgpr = sgpr;
|
||||
break;
|
||||
case 3: // V3 is always the instance offset
|
||||
data.instance_offset_sgpr = sgpr;
|
||||
break;
|
||||
default:
|
||||
UNREACHABLE();
|
||||
}
|
||||
}
|
||||
|
||||
if (inst.inst_class == InstClass::VectorMemBufFmt) {
|
||||
// SRSRC is in units of 4 SPGRs while SBASE is in pairs of SGPRs
|
||||
const u32 base_sgpr = inst.src[2].code * 4;
|
||||
|
@ -68,7 +84,7 @@ std::vector<VertexAttribute> ParseFetchShader(const u32* code, u32* out_size) {
|
|||
const auto it = std::ranges::find_if(
|
||||
loads, [&](VsharpLoad& load) { return load.dst_reg == base_sgpr; });
|
||||
|
||||
auto& attrib = attributes.emplace_back();
|
||||
auto& attrib = data.attributes.emplace_back();
|
||||
attrib.semantic = semantic_index++;
|
||||
attrib.dest_vgpr = inst.src[1].code;
|
||||
attrib.num_elements = inst.control.mubuf.count;
|
||||
|
@ -83,7 +99,7 @@ std::vector<VertexAttribute> ParseFetchShader(const u32* code, u32* out_size) {
|
|||
}
|
||||
}
|
||||
|
||||
return attributes;
|
||||
return data;
|
||||
}
|
||||
|
||||
} // namespace Shader::Gcn
|
||||
|
|
|
@ -17,6 +17,12 @@ struct VertexAttribute {
|
|||
u8 instance_data; ///< Indicates that the buffer will be accessed in instance rate
|
||||
};
|
||||
|
||||
std::vector<VertexAttribute> ParseFetchShader(const u32* code, u32* out_size);
|
||||
struct FetchShaderData {
|
||||
std::vector<VertexAttribute> attributes;
|
||||
s8 vertex_offset_sgpr = -1; ///< SGPR of vertex offset from VADDR
|
||||
s8 instance_offset_sgpr = -1; ///< SGPR of instance offset from VADDR
|
||||
};
|
||||
|
||||
FetchShaderData ParseFetchShader(const u32* code, u32* out_size);
|
||||
|
||||
} // namespace Shader::Gcn
|
||||
|
|
|
@ -346,7 +346,7 @@ void Translator::EmitFetch(const GcnInst& inst) {
|
|||
|
||||
// Parse the assembly to generate a list of attributes.
|
||||
u32 fetch_size{};
|
||||
const auto attribs = ParseFetchShader(code, &fetch_size);
|
||||
const auto fetch_data = ParseFetchShader(code, &fetch_size);
|
||||
|
||||
if (Config::dumpShaders()) {
|
||||
using namespace Common::FS;
|
||||
|
@ -359,7 +359,10 @@ void Translator::EmitFetch(const GcnInst& inst) {
|
|||
file.WriteRaw<u8>(code, fetch_size);
|
||||
}
|
||||
|
||||
for (const auto& attrib : attribs) {
|
||||
info.vertex_offset_sgpr = fetch_data.vertex_offset_sgpr;
|
||||
info.instance_offset_sgpr = fetch_data.instance_offset_sgpr;
|
||||
|
||||
for (const auto& attrib : fetch_data.attributes) {
|
||||
const IR::Attribute attr{IR::Attribute::Param0 + attrib.semantic};
|
||||
IR::VectorReg dst_reg{attrib.dest_vgpr};
|
||||
|
||||
|
|
|
@ -175,6 +175,9 @@ struct Info {
|
|||
AttributeFlags stores{};
|
||||
boost::container::static_vector<VsOutputMap, 3> vs_outputs;
|
||||
|
||||
s8 vertex_offset_sgpr = -1;
|
||||
s8 instance_offset_sgpr = -1;
|
||||
|
||||
BufferResourceList buffers;
|
||||
ImageResourceList images;
|
||||
SamplerResourceList samplers;
|
||||
|
@ -210,6 +213,18 @@ struct Info {
|
|||
std::memcpy(&data, base + dword_offset, sizeof(T));
|
||||
return data;
|
||||
}
|
||||
|
||||
[[nodiscard]] std::pair<u32, u32> GetDrawOffsets() const noexcept {
|
||||
u32 vertex_offset = 0;
|
||||
u32 instance_offset = 0;
|
||||
if (vertex_offset_sgpr != -1) {
|
||||
vertex_offset = user_data[vertex_offset_sgpr];
|
||||
}
|
||||
if (instance_offset_sgpr != -1) {
|
||||
instance_offset = user_data[instance_offset_sgpr];
|
||||
}
|
||||
return {vertex_offset, instance_offset};
|
||||
}
|
||||
};
|
||||
|
||||
constexpr AmdGpu::Buffer BufferResource::GetVsharp(const Info& info) const noexcept {
|
||||
|
|
|
@ -52,13 +52,17 @@ void Rasterizer::Draw(bool is_indexed, u32 index_offset) {
|
|||
BeginRendering();
|
||||
UpdateDynamicState(*pipeline);
|
||||
|
||||
const auto [vertex_offset, instance_offset] = vs_info.GetDrawOffsets();
|
||||
|
||||
if (is_indexed) {
|
||||
cmdbuf.drawIndexed(num_indices, regs.num_instances.NumInstances(), 0, 0, 0);
|
||||
cmdbuf.drawIndexed(num_indices, regs.num_instances.NumInstances(), 0, s32(vertex_offset),
|
||||
instance_offset);
|
||||
} else {
|
||||
const u32 num_vertices = regs.primitive_type == AmdGpu::Liverpool::PrimitiveType::RectList
|
||||
? 4
|
||||
: regs.num_indices;
|
||||
cmdbuf.draw(num_vertices, regs.num_instances.NumInstances(), 0, 0);
|
||||
cmdbuf.draw(num_vertices, regs.num_instances.NumInstances(), vertex_offset,
|
||||
instance_offset);
|
||||
}
|
||||
}
|
||||
|
||||
|
|
Loading…
Reference in New Issue