[Mesa-dev] [PATCH v4 18/44] i965/fs: Add byte scattered write message and fs support

Chema Casanova jmcasanova at igalia.com
Fri Dec 1 21:21:55 UTC 2017


On 30/11/17 21:42, Jason Ekstrand wrote:
> On Wed, Nov 29, 2017 at 6:08 PM, Jose Maria Casanova Crespo
> <jmcasanova at igalia.com <mailto:jmcasanova at igalia.com>> wrote:
> 
>     v2: (Jason Ekstrand)
>         - Enable bit_size parameter to scattered messages to enable
>     different
>           bitsizes byte/word/dword.
>         - Remove use of brw_send_indirect_scattered_message in favor of
>           brw_send_indirect_surface_message.
>         - Move scattered messages to surface messages namespace.
>         - Assert align1 for scattered messages and assume Gen8+.
>         - Inline brw_set_dp_byte_scattered_write.
> 
>     Signed-off-by: Jose Maria Casanova Crespo <jmcasanova at igalia.com
>     <mailto:jmcasanova at igalia.com>>
>     Signed-off-by: Alejandro Piñeiro <apinheiro at igalia.com
>     <mailto:apinheiro at igalia.com>>
>     ---
>      src/intel/compiler/brw_eu.h                    |  7 +++++
>      src/intel/compiler/brw_eu_defines.h            | 17 +++++++++++
>      src/intel/compiler/brw_eu_emit.c               | 42
>     ++++++++++++++++++++++++++
>      src/intel/compiler/brw_fs.cpp                  | 14 +++++++++
>      src/intel/compiler/brw_fs_copy_propagation.cpp |  2 ++
>      src/intel/compiler/brw_fs_generator.cpp        |  6 ++++
>      src/intel/compiler/brw_fs_surface_builder.cpp  | 11 +++++++
>      src/intel/compiler/brw_fs_surface_builder.h    |  7 +++++
>      src/intel/compiler/brw_shader.cpp              |  7 +++++
>      9 files changed, 113 insertions(+)
> 
>     diff --git a/src/intel/compiler/brw_eu.h b/src/intel/compiler/brw_eu.h
>     index 343dcd867d..3ac3b4342a 100644
>     --- a/src/intel/compiler/brw_eu.h
>     +++ b/src/intel/compiler/brw_eu.h
>     @@ -485,6 +485,13 @@ brw_typed_surface_write(struct brw_codegen *p,
>                              unsigned msg_length,
>                              unsigned num_channels);
> 
>     +void
>     +brw_byte_scattered_write(struct brw_codegen *p,
>     +                         struct brw_reg payload,
>     +                         struct brw_reg surface,
>     +                         unsigned msg_length,
>     +                         unsigned bit_size);
>     +
>      void
>      brw_memory_fence(struct brw_codegen *p,
>                       struct brw_reg dst);
>     diff --git a/src/intel/compiler/brw_eu_defines.h
>     b/src/intel/compiler/brw_eu_defines.h
>     index 9d5cf05c86..de6330ee54 100644
>     --- a/src/intel/compiler/brw_eu_defines.h
>     +++ b/src/intel/compiler/brw_eu_defines.h
>     @@ -402,6 +402,16 @@ enum opcode {
> 
>         SHADER_OPCODE_RND_MODE,
> 
>     +   /**
>     +    * Byte scattered write/read opcodes.
>     +    *
>     +    * LOGICAL opcodes are eventually translated to the matching
>     non-LOGICAL
>     +    * opcode, but instead of taking a single payload blog they
>     expect their
>     +    * arguments separately as individual sources, like untyped
>     write/read.
>     +    */
>     +   SHADER_OPCODE_BYTE_SCATTERED_WRITE,
>     +   SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL,
>     +
>         SHADER_OPCODE_MEMORY_FENCE,
> 
>         SHADER_OPCODE_GEN4_SCRATCH_READ,
>     @@ -1255,4 +1265,11 @@ enum PACKED brw_rnd_mode {
>         BRW_RND_MODE_UNSPECIFIED,  /* Unspecified rounding mode */
>      };
> 
>     +/* MDC_DS - Data Size Message Descriptor Control Field */
>     +enum PACKED brw_data_size {
> 
> 
> I'm not sure how I feel about this being an enum with such a generic name.

