.next_clause_type = next ? next->clause_type : 0,
};
+ header.branch_cond |= header.back_to_back;
+
uint64_t u = 0;
memcpy(&u, &header, sizeof(header));
return u;
/* Only check top 60-bits since that's what's actually embedded
* in the clause, the bottom 4-bits are bundle-inline */
- unsigned candidates[2] = {
+ uint64_t candidates[2] = {
clause->constants[i] >> 4,
clause->constants[i] >> 36
};
+ /* For <64-bit mode, we treat lo/hi separately */
+
if (!b64)
- candidates[0] &= 0xFFFFFFFF;
+ candidates[0] &= (0xFFFFFFFF >> 4);
if (candidates[0] == want)
return i;
if (!ins)
return assigned;
+ if (ins->type == BI_BLEND) {
+ assert(!assigned);
+ regs->uniform_constant = 0x8;
+ return true;
+ }
+
bi_foreach_src(ins, s) {
if (s == 0 && (ins->type == BI_LOAD_VAR_ADDRESS || ins->type == BI_LOAD_ATTR)) continue;
if (ins->src[s] & BIR_INDEX_CONSTANT) {
bool hi = false;
bool b64 = nir_alu_type_get_type_size(ins->src_types[s]) > 32;
- uint64_t cons = bi_get_immediate(ins, ins->src[s]);
+ uint64_t cons = bi_get_immediate(ins, s);
unsigned idx = bi_lookup_constant(clause, cons, &hi, b64);
- unsigned f = bi_constant_field(idx) | (cons & 0xF);
+ unsigned lo = clause->constants[idx] & 0xF;
+ unsigned f = bi_constant_field(idx) | lo;
if (assigned && regs->uniform_constant != f)
unreachable("Mismatched uniform/const field: imm");
if (!bi_writes_component(ins, src)) continue;
unsigned k = ins->swizzle[src][c];
- assert(k < 1);
+ assert(k <= 1);
swizzle |= (k << c);
}
/* (-a)(-b) = ab, so we only need one negate bit */
bool negate_mul = ins->src_neg[0] ^ ins->src_neg[1];
- if (ins->dest_type == nir_type_float32) {
+ if (ins->op.mscale) {
+ assert(!(ins->src_abs[0] && ins->src_abs[1]));
+ assert(!ins->src_abs[2] || !ins->src_neg[3] || !ins->src_abs[3]);
+
+ /* We can have exactly one abs, and can flip the multiplication
+ * to make it fit if we have to */
+ bool flip_ab = ins->src_abs[1];
+
+ struct bifrost_fma_mscale pack = {
+ .src0 = bi_get_src(ins, regs, flip_ab ? 1 : 0, true),
+ .src1 = bi_get_src(ins, regs, flip_ab ? 0 : 1, true),
+ .src2 = bi_get_src(ins, regs, 2, true),
+ .src3 = bi_get_src(ins, regs, 3, true),
+ .mscale_mode = 0,
+ .mode = ins->outmod,
+ .src0_abs = ins->src_abs[0] || ins->src_abs[1],
+ .src1_neg = negate_mul,
+ .src2_neg = ins->src_neg[2],
+ .op = BIFROST_FMA_OP_MSCALE,
+ };
+
+ RETURN_PACKED(pack);
+ } else if (ins->dest_type == nir_type_float32) {
struct bifrost_fma_fma pack = {
.src0 = bi_get_src(ins, regs, 0, true),
.src1 = bi_get_src(ins, regs, 1, true),
RETURN_PACKED(pack);
}
-static unsigned
-bi_pack_fma_addmin_f16(bi_instruction *ins, struct bi_registers *regs)
+static bool
+bi_pack_fp16_abs(bi_instruction *ins, struct bi_registers *regs, bool *flip)
{
- unsigned op =
- (ins->type == BI_ADD) ? BIFROST_FMA_OP_FADD16 :
- (ins->op.minmax == BI_MINMAX_MIN) ? BIFROST_FMA_OP_FMIN16 :
- BIFROST_FMA_OP_FMAX16;
-
/* Absolute values are packed in a quirky way. Let k = src1 < src0. Let
* l be an auxiliary bit we encode. Then the hardware determines:
*
unsigned abs_0 = ins->src_abs[0], abs_1 = ins->src_abs[1];
unsigned src_0 = bi_get_src(ins, regs, 0, true);
unsigned src_1 = bi_get_src(ins, regs, 1, true);
