gpu: implement polygon offset

Co-authored-by: Pouare514 <julienaulycee@gmail.com>
This commit is contained in:
nmzik
2026-08-06 22:02:19 +02:00
parent 3052dac82e
commit e9f5da27f8
9 changed files with 169 additions and 49 deletions
@@ -18,7 +18,9 @@
#include <algorithm>
#include <array>
#include <atomic>
#include <bit>
#include <cstdio>
#include <cstring>
#include <vector>
#define KYTY_HW_CTX_PARSER_ARGS \
@@ -895,23 +897,44 @@ KYTY_HW_CTX_PARSER(HwCtxSetModeControl) {
return 1;
}
static void HwCtxIgnorePolyOffsetRegister(uint32_t cmd_offset, uint32_t value) {
static std::atomic<uint32_t> log_count = 0;
static void HwCtxSetPolyOffsetRegister(CommandProcessor& cp, uint32_t cmd_offset, uint32_t value) {
HW::PolyOffset offset = cp.GetCtx().GetPolyOffset();
auto count = log_count.fetch_add(1);
if (count < 8) {
LOGF_COLOR(Log::Color::Red,
"\t temporary: ignoring polygon offset context register 0x%03" PRIx32
" = 0x%08" PRIx32 " (depth bias not implemented)\n",
cmd_offset, value);
switch (cmd_offset) {
case Pm4::PA_SU_POLY_OFFSET_DB_FMT_CNTL: {
const auto neg_bits =
KYTY_PM4_GET(value, PA_SU_POLY_OFFSET_DB_FMT_CNTL, NEG_NUM_DB_BITS);
offset.neg_num_db_bits = std::bit_cast<int8_t>(static_cast<uint8_t>(neg_bits));
offset.db_is_float_fmt =
KYTY_PM4_GET(value, PA_SU_POLY_OFFSET_DB_FMT_CNTL, DB_IS_FLOAT_FMT) != 0;
break;
}
case Pm4::PA_SU_POLY_OFFSET_CLAMP: offset.clamp = std::bit_cast<float>(value); break;
case Pm4::PA_SU_POLY_OFFSET_FRONT_SCALE:
offset.front_scale = std::bit_cast<float>(value);
break;
case Pm4::PA_SU_POLY_OFFSET_FRONT_OFFSET:
offset.front_offset = std::bit_cast<float>(value);
break;
case Pm4::PA_SU_POLY_OFFSET_BACK_SCALE:
offset.back_scale = std::bit_cast<float>(value);
break;
case Pm4::PA_SU_POLY_OFFSET_BACK_OFFSET:
offset.back_offset = std::bit_cast<float>(value);
break;
default:
EXIT("unknown polygon offset context register 0x%03" PRIx32 " = 0x%08" PRIx32 "\n",
cmd_offset, value);
}
cp.GetCtx().SetPolyOffset(offset);
}
KYTY_HW_CTX_PARSER(HwCtxSetPolyOffsetRegisters) {
auto num_values = KYTY_PM4_LEN(cmd_id) - 2u;
for (uint32_t i = 0; i < num_values; i++) {
HwCtxIgnorePolyOffsetRegister(cmd_offset + i, buffer[i]);
HwCtxSetPolyOffsetRegister(cp, cmd_offset + i, buffer[i]);
}
return num_values;
@@ -3705,22 +3728,22 @@ void GraphicsInitJmpTablesCxIndirect() {
};
g_hw_ctx_indirect_func[Pm4::PA_SU_POLY_OFFSET_DB_FMT_CNTL] = [](KYTY_HW_CTX_INDIRECT_ARGS) {
HwCtxIgnorePolyOffsetRegister(cmd_offset, value);