Right, PRM use a more exactly "Data Elements" but this field only used
byte_scattered/scaled writes/reads. As I will follow your next
suggestion of using #define, I'm chaging the name to:

#define GEN7_BYTE_SCATTERED_DATA_ELEMENT_BYTE  0
#define GEN7_BYTE_SCATTERED_DATA_ELEMENT_WORD  1
#define GEN7_BYTE_SCATTERED_DATA_ELEMENT_DWORD 2

I'll include in the comment about MSC_DS
"Specifies the number of Bytes to be read or written per Dword used at
byte_scattered read/write and byte_scaled read/write messages."

>  
> 
>     +   GEN7_BYTE_SCATTERED_DATA_SIZE_BYTE = 0,
>     +   GEN7_BYTE_SCATTERED_DATA_SIZE_WORD = 1,
>     +   GEN7_BYTE_SCATTERED_DATA_SIZE_DWORD = 2
>     +};
>     +
>      #endif /* BRW_EU_DEFINES_H */
>     diff --git a/src/intel/compiler/brw_eu_emit.c
>     b/src/intel/compiler/brw_eu_emit.c
>     index ca97ff7325..ded7e228cf 100644
>     --- a/src/intel/compiler/brw_eu_emit.c
>     +++ b/src/intel/compiler/brw_eu_emit.c
>     @@ -2580,6 +2580,7 @@ brw_send_indirect_surface_message(struct
>     brw_codegen *p,
>         return insn;
>      }
> 
>     +
>      static bool
>      while_jumps_before_offset(const struct gen_device_info *devinfo,
>                                brw_inst *insn, int while_offset, int
>     start_offset)
>     @@ -2983,6 +2984,47 @@ brw_untyped_surface_write(struct brw_codegen *p,
>            p, insn, num_channels);
>      }
> 
>     +static enum brw_data_size brw_data_size_from_bit_size(unsigned
>     bit_size)
> 
> 
> Please put the return type on it's own line.

OK. It will be unsigned. and I'm changing the name of the function
according to brw_scattered_data_element_from_bit_size to maintain
coherence with "enum" values.


>     +{
>     +   switch (bit_size) {
>     +   case 8:
>     +      return GEN7_BYTE_SCATTERED_DATA_SIZE_BYTE;
>     +   case 16:
>     +      return GEN7_BYTE_SCATTERED_DATA_SIZE_WORD;
>     +   case 32:
>     +      return GEN7_BYTE_SCATTERED_DATA_SIZE_DWORD;
>     +   default:
>     +      unreachable("Unsupported bit_size for byte scattered messages");
>     +   }
>     +}
>     +
>     +void
>     +brw_byte_scattered_write(struct brw_codegen *p,
>     +                         struct brw_reg payload,
>     +                         struct brw_reg surface,
>     +                         unsigned msg_length,
>     +                         unsigned bit_size)
>     +{
>     +   assert(brw_inst_access_mode(p->devinfo, p->current) == BRW_ALIGN_1);
>     +   const struct gen_device_info *devinfo = p->devinfo;
>     +   const unsigned sfid = GEN7_SFID_DATAPORT_DATA_CACHE;
>     +
>     +   struct brw_inst *insn = brw_send_indirect_surface_message(
>     +      p, sfid, brw_writemask(brw_null_reg(), WRITEMASK_XYZW),
>     +      payload, surface, msg_length, 0, true);
>     +
>     +   unsigned msg_control = brw_data_size_from_bit_size(bit_size) << 2;
> 
> 
> I have no idea what C is going to do with types when you have a packed
> enum and you then provide a shift.  I think I would be more comfortable
> if you used #defines and had brw_data_size_from_bit_size return an unsigned.

If packed enum uses a char as type, it this emum has only 0,1,2 values
the shiff would fit in the char, and then it would be converted to
uint32. But I can not confirm what each different compiler would do. So
let's go with the simple choice and use #defines, what about changing
the name of the function brw_byte_scattered_data_element_from_bit_size.

>     +
>     +   if (brw_inst_exec_size(devinfo, p->current) == BRW_EXECUTE_16)
>     +      msg_control |= 1;
>     +   else
>     +      msg_control |= 0;
>     +
>     +   brw_inst_set_dp_msg_type(devinfo, insn,
>     +                           
>     HSW_DATAPORT_DC_PORT0_BYTE_SCATTERED_WRITE);
> 
> 
> Should we assert at the top that devinfo->gen > 7 || devinfo->is_haswell?

Yes, I thought that when asserting for ALIGN1 we were already limitting
that really. But yes, I'll include it.

