August 2018 - Wine-Devel - list.winehq.org

[PATCH vkd3d 05/14] libs/vkd3d-shader: Translate sincos instructions.

by Józef Kucia

From: Józef Kucia <jkucia(a)codeweavers.com> Signed-off-by: Józef Kucia <jkucia(a)codeweavers.com> --- libs/vkd3d-shader/spirv.c | 68 ++++++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 65 insertions(+), 3 deletions(-) diff --git a/libs/vkd3d-shader/spirv.c b/libs/vkd3d-shader/spirv.c index a0f084cea366..21d6a1a33288 100644 --- a/libs/vkd3d-shader/spirv.c +++ b/libs/vkd3d-shader/spirv.c @@ -1427,12 +1427,29 @@ static void vkd3d_spirv_build_op_memory_barrier(struct vkd3d_spirv_builder *buil SpvOpMemoryBarrier, memory_id, memory_semantics_id); } +static uint32_t vkd3d_spirv_build_op_glsl_std450_tr1(struct vkd3d_spirv_builder *builder, + enum GLSLstd450 op, uint32_t result_type, uint32_t operand) +{ + uint32_t id = vkd3d_spirv_get_glsl_std450_instr_set(builder); + return vkd3d_spirv_build_op_ext_inst(builder, result_type, id, op, &operand, 1); +} + static uint32_t vkd3d_spirv_build_op_glsl_std450_fabs(struct vkd3d_spirv_builder *builder, uint32_t result_type, uint32_t operand) { - uint32_t glsl_std450_id = vkd3d_spirv_get_glsl_std450_instr_set(builder); - return vkd3d_spirv_build_op_ext_inst(builder, result_type, glsl_std450_id, - GLSLstd450FAbs, &operand, 1); + return vkd3d_spirv_build_op_glsl_std450_tr1(builder, GLSLstd450FAbs, result_type, operand); +} + +static uint32_t vkd3d_spirv_build_op_glsl_std450_sin(struct vkd3d_spirv_builder *builder, + uint32_t result_type, uint32_t operand) +{ + return vkd3d_spirv_build_op_glsl_std450_tr1(builder, GLSLstd450Sin, result_type, operand); +} + +static uint32_t vkd3d_spirv_build_op_glsl_std450_cos(struct vkd3d_spirv_builder *builder, + uint32_t result_type, uint32_t operand) +{ + return vkd3d_spirv_build_op_glsl_std450_tr1(builder, GLSLstd450Cos, result_type, operand); } static uint32_t vkd3d_spirv_build_op_glsl_std450_nclamp(struct vkd3d_spirv_builder *builder, @@ -4284,6 +4301,48 @@ static void vkd3d_dxbc_compiler_emit_rcp(struct vkd3d_dxbc_compiler *compiler, vkd3d_dxbc_compiler_emit_store_dst(compiler, dst, val_id); } +static void vkd3d_dxbc_compiler_emit_sincos(struct vkd3d_dxbc_compiler *compiler, + const struct vkd3d_shader_instruction *instruction) +{ + const struct vkd3d_shader_dst_param *dst_sin = &instruction->dst[0]; + const struct vkd3d_shader_dst_param *dst_cos = &instruction->dst[1]; + struct vkd3d_spirv_builder *builder = &compiler->spirv_builder; + const struct vkd3d_shader_src_param *src = instruction->src; + uint32_t type_id, src_id, sin_id = 0, cos_id = 0; + unsigned int component_count; + + if (dst_sin->reg.type != VKD3DSPR_NULL) + { + component_count = vkd3d_write_mask_component_count(dst_sin->write_mask); + type_id = vkd3d_spirv_get_type_id(builder, + vkd3d_component_type_from_data_type(dst_sin->reg.data_type), component_count); + + src_id = vkd3d_dxbc_compiler_emit_load_src(compiler, src, dst_sin->write_mask); + + sin_id = vkd3d_spirv_build_op_glsl_std450_sin(builder, type_id, src_id); + } + + if (dst_cos->reg.type != VKD3DSPR_NULL) + { + if (dst_sin->reg.type == VKD3DSPR_NULL || dst_cos->write_mask != dst_sin->write_mask) + { + component_count = vkd3d_write_mask_component_count(dst_cos->write_mask); + type_id = vkd3d_spirv_get_type_id(builder, + vkd3d_component_type_from_data_type(dst_cos->reg.data_type), component_count); + + src_id = vkd3d_dxbc_compiler_emit_load_src(compiler, src, dst_cos->write_mask); + } + + cos_id = vkd3d_spirv_build_op_glsl_std450_cos(builder, type_id, src_id); + } + + if (sin_id) + vkd3d_dxbc_compiler_emit_store_dst(compiler, dst_sin, sin_id); + + if (cos_id) + vkd3d_dxbc_compiler_emit_store_dst(compiler, dst_cos, cos_id); +} + static void vkd3d_dxbc_compiler_emit_imul(struct vkd3d_dxbc_compiler *compiler, const struct vkd3d_shader_instruction *instruction) { @@ -5951,6 +6010,9 @@ void vkd3d_dxbc_compiler_handle_instruction(struct vkd3d_dxbc_compiler *compiler case VKD3DSIH_RCP: vkd3d_dxbc_compiler_emit_rcp(compiler, instruction); break; + case VKD3DSIH_SINCOS: + vkd3d_dxbc_compiler_emit_sincos(compiler, instruction); + break; case VKD3DSIH_IMUL: vkd3d_dxbc_compiler_emit_imul(compiler, instruction); break; -- 2.16.4