- bool l = false;
- bool flip = false;
+
+ assert(!(abs_0 && abs_1 && src_0 == src_1));
if (!abs_0 && !abs_1) {
/* Force k = 0 <===> NOT(src1 < src0) */
- flip = (src_1 < src_0);
+ *flip = (src_1 < src_0);
+ return false;
} else if (abs_0 && !abs_1) {
- l = src_1 >= src_0;
+ return src_1 >= src_0;
} else if (abs_1 && !abs_0) {
- flip = true;
- l = src_0 >= src_1;
+ *flip = true;
+ return src_0 >= src_1;
} else {
- flip = (src_0 >= src_1);
- l = true;
+ *flip = (src_0 >= src_1);
+ return true;
}
+}
- struct bifrost_fma_add_minmax16 pack = {
- .src0 = flip ? src_1 : src_0,
- .src1 = flip ? src_0 : src_1,
- .src0_neg = ins->src_neg[flip ? 1 : 0],
- .src1_neg = ins->src_neg[flip ? 0 : 1],
- .abs1 = l,
- .outmod = ins->outmod,
- .mode = (ins->type == BI_ADD) ? ins->roundmode : ins->minmax,
- .op = op
- };
+static unsigned
+bi_pack_fmadd_min_f16(bi_instruction *ins, struct bi_registers *regs, bool FMA)
+{
+ unsigned op =
+ (!FMA) ? ((ins->op.minmax == BI_MINMAX_MIN) ?
+ BIFROST_ADD_OP_FMIN16 : BIFROST_ADD_OP_FMAX16) :
+ (ins->type == BI_ADD) ? BIFROST_FMA_OP_FADD16 :
+ (ins->op.minmax == BI_MINMAX_MIN) ? BIFROST_FMA_OP_FMIN16 :
+ BIFROST_FMA_OP_FMAX16;
- RETURN_PACKED(pack);
+ bool flip = false;
+ bool l = bi_pack_fp16_abs(ins, regs, &flip);
+ unsigned src_0 = bi_get_src(ins, regs, 0, true);
+ unsigned src_1 = bi_get_src(ins, regs, 1, true);
+
+ if (FMA) {
+ struct bifrost_fma_add_minmax16 pack = {
+ .src0 = flip ? src_1 : src_0,
+ .src1 = flip ? src_0 : src_1,
+ .src0_neg = ins->src_neg[flip ? 1 : 0],
+ .src1_neg = ins->src_neg[flip ? 0 : 1],
+ .abs1 = l,
+ .outmod = ins->outmod,
+ .mode = (ins->type == BI_ADD) ? ins->roundmode : ins->minmax,
+ .op = op
+ };
+
+ RETURN_PACKED(pack);
+ } else {
+ /* Can't have modes for fp16 */
+ assert(ins->outmod == 0);
+
+ struct bifrost_add_fmin16 pack = {
+ .src0 = flip ? src_1 : src_0,
+ .src1 = flip ? src_0 : src_1,
+ .src0_neg = ins->src_neg[flip ? 1 : 0],
+ .src1_neg = ins->src_neg[flip ? 0 : 1],
+ .abs1 = l,
+ .src0_swizzle = bi_swiz16(ins, 0),
+ .src1_swizzle = bi_swiz16(ins, 1),
+ .mode = ins->minmax,
+ .op = op
+ };
+
+ RETURN_PACKED(pack);
+ }
}
static unsigned
if (ins->dest_type == nir_type_float32)
return bi_pack_fma_addmin_f32(ins, regs);
else if(ins->dest_type == nir_type_float16)
- return bi_pack_fma_addmin_f16(ins, regs);
+ return bi_pack_fmadd_min_f16(ins, regs, true);
else
unreachable("Unknown FMA/ADD type");
}
RETURN_PACKED(pack);
}
+static unsigned
+bi_pack_fma_2src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
+{
+ struct bifrost_fma_2src pack = {
+ .src0 = bi_get_src(ins, regs, 0, true),
+ .src1 = bi_get_src(ins, regs, 1, true),
+ .op = op
+ };
+
+ RETURN_PACKED(pack);
+}
+
static unsigned
bi_pack_add_1src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
{
bool flip = false, invert = false;
enum bifrost_csel_cond cond =
- bi_cond_to_csel(ins->csel_cond, &flip, &invert, ins->src_types[0]);
+ bi_cond_to_csel(ins->cond, &flip, &invert, ins->src_types[0]);
unsigned size = nir_alu_type_get_type_size(ins->dest_type);
RETURN_PACKED(pack);
}
+static unsigned
+bi_pack_fma_frexp(bi_instruction *ins, struct bi_registers *regs)
+{
+ unsigned op = BIFROST_FMA_OP_FREXPE_LOG;
