[Mesa-dev] [RFC PATCH 4/4] radv: add support for push constants inlining when possible

Samuel Pitoiset samuel.pitoiset at gmail.com
Tue Feb 5 10:44:04 UTC 2019


On 2/5/19 11:29 AM, Bas Nieuwenhuizen wrote:
> On Tue, Feb 5, 2019 at 11:07 AM Samuel Pitoiset
> <samuel.pitoiset at gmail.com> wrote:
>>
>> On 2/5/19 10:58 AM, Bas Nieuwenhuizen wrote:
>>> On Fri, Jan 25, 2019 at 5:27 PM Samuel Pitoiset
>>> <samuel.pitoiset at gmail.com> wrote:
>>>> This removes some scalar loads from shaders, but it increases
>>>> the number of SET_SH_REG packets. This is currently basic but
>>>> it could be improved if needed. Inlining dynamic offsets might
>>>> also help.
>>>>
>>>> Original idea from Dave Airlie.
>>>>
>>>> 29164 shaders in 15096 tests
>>>> Totals:
>>>> SGPRS: 1336072 -> 1365241 (2.18 %)
>>>> VGPRS: 937784 -> 934592 (-0.34 %)
>>>> Spilled SGPRs: 24751 -> 24796 (0.18 %)
>>>> Code Size: 50001672 -> 49815524 (-0.37 %) bytes
>>>> Max Waves: 208755 -> 208830 (0.04 %)
>>>>
>>>> Totals from affected shaders:
>>>> SGPRS: 295018 -> 324187 (9.89 %)
>>>> VGPRS: 243108 -> 239916 (-1.31 %)
>>>> Spilled SGPRs: 1464 -> 1509 (3.07 %)
>>>> Code Size: 8028188 -> 7842040 (-2.32 %) bytes
>>>> Max Waves: 69580 -> 69655 (0.11 %)
>>>>
>>>> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
>>>> ---
>>>>    src/amd/common/ac_nir_to_llvm.c   | 23 +++++++--
>>>>    src/amd/common/ac_shader_abi.h    |  4 ++
>>>>    src/amd/vulkan/radv_cmd_buffer.c  | 78 ++++++++++++++++++++++---------
>>>>    src/amd/vulkan/radv_nir_to_llvm.c | 54 +++++++++++++++++++++
>>>>    src/amd/vulkan/radv_shader.h      | 10 ++--
>>>>    src/amd/vulkan/radv_shader_info.c |  4 ++
>>>>    6 files changed, 145 insertions(+), 28 deletions(-)
>>>>
>>>> diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
>>>> index f509fc31dff..db1574b5b35 100644
>>>> --- a/src/amd/common/ac_nir_to_llvm.c
>>>> +++ b/src/amd/common/ac_nir_to_llvm.c
>>>> @@ -1392,10 +1392,27 @@ static LLVMValueRef visit_load_push_constant(struct ac_nir_context *ctx,
>>>>                                                 nir_intrinsic_instr *instr)
>>>>    {
>>>>           LLVMValueRef ptr, addr;
>>>> +       LLVMValueRef src0 = get_src(ctx, instr->src[0]);
>>>> +       unsigned index = nir_intrinsic_base(instr);
>>>>
>>>> -       addr = LLVMConstInt(ctx->ac.i32, nir_intrinsic_base(instr), 0);
>>>> -       addr = LLVMBuildAdd(ctx->ac.builder, addr,
>>>> -                           get_src(ctx, instr->src[0]), "");
>>>> +       addr = LLVMConstInt(ctx->ac.i32, index, 0);
>>>> +       addr = LLVMBuildAdd(ctx->ac.builder, addr, src0, "");
>>>> +
>>>> +       /* Load constant values from user SGPRS when possible, otherwise
>>>> +        * fallback to the default path that loads directly from memory.
