tint: Preserve padding in matrices with three rows

The PreservePadding transform now decomposes writes to matrices with
three rows into separate column vector writes, to avoid modifying
padding between columns.

Bug: tint:1571
Change-Id: If575f79bb87f52810783fd3338e2f3ce3228ab2e
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/121600
Auto-Submit: James Price <jrprice@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: James Price <jrprice@google.com>
This commit is contained in:
James Price
2023-02-27 16:06:54 +00:00
committed by Dawn LUCI CQ
parent fe19fee3ea
commit 6176c85be8
84 changed files with 3733 additions and 1968 deletions

View File

@@ -16,6 +16,22 @@ layout(binding = 1, std430) buffer u_block_ssbo {
f16mat4x3 inner[4];
} s;
void assign_and_preserve_padding_1_s_X(uint dest[1], f16mat4x3 value) {
s.inner[dest[0]][0] = value[0u];
s.inner[dest[0]][1] = value[1u];
s.inner[dest[0]][2] = value[2u];
s.inner[dest[0]][3] = value[3u];
}
void assign_and_preserve_padding_s(f16mat4x3 value[4]) {
{
for(uint i = 0u; (i < 4u); i = (i + 1u)) {
uint tint_symbol[1] = uint[1](i);
assign_and_preserve_padding_1_s_X(tint_symbol, value[i]);
}
}
}
f16mat4x3 conv_mat4x3_f16(mat4x3_f16 val) {
return f16mat4x3(val.col0, val.col1, val.col2, val.col3);
}
@@ -31,8 +47,9 @@ f16mat4x3[4] conv_arr4_mat4x3_f16(mat4x3_f16 val[4]) {
}
void f() {
s.inner = conv_arr4_mat4x3_f16(u.inner);
s.inner[1] = conv_mat4x3_f16(u.inner[2u]);
assign_and_preserve_padding_s(conv_arr4_mat4x3_f16(u.inner));
uint tint_symbol_1[1] = uint[1](1u);
assign_and_preserve_padding_1_s_X(tint_symbol_1, conv_mat4x3_f16(u.inner[2u]));
s.inner[1][0] = u.inner[0u].col1.zxy;
s.inner[1][0].x = u.inner[0u].col1[0u];
}

View File

@@ -14,9 +14,22 @@ struct tint_array {
T elements[N];
};
void assign_and_preserve_padding_1(device half4x3* const dest, half4x3 value) {
(*(dest))[0u] = value[0u];
(*(dest))[1u] = value[1u];
(*(dest))[2u] = value[2u];
(*(dest))[3u] = value[3u];
}
void assign_and_preserve_padding(device tint_array<half4x3, 4>* const dest, tint_array<half4x3, 4> value) {
for(uint i = 0u; (i < 4u); i = (i + 1u)) {
assign_and_preserve_padding_1(&((*(dest))[i]), value[i]);
}
}
kernel void f(device tint_array<half4x3, 4>* tint_symbol [[buffer(1)]], const constant tint_array<half4x3, 4>* tint_symbol_1 [[buffer(0)]]) {
*(tint_symbol) = *(tint_symbol_1);
(*(tint_symbol))[1] = (*(tint_symbol_1))[2];
assign_and_preserve_padding(tint_symbol, *(tint_symbol_1));
assign_and_preserve_padding_1(&((*(tint_symbol))[1]), (*(tint_symbol_1))[2]);
(*(tint_symbol))[1][0] = half3((*(tint_symbol_1))[0][1]).zxy;
(*(tint_symbol))[1][0][0] = (*(tint_symbol_1))[0][1][0];
return;