HwCtxSetPolyOffsetRegister(cp, cmd_offset, value);
};
g_hw_ctx_indirect_func[Pm4::PA_SU_POLY_OFFSET_CLAMP] = [](KYTY_HW_CTX_INDIRECT_ARGS) {
HwCtxIgnorePolyOffsetRegister(cmd_offset, value);
HwCtxSetPolyOffsetRegister(cp, cmd_offset, value);
};
g_hw_ctx_indirect_func[Pm4::PA_SU_POLY_OFFSET_FRONT_SCALE] = [](KYTY_HW_CTX_INDIRECT_ARGS) {
HwCtxIgnorePolyOffsetRegister(cmd_offset, value);
HwCtxSetPolyOffsetRegister(cp, cmd_offset, value);
};
g_hw_ctx_indirect_func[Pm4::PA_SU_POLY_OFFSET_FRONT_OFFSET] = [](KYTY_HW_CTX_INDIRECT_ARGS) {
HwCtxIgnorePolyOffsetRegister(cmd_offset, value);
HwCtxSetPolyOffsetRegister(cp, cmd_offset, value);
};
g_hw_ctx_indirect_func[Pm4::PA_SU_POLY_OFFSET_BACK_SCALE] = [](KYTY_HW_CTX_INDIRECT_ARGS) {
HwCtxIgnorePolyOffsetRegister(cmd_offset, value);
HwCtxSetPolyOffsetRegister(cp, cmd_offset, value);
};
g_hw_ctx_indirect_func[Pm4::PA_SU_POLY_OFFSET_BACK_OFFSET] = [](KYTY_HW_CTX_INDIRECT_ARGS) {
HwCtxIgnorePolyOffsetRegister(cmd_offset, value);
HwCtxSetPolyOffsetRegister(cp, cmd_offset, value);
};
g_hw_ctx_indirect_func[Pm4::DB_Z_INFO] = [](KYTY_HW_CTX_INDIRECT_ARGS) {
+13
View File
@@ -348,6 +348,16 @@ struct ModeControl {
bool persp_corr_dis = false;
};
struct PolyOffset {
int8_t neg_num_db_bits = -23;
bool db_is_float_fmt = true;
float clamp = 0.0f;
float front_scale = 0.0f;
float front_offset = 0.0f;
float back_scale = 0.0f;
float back_offset = 0.0f;
};
struct BlendControl {
uint8_t color_srcblend = 1;
uint8_t color_comb_fcn = 0;
@@ -905,6 +915,8 @@ public:
void SetDepthControl(const DepthControl& control) { m_depth_control = control; }
[[nodiscard]] const ModeControl& GetModeControl() const { return m_mode_control; }
void SetModeControl(const ModeControl& control) { m_mode_control = control; }
[[nodiscard]] const PolyOffset& GetPolyOffset() const { return m_poly_offset; }
void SetPolyOffset(const PolyOffset& offset) { m_poly_offset = offset; }
[[nodiscard]] const EqaaControl& GetEqaaControl() const { return m_eqaa_control; }
void SetEqaaControl(const EqaaControl& control) { m_eqaa_control = control; }
[[nodiscard]] const StencilControl& GetStencilControl() const { return m_stencil_control; }
@@ -1002,6 +1014,7 @@ private:
uint8_t m_stencil_clear_value = 0;
ModeControl m_mode_control;
PolyOffset m_poly_offset;
EqaaControl m_eqaa_control;
ShaderRegisters m_sh_regs;
+22 -18
View File
@@ -578,24 +578,28 @@ constexpr uint32_t DB_HTILE_SURFACE_PREFETCH_HEIGHT_MASK = 0x3F;
