From: Józef Kucia jkucia@codeweavers.com
Signed-off-by: Józef Kucia jkucia@codeweavers.com --- libs/vkd3d-shader/spirv.c | 11 +++++++---- tests/d3d12.c | 28 ++++++++++++++++++++++++++++ 2 files changed, 35 insertions(+), 4 deletions(-)
diff --git a/libs/vkd3d-shader/spirv.c b/libs/vkd3d-shader/spirv.c index 1d97a98c2e80..9c51df14316e 100644 --- a/libs/vkd3d-shader/spirv.c +++ b/libs/vkd3d-shader/spirv.c @@ -4732,15 +4732,16 @@ static void vkd3d_dxbc_compiler_emit_f16tof32(struct vkd3d_dxbc_compiler *compil struct vkd3d_spirv_builder *builder = &compiler->spirv_builder; const struct vkd3d_shader_dst_param *dst = instruction->dst; const struct vkd3d_shader_src_param *src = instruction->src; + uint32_t components[VKD3D_VEC4_SIZE]; + unsigned int i, j; DWORD write_mask; - unsigned int i;
instr_set_id = vkd3d_spirv_get_glsl_std450_instr_set(builder); type_id = vkd3d_spirv_get_type_id(builder, VKD3D_TYPE_FLOAT, 2); scalar_type_id = vkd3d_spirv_get_type_id(builder, VKD3D_TYPE_FLOAT, 1);
/* FIXME: Consider a single UnpackHalf2x16 intruction per 2 components. */ - for (i = 0; i < VKD3D_VEC4_SIZE; ++i) + for (i = 0, j = 0; i < VKD3D_VEC4_SIZE; ++i) { if (!(write_mask = dst->write_mask & (VKD3DSP_WRITEMASK_0 << i))) continue; @@ -4748,10 +4749,12 @@ static void vkd3d_dxbc_compiler_emit_f16tof32(struct vkd3d_dxbc_compiler *compil src_id = vkd3d_dxbc_compiler_emit_load_src(compiler, src, write_mask); result_id = vkd3d_spirv_build_op_ext_inst(builder, type_id, instr_set_id, GLSLstd450UnpackHalf2x16, &src_id, 1); - result_id = vkd3d_spirv_build_op_composite_extract1(builder, + components[j++] = vkd3d_spirv_build_op_composite_extract1(builder, scalar_type_id, result_id, 0); - vkd3d_dxbc_compiler_emit_store_reg(compiler, &dst->reg, write_mask, result_id); } + + vkd3d_dxbc_compiler_emit_store_dst_components(compiler, + dst, vkd3d_component_type_from_data_type(dst->reg.data_type), components); }
static void vkd3d_dxbc_compiler_emit_f32tof16(struct vkd3d_dxbc_compiler *compiler, diff --git a/tests/d3d12.c b/tests/d3d12.c index 34c4802f7fde..0ef382f46103 100644 --- a/tests/d3d12.c +++ b/tests/d3d12.c @@ -6567,6 +6567,29 @@ static void test_shader_instructions(void) 0x0500001c, 0x001020f2, 0x00000000, 0x00100e46, 0x00000000, 0x0100003e, }; static const D3D12_SHADER_BYTECODE ps_f16tof32 = {ps_f16tof32_code, sizeof(ps_f16tof32_code)}; + static const DWORD ps_f16tof32_2_code[] = + { +#if 0 + ps_5_0 + dcl_globalFlags refactoringAllowed + dcl_constantbuffer cb0[1], immediateIndexed + dcl_output o0.xyzw + dcl_temps 1 + mov r0.xyzw, cb0[0].xyzw + f16tof32 r0.xyzw, r0.wzyx + ftou o0.xyzw, r0.xyzw + ret +#endif + 0x43425844, 0x38472f03, 0x2c49b7dd, 0xc2d76bbf, 0xfc093e1d, 0x00000001, 0x000000ec, 0x00000003, + 0x0000002c, 0x0000003c, 0x00000070, 0x4e475349, 0x00000008, 0x00000000, 0x00000008, 0x4e47534f, + 0x0000002c, 0x00000001, 0x00000008, 0x00000020, 0x00000000, 0x00000000, 0x00000001, 0x00000000, + 0x0000000f, 0x545f5653, 0x65677261, 0xabab0074, 0x58454853, 0x00000074, 0x00000050, 0x0000001d, + 0x0100086a, 0x04000059, 0x00208e46, 0x00000000, 0x00000001, 0x03000065, 0x001020f2, 0x00000000, + 0x02000068, 0x00000001, 0x06000036, 0x001000f2, 0x00000000, 0x00208e46, 0x00000000, 0x00000000, + 0x05000083, 0x001000f2, 0x00000000, 0x001001b6, 0x00000000, 0x0500001c, 0x001020f2, 0x00000000, + 0x00100e46, 0x00000000, 0x0100003e, + }; + static const D3D12_SHADER_BYTECODE ps_f16tof32_2 = {ps_f16tof32_2_code, sizeof(ps_f16tof32_2_code)}; static const DWORD ps_f32tof16_code[] = { #if 0 @@ -7325,6 +7348,11 @@ static void test_shader_instructions(void) {&ps_f16tof32, {{{0x000f0000, 0x000f3c00, 0x000f5640, 0x000f5bd0}}}, {{0, 1, 100, 250}}}, {&ps_f16tof32, {{{0xffff0000, 0xffff3c00, 0xffff5640, 0xffff5bd0}}}, {{0, 1, 100, 250}}},
+ {&ps_f16tof32_2, {{{0x00000000, 0x00003c00, 0x00005640, 0x00005bd0}}}, {{250, 100, 1, 0}}}, + {&ps_f16tof32_2, {{{0x00010000, 0x00013c00, 0x00015640, 0x00015bd0}}}, {{250, 100, 1, 0}}}, + {&ps_f16tof32_2, {{{0x000f0000, 0x000f3c00, 0x000f5640, 0x000f5bd0}}}, {{250, 100, 1, 0}}}, + {&ps_f16tof32_2, {{{0xffff0000, 0xffff3c00, 0xffff5640, 0xffff5bd0}}}, {{250, 100, 1, 0}}}, + {&ps_f32tof16, {.f = {{0.0f, 1.0f, -1.0f, 666.0f}}}, {{0, 0x3c00, 0xbc00, 0x6134}}},
{&ps_imad, {{{0, 0, 0, 0}, {0, 0, 0, 0}, {0, 0, 0, 0}}}, {{ 0, 0, 0, 0}}},
-
Henri Verbeet -
Józef Kucia