View File

@@ -1,7 +1,7 @@
; SPIR-V
; Version: 1.3
; Generator: Google Tint Compiler; 0
; Bound: 91
; Bound: 134
; Schema: 0
OpCapability Shader
OpCapability Float16
@@ -22,13 +22,20 @@
OpName %u_block "u_block"
OpMemberName %u_block 0 "inner"
OpName %s "s"
OpName %assign_and_preserve_padding_1_s_X "assign_and_preserve_padding_1_s_X"
OpName %dest "dest"
OpName %value "value"
OpName %assign_and_preserve_padding_s "assign_and_preserve_padding_s"
OpName %value_0 "value"
OpName %i "i"
OpName %var_for_index "var_for_index"
OpName %conv_mat4x3_f16 "conv_mat4x3_f16"
OpName %val "val"
OpName %conv_arr4_mat4x3_f16 "conv_arr4_mat4x3_f16"
OpName %val_0 "val"
OpName %arr "arr"
OpName %i "i"
OpName %var_for_index "var_for_index"
OpName %i_0 "i"
OpName %var_for_index_1 "var_for_index_1"
OpName %f "f"
OpDecorate %u_block_std140 Block
OpMemberDecorate %u_block_std140 0 Offset 0
@@ -47,6 +54,7 @@
OpDecorate %_arr_mat4v3half_uint_4 ArrayStride 32
OpDecorate %s DescriptorSet 0
OpDecorate %s Binding 1
OpDecorate %_arr_uint_uint_1 ArrayStride 4
%half = OpTypeFloat 16
%v3half = OpTypeVector %half 3
%mat4x3_f16 = OpTypeStruct %v3half %v3half %v3half %v3half
@@ -61,100 +69,160 @@
%u_block = OpTypeStruct %_arr_mat4v3half_uint_4
%_ptr_StorageBuffer_u_block = OpTypePointer StorageBuffer %u_block
%s = OpVariable %_ptr_StorageBuffer_u_block StorageBuffer
%15 = OpTypeFunction %mat4v3half %mat4x3_f16
%24 = OpTypeFunction %_arr_mat4v3half_uint_4 %_arr_mat4x3_f16_uint_4
%_ptr_Function__arr_mat4v3half_uint_4 = OpTypePointer Function %_arr_mat4v3half_uint_4
%30 = OpConstantNull %_arr_mat4v3half_uint_4
%_ptr_Function_uint = OpTypePointer Function %uint
%33 = OpConstantNull %uint
%bool = OpTypeBool
%_ptr_Function__arr_mat4x3_f16_uint_4 = OpTypePointer Function %_arr_mat4x3_f16_uint_4
%46 = OpConstantNull %_arr_mat4x3_f16_uint_4
%_ptr_Function_mat4v3half = OpTypePointer Function %mat4v3half
%_ptr_Function_mat4x3_f16 = OpTypePointer Function %mat4x3_f16
%uint_1 = OpConstant %uint 1
%void = OpTypeVoid
%59 = OpTypeFunction %void
%uint_1 = OpConstant %uint 1
%_arr_uint_uint_1 = OpTypeArray %uint %uint_1
%15 = OpTypeFunction %void %_arr_uint_uint_1 %mat4v3half
%uint_0 = OpConstant %uint 0
%_ptr_StorageBuffer__arr_mat4v3half_uint_4 = OpTypePointer StorageBuffer %_arr_mat4v3half_uint_4
%_ptr_Uniform__arr_mat4x3_f16_uint_4 = OpTypePointer Uniform %_arr_mat4x3_f16_uint_4
%int = OpTypeInt 32 1
%int_1 = OpConstant %int 1
%_ptr_StorageBuffer_mat4v3half = OpTypePointer StorageBuffer %mat4v3half
%uint_2 = OpConstant %uint 2
%_ptr_Uniform_mat4x3_f16 = OpTypePointer Uniform %mat4x3_f16
%79 = OpConstantNull %int
%25 = OpConstantNull %int
%_ptr_StorageBuffer_v3half = OpTypePointer StorageBuffer %v3half
%29 = OpConstantNull %uint
%int_1 = OpConstant %int 1
%int_2 = OpConstant %int 2
%uint_2 = OpConstant %uint 2
%int_3 = OpConstant %int 3
%uint_3 = OpConstant %uint 3
%45 = OpTypeFunction %void %_arr_mat4v3half_uint_4
%_ptr_Function_uint = OpTypePointer Function %uint
%bool = OpTypeBool
%_ptr_Function__arr_mat4v3half_uint_4 = OpTypePointer Function %_arr_mat4v3half_uint_4
%63 = OpConstantNull %_arr_mat4v3half_uint_4
%_ptr_Function_mat4v3half = OpTypePointer Function %mat4v3half
%73 = OpTypeFunction %mat4v3half %mat4x3_f16
%82 = OpTypeFunction %_arr_mat4v3half_uint_4 %_arr_mat4x3_f16_uint_4
%_ptr_Function__arr_mat4x3_f16_uint_4 = OpTypePointer Function %_arr_mat4x3_f16_uint_4
%99 = OpConstantNull %_arr_mat4x3_f16_uint_4
%_ptr_Function_mat4x3_f16 = OpTypePointer Function %mat4x3_f16