Aug. 1, 2018

[PATCH vkd3d 04/14] tests: Add test for round_ne instruction.

by Józef Kucia

From: Józef Kucia <jkucia(a)codeweavers.com> Signed-off-by: Józef Kucia <jkucia(a)codeweavers.com> --- tests/d3d12.c | 28 ++++++++++++++++++++++++++++ 1 file changed, 28 insertions(+) diff --git a/tests/d3d12.c b/tests/d3d12.c index 608ff68c3af2..83ec6c83b6b4 100644 --- a/tests/d3d12.c +++ b/tests/d3d12.c @@ -5540,6 +5540,24 @@ static void test_shader_instructions(void) 0x00000000, 0x00000000, 0x05000036, 0x00102082, 0x00000000, 0x00004001, 0x00000000, 0x0100003e, }; static const D3D12_SHADER_BYTECODE ps_round = {ps_round_code, sizeof(ps_round_code)}; + static const DWORD ps_round_ne_code[] = + { +#if 0 + float4 src0; + + void main(out float4 dst : SV_Target) + { + dst = round(src0); + } +#endif + 0x43425844, 0xa2be1ad3, 0xf1389007, 0xc8221829, 0xcbef8ed0, 0x00000001, 0x000000bc, 0x00000003, + 0x0000002c, 0x0000003c, 0x00000070, 0x4e475349, 0x00000008, 0x00000000, 0x00000008, 0x4e47534f, + 0x0000002c, 0x00000001, 0x00000008, 0x00000020, 0x00000000, 0x00000000, 0x00000003, 0x00000000, + 0x0000000f, 0x545f5653, 0x65677261, 0xabab0074, 0x58454853, 0x00000044, 0x00000050, 0x00000011, + 0x0100086a, 0x04000059, 0x00208e46, 0x00000000, 0x00000001, 0x03000065, 0x001020f2, 0x00000000, + 0x06000040, 0x001020f2, 0x00000000, 0x00208e46, 0x00000000, 0x00000000, 0x0100003e, + }; + static const D3D12_SHADER_BYTECODE ps_round_ne = {ps_round_ne_code, sizeof(ps_round_ne_code)}; static const DWORD ps_frc_code[] = { #if 0 @@ -6427,6 +6445,16 @@ static void test_shader_instructions(void) {&ps_round, {{ INFINITY}}, {{ INFINITY, INFINITY, INFINITY}}}, {&ps_round, {{-INFINITY}}, {{-INFINITY, -INFINITY, -INFINITY}}}, + {&ps_round_ne, {{ 0.0f, -0.0f, 0.5f, -0.5f}}, {{ 0.0f, -0.0f, 0.0f, -0.0f}}}, + {&ps_round_ne, {{ 2.0f, 3.0f, 4.0f, 5.0f}}, {{ 2.0f, 3.0f, 4.0f, 5.0f}}}, + {&ps_round_ne, {{ 2.4f, 3.4f, 4.4f, 5.4f}}, {{ 2.0f, 3.0f, 4.0f, 5.0f}}}, + {&ps_round_ne, {{ 2.5f, 3.5f, 4.5f, 5.5f}}, {{ 2.0f, 4.0f, 4.0f, 6.0f}}}, + {&ps_round_ne, {{ 2.6f, 3.6f, 4.6f, 5.6f}}, {{ 3.0f, 4.0f, 5.0f, 6.0f}}}, + {&ps_round_ne, {{-2.5f, -3.5f, -4.5f, -5.5f}}, {{-2.0f, -4.0f, -4.0f, -6.0f}}}, + {&ps_round_ne, {{-2.4f, -3.4f, -4.4f, -5.4f}}, {{-2.0f, -3.0f, -4.0f, -5.0f}}}, + {&ps_round_ne, {{ INFINITY}}, {{ INFINITY}}}, + {&ps_round_ne, {{-INFINITY}}, {{-INFINITY}}}, + {&ps_frc, {{ 0.0f}}, {{0.0f, 0.0f}}}, {&ps_frc, {{-0.0f}}, {{0.0f, 0.0f}}}, {&ps_frc, {{ 1.0f}}, {{0.0f, 0.0f}}}, -- 2.16.4

Aug. 1, 2018

[PATCH vkd3d 03/14] libs/vkd3d-shader: Translate round_ne instructions.