>     +   brw_inst_set_dp_msg_control(devinfo, insn, msg_control);
>     +}
>     +
>      static void
>      brw_set_dp_typed_atomic_message(struct brw_codegen *p,
>                                      struct brw_inst *insn,
>     diff --git a/src/intel/compiler/brw_fs.cpp
>     b/src/intel/compiler/brw_fs.cpp
>     index 36fb337c62..32f1d757f0 100644
>     --- a/src/intel/compiler/brw_fs.cpp
>     +++ b/src/intel/compiler/brw_fs.cpp
>     @@ -250,6 +250,7 @@ fs_inst::is_send_from_grf() const
>         case SHADER_OPCODE_UNTYPED_ATOMIC:
>         case SHADER_OPCODE_UNTYPED_SURFACE_READ:
>         case SHADER_OPCODE_UNTYPED_SURFACE_WRITE:
>     +   case SHADER_OPCODE_BYTE_SCATTERED_WRITE:
>         case SHADER_OPCODE_TYPED_ATOMIC:
>         case SHADER_OPCODE_TYPED_SURFACE_READ:
>         case SHADER_OPCODE_TYPED_SURFACE_WRITE:
>     @@ -749,6 +750,11 @@ fs_inst::components_read(unsigned i) const
>            else
>               return 1;
> 
>     +   case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL:
>     +      assert(src[3].file == IMM &&
>     +             src[4].file == IMM);
>     +      return 1;
>     +
>         case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL:
>         case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL: {
>            assert(src[3].file == IMM &&
>     @@ -791,6 +797,7 @@ fs_inst::size_read(int arg) const
>         case SHADER_OPCODE_TYPED_SURFACE_READ:
>         case SHADER_OPCODE_TYPED_SURFACE_WRITE:
>         case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET:
>     +   case SHADER_OPCODE_BYTE_SCATTERED_WRITE:
>            if (arg == 0)
>               return mlen * REG_SIZE;
>            break;
>     @@ -4538,6 +4545,12 @@ fs_visitor::lower_logical_sends()
>                                          ibld.sample_mask_reg());
>               break;
> 
>     +      case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL:
>     +         lower_surface_logical_send(ibld, inst,
>     +                                    SHADER_OPCODE_BYTE_SCATTERED_WRITE,
>     +                                    ibld.sample_mask_reg());
>     +         break;
>     +
>            case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL:
>               lower_surface_logical_send(ibld, inst,
>                                          SHADER_OPCODE_UNTYPED_ATOMIC,
>     @@ -5022,6 +5035,7 @@ get_lowered_simd_width(const struct
>     gen_device_info *devinfo,
>         case SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL:
>         case SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL:
>         case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL:
>     +   case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL:
>            return MIN2(16, inst->exec_size);
> 
>         case SHADER_OPCODE_URB_READ_SIMD8:
>     diff --git a/src/intel/compiler/brw_fs_copy_propagation.cpp
>     b/src/intel/compiler/brw_fs_copy_propagation.cpp
>     index cb11739608..fcf4706b7a 100644
>     --- a/src/intel/compiler/brw_fs_copy_propagation.cpp
>     +++ b/src/intel/compiler/brw_fs_copy_propagation.cpp
>     @@ -655,6 +655,7 @@ fs_visitor::try_constant_propagate(fs_inst
>     *inst, acp_entry *entry)
>            case SHADER_OPCODE_TYPED_ATOMIC:
>            case SHADER_OPCODE_TYPED_SURFACE_READ:
>            case SHADER_OPCODE_TYPED_SURFACE_WRITE:
>     +      case SHADER_OPCODE_BYTE_SCATTERED_WRITE:
>               /* We only propagate into the surface argument of the
>                * instruction. Everything else goes through LOAD_PAYLOAD.
>                */
>     @@ -694,6 +695,7 @@ fs_visitor::try_constant_propagate(fs_inst
>     *inst, acp_entry *entry)
>            case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL:
>            case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL:
>            case SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL:
>     +      case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL:
>               inst->src[i] = val;
>               progress = true;
>               break;
>     diff --git a/src/intel/compiler/brw_fs_generator.cpp
>     b/src/intel/compiler/brw_fs_generator.cpp
>     index 1835c4bf72..fedc9acf97 100644
>     --- a/src/intel/compiler/brw_fs_generator.cpp
>     +++ b/src/intel/compiler/brw_fs_generator.cpp