%110 = OpTypeFunction %void
%_ptr_Uniform__arr_mat4x3_f16_uint_4 = OpTypePointer Uniform %_arr_mat4x3_f16_uint_4
%119 = OpConstantComposite %_arr_uint_uint_1 %uint_1
%_ptr_Uniform_mat4x3_f16 = OpTypePointer Uniform %mat4x3_f16
%_ptr_Uniform_v3half = OpTypePointer Uniform %v3half
%_ptr_StorageBuffer_half = OpTypePointer StorageBuffer %half
%_ptr_Uniform_half = OpTypePointer Uniform %half
%conv_mat4x3_f16 = OpFunction %mat4v3half None %15
%val = OpFunctionParameter %mat4x3_f16
%18 = OpLabel
%19 = OpCompositeExtract %v3half %val 0
%20 = OpCompositeExtract %v3half %val 1
%21 = OpCompositeExtract %v3half %val 2
%22 = OpCompositeExtract %v3half %val 3
%23 = OpCompositeConstruct %mat4v3half %19 %20 %21 %22
OpReturnValue %23
OpFunctionEnd
%conv_arr4_mat4x3_f16 = OpFunction %_arr_mat4v3half_uint_4 None %24
%val_0 = OpFunctionParameter %_arr_mat4x3_f16_uint_4
%27 = OpLabel
%arr = OpVariable %_ptr_Function__arr_mat4v3half_uint_4 Function %30
%i = OpVariable %_ptr_Function_uint Function %33
%var_for_index = OpVariable %_ptr_Function__arr_mat4x3_f16_uint_4 Function %46
OpBranch %34
%34 = OpLabel
OpLoopMerge %35 %36 None
OpBranch %37
%37 = OpLabel
%39 = OpLoad %uint %i
%40 = OpULessThan %bool %39 %uint_4
%38 = OpLogicalNot %bool %40
OpSelectionMerge %42 None
OpBranchConditional %38 %43 %42
%43 = OpLabel
OpBranch %35
%42 = OpLabel
OpStore %var_for_index %val_0
%47 = OpLoad %uint %i
%49 = OpAccessChain %_ptr_Function_mat4v3half %arr %47
%51 = OpLoad %uint %i
%53 = OpAccessChain %_ptr_Function_mat4x3_f16 %var_for_index %51
%54 = OpLoad %mat4x3_f16 %53
%50 = OpFunctionCall %mat4v3half %conv_mat4x3_f16 %54
OpStore %49 %50
OpBranch %36
%36 = OpLabel
%55 = OpLoad %uint %i
%57 = OpIAdd %uint %55 %uint_1
OpStore %i %57
OpBranch %34
%35 = OpLabel
%58 = OpLoad %_arr_mat4v3half_uint_4 %arr
OpReturnValue %58
OpFunctionEnd
%f = OpFunction %void None %59
%62 = OpLabel
%65 = OpAccessChain %_ptr_StorageBuffer__arr_mat4v3half_uint_4 %s %uint_0
%68 = OpAccessChain %_ptr_Uniform__arr_mat4x3_f16_uint_4 %u %uint_0
%69 = OpLoad %_arr_mat4x3_f16_uint_4 %68
%66 = OpFunctionCall %_arr_mat4v3half_uint_4 %conv_arr4_mat4x3_f16 %69
OpStore %65 %66
%73 = OpAccessChain %_ptr_StorageBuffer_mat4v3half %s %uint_0 %int_1
%77 = OpAccessChain %_ptr_Uniform_mat4x3_f16 %u %uint_0 %uint_2
%78 = OpLoad %mat4x3_f16 %77
%74 = OpFunctionCall %mat4v3half %conv_mat4x3_f16 %78
OpStore %73 %74
%81 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %int_1 %79
%83 = OpAccessChain %_ptr_Uniform_v3half %u %uint_0 %33 %uint_1
%84 = OpLoad %v3half %83
%85 = OpVectorShuffle %v3half %84 %84 2 0 1
OpStore %81 %85
%87 = OpAccessChain %_ptr_StorageBuffer_half %s %uint_0 %int_1 %79 %uint_0
%89 = OpAccessChain %_ptr_Uniform_half %u %uint_0 %33 %uint_1 %33
%90 = OpLoad %half %89
OpStore %87 %90
%assign_and_preserve_padding_1_s_X = OpFunction %void None %15
%dest = OpFunctionParameter %_arr_uint_uint_1
%value = OpFunctionParameter %mat4v3half
%22 = OpLabel
%26 = OpCompositeExtract %uint %dest 0
%28 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %26 %25
%30 = OpCompositeExtract %v3half %value 0
OpStore %28 %30
%31 = OpCompositeExtract %uint %dest 0
%33 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %31 %int_1
%34 = OpCompositeExtract %v3half %value 1
OpStore %33 %34
%35 = OpCompositeExtract %uint %dest 0
%37 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %35 %int_2
%39 = OpCompositeExtract %v3half %value 2
OpStore %37 %39
%40 = OpCompositeExtract %uint %dest 0
%42 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %40 %int_3