by Józef Kucia

From: Józef Kucia <jkucia(a)codeweavers.com> Signed-off-by: Józef Kucia <jkucia(a)codeweavers.com> --- libs/vkd3d-shader/spirv.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/libs/vkd3d-shader/spirv.c b/libs/vkd3d-shader/spirv.c index d236db9dc23c..a0f084cea366 100644 --- a/libs/vkd3d-shader/spirv.c +++ b/libs/vkd3d-shader/spirv.c @@ -4063,6 +4063,7 @@ static enum GLSLstd450 vkd3d_dxbc_compiler_map_ext_glsl_instruction( {VKD3DSIH_MAD, GLSLstd450Fma}, {VKD3DSIH_MAX, GLSLstd450NMax}, {VKD3DSIH_MIN, GLSLstd450NMin}, + {VKD3DSIH_ROUND_NE, GLSLstd450RoundEven}, {VKD3DSIH_ROUND_NI, GLSLstd450Floor}, {VKD3DSIH_ROUND_PI, GLSLstd450Ceil}, {VKD3DSIH_ROUND_Z, GLSLstd450Trunc}, @@ -5932,6 +5933,7 @@ void vkd3d_dxbc_compiler_handle_instruction(struct vkd3d_dxbc_compiler *compiler case VKD3DSIH_MAD: case VKD3DSIH_MAX: case VKD3DSIH_MIN: + case VKD3DSIH_ROUND_NE: case VKD3DSIH_ROUND_NI: case VKD3DSIH_ROUND_PI: case VKD3DSIH_ROUND_Z: -- 2.16.4

Aug. 1, 2018

[PATCH vkd3d 02/14] libs/vkd3d: Add support for more formats.