>>>> +        */
>>>> +       if (LLVMIsConstant(src0) &&
>>>> +           index == 0 &&
>>>> +           instr->dest.ssa.bit_size == 32) {
>>>> +               unsigned offset = LLVMConstIntGetZExtValue(src0) / 4;
>>>> +               unsigned count = instr->dest.ssa.num_components;
>>>> +
>>>> +               if (offset + count <= ctx->abi->num_inline_push_consts) {
>>>> +                       return ac_build_gather_values(&ctx->ac,
>>>> +                                                     ctx->abi->inline_push_consts + offset,
>>>> +                                                     count);
>>>> +               }
>>>> +       }
>>>>
>>>>           ptr = ac_build_gep0(&ctx->ac, ctx->abi->push_constants, addr);
>>>>
>>>> diff --git a/src/amd/common/ac_shader_abi.h b/src/amd/common/ac_shader_abi.h
>>>> index ee18e6c1923..704c3d107c2 100644
>>>> --- a/src/amd/common/ac_shader_abi.h
>>>> +++ b/src/amd/common/ac_shader_abi.h
>>>> @@ -32,6 +32,8 @@ struct nir_variable;
>>>>
>>>>    #define AC_LLVM_MAX_OUTPUTS (VARYING_SLOT_VAR31 + 1)
>>>>
>>>> +#define AC_MAX_INLINE_PUSH_CONSTS 8
>>>> +
>>>>    enum ac_descriptor_type {
>>>>           AC_DESC_IMAGE,
>>>>           AC_DESC_FMASK,
>>>> @@ -66,6 +68,8 @@ struct ac_shader_abi {
>>>>
>>>>           /* Vulkan only */
>>>>           LLVMValueRef push_constants;
>>>> +       LLVMValueRef inline_push_consts[AC_MAX_INLINE_PUSH_CONSTS];
>>>> +       unsigned num_inline_push_consts;
>>>>           LLVMValueRef view_index;
>>>>
>>>>           LLVMValueRef outputs[AC_LLVM_MAX_OUTPUTS * 4];
>>>> diff --git a/src/amd/vulkan/radv_cmd_buffer.c b/src/amd/vulkan/radv_cmd_buffer.c
>>>> index aae90290841..f80e2078da0 100644
>>>> --- a/src/amd/vulkan/radv_cmd_buffer.c
>>>> +++ b/src/amd/vulkan/radv_cmd_buffer.c
>>>> @@ -627,6 +627,23 @@ radv_emit_descriptor_pointers(struct radv_cmd_buffer *cmd_buffer,
>>>>           }
>>>>    }
>>>>
>>>> +static void
>>>> +radv_emit_inline_push_consts(struct radv_cmd_buffer *cmd_buffer,
>>>> +                            struct radv_pipeline *pipeline,
>>>> +                            gl_shader_stage stage,
>>>> +                            int idx, int count, uint32_t *values)
>>>> +{
>>>> +       struct radv_userdata_info *loc = radv_lookup_user_sgpr(pipeline, stage, idx);
>>>> +       uint32_t base_reg = pipeline->user_data_0[stage];
>>>> +       if (loc->sgpr_idx == -1)
>>>> +               return;
>>>> +
>>>> +       assert(loc->num_sgprs == count);
>>>> +
>>>> +       radeon_set_sh_reg_seq(cmd_buffer->cs, base_reg + loc->sgpr_idx * 4, count);
>>>> +       radeon_emit_array(cmd_buffer->cs, values, count);
>>>> +}
>>>> +
>>>>    static void
>>>>    radv_update_multisample_state(struct radv_cmd_buffer *cmd_buffer,
>>>>                                 struct radv_pipeline *pipeline)
>>>> @@ -1900,6 +1917,7 @@ radv_flush_constants(struct radv_cmd_buffer *cmd_buffer,
>>>>                   radv_get_descriptors_state(cmd_buffer, bind_point);
>>>>           struct radv_pipeline_layout *layout = pipeline->layout;
>>>>           struct radv_shader_variant *shader, *prev_shader;
>>>> +       bool need_push_constants = false;
>>>>           unsigned offset;
>>>>           void *ptr;
>>>>           uint64_t va;
>>>> @@ -1909,37 +1927,55 @@ radv_flush_constants(struct radv_cmd_buffer *cmd_buffer,
>>>>               (!layout->push_constant_size && !layout->dynamic_offset_count))
>>>>                   return;
>>>>
>>>> -       if (!radv_cmd_buffer_upload_alloc(cmd_buffer, layout->push_constant_size +
>>>> -                                         16 * layout->dynamic_offset_count,
>>>> -                                         256, &offset, &ptr))
>>>> -               return;
>>>> +       radv_foreach_stage(stage, stages) {
>>>> +               if (!pipeline->shaders[stage])
>>>> +                       continue;
>>>>
>>>> -       memcpy(ptr, cmd_buffer->push_constants, layout->push_constant_size);
>>>> -       memcpy((char*)ptr + layout->push_constant_size,
>>>> -              descriptors_state->dynamic_buffers,
>>>> -              16 * layout->dynamic_offset_count);
>>>> +               need_push_constants |= pipeline->shaders[stage]->info.info.loads_push_constants;
>>>> +               need_push_constants |= pipeline->shaders[stage]->info.info.loads_dynamic_offsets;
>>>>
>>>> -       va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
>>>> -       va += offset;
