[D3D12] ROV: 32-bit loading, except for 7e3

This commit is contained in:
Triang3l
2018-10-12 20:06:40 +03:00
parent 3dc15dbb44
commit 10a3dc6d94
7 changed files with 809 additions and 482 deletions

View File

@@ -1793,7 +1793,8 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
dirty |= system_constants_.alpha_test != alpha_test;
system_constants_.alpha_test = alpha_test;
// Color exponent bias and output index mapping.
// Color exponent bias and output index mapping or ROV writing.
uint32_t rb_color_mask = regs[XE_GPU_REG_RB_COLOR_MASK].u32;
for (uint32_t i = 0; i < 4; ++i) {
uint32_t color_info;
switch (i) {
@@ -1841,9 +1842,60 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
80;
dirty |= system_constants_.edram_pitch_tiles != edram_pitch_tiles;
system_constants_.edram_pitch_tiles = edram_pitch_tiles;
static const uint32_t kRTFormatFlags[16] = {
// k_8_8_8_8
DxbcShaderTranslator::kRTFlag_FormatFixed,
// k_8_8_8_8_GAMMA
DxbcShaderTranslator::kRTFlag_FormatFixed,
// k_2_10_10_10
DxbcShaderTranslator::kRTFlag_FormatFixed,
// k_2_10_10_10_FLOAT
DxbcShaderTranslator::kRTFlag_FormatFloat10,
// k_16_16
DxbcShaderTranslator::kRTFlag_FormatFixed,
// k_16_16_16_16
DxbcShaderTranslator::kRTFlag_Format64bpp |
DxbcShaderTranslator::kRTFlag_FormatFixed,
// k_16_16_FLOAT
DxbcShaderTranslator::kRTFlag_FormatFloat16,
// k_16_16_16_16_FLOAT
DxbcShaderTranslator::kRTFlag_Format64bpp |
DxbcShaderTranslator::kRTFlag_FormatFloat16,
// Unused
0,
// Unused
0,
// k_2_10_10_10_AS_16_16_16_16
DxbcShaderTranslator::kRTFlag_FormatFixed,
// Unused.
0,
// k_2_10_10_10_FLOAT_AS_16_16_16_16
DxbcShaderTranslator::kRTFlag_FormatFloat10,
// Unused.
0,
// k_32_FLOAT
0,
// k_32_32_FLOAT
DxbcShaderTranslator::kRTFlag_Format64bpp,
};
static const uint32_t kRTFormatAllComponentsMask[16] = {
0b1111, 0b1111, 0b1111, 0b1111, 0b0011, 0b1111, 0b0011, 0b1111,
0b0000, 0b0000, 0b1111, 0b0000, 0b1111, 0b0000, 0b0001, 0b0011,
};
uint32_t rt_mask_all = kRTFormatAllComponentsMask[uint32_t(color_format)];
uint32_t rt_mask = (rb_color_mask >> (i * 4)) & rt_mask_all;
uint32_t rt_flags = kRTFormatFlags[uint32_t(color_format)];
if (rt_mask != 0) {
rt_flags |= DxbcShaderTranslator::kRTFlag_Used;
if (rt_mask != rt_mask_all) {
rt_flags |= DxbcShaderTranslator::kRTFlag_LoadingNeeded;
}
}
dirty |= system_constants_.edram_rt_flags[i] != rt_flags;
system_constants_.edram_rt_flags[i] = rt_flags;
if (system_constants_color_formats_[i] != color_format) {
dirty = true;
uint32_t rt_flags = 0;
uint32_t color_mask = UINT32_MAX, alpha_mask = UINT32_MAX;
// Initialize min/max to Infinity.
uint32_t color_min = 0xFF800000u, alpha_min = 0xFF800000u;
uint32_t color_max = 0x7F800000u, alpha_max = 0x7F800000u;
@@ -1851,7 +1903,6 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
switch (color_format) {
case ColorRenderTargetFormat::k_8_8_8_8:
case ColorRenderTargetFormat::k_8_8_8_8_GAMMA:
rt_flags |= DxbcShaderTranslator::kRTFlag_FormatFixed;
system_constants_.edram_rt_pack_width_low[i][0] = 8;
system_constants_.edram_rt_pack_width_low[i][1] = 8;
system_constants_.edram_rt_pack_width_low[i][2] = 8;
@@ -1860,13 +1911,13 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
system_constants_.edram_rt_pack_offset_low[i][1] = 8;
system_constants_.edram_rt_pack_offset_low[i][2] = 16;
system_constants_.edram_rt_pack_offset_low[i][3] = 24;
color_mask = alpha_mask = 255;
color_min = alpha_min = 0;
color_max = alpha_max = 0x3F800000;