by Józef Kucia

From: Józef Kucia <jkucia(a)codeweavers.com> Signed-off-by: Józef Kucia <jkucia(a)codeweavers.com> --- libs/vkd3d/utils.c | 103 ++++++++++++++++++++++++++++------------------------- 1 file changed, 54 insertions(+), 49 deletions(-) diff --git a/libs/vkd3d/utils.c b/libs/vkd3d/utils.c index 87688f7752ad..98b764a169d1 100644 --- a/libs/vkd3d/utils.c +++ b/libs/vkd3d/utils.c @@ -24,55 +24,60 @@ #define DEPTH_STENCIL (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT) static const struct vkd3d_format vkd3d_formats[] = { - {DXGI_FORMAT_R32G32B32A32_FLOAT, VK_FORMAT_R32G32B32A32_SFLOAT, 16, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R32G32B32A32_UINT, VK_FORMAT_R32G32B32A32_UINT, 16, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R32G32B32A32_SINT, VK_FORMAT_R32G32B32A32_SINT, 16, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R32G32B32_FLOAT, VK_FORMAT_R32G32B32_SFLOAT, 12, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R32G32_FLOAT, VK_FORMAT_R32G32_SFLOAT, 8, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R16G16B16A16_FLOAT, VK_FORMAT_R16G16B16A16_SFLOAT, 8, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R16G16B16A16_UINT, VK_FORMAT_R16G16B16A16_UINT, 8, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R10G10B10A2_UNORM, VK_FORMAT_A2B10G10R10_UNORM_PACK32, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R11G11B10_FLOAT, VK_FORMAT_B10G11R11_UFLOAT_PACK32, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R8G8B8A8_TYPELESS, VK_FORMAT_R8G8B8A8_UNORM, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R8G8B8A8_UNORM, VK_FORMAT_R8G8B8A8_UNORM, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R8G8B8A8_UNORM_SRGB, VK_FORMAT_R8G8B8A8_SRGB, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R16G16_FLOAT, VK_FORMAT_R16G16_SFLOAT, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R16G16_UNORM, VK_FORMAT_R16G16_UNORM, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R32_TYPELESS, VK_FORMAT_R32_UINT, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_D32_FLOAT, VK_FORMAT_D32_SFLOAT, 4, 1, 1, 1, DEPTH}, - {DXGI_FORMAT_R32_FLOAT, VK_FORMAT_R32_SFLOAT, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R32_UINT, VK_FORMAT_R32_UINT, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R32_SINT, VK_FORMAT_R32_SINT, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R16_TYPELESS, VK_FORMAT_R16_UINT, 2, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R16_FLOAT, VK_FORMAT_R16_SFLOAT, 2, 1, 1, 1, COLOR}, - {DXGI_FORMAT_D16_UNORM, VK_FORMAT_D16_UNORM, 2, 1, 1, 1, DEPTH}, - {DXGI_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, 2, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, 1, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R8_UINT, VK_FORMAT_R8_UINT, 1, 1, 1, 1, COLOR}, - {DXGI_FORMAT_R8_SNORM, VK_FORMAT_R8_SNORM, 1, 1, 1, 1, COLOR}, - {DXGI_FORMAT_A8_UNORM, VK_FORMAT_R8_UNORM, 1, 1, 1, 1, COLOR}, - {DXGI_FORMAT_B8G8R8A8_UNORM, VK_FORMAT_B8G8R8A8_UNORM, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_B8G8R8A8_TYPELESS, VK_FORMAT_B8G8R8A8_UNORM, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_B8G8R8A8_UNORM_SRGB, VK_FORMAT_B8G8R8A8_SRGB, 4, 1, 1, 1, COLOR}, - {DXGI_FORMAT_BC1_TYPELESS, VK_FORMAT_BC1_RGBA_UNORM_BLOCK, 1, 4, 4, 8, COLOR}, - {DXGI_FORMAT_BC1_UNORM, VK_FORMAT_BC1_RGBA_UNORM_BLOCK, 1, 4, 4, 8, COLOR}, - {DXGI_FORMAT_BC1_UNORM_SRGB, VK_FORMAT_BC1_RGBA_SRGB_BLOCK, 1, 4, 4, 8, COLOR}, - {DXGI_FORMAT_BC2_TYPELESS, VK_FORMAT_BC2_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC2_UNORM, VK_FORMAT_BC2_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC2_UNORM_SRGB, VK_FORMAT_BC2_SRGB_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC3_TYPELESS, VK_FORMAT_BC3_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC3_UNORM, VK_FORMAT_BC3_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC3_UNORM_SRGB, VK_FORMAT_BC3_SRGB_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC4_TYPELESS, VK_FORMAT_BC4_UNORM_BLOCK, 1, 4, 4, 8, COLOR}, - {DXGI_FORMAT_BC4_UNORM, VK_FORMAT_BC4_UNORM_BLOCK, 1, 4, 4, 8, COLOR}, - {DXGI_FORMAT_BC4_SNORM, VK_FORMAT_BC4_SNORM_BLOCK, 1, 4, 4, 8, COLOR}, - {DXGI_FORMAT_BC5_TYPELESS, VK_FORMAT_BC5_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC5_UNORM, VK_FORMAT_BC5_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC5_SNORM, VK_FORMAT_BC5_SNORM_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC6H_UF16, VK_FORMAT_BC6H_UFLOAT_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC6H_SF16, VK_FORMAT_BC6H_SFLOAT_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC7_UNORM, VK_FORMAT_BC7_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, - {DXGI_FORMAT_BC7_UNORM_SRGB, VK_FORMAT_BC7_SRGB_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_R32G32B32A32_TYPELESS, VK_FORMAT_R32G32B32A32_SFLOAT, 16, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R32G32B32A32_FLOAT, VK_FORMAT_R32G32B32A32_SFLOAT, 16, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R32G32B32A32_UINT, VK_FORMAT_R32G32B32A32_UINT, 16, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R32G32B32A32_SINT, VK_FORMAT_R32G32B32A32_SINT, 16, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R32G32B32_FLOAT, VK_FORMAT_R32G32B32_SFLOAT, 12, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R32G32_FLOAT, VK_FORMAT_R32G32_SFLOAT, 8, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R16G16B16A16_TYPELESS, VK_FORMAT_R16G16B16A16_SFLOAT, 8, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R16G16B16A16_FLOAT, VK_FORMAT_R16G16B16A16_SFLOAT, 8, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R16G16B16A16_UINT, VK_FORMAT_R16G16B16A16_UINT, 8, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R10G10B10A2_UNORM, VK_FORMAT_A2B10G10R10_UNORM_PACK32, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R11G11B10_FLOAT, VK_FORMAT_B10G11R11_UFLOAT_PACK32, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R8G8_UNORM, VK_FORMAT_R8G8_UNORM, 2, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R8G8B8A8_TYPELESS, VK_FORMAT_R8G8B8A8_UNORM, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R8G8B8A8_UNORM, VK_FORMAT_R8G8B8A8_UNORM, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R8G8B8A8_UNORM_SRGB, VK_FORMAT_R8G8B8A8_SRGB, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R8G8B8A8_UINT, VK_FORMAT_R8G8B8A8_UINT, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R16G16_TYPELESS, VK_FORMAT_R16G16_SFLOAT, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R16G16_FLOAT, VK_FORMAT_R16G16_SFLOAT, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R16G16_UNORM, VK_FORMAT_R16G16_UNORM, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R32_TYPELESS, VK_FORMAT_R32_UINT, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_D32_FLOAT, VK_FORMAT_D32_SFLOAT, 4, 1, 1, 1, DEPTH}, + {DXGI_FORMAT_R32_FLOAT, VK_FORMAT_R32_SFLOAT, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R32_UINT, VK_FORMAT_R32_UINT, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R32_SINT, VK_FORMAT_R32_SINT, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R16_TYPELESS, VK_FORMAT_R16_UINT, 2, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R16_FLOAT, VK_FORMAT_R16_SFLOAT, 2, 1, 1, 1, COLOR}, + {DXGI_FORMAT_D16_UNORM, VK_FORMAT_D16_UNORM, 2, 1, 1, 1, DEPTH}, + {DXGI_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, 2, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, 1, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R8_UINT, VK_FORMAT_R8_UINT, 1, 1, 1, 1, COLOR}, + {DXGI_FORMAT_R8_SNORM, VK_FORMAT_R8_SNORM, 1, 1, 1, 1, COLOR}, + {DXGI_FORMAT_A8_UNORM, VK_FORMAT_R8_UNORM, 1, 1, 1, 1, COLOR}, + {DXGI_FORMAT_B8G8R8A8_UNORM, VK_FORMAT_B8G8R8A8_UNORM, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_B8G8R8A8_TYPELESS, VK_FORMAT_B8G8R8A8_UNORM, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_B8G8R8A8_UNORM_SRGB, VK_FORMAT_B8G8R8A8_SRGB, 4, 1, 1, 1, COLOR}, + {DXGI_FORMAT_BC1_TYPELESS, VK_FORMAT_BC1_RGBA_UNORM_BLOCK, 1, 4, 4, 8, COLOR}, + {DXGI_FORMAT_BC1_UNORM, VK_FORMAT_BC1_RGBA_UNORM_BLOCK, 1, 4, 4, 8, COLOR}, + {DXGI_FORMAT_BC1_UNORM_SRGB, VK_FORMAT_BC1_RGBA_SRGB_BLOCK, 1, 4, 4, 8, COLOR}, + {DXGI_FORMAT_BC2_TYPELESS, VK_FORMAT_BC2_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC2_UNORM, VK_FORMAT_BC2_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC2_UNORM_SRGB, VK_FORMAT_BC2_SRGB_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC3_TYPELESS, VK_FORMAT_BC3_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC3_UNORM, VK_FORMAT_BC3_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC3_UNORM_SRGB, VK_FORMAT_BC3_SRGB_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC4_TYPELESS, VK_FORMAT_BC4_UNORM_BLOCK, 1, 4, 4, 8, COLOR}, + {DXGI_FORMAT_BC4_UNORM, VK_FORMAT_BC4_UNORM_BLOCK, 1, 4, 4, 8, COLOR}, + {DXGI_FORMAT_BC4_SNORM, VK_FORMAT_BC4_SNORM_BLOCK, 1, 4, 4, 8, COLOR}, + {DXGI_FORMAT_BC5_TYPELESS, VK_FORMAT_BC5_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC5_UNORM, VK_FORMAT_BC5_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC5_SNORM, VK_FORMAT_BC5_SNORM_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC6H_UF16, VK_FORMAT_BC6H_UFLOAT_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC6H_SF16, VK_FORMAT_BC6H_SFLOAT_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC7_UNORM, VK_FORMAT_BC7_UNORM_BLOCK, 1, 4, 4, 16, COLOR}, + {DXGI_FORMAT_BC7_UNORM_SRGB, VK_FORMAT_BC7_SRGB_BLOCK, 1, 4, 4, 16, COLOR}, }; /* Each depth/stencil format is only compatible with itself in Vulkan. */ -- 2.16.4