+ return bi_pack_fma_1src(ins, regs, op);
+}
+
+static unsigned
+bi_pack_fma_reduce(bi_instruction *ins, struct bi_registers *regs)
+{
+ if (ins->op.reduce == BI_REDUCE_ADD_FREXPM) {
+ return bi_pack_fma_2src(ins, regs, BIFROST_FMA_OP_ADD_FREXPM);
+ } else {
+ unreachable("Invalid reduce op");
+ }
+}
+
/* We have a single convert opcode in the IR but a number of opcodes that could
* come out. In particular we have native opcodes for:
*
*/
static unsigned
-bi_pack_fma_convert(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_convert(bi_instruction *ins, struct bi_registers *regs, bool FMA)
{
nir_alu_type from_base = nir_alu_type_get_base_type(ins->src_types[0]);
unsigned from_size = nir_alu_type_get_type_size(ins->src_types[0]);
nir_alu_type to_base = nir_alu_type_get_base_type(ins->dest_type);
unsigned to_size = nir_alu_type_get_type_size(ins->dest_type);
bool to_unsigned = to_base == nir_type_uint;
+ bool to_float = to_base == nir_type_float;
/* Sanity check */
assert((from_base != to_base) || (from_size != to_size));
assert((MAX2(from_size, to_size) / MIN2(from_size, to_size)) <= 2);
- if (from_size == 16 && to_size == 16) {
- /* f2i_i2f16 */
- unreachable("i16 not yet implemented");
- } else if (from_size == 32 && to_size == 32) {
- unsigned op = 0;
+ /* f32 to f16 is special */
+ if (from_size == 32 && to_size == 16 && from_base == nir_type_float && to_base == from_base) {
+ /* TODO: second vectorized source? */
+ struct bifrost_fma_2src pfma = {
+ .src0 = bi_get_src(ins, regs, 0, true),
+ .src1 = BIFROST_SRC_STAGE, /* 0 */
+ .op = BIFROST_FMA_FLOAT32_TO_16
+ };
- if (from_base == nir_type_float) {
- op = BIFROST_FMA_FLOAT32_TO_INT(to_unsigned);
+ struct bifrost_add_2src padd = {
+ .src0 = bi_get_src(ins, regs, 0, true),
+ .src1 = BIFROST_SRC_STAGE, /* 0 */
+ .op = BIFROST_ADD_FLOAT32_TO_16
+ };
+
+ if (FMA) {
+ RETURN_PACKED(pfma);
} else {
- op = BIFROST_FMA_INT_TO_FLOAT32(from_unsigned);
+ RETURN_PACKED(padd);
}
+ }
+
+ /* Otherwise, figure out the mode */
+ unsigned op = 0;
- return bi_pack_fma_1src(ins, regs, op);
- } else if (from_size == 16 && to_size == 32) {
- bool from_y = ins->swizzle[0][0];
+ if (from_size == 16 && to_size == 32) {
+ unsigned component = ins->swizzle[0][0];
+ assert(component <= 1);
+
+ if (from_base == nir_type_float)
+ op = BIFROST_CONVERT_5(component);
+ else
+ op = BIFROST_CONVERT_4(from_unsigned, component, to_float);
+ } else {
+ unsigned mode = 0;
+ unsigned swizzle = (from_size == 16) ? bi_swiz16(ins, 0) : 0;
+ bool is_unsigned = from_unsigned;
if (from_base == nir_type_float) {
- return bi_pack_fma_1src(ins, regs,
- BIFROST_FMA_FLOAT16_TO_32(from_y));
+ assert(to_base != nir_type_float);
+ is_unsigned = to_unsigned;
+
+ if (from_size == 32 && to_size == 32)
+ mode = BIFROST_CONV_F32_TO_I32;
+ else if (from_size == 16 && to_size == 16)
+ mode = BIFROST_CONV_F16_TO_I16;
+ else
+ unreachable("Invalid float conversion");
} else {
- unreachable("i16 not yet implemented");
+ assert(to_base == nir_type_float);
+ assert(from_size == to_size);
+
+ if (to_size == 32)
+ mode = BIFROST_CONV_I32_TO_F32;
+ else if (to_size == 16)
+ mode = BIFROST_CONV_I16_TO_F16;
+ else