>     @@ -2073,6 +2073,12 @@ fs_generator::generate_code(const cfg_t *cfg,
>     int dispatch_width)
>                                         inst->mlen, src[2].ud);
>               break;
> 
>     +      case SHADER_OPCODE_BYTE_SCATTERED_WRITE:
>     +         assert(src[2].file == BRW_IMMEDIATE_VALUE);
>     +         brw_byte_scattered_write(p, src[0], src[1],
>     +                                  inst->mlen, src[2].ud);
>     +         break;
>     +
>            case SHADER_OPCODE_TYPED_ATOMIC:
>               assert(src[2].file == BRW_IMMEDIATE_VALUE);
>               brw_typed_atomic(p, dst, src[0], src[1],
>     diff --git a/src/intel/compiler/brw_fs_surface_builder.cpp
>     b/src/intel/compiler/brw_fs_surface_builder.cpp
>     index d00d8920b2..37cc29e361 100644
>     --- a/src/intel/compiler/brw_fs_surface_builder.cpp
>     +++ b/src/intel/compiler/brw_fs_surface_builder.cpp
>     @@ -160,6 +160,16 @@ namespace brw {
>               return emit_send(bld, SHADER_OPCODE_TYPED_ATOMIC_LOGICAL,
>                                addr, tmp, surface, dims, op, rsize);
>            }
>     +
>     +      void
>     +      emit_byte_scattered_write(const fs_builder &bld, const fs_reg
>     &surface,
>     +                                const fs_reg &addr, const fs_reg &src,
>     +                                unsigned dims, unsigned size,
>     +                                unsigned bit_size, brw_predicate pred)
>     +      {
>     +         emit_send(bld, SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL,
>     +                   addr, src, surface, dims, bit_size, 0, pred);
>     +      }
>         }
>      }
> 
>     @@ -1192,3 +1202,4 @@ namespace brw {
>            }
>         }
>      }
>     +
>     diff --git a/src/intel/compiler/brw_fs_surface_builder.h
>     b/src/intel/compiler/brw_fs_surface_builder.h
>     index 32b56d387f..bf9a8c68c8 100644
>     --- a/src/intel/compiler/brw_fs_surface_builder.h
>     +++ b/src/intel/compiler/brw_fs_surface_builder.h
>     @@ -63,6 +63,13 @@ namespace brw {
>                              const fs_reg &src0, const fs_reg &src1,
>                              unsigned dims, unsigned rsize, unsigned op,
>                              brw_predicate pred = BRW_PREDICATE_NONE);
>     +
>     +      void
>     +      emit_byte_scattered_write(const fs_builder &bld, const fs_reg
>     &surface,
>     +                                const fs_reg &addr, const fs_reg &src,
>     +                                unsigned dims, unsigned size,
>     +                                unsigned bit_size,
>     +                                brw_predicate pred =
>     BRW_PREDICATE_NONE);
>         }
> 
>         namespace image_access {
>     diff --git a/src/intel/compiler/brw_shader.cpp
>     b/src/intel/compiler/brw_shader.cpp
>     index d7d7616cf4..209552e1b2 100644
>     --- a/src/intel/compiler/brw_shader.cpp
>     +++ b/src/intel/compiler/brw_shader.cpp
>     @@ -293,6 +293,11 @@ brw_instruction_name(const struct
>     gen_device_info *devinfo, enum opcode op)
>         case SHADER_OPCODE_MEMORY_FENCE:
>            return "memory_fence";
> 
>     +   case SHADER_OPCODE_BYTE_SCATTERED_WRITE:
>     +      return "byte_scattered_write";
>     +   case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL:
>     +      return "byte_scattered_write_logical";
>     +
>         case SHADER_OPCODE_LOAD_PAYLOAD:
>            return "load_payload";
>         case FS_OPCODE_PACK:
>     @@ -963,6 +968,8 @@ backend_instruction::has_side_effects() const
>         case SHADER_OPCODE_GEN4_SCRATCH_WRITE:
>         case SHADER_OPCODE_UNTYPED_SURFACE_WRITE:
>         case SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL:
>     +   case SHADER_OPCODE_BYTE_SCATTERED_WRITE:
>     +   case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL:
>         case SHADER_OPCODE_TYPED_ATOMIC:
>         case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL:
>         case SHADER_OPCODE_TYPED_SURFACE_WRITE:
>     --
>     2.14.3
> 
>     _______________________________________________
>     mesa-dev mailing list
>     mesa-dev at lists.freedesktop.org <mailto:mesa-dev at lists.freedesktop.org>
>     https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>     <https://lists.freedesktop.org/mailman/listinfo/mesa-dev>
> 
> 


More information about the mesa-dev mailing list