Invert component names in RG11B10 and RGB10A2

The CL that introduced this formats reordered their components based on
the name of the Vulkan formats, but it turns out that WebGPU lists
components from low bits to high-bits while Vulkan PACK32 formats are
listed from high to low. So the format component orders match, except
for RGB10A2 which is actually BGR10A2 in Vulkan. Instead the Vulkan
backend is updated to correctly use the RGB10A2 format.

BUG=dawn:128

Change-Id: If7f045f020429c44c84b9aed34a4a80107208d5c
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/8601
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Jiawei Shao <jiawei.shao@intel.com>
This commit is contained in:
Corentin Wallez 2019-07-08 10:12:16 +00:00 committed by Commit Bot service account
parent ec05355c72
commit 6be313225e
5 changed files with 25 additions and 35 deletions

View File

@ -1075,8 +1075,8 @@
{"value": 25, "name": "RGBA8 sint"},
{"value": 26, "name": "BGRA8 unorm"},
{"value": 27, "name": "BGRA8 unorm srgb"},
{"value": 28, "name": "A2 RGB10 unorm"},
{"value": 29, "name": "B10 GR11 float"},
{"value": 28, "name": "RGB10 A2 unorm"},
{"value": 29, "name": "RG11 B10 float"},
{"value": 30, "name": "RG32 float"},
{"value": 31, "name": "RG32 uint"},

View File

@ -160,8 +160,8 @@ namespace dawn_native {
case dawn::TextureFormat::RGBA8Sint:
case dawn::TextureFormat::BGRA8Unorm:
case dawn::TextureFormat::BGRA8UnormSrgb:
case dawn::TextureFormat::A2RGB10Unorm:
case dawn::TextureFormat::B10GR11Float:
case dawn::TextureFormat::RGB10A2Unorm:
case dawn::TextureFormat::RG11B10Float:
return MakeColorFormat(format, true, 4);
case dawn::TextureFormat::RG32Uint:

View File

@ -179,12 +179,9 @@ namespace dawn_native { namespace metal {
return MTLPixelFormatBGRA8Unorm;
case dawn::TextureFormat::BGRA8UnormSrgb:
return MTLPixelFormatBGRA8Unorm_sRGB;
case dawn::TextureFormat::A2RGB10Unorm:
// TODO(cwallez@chromium.org): The format expectations are inverted compared to the
// implementation of the format in Metal.
UNREACHABLE();
case dawn::TextureFormat::RGB10A2Unorm:
return MTLPixelFormatRGB10A2Unorm;
case dawn::TextureFormat::B10GR11Float:
case dawn::TextureFormat::RG11B10Float:
return MTLPixelFormatRG11B10Float;
case dawn::TextureFormat::RG32Uint:

View File

@ -269,9 +269,9 @@ namespace dawn_native { namespace vulkan {
return VK_FORMAT_B8G8R8A8_UNORM;
case dawn::TextureFormat::BGRA8UnormSrgb:
return VK_FORMAT_B8G8R8A8_SRGB;
case dawn::TextureFormat::A2RGB10Unorm:
return VK_FORMAT_A2R10G10B10_UNORM_PACK32;
case dawn::TextureFormat::B10GR11Float:
case dawn::TextureFormat::RGB10A2Unorm:
return VK_FORMAT_A2B10G10R10_UNORM_PACK32;
case dawn::TextureFormat::RG11B10Float:
return VK_FORMAT_B10G11R11_UFLOAT_PACK32;
case dawn::TextureFormat::RG32Uint:

View File

@ -615,23 +615,18 @@ TEST_P(TextureFormatTest, BGRA8UnormSrgb) {
1.0e-3);
}
// Test the A2RGB10Unorm format
// TODO(cwallez@chromium.org): This is actually RGB10A2 in WebGPU but Vulkan doesn't support that
// format. Do all platforms have A in the high bits?
TEST_P(TextureFormatTest, A2RGB10Unorm) {
// TODO(cwallez@chromium.org): The format R and B channel are inverted compared to what Metal
// does.
DAWN_SKIP_TEST_IF(IsMetal());
auto MakeA2RGB10 = [](uint32_t r, uint32_t g, uint32_t b, uint32_t a) -> uint32_t {
// Test the RGB10A2Unorm format
TEST_P(TextureFormatTest, RGB10A2Unorm) {
auto MakeRGB10A2 = [](uint32_t r, uint32_t g, uint32_t b, uint32_t a) -> uint32_t {
ASSERT((r & 0x3FF) == r);
ASSERT((g & 0x3FF) == g);
ASSERT((b & 0x3FF) == b);
ASSERT((a & 0x3) == a);
return a << 30 | r << 20 | g << 10 | b;
return r | g << 10 | b << 20 | a << 30;
};
std::vector<uint32_t> textureData = {MakeA2RGB10(0, 0, 0, 0), MakeA2RGB10(1023, 1023, 1023, 1),
MakeA2RGB10(243, 578, 765, 2), MakeA2RGB10(0, 0, 0, 3)};
std::vector<uint32_t> textureData = {MakeRGB10A2(0, 0, 0, 0), MakeRGB10A2(1023, 1023, 1023, 1),
MakeRGB10A2(243, 578, 765, 2), MakeRGB10A2(0, 0, 0, 3)};
// clang-format off
std::vector<float> expectedData = {
0.0f, 0.0f, 0.0f, 0.0f,
@ -641,14 +636,12 @@ TEST_P(TextureFormatTest, A2RGB10Unorm) {
};
// clang-format on
DoSampleTest({dawn::TextureFormat::A2RGB10Unorm, 4, Float, 4}, textureData, expectedData,
DoSampleTest({dawn::TextureFormat::RGB10A2Unorm, 4, Float, 4}, textureData, expectedData,
2.0e-3);
}
// Test the B10GR11Float format
// TODO(cwallez@chromium.org): This is actually GR11B10 in WebGPU but Vulkan doesn't support that
// format. Do all platforms have it reversed?
TEST_P(TextureFormatTest, B10GR11Float) {
// Test the RG11B10Float format
TEST_P(TextureFormatTest, RG11B10Float) {
constexpr uint32_t kFloat11Zero = 0;
constexpr uint32_t kFloat11Infinity = 0x7C0;
constexpr uint32_t kFloat11Nan = 0x7C1;
@ -659,20 +652,20 @@ TEST_P(TextureFormatTest, B10GR11Float) {
constexpr uint32_t kFloat10Nan = 0x3E1;
constexpr uint32_t kFloat10One = 0x1E0;
auto MakeB10GR11 = [](uint32_t r, uint32_t g, uint32_t b) {
auto MakeRG11B10 = [](uint32_t r, uint32_t g, uint32_t b) {
ASSERT((r & 0x7FF) == r);
ASSERT((g & 0x7FF) == g);
ASSERT((b & 0x3FF) == b);
return b << 22 | g << 11 | r;
return r | g << 11 | b << 22;
};
// Test each of (0, 1, INFINITY, NaN) for each component but never two with the same value at a
// time.
std::vector<uint32_t> textureData = {
MakeB10GR11(kFloat11Zero, kFloat11Infinity, kFloat10Nan),
MakeB10GR11(kFloat11Infinity, kFloat11Nan, kFloat10One),
MakeB10GR11(kFloat11Nan, kFloat11One, kFloat10Zero),
MakeB10GR11(kFloat11One, kFloat11Zero, kFloat10Infinity),
MakeRG11B10(kFloat11Zero, kFloat11Infinity, kFloat10Nan),
MakeRG11B10(kFloat11Infinity, kFloat11Nan, kFloat10One),
MakeRG11B10(kFloat11Nan, kFloat11One, kFloat10Zero),
MakeRG11B10(kFloat11One, kFloat11Zero, kFloat10Infinity),
};
// This is one of the only 3-channel formats, so we don't have specific testing for them. Alpha
@ -686,7 +679,7 @@ TEST_P(TextureFormatTest, B10GR11Float) {
};
// clang-format on
DoSampleTest({dawn::TextureFormat::B10GR11Float, 4, Float, 4}, textureData, expectedData);
DoSampleTest({dawn::TextureFormat::RG11B10Float, 4, Float, 4}, textureData, expectedData);
}
// TODO(cwallez@chromium.org): Add tests for depth-stencil formats when we know if they are copyable