>>>> +               uint8_t count = pipeline->shaders[stage]->info.info.num_inline_push_consts;
>>>>
>>>> -       MAYBE_UNUSED unsigned cdw_max = radeon_check_space(cmd_buffer->device->ws,
>>>> -                                                          cmd_buffer->cs, MESA_SHADER_STAGES * 4);
>>>> +               radv_emit_inline_push_consts(cmd_buffer, pipeline, stage,
>>>> +                                            AC_UD_INLINE_PUSH_CONSTANTS,
>>>> +                                            count,
>>>> +                                            (uint32_t *)&cmd_buffer->push_constants[0]);
>>>> +       }
>>>>
>>>> -       prev_shader = NULL;
>>>> -       radv_foreach_stage(stage, stages) {
>>>> -               shader = radv_get_shader(pipeline, stage);
>>>> +       if (need_push_constants) {
>>>> +               if (!radv_cmd_buffer_upload_alloc(cmd_buffer, layout->push_constant_size +
>>>> +                                                 16 * layout->dynamic_offset_count,
>>>> +                                                 256, &offset, &ptr))
>>>> +                       return;
>>>> +
>>>> +               memcpy(ptr, cmd_buffer->push_constants, layout->push_constant_size);
>>>> +               memcpy((char*)ptr + layout->push_constant_size,
>>>> +                      descriptors_state->dynamic_buffers,
>>>> +                      16 * layout->dynamic_offset_count);
>>>> +
>>>> +               va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
>>>> +               va += offset;
>>>> +
>>>> +               MAYBE_UNUSED unsigned cdw_max =
>>>> +                       radeon_check_space(cmd_buffer->device->ws,
>>>> +                                          cmd_buffer->cs, MESA_SHADER_STAGES * 4);
>>>> +
>>>> +               prev_shader = NULL;
>>>> +               radv_foreach_stage(stage, stages) {
>>>> +                       shader = radv_get_shader(pipeline, stage);
>>>>
>>>> -               /* Avoid redundantly emitting the address for merged stages. */
>>>> -               if (shader && shader != prev_shader) {
>>>> -                       radv_emit_userdata_address(cmd_buffer, pipeline, stage,
>>>> -                                                  AC_UD_PUSH_CONSTANTS, va);
>>>> +                       /* Avoid redundantly emitting the address for merged stages. */
>>>> +                       if (shader && shader != prev_shader) {
>>>> +                               radv_emit_userdata_address(cmd_buffer, pipeline, stage,
>>>> +                                                          AC_UD_PUSH_CONSTANTS, va);
>>>>
>>>> -                       prev_shader = shader;
>>>> +                               prev_shader = shader;
>>>> +                       }
>>>>                   }
>>>> +               assert(cmd_buffer->cs->cdw <= cdw_max);
>>>>           }
>>>>
>>>>           cmd_buffer->push_constant_stages &= ~stages;
>>>> -       assert(cmd_buffer->cs->cdw <= cdw_max);
>>>>    }
>>>>
>>>>    static void
>>>> diff --git a/src/amd/vulkan/radv_nir_to_llvm.c b/src/amd/vulkan/radv_nir_to_llvm.c
>>>> index 11417c5991b..1cffa748d15 100644
>>>> --- a/src/amd/vulkan/radv_nir_to_llvm.c
>>>> +++ b/src/amd/vulkan/radv_nir_to_llvm.c
>>>> @@ -627,6 +627,47 @@ count_vs_user_sgprs(struct radv_shader_context *ctx)
>>>>           return count;
>>>>    }
>>>>
>>>> +static void allocate_inline_push_consts(struct radv_shader_context *ctx,
>>>> +                                       struct user_sgpr_info *user_sgpr_info)
>>>> +{
>>>> +       uint8_t remaining_sgprs = user_sgpr_info->remaining_sgprs;
>>>> +
>>>> +       /* Only supported if shaders don't have indirect push constants. */
>>>> +       if (ctx->shader_info->info.has_indirect_push_constants)
>>>> +               return;
>>>> +
>>>> +       /* Only supported for 32-bit push constants. */
>>>> +       if (!ctx->shader_info->info.has_32bit_push_constants)
>>>> +               return;
>>>> +
>>>> +       /* Only supported if base starts from 0. */
>>>> +       if (ctx->shader_info->info.min_push_constant_used != 0)
>>>> +               return;
>>>> +
>>>> +       uint8_t num_push_consts =
>>>> +               (ctx->shader_info->info.max_push_constant_used -
>>>> +                ctx->shader_info->info.min_push_constant_used) / 4;
>>> I think you need +1? if min_push_constant_used ==
>>> max_push_constant_used you still use 1 push constant. (The code in the
>>> info pass adds range, but that does not include the size of the loaded
>>> thing?)
