diff options
| author | Samuel Pitoiset <samuel.pitoiset@gmail.com> | 2020-02-07 12:53:31 +0100 |
|---|---|---|
| committer | Samuel Pitoiset <samuel.pitoiset@gmail.com> | 2020-02-07 14:17:06 +0100 |
| commit | 4b978cd950cef844afce07993ddb697779e5648d (patch) | |
| tree | 91855bcfb739bcee95de102019efad0d851a3a7e /src/amd | |
| parent | da76dfb5159c2ca8ee24d64a5f85a68f28b70c65 (diff) | |
aco: do not use ds_{read,write}2 on GFX6
According to LLVM, these instructions have a bounds checking bug.
LLVM only uses them on GFX7+.
This fixes broken geometry in Assassins Creed Origins.
Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/2489
Fixes: 4a553212fa1 ("radv: enable ACO support for GFX6")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3746>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3746>
Diffstat (limited to 'src/amd')
| -rw-r--r-- | src/amd/compiler/aco_instruction_selection.cpp | 10 |
1 files changed, 6 insertions, 4 deletions
diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp index d885f79bd63..24be495dbb5 100644 --- a/src/amd/compiler/aco_instruction_selection.cpp +++ b/src/amd/compiler/aco_instruction_selection.cpp @@ -2657,6 +2657,7 @@ void load_lds(isel_context *ctx, unsigned elem_size_bytes, Temp dst, unsigned total_bytes = num_components * elem_size_bytes; std::array<Temp, NIR_MAX_VEC_COMPONENTS> result; bool large_ds_read = ctx->options->chip_class >= GFX7; + bool usable_read2 = ctx->options->chip_class >= GFX7; while (bytes_read < total_bytes) { unsigned todo = total_bytes - bytes_read; @@ -2668,7 +2669,7 @@ void load_lds(isel_context *ctx, unsigned elem_size_bytes, Temp dst, if (todo >= 16 && aligned16 && large_ds_read) { op = aco_opcode::ds_read_b128; todo = 16; - } else if (todo >= 16 && aligned8) { + } else if (todo >= 16 && aligned8 && usable_read2) { op = aco_opcode::ds_read2_b64; read2 = true; todo = 16; @@ -2678,7 +2679,7 @@ void load_lds(isel_context *ctx, unsigned elem_size_bytes, Temp dst, } else if (todo >= 8 && aligned8) { op = aco_opcode::ds_read_b64; todo = 8; - } else if (todo >= 8) { + } else if (todo >= 8 && usable_read2) { op = aco_opcode::ds_read2_b32; read2 = true; todo = 8; @@ -2783,6 +2784,7 @@ void ds_write_helper(isel_context *ctx, Operand m, Temp address, Temp data, unsi Builder bld(ctx->program, ctx->block); unsigned bytes_written = 0; bool large_ds_write = ctx->options->chip_class >= GFX7; + bool usable_write2 = ctx->options->chip_class >= GFX7; while (bytes_written < total_size * 4) { unsigned todo = total_size * 4 - bytes_written; @@ -2795,7 +2797,7 @@ void ds_write_helper(isel_context *ctx, Operand m, Temp address, Temp data, unsi if (todo >= 16 && aligned16 && large_ds_write) { op = aco_opcode::ds_write_b128; size = 4; - } else if (todo >= 16 && aligned8) { + } else if (todo >= 16 && aligned8 && usable_write2) { op = aco_opcode::ds_write2_b64; write2 = true; size = 4; @@ -2805,7 +2807,7 @@ void ds_write_helper(isel_context *ctx, Operand m, Temp address, Temp data, unsi } else if (todo >= 8 && aligned8) { op = aco_opcode::ds_write_b64; size = 2; - } else if (todo >= 8) { + } else if (todo >= 8 && usable_write2) { op = aco_opcode::ds_write2_b32; write2 = true; size = 2; |