constexpr uint32_t DB_HTILE_SURFACE_DST_OUTSIDE_ZERO_TO_ONE_SHIFT = 16;
constexpr uint32_t DB_HTILE_SURFACE_DST_OUTSIDE_ZERO_TO_ONE_MASK = 0x1;
constexpr uint32_t DB_SRESULTS_COMPARE_STATE0 = 0x2B0;
constexpr uint32_t DB_SRESULTS_COMPARE_STATE1 = 0x2B1;
constexpr uint32_t VGT_GS_MAX_VERT_OUT = 0x2CE;
constexpr uint32_t GE_NGG_SUBGRP_CNTL = 0x2D3;
constexpr uint32_t VGT_TESS_DISTRIBUTION = 0x2D4;
constexpr uint32_t VGT_SHADER_STAGES_EN = 0x2D5;
constexpr uint32_t VGT_LS_HS_CONFIG = 0x2D6;
constexpr uint32_t VGT_TF_PARAM = 0x2DB;
constexpr uint32_t DB_ALPHA_TO_MASK = 0x2DC;
constexpr uint32_t PA_SU_POLY_OFFSET_DB_FMT_CNTL = 0x2DE;
constexpr uint32_t PA_SU_POLY_OFFSET_CLAMP = 0x2DF;
constexpr uint32_t PA_SU_POLY_OFFSET_FRONT_SCALE = 0x2E0;
constexpr uint32_t PA_SU_POLY_OFFSET_FRONT_OFFSET = 0x2E1;
constexpr uint32_t PA_SU_POLY_OFFSET_BACK_SCALE = 0x2E2;
constexpr uint32_t PA_SU_POLY_OFFSET_BACK_OFFSET = 0x2E3;
constexpr uint32_t VGT_GS_INSTANCE_CNT = 0x2E4;
constexpr uint32_t PA_SC_CENTROID_PRIORITY_0 = 0x2F5;
constexpr uint32_t PA_SC_CENTROID_PRIORITY_1 = 0x2F6;
constexpr uint32_t DB_SRESULTS_COMPARE_STATE0 = 0x2B0;
constexpr uint32_t DB_SRESULTS_COMPARE_STATE1 = 0x2B1;
constexpr uint32_t VGT_GS_MAX_VERT_OUT = 0x2CE;
constexpr uint32_t GE_NGG_SUBGRP_CNTL = 0x2D3;
constexpr uint32_t VGT_TESS_DISTRIBUTION = 0x2D4;
constexpr uint32_t VGT_SHADER_STAGES_EN = 0x2D5;
constexpr uint32_t VGT_LS_HS_CONFIG = 0x2D6;
constexpr uint32_t VGT_TF_PARAM = 0x2DB;
constexpr uint32_t DB_ALPHA_TO_MASK = 0x2DC;
constexpr uint32_t PA_SU_POLY_OFFSET_DB_FMT_CNTL = 0x2DE;
constexpr uint32_t PA_SU_POLY_OFFSET_DB_FMT_CNTL_NEG_NUM_DB_BITS_SHIFT = 0;
constexpr uint32_t PA_SU_POLY_OFFSET_DB_FMT_CNTL_NEG_NUM_DB_BITS_MASK = 0xFF;
constexpr uint32_t PA_SU_POLY_OFFSET_DB_FMT_CNTL_DB_IS_FLOAT_FMT_SHIFT = 8;
constexpr uint32_t PA_SU_POLY_OFFSET_DB_FMT_CNTL_DB_IS_FLOAT_FMT_MASK = 0x1;
constexpr uint32_t PA_SU_POLY_OFFSET_CLAMP = 0x2DF;
constexpr uint32_t PA_SU_POLY_OFFSET_FRONT_SCALE = 0x2E0;
constexpr uint32_t PA_SU_POLY_OFFSET_FRONT_OFFSET = 0x2E1;
constexpr uint32_t PA_SU_POLY_OFFSET_BACK_SCALE = 0x2E2;
constexpr uint32_t PA_SU_POLY_OFFSET_BACK_OFFSET = 0x2E3;
constexpr uint32_t VGT_GS_INSTANCE_CNT = 0x2E4;
constexpr uint32_t PA_SC_CENTROID_PRIORITY_0 = 0x2F5;
constexpr uint32_t PA_SC_CENTROID_PRIORITY_1 = 0x2F6;
constexpr uint32_t PA_SC_AA_CONFIG = 0x2F8;
constexpr uint32_t PA_SC_AA_CONFIG_MSAA_NUM_SAMPLES_SHIFT = 0;