Aug. 1, 2018

[PATCH v2 vkd3d 01/14] tests: Add test for creating heaps.

by Józef Kucia

From: Józef Kucia <jkucia(a)codeweavers.com> Signed-off-by: Józef Kucia <jkucia(a)codeweavers.com> --- Version 2: Use "line" in ok() calls. --- tests/d3d12.c | 173 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 173 insertions(+) diff --git a/tests/d3d12.c b/tests/d3d12.c index 83dc8f903c81..608ff68c3af2 100644 --- a/tests/d3d12.c +++ b/tests/d3d12.c @@ -157,6 +157,53 @@ static void check_interface_(unsigned int line, IUnknown *iface, REFIID riid, bo IUnknown_Release(unk); } +#define check_heap_properties(a, b) check_heap_properties_(__LINE__, a, b) +static void check_heap_properties_(unsigned int line, + const D3D12_HEAP_PROPERTIES *properties, const D3D12_HEAP_PROPERTIES *expected_properties) +{ + D3D12_HEAP_PROPERTIES expected = *expected_properties; + + if (!expected.CreationNodeMask) + expected.CreationNodeMask = 0x1; + if (!expected.VisibleNodeMask) + expected.VisibleNodeMask = 0x1; + + ok_(line)(properties->Type == expected.Type, + "Got type %#x, expected %#x.\n", properties->Type, expected.Type); + ok_(line)(properties->CPUPageProperty == expected.CPUPageProperty, + "Got CPU page properties %#x, expected %#x.\n", + properties->CPUPageProperty, expected.CPUPageProperty); + ok_(line)(properties->MemoryPoolPreference == expected.MemoryPoolPreference, + "Got memory pool %#x, expected %#x.\n", + properties->MemoryPoolPreference, expected.MemoryPoolPreference); + ok_(line)(properties->CreationNodeMask == expected.CreationNodeMask, + "Got creation node mask %#x, expected %#x.\n", + properties->CreationNodeMask, expected.CreationNodeMask); + ok_(line)(properties->VisibleNodeMask == expected.VisibleNodeMask, + "Got visible node mask %#x, expected %#x.\n", + properties->VisibleNodeMask, expected.VisibleNodeMask); +} + +#define check_heap_desc(a, b) check_heap_desc_(__LINE__, a, b) +static void check_heap_desc_(unsigned int line, const D3D12_HEAP_DESC *desc, + const D3D12_HEAP_DESC *expected_desc) +{ + D3D12_HEAP_DESC expected = *expected_desc; + + if (!expected.Alignment) + expected.Alignment = D3D12_DEFAULT_RESOURCE_PLACEMENT_ALIGNMENT; + + ok_(line)(desc->SizeInBytes == expected.SizeInBytes, + "Got size %"PRIu64", expected %"PRIu64".\n", + desc->SizeInBytes, expected.SizeInBytes); + check_heap_properties_(line, &desc->Properties, &expected.Properties); + ok_(line)(desc->Alignment == expected.Alignment, + "Got alignment %"PRIu64", expected %"PRIu64".\n", + desc->Alignment, expected.Alignment); + ok_(line)(desc->Flags == expected.Flags, + "Got flags %#x, expected %#x.\n", desc->Flags, expected.Flags); +} + static void uav_barrier(ID3D12GraphicsCommandList *list, ID3D12Resource *resource) { D3D12_RESOURCE_BARRIER barrier; @@ -1731,6 +1778,131 @@ static void test_create_committed_resource(void) ok(!refcount, "ID3D12Device has %u references left.\n", (unsigned int)refcount); } +static void test_create_heap(void) +{ + D3D12_FEATURE_DATA_D3D12_OPTIONS options; + D3D12_HEAP_DESC desc, result_desc; + ID3D12Device *device, *tmp_device; + unsigned int i, j; + ID3D12Heap *heap; + ULONG refcount; + HRESULT hr; + + static const struct + { + UINT64 alignment; + HRESULT expected_hr; + } + tests[] = + { + {D3D12_DEFAULT_MSAA_RESOURCE_PLACEMENT_ALIGNMENT, S_OK}, + {D3D12_DEFAULT_RESOURCE_PLACEMENT_ALIGNMENT, S_OK}, + {2 * D3D12_DEFAULT_MSAA_RESOURCE_PLACEMENT_ALIGNMENT, E_INVALIDARG}, + {2 * D3D12_DEFAULT_RESOURCE_PLACEMENT_ALIGNMENT, E_INVALIDARG}, + {D3D12_SMALL_RESOURCE_PLACEMENT_ALIGNMENT, E_INVALIDARG}, + }; + static const struct + { + D3D12_HEAP_FLAGS flags; + const char *name; + } + heap_flags[] = + { + {D3D12_HEAP_FLAG_ALLOW_ONLY_BUFFERS, "buffers"}, + {D3D12_HEAP_FLAG_ALLOW_ONLY_NON_RT_DS_TEXTURES, "textures"}, + {D3D12_HEAP_FLAG_ALLOW_ONLY_RT_DS_TEXTURES, "rt_ds_textures"}, + }; + + if (!(device = create_device())) + { + skip("Failed to create device.