color_store_scale = alpha_store_scale = 255.0f;
break;
case ColorRenderTargetFormat::k_2_10_10_10:
case ColorRenderTargetFormat::k_2_10_10_10_AS_16_16_16_16:
rt_flags |= DxbcShaderTranslator::kRTFlag_FormatFixed;
system_constants_.edram_rt_pack_width_low[i][0] = 10;
system_constants_.edram_rt_pack_width_low[i][1] = 10;
system_constants_.edram_rt_pack_width_low[i][2] = 10;
@@ -1875,6 +1926,8 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
system_constants_.edram_rt_pack_offset_low[i][1] = 10;
system_constants_.edram_rt_pack_offset_low[i][2] = 20;
system_constants_.edram_rt_pack_offset_low[i][3] = 30;
color_mask = 1023;
alpha_mask = 3;
color_min = alpha_min = 0;
color_max = alpha_max = 0x3F800000;
// 1023.0.
@@ -1883,7 +1936,6 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
break;
case ColorRenderTargetFormat::k_2_10_10_10_FLOAT:
case ColorRenderTargetFormat::k_2_10_10_10_FLOAT_AS_16_16_16_16:
rt_flags |= DxbcShaderTranslator::kRTFlag_FormatFloat10;
system_constants_.edram_rt_pack_width_low[i][0] = 10;
system_constants_.edram_rt_pack_width_low[i][1] = 10;
system_constants_.edram_rt_pack_width_low[i][2] = 10;
@@ -1892,6 +1944,8 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
system_constants_.edram_rt_pack_offset_low[i][1] = 10;
system_constants_.edram_rt_pack_offset_low[i][2] = 20;
system_constants_.edram_rt_pack_offset_low[i][3] = 30;
color_mask = 1023;
alpha_mask = 3;
color_min = alpha_min = 0;
// 31.875.
color_max = 0x41FF0000;
@@ -1900,7 +1954,6 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
break;
case ColorRenderTargetFormat::k_16_16:
case ColorRenderTargetFormat::k_16_16_16_16:
rt_flags |= DxbcShaderTranslator::kRTFlag_FormatFixed;
system_constants_.edram_rt_pack_width_low[i][0] = 16;
system_constants_.edram_rt_pack_width_low[i][1] = 16;
system_constants_.edram_rt_pack_width_low[i][2] = 0;
@@ -1910,6 +1963,7 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
system_constants_.edram_rt_pack_offset_low[i][2] = 0;
system_constants_.edram_rt_pack_offset_low[i][3] = 0;
// TODO(Triang3l): 64bpp variant.
// Color and alpha mask UINT32_MAX because the format is signed.
// -32.0.
color_min = alpha_min = 0xC2000000u;
// 32.0.
@@ -1918,7 +1972,6 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
break;
case ColorRenderTargetFormat::k_16_16_FLOAT:
case ColorRenderTargetFormat::k_16_16_16_16_FLOAT:
rt_flags |= DxbcShaderTranslator::kRTFlag_FormatFloat16;
system_constants_.edram_rt_pack_width_low[i][0] = 16;
system_constants_.edram_rt_pack_width_low[i][1] = 16;
system_constants_.edram_rt_pack_width_low[i][2] = 0;
@@ -1928,6 +1981,7 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
system_constants_.edram_rt_pack_offset_low[i][2] = 0;
system_constants_.edram_rt_pack_offset_low[i][3] = 0;
// TODO(Triang3l): 64bpp variant.
color_mask = alpha_mask = 0xFFFF;
break;
case ColorRenderTargetFormat::k_32_FLOAT:
case ColorRenderTargetFormat::k_32_32_FLOAT:
@@ -1945,9 +1999,14 @@ void D3D12CommandProcessor::UpdateSystemConstantValues(
assert_always();
break;
}
system_constants_.edram_rt_flags[i] = rt_flags;
uint32_t rt_pair_index = i >> 1;
uint32_t rt_pair_comp = (i & 1) << 1;
system_constants_
.edram_load_mask_rt01_rt23[rt_pair_index][rt_pair_comp] =
color_mask;
system_constants_
.edram_load_mask_rt01_rt23[rt_pair_index][rt_pair_comp + 1] =
alpha_mask;
system_constants_
.edram_store_min_rt01_rt23[rt_pair_index][rt_pair_comp] = color_min;
system_constants_