%44 = OpCompositeExtract %v3half %value 3
OpStore %42 %44
OpReturn
OpFunctionEnd
%assign_and_preserve_padding_s = OpFunction %void None %45
%value_0 = OpFunctionParameter %_arr_mat4v3half_uint_4
%48 = OpLabel
%i = OpVariable %_ptr_Function_uint Function %29
%var_for_index = OpVariable %_ptr_Function__arr_mat4v3half_uint_4 Function %63
OpStore %i %29
OpBranch %51
%51 = OpLabel
OpLoopMerge %52 %53 None
OpBranch %54
%54 = OpLabel
%56 = OpLoad %uint %i
%57 = OpULessThan %bool %56 %uint_4
%55 = OpLogicalNot %bool %57
OpSelectionMerge %59 None
OpBranchConditional %55 %60 %59
%60 = OpLabel
OpBranch %52
%59 = OpLabel
OpStore %var_for_index %value_0
%65 = OpLoad %uint %i
%66 = OpCompositeConstruct %_arr_uint_uint_1 %65
%67 = OpLoad %uint %i
%69 = OpAccessChain %_ptr_Function_mat4v3half %var_for_index %67
%70 = OpLoad %mat4v3half %69
%64 = OpFunctionCall %void %assign_and_preserve_padding_1_s_X %66 %70
OpBranch %53
%53 = OpLabel
%71 = OpLoad %uint %i
%72 = OpIAdd %uint %71 %uint_1
OpStore %i %72
OpBranch %51
%52 = OpLabel
OpReturn
OpFunctionEnd
%conv_mat4x3_f16 = OpFunction %mat4v3half None %73
%val = OpFunctionParameter %mat4x3_f16
%76 = OpLabel
%77 = OpCompositeExtract %v3half %val 0
%78 = OpCompositeExtract %v3half %val 1
%79 = OpCompositeExtract %v3half %val 2
%80 = OpCompositeExtract %v3half %val 3
%81 = OpCompositeConstruct %mat4v3half %77 %78 %79 %80
OpReturnValue %81
OpFunctionEnd
%conv_arr4_mat4x3_f16 = OpFunction %_arr_mat4v3half_uint_4 None %82
%val_0 = OpFunctionParameter %_arr_mat4x3_f16_uint_4
%85 = OpLabel
%arr = OpVariable %_ptr_Function__arr_mat4v3half_uint_4 Function %63
%i_0 = OpVariable %_ptr_Function_uint Function %29
%var_for_index_1 = OpVariable %_ptr_Function__arr_mat4x3_f16_uint_4 Function %99
OpBranch %88
%88 = OpLabel
OpLoopMerge %89 %90 None
OpBranch %91
%91 = OpLabel
%93 = OpLoad %uint %i_0
%94 = OpULessThan %bool %93 %uint_4
%92 = OpLogicalNot %bool %94
OpSelectionMerge %95 None
OpBranchConditional %92 %96 %95
%96 = OpLabel
OpBranch %89
%95 = OpLabel
OpStore %var_for_index_1 %val_0
%100 = OpLoad %uint %i_0
%101 = OpAccessChain %_ptr_Function_mat4v3half %arr %100
%103 = OpLoad %uint %i_0
%105 = OpAccessChain %_ptr_Function_mat4x3_f16 %var_for_index_1 %103
%106 = OpLoad %mat4x3_f16 %105
%102 = OpFunctionCall %mat4v3half %conv_mat4x3_f16 %106
OpStore %101 %102
OpBranch %90
%90 = OpLabel
%107 = OpLoad %uint %i_0
%108 = OpIAdd %uint %107 %uint_1
OpStore %i_0 %108
OpBranch %88
%89 = OpLabel
%109 = OpLoad %_arr_mat4v3half_uint_4 %arr
OpReturnValue %109
OpFunctionEnd
%f = OpFunction %void None %110
%112 = OpLabel
%116 = OpAccessChain %_ptr_Uniform__arr_mat4x3_f16_uint_4 %u %uint_0
%117 = OpLoad %_arr_mat4x3_f16_uint_4 %116
%114 = OpFunctionCall %_arr_mat4v3half_uint_4 %conv_arr4_mat4x3_f16 %117
%113 = OpFunctionCall %void %assign_and_preserve_padding_s %114
%122 = OpAccessChain %_ptr_Uniform_mat4x3_f16 %u %uint_0 %uint_2
%123 = OpLoad %mat4x3_f16 %122
%120 = OpFunctionCall %mat4v3half %conv_mat4x3_f16 %123
%118 = OpFunctionCall %void %assign_and_preserve_padding_1_s_X %119 %120
%124 = OpAccessChain %_ptr_StorageBuffer_v3half %s %uint_0 %int_1 %25
%126 = OpAccessChain %_ptr_Uniform_v3half %u %uint_0 %29 %uint_1
%127 = OpLoad %v3half %126
%128 = OpVectorShuffle %v3half %127 %127 2 0 1
OpStore %124 %128
%130 = OpAccessChain %_ptr_StorageBuffer_half %s %uint_0 %int_1 %25 %uint_0
%132 = OpAccessChain %_ptr_Uniform_half %u %uint_0 %29 %uint_1 %29
%133 = OpLoad %half %132
OpStore %130 %133
OpReturn
OpFunctionEnd