+ unreachable("Invalid int conversion");
}
- } else if (from_size == 32 && to_size == 16) {
- if (from_base == nir_type_float) {
- /* TODO: second vectorized source? */
- struct bifrost_fma_2src pack = {
- .src0 = bi_get_src(ins, regs, 0, true),
- .src1 = BIFROST_SRC_STAGE, /* 0 */
- .op = BIFROST_FMA_FLOAT32_TO_16
- };
-
- RETURN_PACKED(pack);
- } else {
- unreachable("i16 not yet implemented");
+
+ /* Fixup swizzle for 32-bit only modes */
+
+ if (mode == BIFROST_CONV_I32_TO_F32)
+ swizzle = 0b11;
+ else if (mode == BIFROST_CONV_F32_TO_I32)
+ swizzle = 0b10;
+
+ op = BIFROST_CONVERT(is_unsigned, ins->roundmode, swizzle, mode);
+
+ /* Unclear what the top bit is for... maybe 16-bit related */
+ bool mode2 = mode == BIFROST_CONV_F16_TO_I16;
+ bool mode6 = mode == BIFROST_CONV_I16_TO_F16;
+
+ if (!(mode2 || mode6))
+ op |= 0x100;
+ }
+
+ if (FMA)
+ return bi_pack_fma_1src(ins, regs, BIFROST_FMA_CONVERT | op);
+ else
+ return bi_pack_add_1src(ins, regs, BIFROST_ADD_CONVERT | op);
+}
+
+static unsigned
+bi_pack_fma_select(bi_instruction *ins, struct bi_registers *regs)
+{
+ unsigned size = nir_alu_type_get_type_size(ins->src_types[0]);
+
+ if (size == 16) {
+ unsigned swiz = (ins->swizzle[0][0] | (ins->swizzle[1][0] << 1));
+ unsigned op = BIFROST_FMA_SEL_16(swiz);
+ return bi_pack_fma_2src(ins, regs, op);
+ } else if (size == 8) {
+ unsigned swiz = 0;
+
+ for (unsigned c = 0; c < 4; ++c) {
+ if (ins->swizzle[c][0]) {
+ /* Ensure lowering restriction is met */
+ assert(ins->swizzle[c][0] == 2);
+ swiz |= (1 << c);
+ }
}
+
+ struct bifrost_fma_sel8 pack = {
+ .src0 = bi_get_src(ins, regs, 0, true),
+ .src1 = bi_get_src(ins, regs, 1, true),
+ .src2 = bi_get_src(ins, regs, 2, true),
+ .src3 = bi_get_src(ins, regs, 3, true),
+ .swizzle = swiz,
+ .op = BIFROST_FMA_OP_SEL8
+ };
+
+ RETURN_PACKED(pack);
+ } else {
+ unreachable("Unimplemented");
}
+}
+
+static enum bifrost_fcmp_cond
+bi_fcmp_cond(enum bi_cond cond)
+{
+ switch (cond) {
+ case BI_COND_LT: return BIFROST_OLT;
+ case BI_COND_LE: return BIFROST_OLE;
+ case BI_COND_GE: return BIFROST_OGE;
+ case BI_COND_GT: return BIFROST_OGT;
+ case BI_COND_EQ: return BIFROST_OEQ;
+ case BI_COND_NE: return BIFROST_UNE;
+ default: unreachable("Unknown bi_cond");
+ }
+}
+
+/* a <?> b <==> b <flip(?)> a (TODO: NaN behaviour?) */
+
+static enum bifrost_fcmp_cond
+bi_flip_fcmp(enum bifrost_fcmp_cond cond)
+{
+ switch (cond) {
+ case BIFROST_OGT:
+ return BIFROST_OLT;
+ case BIFROST_OGE:
+ return BIFROST_OLE;
+ case BIFROST_OLT:
+ return BIFROST_OGT;
+ case BIFROST_OLE:
+ return BIFROST_OGE;
+ case BIFROST_OEQ:
+ case BIFROST_UNE:
+ return cond;
+ default:
+ unreachable("Unknown fcmp cond");
+ }
+}
+
+static unsigned
+bi_pack_fma_cmp(bi_instruction *ins, struct bi_registers *regs)
+{
+ nir_alu_type Tl = ins->src_types[0];
+ nir_alu_type Tr = ins->src_types[1];
+
+ if (Tl == nir_type_float32 || Tr == nir_type_float32) {
+ /* TODO: Mixed 32/16 cmp */
+ assert(Tl == Tr);
+
+ enum bifrost_fcmp_cond cond = bi_fcmp_cond(ins->cond);
+
+ /* Only src1 has neg, so we arrange:
+ * a < b --- native
+ * a < -b --- native
+ * -a < -b <===> a > b
+ * -a < b <===> a > -b
+ * TODO: Is this NaN-precise?