\n"); + return; + } + + desc.SizeInBytes = D3D12_DEFAULT_RESOURCE_PLACEMENT_ALIGNMENT; + memset(&desc.Properties, 0, sizeof(desc.Properties)); + desc.Properties.Type = D3D12_HEAP_TYPE_DEFAULT; + desc.Alignment = 0; + desc.Flags = D3D12_HEAP_FLAG_ALLOW_ONLY_NON_RT_DS_TEXTURES; + hr = ID3D12Device_CreateHeap(device, &desc, &IID_ID3D12Heap, (void **)&heap); + ok(hr == S_OK, "Failed to create heap, hr %#x.\n", hr); + + refcount = get_refcount(device); + ok(refcount == 2, "Got unexpected refcount %u.\n", (unsigned int)refcount); + hr = ID3D12Heap_GetDevice(heap, &IID_ID3D12Device, (void **)&tmp_device); + ok(hr == S_OK, "Failed to get device, hr %#x.\n", hr); + refcount = get_refcount(device); + ok(refcount == 3, "Got unexpected refcount %u.\n", (unsigned int)refcount); + refcount = ID3D12Device_Release(tmp_device); + ok(refcount == 2, "Got unexpected refcount %u.\n", (unsigned int)refcount); + + check_interface(heap, &IID_ID3D12Object, TRUE); + check_interface(heap, &IID_ID3D12DeviceChild, TRUE); + check_interface(heap, &IID_ID3D12Pageable, TRUE); + check_interface(heap, &IID_ID3D12Heap, TRUE); + + result_desc = ID3D12Heap_GetDesc(heap); + check_heap_desc(&result_desc, &desc); + + refcount = ID3D12Heap_Release(heap); + ok(!refcount, "ID3D12Heap has %u references left.\n", (unsigned int)refcount); + + desc.SizeInBytes = 0; + hr = ID3D12Device_CreateHeap(device, &desc, &IID_ID3D12Heap, (void **)&heap); + ok(hr == E_INVALIDARG, "Got unexpected hr %#x.\n", hr); + + desc.SizeInBytes = D3D12_DEFAULT_RESOURCE_PLACEMENT_ALIGNMENT; + desc.Flags = D3D12_HEAP_FLAG_ALLOW_ONLY_NON_RT_DS_TEXTURES | D3D12_HEAP_FLAG_ALLOW_DISPLAY; + hr = ID3D12Device_CreateHeap(device, &desc, &IID_ID3D12Heap, (void **)&heap); + ok(hr == E_INVALIDARG, "Got unexpected hr %#x.\n", hr); + desc.Flags = D3D12_HEAP_FLAG_ALLOW_ONLY_RT_DS_TEXTURES | D3D12_HEAP_FLAG_ALLOW_DISPLAY; + hr = ID3D12Device_CreateHeap(device, &desc, &IID_ID3D12Heap, (void **)&heap); + ok(hr == E_INVALIDARG, "Got unexpected hr %#x.\n", hr); + + for (i = 0; i < ARRAY_SIZE(tests); ++i) + { + for (j = 0; j < ARRAY_SIZE(heap_flags); ++j) + { + desc.SizeInBytes = 10 * tests[i].alignment; + desc.Alignment = tests[i].alignment; + desc.Flags = heap_flags[j].flags; + hr = ID3D12Device_CreateHeap(device, &desc, &IID_ID3D12Heap, (void **)&heap); + ok(hr == tests[i].expected_hr, "Test %u, %s: Got hr %#x, expected %#x.\n", + i, heap_flags[j].name, hr, tests[i].expected_hr); + if (FAILED(hr)) + continue; + + result_desc = ID3D12Heap_GetDesc(heap); + check_heap_desc(&result_desc, &desc); + + refcount = ID3D12Heap_Release(heap); + ok(!refcount, "ID3D12Heap has %u references left.\n", (unsigned int)refcount); + } + } + + hr = ID3D12Device_CheckFeatureSupport(device, D3D12_FEATURE_D3D12_OPTIONS, &options, sizeof(options)); + ok(hr == S_OK, "Failed to check feature support, hr %#x.\n", hr); + if (options.ResourceHeapTier < D3D12_RESOURCE_HEAP_TIER_2) + { + skip("Resource heap tier %u.\n", options.ResourceHeapTier); + goto done; + } + + desc.SizeInBytes = D3D12_DEFAULT_RESOURCE_PLACEMENT_ALIGNMENT; + desc.Alignment = D3D12_DEFAULT_RESOURCE_PLACEMENT_ALIGNMENT; + desc.Flags = D3D12_HEAP_FLAG_ALLOW_ALL_BUFFERS_AND_TEXTURES; + hr = ID3D12Device_CreateHeap(device, &desc, &IID_ID3D12Heap, (void **)&heap); + ok(hr == S_OK, "Got unexpected hr %#x.\n", hr); + result_desc = ID3D12Heap_GetDesc(heap); + check_heap_desc(&result_desc, &desc); + refcount = ID3D12Heap_Release(heap); + ok(!refcount, "ID3D12Heap has %u references left.\n", (unsigned int)refcount); + +done: + refcount = ID3D12Device_Release(device); + ok(!refcount, "ID3D12Device has %u references left.\n", (unsigned int)refcount); +} + static void test_create_descriptor_heap(void) { D3D12_DESCRIPTOR_HEAP_DESC heap_desc; @@ -17652,6 +17824,7 @@ START_TEST(d3d12) run_test(test_create_command_list); run_test(test_create_command_queue); run_test(test_create_committed_resource); + run_test(test_create_heap); run_test(test_create_descriptor_heap); run_test(test_create_sampler); run_test(test_create_unordered_access_view); -- 2.16.4