>> No, as min is 'base' and max is 'base + range'.
>>
>>> Now that I think about it, the shader info pass does not take into
>>> account the size of the variable at all right? What happens with a
>>> vec4? The range does not include the object itself right?
>> If it's a vec4 starting from base=0, range should be 16, that works for
>> vec4.
> hmm, so it includes the size of type. If I read the spir-v parser
> correctly we set range very large though (looks like range is always
> the total size of all push constants?)

Yes, it is. This implementation doesn't yet support holes. If the range 
is too large, the driver won't inline.

I assume this is something that can be improved.

>
>>>> +
>>>> +       /* Check if the number of user SGPRs is large enough. */
>>>> +       if (num_push_consts < remaining_sgprs) {
>>>> +               ctx->shader_info->info.num_inline_push_consts = num_push_consts;
>>>> +       } else {
>>>> +               ctx->shader_info->info.num_inline_push_consts = remaining_sgprs;
>>>> +       }
>>>> +
>>>> +       /* Clamp to the maximum number of allowed inlined push constants. */
>>>> +       if (ctx->shader_info->info.num_inline_push_consts > AC_MAX_INLINE_PUSH_CONSTS)
>>>> +               ctx->shader_info->info.num_inline_push_consts = AC_MAX_INLINE_PUSH_CONSTS;
>>>> +
>>>> +       if (ctx->shader_info->info.num_inline_push_consts == num_push_consts &&
>>>> +           !ctx->shader_info->info.loads_dynamic_offsets) {
>>> This won't work, as we can reach here even if there are some accesses
>>> which don't use inline push constants, e.g. some non 32-bit accesses.
>> Good catch! One possible fix is to change 'has_32bit_push_constants' to
>> 'has_only_32bit_push_constants' and bail out if it's false.
>>
>> What do you think?
> That works.
>
>>>> +               /* Disable the default push constants path if all constants are
>>>> +                * inlined and if shaders don't use dynamic descriptors.
