[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