+ */
+
+ bool flip = ins->src_neg[0];
+ bool neg = ins->src_neg[0] ^ ins->src_neg[1];
+
+ if (flip)
+ cond = bi_flip_fcmp(cond);
+
+ struct bifrost_fma_fcmp pack = {
+ .src0 = bi_get_src(ins, regs, 0, true),
+ .src1 = bi_get_src(ins, regs, 1, true),
+ .src0_abs = ins->src_abs[0],
+ .src1_abs = ins->src_abs[1],
+ .src1_neg = neg,
+ .src_expand = 0,
+ .unk1 = 0,
+ .cond = cond,
+ .op = BIFROST_FMA_OP_FCMP_GL
+ };
+
+ RETURN_PACKED(pack);
+ } else if (Tl == nir_type_float16 && Tr == nir_type_float16) {
+ bool flip = false;
+ bool l = bi_pack_fp16_abs(ins, regs, &flip);
+ enum bifrost_fcmp_cond cond = bi_fcmp_cond(ins->cond);
+
+ if (flip)
+ cond = bi_flip_fcmp(cond);
+
+ struct bifrost_fma_fcmp16 pack = {
+ .src0 = bi_get_src(ins, regs, flip ? 1 : 0, true),
+ .src1 = bi_get_src(ins, regs, flip ? 0 : 1, true),
+ .src0_swizzle = bi_swiz16(ins, flip ? 1 : 0),
+ .src1_swizzle = bi_swiz16(ins, flip ? 0 : 1),
+ .abs1 = l,
+ .unk = 0,
+ .cond = cond,
+ .op = BIFROST_FMA_OP_FCMP_GL_16,
+ };
- unreachable("Unknown convert");
+ RETURN_PACKED(pack);
+ } else {
+ unreachable("Unknown cmp type");
+ }
}
+static unsigned
+bi_fma_bitwise_op(enum bi_bitwise_op op, bool rshift)
+{
+ switch (op) {
+ case BI_BITWISE_OR:
+ /* Via De Morgan's */
+ return rshift ?
+ BIFROST_FMA_OP_RSHIFT_NAND :
+ BIFROST_FMA_OP_LSHIFT_NAND;
+ case BI_BITWISE_AND:
+ return rshift ?
