From: Alyssa Rosenzweig Date: Wed, 24 Apr 2019 02:18:28 +0000 (+0000) Subject: panfrost/midgard: Add "op commutes?" property X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=bb1aff3007f4dcb0149e82e205e7451bcd4942de;p=mesa.git panfrost/midgard: Add "op commutes?" property Signed-off-by: Alyssa Rosenzweig --- diff --git a/src/gallium/drivers/panfrost/midgard/helpers.h b/src/gallium/drivers/panfrost/midgard/helpers.h index 40923f287d0..5912963ccdd 100644 --- a/src/gallium/drivers/panfrost/midgard/helpers.h +++ b/src/gallium/drivers/panfrost/midgard/helpers.h @@ -68,6 +68,9 @@ #define QUIRK_FLIPPED_R24 (1 << 2) +/* Is the op commutative? */ +#define OP_COMMUTES (1 << 3) + /* Vector-independant shorthands for the above; these numbers are arbitrary and * not from the ISA. Convert to the above with unit_enum_to_midgard */ @@ -151,14 +154,14 @@ static struct { const char *name; unsigned props; } alu_opcode_props[256] = { - [midgard_alu_op_fadd] = {"fadd", UNITS_ADD}, - [midgard_alu_op_fmul] = {"fmul", UNITS_MUL | UNIT_VLUT}, - [midgard_alu_op_fmin] = {"fmin", UNITS_MUL | UNITS_ADD}, - [midgard_alu_op_fmax] = {"fmax", UNITS_MUL | UNITS_ADD}, - [midgard_alu_op_imin] = {"imin", UNITS_MOST}, - [midgard_alu_op_imax] = {"imax", UNITS_MOST}, - [midgard_alu_op_umin] = {"umin", UNITS_MOST}, - [midgard_alu_op_umax] = {"umax", UNITS_MOST}, + [midgard_alu_op_fadd] = {"fadd", UNITS_ADD | OP_COMMUTES}, + [midgard_alu_op_fmul] = {"fmul", UNITS_MUL | UNIT_VLUT | OP_COMMUTES}, + [midgard_alu_op_fmin] = {"fmin", UNITS_MUL | UNITS_ADD | OP_COMMUTES}, + [midgard_alu_op_fmax] = {"fmax", UNITS_MUL | UNITS_ADD | OP_COMMUTES}, + [midgard_alu_op_imin] = {"imin", UNITS_MOST | OP_COMMUTES}, + [midgard_alu_op_imax] = {"imax", UNITS_MOST | OP_COMMUTES}, + [midgard_alu_op_umin] = {"umin", UNITS_MOST | OP_COMMUTES}, + [midgard_alu_op_umax] = {"umax", UNITS_MOST | OP_COMMUTES}, [midgard_alu_op_fmov] = {"fmov", UNITS_ALL | QUIRK_FLIPPED_R24}, [midgard_alu_op_fround] = {"fround", UNITS_ADD}, [midgard_alu_op_froundeven] = {"froundeven", UNITS_ADD}, @@ -169,24 +172,24 @@ static struct { /* Though they output a scalar, they need to run on a vector unit * since they process vectors */ - [midgard_alu_op_fdot3] = {"fdot3", UNIT_VMUL | OP_CHANNEL_COUNT(3)}, - [midgard_alu_op_fdot3r] = {"fdot3r", UNIT_VMUL | OP_CHANNEL_COUNT(3)}, - [midgard_alu_op_fdot4] = {"fdot4", UNIT_VMUL | OP_CHANNEL_COUNT(4)}, + [midgard_alu_op_fdot3] = {"fdot3", UNIT_VMUL | OP_CHANNEL_COUNT(3) | OP_COMMUTES}, + [midgard_alu_op_fdot3r] = {"fdot3r", UNIT_VMUL | OP_CHANNEL_COUNT(3) | OP_COMMUTES}, + [midgard_alu_op_fdot4] = {"fdot4", UNIT_VMUL | OP_CHANNEL_COUNT(4) | OP_COMMUTES}, /* Incredibly, iadd can run on vmul, etc */ - [midgard_alu_op_iadd] = {"iadd", UNITS_MOST}, + [midgard_alu_op_iadd] = {"iadd", UNITS_MOST | OP_COMMUTES}, [midgard_alu_op_iabs] = {"iabs", UNITS_MOST}, [midgard_alu_op_isub] = {"isub", UNITS_MOST}, - [midgard_alu_op_imul] = {"imul", UNITS_MUL}, + [midgard_alu_op_imul] = {"imul", UNITS_MUL | OP_COMMUTES}, [midgard_alu_op_imov] = {"imov", UNITS_MOST | QUIRK_FLIPPED_R24}, /* For vector comparisons, use ball etc */ - [midgard_alu_op_feq] = {"feq", UNITS_MOST}, - [midgard_alu_op_fne] = {"fne", UNITS_MOST}, + [midgard_alu_op_feq] = {"feq", UNITS_MOST | OP_COMMUTES}, + [midgard_alu_op_fne] = {"fne", UNITS_MOST | OP_COMMUTES}, [midgard_alu_op_fle] = {"fle", UNITS_MOST}, [midgard_alu_op_flt] = {"flt", UNITS_MOST}, - [midgard_alu_op_ieq] = {"ieq", UNITS_MOST}, - [midgard_alu_op_ine] = {"ine", UNITS_MOST}, + [midgard_alu_op_ieq] = {"ieq", UNITS_MOST | OP_COMMUTES}, + [midgard_alu_op_ine] = {"ine", UNITS_MOST | OP_COMMUTES}, [midgard_alu_op_ilt] = {"ilt", UNITS_MOST}, [midgard_alu_op_ile] = {"ile", UNITS_MOST}, [midgard_alu_op_ult] = {"ult", UNITS_MOST}, @@ -213,10 +216,10 @@ static struct { [midgard_alu_op_fcos] = {"fcos", UNIT_VLUT}, /* XXX: Test case where it's right on smul but not sadd */ - [midgard_alu_op_iand] = {"iand", UNITS_ADD}, + [midgard_alu_op_iand] = {"iand", UNITS_ADD | OP_COMMUTES}, - [midgard_alu_op_ior] = {"ior", UNITS_ADD}, - [midgard_alu_op_ixor] = {"ixor", UNITS_ADD}, + [midgard_alu_op_ior] = {"ior", UNITS_ADD | OP_COMMUTES}, + [midgard_alu_op_ixor] = {"ixor", UNITS_ADD | OP_COMMUTES}, [midgard_alu_op_ilzcnt] = {"ilzcnt", UNITS_ADD}, [midgard_alu_op_ibitcount8] = {"ibitcount8", UNITS_ADD}, [midgard_alu_op_inot] = {"inot", UNITS_MOST}, @@ -224,12 +227,12 @@ static struct { [midgard_alu_op_iasr] = {"iasr", UNITS_ADD}, [midgard_alu_op_ilsr] = {"ilsr", UNITS_ADD}, - [midgard_alu_op_fball_eq] = {"fball_eq", UNITS_VECTOR}, - [midgard_alu_op_fbany_neq] = {"fbany_neq", UNITS_VECTOR}, - [midgard_alu_op_iball_eq] = {"iball_eq", UNITS_VECTOR}, - [midgard_alu_op_iball_neq] = {"iball_neq", UNITS_VECTOR}, - [midgard_alu_op_ibany_eq] = {"ibany_eq", UNITS_VECTOR}, - [midgard_alu_op_ibany_neq] = {"ibany_neq", UNITS_VECTOR}, + [midgard_alu_op_fball_eq] = {"fball_eq", UNITS_VECTOR | OP_COMMUTES}, + [midgard_alu_op_fbany_neq] = {"fbany_neq", UNITS_VECTOR | OP_COMMUTES}, + [midgard_alu_op_iball_eq] = {"iball_eq", UNITS_VECTOR | OP_COMMUTES}, + [midgard_alu_op_iball_neq] = {"iball_neq", UNITS_VECTOR | OP_COMMUTES}, + [midgard_alu_op_ibany_eq] = {"ibany_eq", UNITS_VECTOR | OP_COMMUTES}, + [midgard_alu_op_ibany_neq] = {"ibany_neq", UNITS_VECTOR | OP_COMMUTES}, /* These instructions are not yet emitted by the compiler, so * don't speculate about units yet */ @@ -246,8 +249,8 @@ static struct { [midgard_alu_op_ibany_lte] = {"ibany_lte", 0}, [midgard_alu_op_freduce] = {"freduce", 0}, - [midgard_alu_op_bball_eq] = {"bball_eq", 0}, - [midgard_alu_op_bbany_neq] = {"bball_eq", 0}, + [midgard_alu_op_bball_eq] = {"bball_eq", 0 | OP_COMMUTES}, + [midgard_alu_op_bbany_neq] = {"bball_eq", 0 | OP_COMMUTES}, [midgard_alu_op_fatan2_pt1] = {"fatan2_pt1", 0}, [midgard_alu_op_fatan_pt2] = {"fatan_pt2", 0}, }; diff --git a/src/gallium/drivers/panfrost/midgard/midgard_compile.c b/src/gallium/drivers/panfrost/midgard/midgard_compile.c index e5b1a1ec3fd..849426f19b4 100644 --- a/src/gallium/drivers/panfrost/midgard/midgard_compile.c +++ b/src/gallium/drivers/panfrost/midgard/midgard_compile.c @@ -2948,34 +2948,21 @@ embedded_to_inline_constant(compiler_context *ctx) int op = ins->alu.op; if (ins->ssa_args.src0 == SSA_FIXED_REGISTER(REGISTER_CONSTANT)) { - /* Flip based on op. Fallthrough intentional */ - switch (op) { - /* These ops require an operational change to flip their arguments TODO */ + /* These ops require an operational change to flip + * their arguments TODO */ case midgard_alu_op_flt: case midgard_alu_op_fle: case midgard_alu_op_ilt: case midgard_alu_op_ile: case midgard_alu_op_fcsel: case midgard_alu_op_icsel: - case midgard_alu_op_isub: DBG("Missed non-commutative flip (%s)\n", alu_opcode_props[op].name); + default: break; + } - /* These ops are commutative and Just Flip */ - case midgard_alu_op_fne: - case midgard_alu_op_fadd: - case midgard_alu_op_fmul: - case midgard_alu_op_fmin: - case midgard_alu_op_fmax: - case midgard_alu_op_iadd: - case midgard_alu_op_imul: - case midgard_alu_op_feq: - case midgard_alu_op_ieq: - case midgard_alu_op_ine: - case midgard_alu_op_iand: - case midgard_alu_op_ior: - case midgard_alu_op_ixor: + if (alu_opcode_props[op].props & OP_COMMUTES) { /* Flip the SSA numbers */ ins->ssa_args.src0 = ins->ssa_args.src1; ins->ssa_args.src1 = SSA_FIXED_REGISTER(REGISTER_CONSTANT); @@ -2987,9 +2974,6 @@ embedded_to_inline_constant(compiler_context *ctx) src_temp = ins->alu.src2; ins->alu.src2 = ins->alu.src1; ins->alu.src1 = src_temp; - - default: - break; } }