[Mesa-dev] [PATCH 3/7] radv: add support for 16bit UBO loads

[Marek Olšák]

BTW, changes to src/amd shouldn't be prefixed with "radv", because the
code is shared with radeonsi and it may confuse people.

Marek

On Thu, Feb 8, 2018 at 4:05 PM, Marek Olšák <maraeo at gmail.com> wrote:
> On Wed, Feb 7, 2018 at 7:47 PM, Daniel Schürmann
> <daniel.schuermann at campus.tu-berlin.de> wrote:
>> Signed-off-by: Daniel Schürmann <daniel.schuermann at campus.tu-berlin.de>
>> ---
>>  src/amd/common/ac_llvm_build.c  | 25 +++++++++++++++++++++++++
>>  src/amd/common/ac_llvm_build.h  |  8 ++++++++
>>  src/amd/common/ac_nir_to_llvm.c | 14 ++++++++++++--
>>  3 files changed, 45 insertions(+), 2 deletions(-)
>>
>> diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
>> index d0b288f51d..197dc40eae 100644
>> --- a/src/amd/common/ac_llvm_build.c
>> +++ b/src/amd/common/ac_llvm_build.c
>> @@ -1046,6 +1046,31 @@ LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
>>                                            can_speculate, true);
>>  }
>>
>> +LLVMValueRef
>> +ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
>> +                           LLVMValueRef rsrc,
>> +                           LLVMValueRef vindex,
>> +                           LLVMValueRef voffset,
>> +                               LLVMValueRef soffset,
>> +                               LLVMValueRef immoffset)
>> +{
>> +       const char *name = "llvm.amdgcn.tbuffer.load.i32";
>> +       LLVMTypeRef type = ctx->i32;
>> +       LLVMValueRef params[] = {
>> +                               rsrc,
>> +                               vindex,
>> +                               voffset,
>> +                               soffset,
>> +                               immoffset,
>> +                               LLVMConstInt(ctx->i32, V_008F0C_BUF_DATA_FORMAT_16, false),
>> +                               LLVMConstInt(ctx->i32, V_008F0C_BUF_NUM_FORMAT_UINT, false),
>> +                               ctx->i1false,
>> +                               ctx->i1false,
>> +       };
>> +       LLVMValueRef res = ac_build_intrinsic(ctx, name, type, params, 9, 0);
>> +       return LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
>> +}
>> +
>>  /**
>>   * Set range metadata on an instruction.  This can only be used on load and
>>   * call instructions.  If you know an instruction can only produce the values
>> diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
>> index 47c843fb4b..da49b06502 100644
>> --- a/src/amd/common/ac_llvm_build.h
>> +++ b/src/amd/common/ac_llvm_build.h
>> @@ -225,6 +225,14 @@ LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
>>                                          bool glc,
>>                                          bool can_speculate);
>>
>> +LLVMValueRef
>> +ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
>> +                           LLVMValueRef rsrc,
>> +                           LLVMValueRef vindex,
>> +                           LLVMValueRef voffset,
>> +                               LLVMValueRef soffset,
>> +                               LLVMValueRef immoffset);
>> +
>>  LLVMValueRef
>>  ac_get_thread_id(struct ac_llvm_context *ctx);
>>
>> diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
>> index 31f16d3acf..2b48e92949 100644
>> --- a/src/amd/common/ac_nir_to_llvm.c
>> +++ b/src/amd/common/ac_nir_to_llvm.c
>> @@ -2697,9 +2697,19 @@ static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx,
>>         if (instr->dest.ssa.bit_size == 64)
>>                 num_components *= 2;
>>
>> -       ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset,
>> +       if (instr->dest.ssa.bit_size == 16) {
>> +               LLVMValueRef results[num_components];
>> +               for (unsigned i = 0; i < num_components; ++i) {
>> +                       results[i] = ac_build_tbuffer_load_short(&ctx->ac, rsrc,
>> +                                                               ctx->ac.i32_0, ctx->ac.i32_0,
>> +                                                               offset, LLVMConstInt(ctx->ac.i32, 2 * i, 0));
>> +               }
>
> For num_components >= 2, it would be better to load it as i32 and then
> bitcast to v2i16, assuming the offset is aligned to 4.
>
> Marek