+ BIFROST_FMA_OP_RSHIFT_AND :
+ BIFROST_FMA_OP_LSHIFT_AND;
+ case BI_BITWISE_XOR:
+ /* Shift direction handled out of band */
+ return BIFROST_FMA_OP_RSHIFT_XOR;
+ default:
+ unreachable("Unknown op");
+ }
+}
+
+static unsigned
+bi_pack_fma_bitwise(bi_instruction *ins, struct bi_registers *regs)
+{
+ unsigned size = nir_alu_type_get_type_size(ins->dest_type);
+ assert(size <= 32);
+
+ bool invert_0 = ins->bitwise.src_invert[0];
+ bool invert_1 = ins->bitwise.src_invert[1];
+
+ if (ins->op.bitwise == BI_BITWISE_OR) {
+ /* Becomes NAND, so via De Morgan's:
+ * f(A) | f(B) = ~(~f(A) & ~f(B))
+ * = NAND(~f(A), ~f(B))
+ */
+
+ invert_0 = !invert_0;
+ invert_1 = !invert_1;
+ } else if (ins->op.bitwise == BI_BITWISE_XOR) {
+ /* ~A ^ ~B = ~(A ^ ~B) = ~(~(A ^ B)) = A ^ B
+ * ~A ^ B = ~(A ^ B) = A ^ ~B
+ */
+
+ invert_0 ^= invert_1;
+ invert_1 = false;
+
+ /* invert_1 ends up specifying shift direction */
+ invert_1 = !ins->bitwise.rshift;
+ }
+
+ struct bifrost_shift_fma pack = {
+ .src0 = bi_get_src(ins, regs, 0, true),
+ .src1 = bi_get_src(ins, regs, 1, true),
+ .src2 = bi_get_src(ins, regs, 2, true),
+ .half = (size == 32) ? 0 : (size == 16) ? 0x7 : (size == 8) ? 0x4 : 0,
+ .unk = 1, /* XXX */
+ .invert_1 = invert_0,
+ .invert_2 = invert_1,
+ .op = bi_fma_bitwise_op(ins->op.bitwise, ins->bitwise.rshift)
+ };
+
+ RETURN_PACKED(pack);
+}
+
static unsigned
bi_pack_fma(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
{
case BI_ADD:
return bi_pack_fma_addmin(bundle.fma, regs);
case BI_CMP:
+ return bi_pack_fma_cmp(bundle.fma, regs);
case BI_BITWISE:
- return BIFROST_FMA_NOP;
+ return bi_pack_fma_bitwise(bundle.fma, regs);
case BI_CONVERT:
- return bi_pack_fma_convert(bundle.fma, regs);
+ return bi_pack_convert(bundle.fma, regs, true);
case BI_CSEL:
return bi_pack_fma_csel(bundle.fma, regs);
case BI_FMA:
return bi_pack_fma_fma(bundle.fma, regs);
case BI_FREXP:
+ return bi_pack_fma_frexp(bundle.fma, regs);
case BI_ISUB:
return BIFROST_FMA_NOP;
case BI_MINMAX:
case BI_MOV:
return bi_pack_fma_1src(bundle.fma, regs, BIFROST_FMA_OP_MOV);
case BI_SHIFT:
- case BI_SWIZZLE:
+ return BIFROST_FMA_NOP;
+ case BI_SELECT:
+ return bi_pack_fma_select(bundle.fma, regs);
case BI_ROUND:
return BIFROST_FMA_NOP;
+ case BI_REDUCE_FMA:
+ return bi_pack_fma_reduce(bundle.fma, regs);
default:
unreachable("Cannot encode class as FMA");
}
BIFROST_ADD_OP_LD_VAR_32 :
BIFROST_ADD_OP_LD_VAR_16;
- unsigned cmask = bi_from_bytemask(ins->writemask, size / 8);
- unsigned channels = util_bitcount(cmask);
- assert(cmask == ((1 << channels) - 1));
-
unsigned packed_addr = 0;
if (ins->src[0] & BIR_INDEX_CONSTANT) {
/* Direct uses address field directly */
- packed_addr = bi_get_immediate(ins, ins->src[0]);
+ packed_addr = bi_get_immediate(ins, 0);
assert(packed_addr < 0b1000);
} else {
/* Indirect gets an extra source */
assert(ins->dest & BIR_INDEX_REGISTER);
clause->data_register = ins->dest & ~BIR_INDEX_REGISTER;
+ unsigned channels = ins->vector_channels;
assert(channels >= 1 && channels <= 4);
struct bifrost_ld_var pack = {
RETURN_PACKED(pack);
}
+static unsigned
+bi_pack_add_add_f16(bi_instruction *ins, struct bi_registers *regs)
+{
+ /* ADD.v2f16 can't have outmod */
+ assert(ins->outmod == BIFROST_NONE);
+
+ struct bifrost_add_faddmin pack = {
+ .src0 = bi_get_src(ins, regs, 0, true),