Aug. 1, 2018

Re: [PATCH 11/13 v2] msvcp90: Add implementation of _Concurrent_vector_Internal_compact.

by Piotr Caban

Hi Hua, On 07/31/18 17:18, Hua Meng wrote: > @@ -1997,12 +2004,40 @@ MSVCP_size_t __thiscall _Concurrent_vector_base_v4__Internal_clear( > /* ?_Internal_compact(a)_Concurrent_vector_base_v4@details(a)Concurrency@@IEAAPEAX_KPEAXP6AX10(a)ZP6AX1PEBX0@Z(a)Z */ > DEFINE_THISCALL_WRAPPER(_Concurrent_vector_base_v4__Internal_compact, 20) > void * __thiscall _Concurrent_vector_base_v4__Internal_compact( > - _Concurrent_vector_base_v4 *this, MSVCP_size_t len, void *v, > + _Concurrent_vector_base_v4 *this, MSVCP_size_t element_size, void *v, > void (__cdecl *clear)(void*, MSVCP_size_t), > void (__cdecl *copy)(void*, const void*, MSVCP_size_t)) > { > - FIXME("(%p %ld %p %p %p) stub\n", this, len, v, clear, copy); > - return NULL; > + compact_block *b; > + MSVCP_size_t size, seg_no, copy_element, clear_element; > + int i; > + > + TRACE("(%p %ld %p %p %p)\n", this, element_size, v, clear, copy); > + > + size = this->early_size; > + seg_no = _vector_base_v4__Segment_index_of(size - 1); It will not work if size==0. > + if(this->first_block == seg_no + 1) return v; The vector still needs to be compacted when there is unused space allocated. You should compare this->first_block with index of last allocated segment. The function is also returning NULL in this case. Thanks, Piotr

Aug. 1, 2018

Re: [PATCH 04/13 v2] msvcp90: Add implementaion of _Concurrent_vector_Internal_reserve.

by Piotr Caban

Hi Hua, On 07/31/18 17:16, Hua Meng wrote: > +/* ?_Internal_reserve(a)_Concurrent_vector_base_v4@details(a)Concurrency@@IAEXIII(a)Z */ > +/* ?_Internal_reserve(a)_Concurrent_vector_base_v4@details(a)Concurrency@@IEAAX_K00(a)Z */ > +DEFINE_THISCALL_WRAPPER(_Concurrent_vector_base_v4__Internal_reserve, 16) > +void __thiscall _Concurrent_vector_base_v4__Internal_reserve( > + _Concurrent_vector_base_v4 *this, MSVCP_size_t size, > + MSVCP_size_t element_size, MSVCP_size_t max_size) > +{ > + MSVCP_size_t block_idx, capacity; > + int i; > + void **new_segment; > + > + TRACE("(%p %ld %ld %ld)\n", this, size, element_size, max_size); > + > + if(size > max_size) _vector_base_v4__Internal_throw_exception(this, 0); > + capacity = _Concurrent_vector_base_v4__Internal_capacity(this); > + if(size <= capacity) return; > + block_idx = _vector_base_v4__Segment_index_of(size - 1); > + if(!this->first_block) > + InterlockedCompareExchangeSizeT(&this->first_block, block_idx + 1, 0); > + i = _vector_base_v4__Segment_index_of(capacity); > + if(this->storage == this->segment) { > + for(; i <= block_idx && i < STORAGE_SIZE; i++) > + concurrent_vector_alloc_segment(this, i, element_size); > + if(block_idx >= STORAGE_SIZE) { > + new_segment = malloc(SEGMENT_SIZE * sizeof(void*)); > + if(new_segment == NULL) _vector_base_v4__Internal_throw_exception(this, 2); > + memset(new_segment, 0, SEGMENT_SIZE * sizeof(void*)); Please change it to: memset(new_segment, 0, SEGMENT_SIZE * sizeof(*new_segment)); > + memcpy(new_segment, this->storage, STORAGE_SIZE * element_size); Copied data size is wrong here. There should be: memcpy(new_segment, this->storage, STORAGE_SIZE * sizeof(*new_segment)); > + if(InterlockedCompareExchangePointer((void*)&this->segment, new_segment, > + this->storage) != this->storage) > + free(new_segment); > + } > + } > + for(; i <= block_idx; i++) > + concurrent_vector_alloc_segment(this, i, element_size); > +} Thanks, Piotr

Aug. 1, 2018

Re: [PATCH] kernel32: Set environment variable %PUBLIC% on the process start-up.

by Alistair Leslie-Hughes

Hi Dmitry, On 01/08/18 13:12, Dmitry Timoshkov wrote: > > Yes, I'm aware of that difference. However since Wine already has the code > path to set ALLUSERSPROFILE=C:\Users\Public I decided to reuse it path to > set PUBLIC=C:\Users\Public. Changing ALLUSERSPROFILE to point to another > place is not the subject of my patch. > Thanks. In other words, a bug with ALLUSERSPROFILE ;) Alistair.

Aug. 1, 2018

Re: [PATCH] kernel32: Set environment variable %PUBLIC% on the process start-up.

by Dmitry Timoshkov

Alistair Leslie-Hughes <leslie_alistair(a)hotmail.com> wrote: > > This patch fixes an installer that has "[%PUBLIC]\Documents" entry > > for some of the targets being copied from a cabinet file. Under > > Windows %PUBLIC% environment variable is set to C:\Users\Public. > > On windows, these are set to different values. > > ALLUSERSPROFILE=C:\ProgramData > > PUBLIC=C:\Users\Public > > > Why have you made them the same? Yes, I'm aware of that difference. However since Wine already has the code path to set ALLUSERSPROFILE=C:\Users\Public I decided to reuse it path to set PUBLIC=C:\Users\Public. Changing ALLUSERSPROFILE to point to another place is not the subject of my patch. -- Dmitry.

Aug. 1, 2018