>>>> +                */
>>>> +               ctx->shader_info->info.loads_push_constants = false;
>>>> +       }
>>>> +}
>>>> +
>>>>    static void allocate_user_sgprs(struct radv_shader_context *ctx,
>>>>                                   gl_shader_stage stage,
>>>>                                   bool has_previous_stage,
>>>> @@ -706,6 +747,8 @@ static void allocate_user_sgprs(struct radv_shader_context *ctx,
>>>>           } else {
>>>>                   user_sgpr_info->remaining_sgprs = remaining_sgprs - num_desc_set;
>>>>           }
>>>> +
>>>> +       allocate_inline_push_consts(ctx, user_sgpr_info);
>>>>    }
>>>>
>>>>    static void
>>>> @@ -735,6 +778,12 @@ declare_global_input_sgprs(struct radv_shader_context *ctx,
>>>>                   add_arg(args, ARG_SGPR, type, &ctx->abi.push_constants);
>>>>           }
>>>>
>>>> +       for (unsigned i = 0; i < ctx->shader_info->info.num_inline_push_consts; i++) {
>>>> +               add_arg(args, ARG_SGPR, ctx->ac.i32,
>>>> +                       &ctx->abi.inline_push_consts[i]);
>>>> +       }
>>>> +       ctx->abi.num_inline_push_consts = ctx->shader_info->info.num_inline_push_consts;
>>>> +
>>>>           if (ctx->shader_info->info.so.num_outputs) {
>>>>                   add_arg(args, ARG_SGPR,
>>>>                           ac_array_in_const32_addr_space(ctx->ac.v4i32),
>>>> @@ -853,6 +902,11 @@ set_global_input_locs(struct radv_shader_context *ctx,
>>>>                   set_loc_shader_ptr(ctx, AC_UD_PUSH_CONSTANTS, user_sgpr_idx);
>>>>           }
>>>>
>>>> +       if (ctx->shader_info->info.num_inline_push_consts) {
>>>> +               set_loc_shader(ctx, AC_UD_INLINE_PUSH_CONSTANTS, user_sgpr_idx,
>>>> +                              ctx->shader_info->info.num_inline_push_consts);
>>>> +       }
>>>> +
>>>>           if (ctx->streamout_buffers) {
>>>>                   set_loc_shader_ptr(ctx, AC_UD_STREAMOUT_BUFFERS,
>>>>                                  user_sgpr_idx);
>>>> diff --git a/src/amd/vulkan/radv_shader.h b/src/amd/vulkan/radv_shader.h
>>>> index 09bc5c2d4a9..d1bc467c5d9 100644
>>>> --- a/src/amd/vulkan/radv_shader.h
>>>> +++ b/src/amd/vulkan/radv_shader.h
>>>> @@ -129,10 +129,11 @@ struct radv_nir_compiler_options {
>>>>    enum radv_ud_index {
>>>>           AC_UD_SCRATCH_RING_OFFSETS = 0,
>>>>           AC_UD_PUSH_CONSTANTS = 1,
>>>> -       AC_UD_INDIRECT_DESCRIPTOR_SETS = 2,
>>>> -       AC_UD_VIEW_INDEX = 3,
>>>> -       AC_UD_STREAMOUT_BUFFERS = 4,
>>>> -       AC_UD_SHADER_START = 5,
>>>> +       AC_UD_INLINE_PUSH_CONSTANTS = 2,
>>>> +       AC_UD_INDIRECT_DESCRIPTOR_SETS = 3,
>>>> +       AC_UD_VIEW_INDEX = 4,
>>>> +       AC_UD_STREAMOUT_BUFFERS = 5,
>>>> +       AC_UD_SHADER_START = 6,
>>>>           AC_UD_VS_VERTEX_BUFFERS = AC_UD_SHADER_START,
>>>>           AC_UD_VS_BASE_VERTEX_START_INSTANCE,
>>>>           AC_UD_VS_MAX_UD,
>>>> @@ -167,6 +168,7 @@ struct radv_shader_info {
>>>>           uint8_t max_push_constant_used;
>>>>           bool has_32bit_push_constants;
>>>>           bool has_indirect_push_constants;
>>>> +       uint8_t num_inline_push_consts;
>>>>           uint32_t desc_set_used_mask;
>>>>           bool needs_multiview_view_index;
>>>>           bool uses_invocation_id;
>>>> diff --git a/src/amd/vulkan/radv_shader_info.c b/src/amd/vulkan/radv_shader_info.c
>>>> index cabe2a470a0..900bca1d097 100644
>>>> --- a/src/amd/vulkan/radv_shader_info.c
>>>> +++ b/src/amd/vulkan/radv_shader_info.c
>>>> @@ -211,6 +211,10 @@ gather_push_constant_info(const nir_shader *nir,
>>>>           if (base < info->min_push_constant_used)
>>>>                   info->min_push_constant_used = base;
>>>>
>>>> +       /* TODO: handle base not 0. */
>>>> +       if (base != 0)
>>>> +               info->min_push_constant_used = -1;
>>>> +
>>>>           info->loads_push_constants = true;
>>>>    }
>>>>
>>>> --
>>>> 2.20.1
>>>>
>>>> _______________________________________________
>>>> mesa-dev mailing list
>>>> mesa-dev at lists.freedesktop.org
>>>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev


More information about the mesa-dev mailing list