+ .src1 = bi_get_src(ins, regs, 1, true),
+ .src0_abs = ins->src_abs[0],
+ .src1_abs = ins->src_abs[1],
+ .src0_neg = ins->src_neg[0],
+ .src1_neg = ins->src_neg[1],
+ .select = bi_swiz16(ins, 0), /* swizzle_0 */
+ .outmod = bi_swiz16(ins, 1), /* swizzle_1 */
+ .mode = ins->roundmode,
+ .op = BIFROST_ADD_OP_FADD16
+ };
+
+ RETURN_PACKED(pack);
+}
+
static unsigned
bi_pack_add_addmin(bi_instruction *ins, struct bi_registers *regs)
{
if (ins->dest_type == nir_type_float32)
return bi_pack_add_addmin_f32(ins, regs);
- else if(ins->dest_type == nir_type_float16)
- unreachable("todo");
- //return bi_pack_add_addmin_f16(ins, regs);
- else
+ else if (ins->dest_type == nir_type_float16) {
+ if (ins->type == BI_ADD)
+ return bi_pack_add_add_f16(ins, regs);
+ else
+ return bi_pack_fmadd_min_f16(ins, regs, false);
+ } else
unreachable("Unknown FMA/ADD type");
}
static unsigned
bi_pack_add_ld_ubo(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
{
- unsigned components = bi_load32_components(ins);
+ assert(ins->vector_channels >= 1 && ins->vector_channels <= 4);
const unsigned ops[4] = {
BIFROST_ADD_OP_LD_UBO_1,
};
bi_write_data_register(clause, ins);
- return bi_pack_add_2src(ins, regs, ops[components - 1]);
+ return bi_pack_add_2src(ins, regs, ops[ins->vector_channels - 1]);
}
static enum bifrost_ldst_type
struct bifrost_ld_var_addr pack = {
.src0 = bi_get_src(ins, regs, 1, false),
.src1 = bi_get_src(ins, regs, 2, false),
- .location = bi_get_immediate(ins, ins->src[0]),
+ .location = bi_get_immediate(ins, 0),
.type = bi_pack_ldst_type(ins->src_types[3]),
.op = BIFROST_ADD_OP_LD_VAR_ADDR
};
static unsigned
bi_pack_add_ld_attr(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
{
+ assert(ins->vector_channels >= 0 && ins->vector_channels <= 4);
+
struct bifrost_ld_attr pack = {
.src0 = bi_get_src(ins, regs, 1, false),
.src1 = bi_get_src(ins, regs, 2, false),
- .location = bi_get_immediate(ins, ins->src[0]),
- .channels = MALI_POSITIVE(bi_load32_components(ins)),
+ .location = bi_get_immediate(ins, 0),
+ .channels = MALI_POSITIVE(ins->vector_channels),
.type = bi_pack_ldst_type(ins->dest_type),
.op = BIFROST_ADD_OP_LD_ATTR
};
static unsigned
bi_pack_add_st_vary(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
{
- assert(ins->store_channels >= 1 && ins->store_channels <= 4);
+ assert(ins->vector_channels >= 1 && ins->vector_channels <= 4);
struct bifrost_st_vary pack = {
.src0 = bi_get_src(ins, regs, 1, false),
.src1 = bi_get_src(ins, regs, 2, false),
.src2 = bi_get_src(ins, regs, 3, false),
- .channels = MALI_POSITIVE(ins->store_channels),
+ .channels = MALI_POSITIVE(ins->vector_channels),
.op = BIFROST_ADD_OP_ST_VAR
};
static unsigned
bi_pack_add_atest(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
{
- /* TODO: fp16 */
- assert(ins->src_types[1] == nir_type_float32);
+ bool fp16 = (ins->src_types[1] == nir_type_float16);
struct bifrost_add_atest pack = {
.src0 = bi_get_src(ins, regs, 0, false),
.src1 = bi_get_src(ins, regs, 1, false),
- .component = 1, /* Set for fp32 */
+ .half = fp16,
+ .component = fp16 ? ins->swizzle[1][0] : 1, /* Set for fp32 */
.op = BIFROST_ADD_OP_ATEST,
};
(Y ? BIFROST_ADD_OP_FRCP_FAST_F16_Y :
BIFROST_ADD_OP_FRCP_FAST_F16_X) :
BIFROST_ADD_OP_FRCP_FAST_F32;
- } else {
+ } else if (ins->op.special == BI_SPECIAL_FRSQ) {
op = fp16 ?
(Y ? BIFROST_ADD_OP_FRSQ_FAST_F16_Y :
BIFROST_ADD_OP_FRSQ_FAST_F16_X) :
BIFROST_ADD_OP_FRSQ_FAST_F32;
+ } else if (ins->op.special == BI_SPECIAL_EXP2_LOW) {
+ assert(!fp16);
+ op = BIFROST_ADD_OP_FEXP2_FAST;
+ } else {
+ unreachable("Unknown special op");
}
return bi_pack_add_1src(ins, regs, op);
}
+static unsigned
+bi_pack_add_table(bi_instruction *ins, struct bi_registers *regs)
+{
+ unsigned op = 0;
+ assert(ins->dest_type == nir_type_float32);
+
+ op = BIFROST_ADD_OP_LOG2_HELP;
+ return bi_pack_add_1src(ins, regs, op);
+}
+static unsigned
+bi_pack_add_tex_compact(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
+{
+ bool f16 = ins->dest_type == nir_type_float16;
+
+ struct bifrost_tex_compact pack = {
+ .src0 = bi_get_src(ins, regs, 0, false),
+ .src1 = bi_get_src(ins, regs, 1, false),
+ .op = f16 ? BIFROST_ADD_OP_TEX_COMPACT_F16 :
+ BIFROST_ADD_OP_TEX_COMPACT_F32,
+ .unknown = 1,
+ .tex_index = 0,
+ .sampler_index = 0
+ };
+
+ bi_write_data_register(clause, ins);
+ RETURN_PACKED(pack);
+}
+
+static unsigned
+bi_pack_add_select(bi_instruction *ins, struct bi_registers *regs)
+{
+ unsigned size = nir_alu_type_get_type_size(ins->src_types[0]);
+ assert(size == 16);
+
+ unsigned swiz = (ins->swizzle[0][0] | (ins->swizzle[1][0] << 1));
+ unsigned op = BIFROST_ADD_SEL_16(swiz);
+ return bi_pack_add_2src(ins, regs, op);
+}
+
static unsigned
bi_pack_add(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
{
case BI_BLEND:
return bi_pack_add_blend(clause, bundle.add, regs);
case BI_BITWISE:
+ return BIFROST_ADD_NOP;
case BI_CONVERT:
+ return bi_pack_convert(bundle.add, regs, false);
case BI_DISCARD:
case BI_FREXP:
case BI_ISUB:
return bi_pack_add_st_vary(clause, bundle.add, regs);
case BI_SPECIAL:
return bi_pack_add_special(bundle.add, regs);
- case BI_SWIZZLE:
+ case BI_TABLE:
+ return bi_pack_add_table(bundle.add, regs);
+ case BI_SELECT:
+ return bi_pack_add_select(bundle.add, regs);
case BI_TEX:
+ if (bundle.add->op.texture == BI_TEX_COMPACT)
+ return bi_pack_add_tex_compact(clause, bundle.add, regs);
+ else
+ unreachable("Unknown tex type");
case BI_ROUND:
return BIFROST_ADD_NOP;
default:
/* TODO: Pos */
assert(index == 0 && clause->bundle_count == 1);
+ assert(only);
+
+ uint64_t hi = clause->constants[index + 0] >> 60ull;
struct bifrost_fmt_constant quad = {
.pos = 0, /* TODO */
.tag = done ? BIFROST_FMTC_FINAL : BIFROST_FMTC_CONSTANTS,
.imm_1 = clause->constants[index + 0] >> 4,
- .imm_2 = only ? 0 : clause->constants[index + 1] >> 4
+ .imm_2 = ((hi < 8) ? (hi << 60ull) : 0) >> 4,
};
/* XXX: On G71, Connor observed that the difference of the top 4 bits
* of the second constant with the first must be less than 8, otherwise
- * we have to swap them. I am not able to reproduce this on G52,
- * further investigation needed. Possibly an errata. XXX */
+ * we have to swap them. On G52, I'm able to reproduce a similar issue
+ * but with a different workaround (modeled above with a single
+ * constant, unclear how to workaround for multiple constants.) Further
+ * investigation needed. Possibly an errata. XXX */
util_dynarray_append(emission, struct bifrost_fmt_constant, quad);