*/
#include "compiler.h"
+#include "bi_print.h"
#define RETURN_PACKED(str) { \
uint64_t temp = 0; \
.scoreboard_index = clause->scoreboard_id,
.clause_type = clause->clause_type,
.next_clause_type = next ? next->clause_type : 0,
+ .suppress_inf = true,
+ .suppress_nan = true,
};
header.branch_cond |= header.back_to_back;
return u;
}
-/* Represents the assignment of ports for a given bundle */
-
-struct bi_registers {
- /* Register to assign to each port */
- unsigned port[4];
-
- /* Read ports can be disabled */
- bool enabled[2];
-
- /* Should we write FMA? what about ADD? If only a single port is
- * enabled it is in port 2, else ADD/FMA is 2/3 respectively */
- bool write_fma, write_add;
-
- /* Should we read with port 3? */
- bool read_port3;
-
- /* Packed uniform/constant */
- uint8_t uniform_constant;
-
- /* Whether writes are actually for the last instruction */
- bool first_instruction;
-};
-
-static inline void
-bi_print_ports(struct bi_registers *regs)
-{
- for (unsigned i = 0; i < 2; ++i) {
- if (regs->enabled[i])
- printf("port %u: %u\n", i, regs->port[i]);
- }
-
- if (regs->write_fma || regs->write_add) {
- printf("port 2 (%s): %u\n",
- regs->write_add ? "ADD" : "FMA",
- regs->port[2]);
- }
-
- if ((regs->write_fma && regs->write_add) || regs->read_port3) {
- printf("port 3 (%s): %u\n",
- regs->read_port3 ? "read" : "FMA",
- regs->port[3]);
- }
-}
-
/* The uniform/constant slot allows loading a contiguous 64-bit immediate or
* pushed uniform per bundle. Figure out which one we need in the bundle (the
* scheduler needs to ensure we only have one type per bundle), validate
static bool
bi_assign_uniform_constant_single(
- struct bi_registers *regs,
+ bi_registers *regs,
bi_clause *clause,
bi_instruction *ins, bool assigned, bool fast_zero)
{
if (s == 0 && (ins->type == BI_LOAD_VAR_ADDRESS || ins->type == BI_LOAD_ATTR)) continue;
if (ins->src[s] & BIR_INDEX_CONSTANT) {
+ /* Let direct addresses through */
+ if (ins->type == BI_LOAD_VAR)
+ continue;
+
bool hi = false;
bool b64 = nir_alu_type_get_type_size(ins->src_types[s]) > 32;
uint64_t cons = bi_get_immediate(ins, s);
regs->uniform_constant = f;
ins->src[s] = BIR_INDEX_PASS | BIFROST_SRC_CONST_LO;
assigned = true;
+ } else if (ins->src[s] & BIR_INDEX_ZERO && fast_zero) {
+ ins->src[s] = BIR_INDEX_PASS | BIFROST_SRC_STAGE;
} else if (s & BIR_INDEX_UNIFORM) {
unreachable("Push uniforms not implemented yet");
}
static void
bi_assign_uniform_constant(
bi_clause *clause,
- struct bi_registers *regs,
+ bi_registers *regs,
bi_bundle bundle)
{
bool assigned =
/* Assigns a port for reading, before anything is written */
static void
-bi_assign_port_read(struct bi_registers *regs, unsigned src)
+bi_assign_port_read(bi_registers *regs, unsigned src)
{
/* We only assign for registers */
if (!(src & BIR_INDEX_REGISTER))
return;
}
- bi_print_ports(regs);
+ bi_print_ports(regs, stderr);
unreachable("Failed to find a free port for src");
}
-static struct bi_registers
-bi_assign_ports(bi_bundle now, bi_bundle prev)
+static bi_registers
+bi_assign_ports(bi_bundle *now, bi_bundle *prev)
{
- struct bi_registers regs = { 0 };
-
/* We assign ports for the main register mechanism. Special ops
* use the data registers, which has its own mechanism entirely
* and thus gets skipped over here. */
- unsigned read_dreg = now.add &&
- bi_class_props[now.add->type] & BI_DATA_REG_SRC;
+ unsigned read_dreg = now->add &&
+ bi_class_props[now->add->type] & BI_DATA_REG_SRC;
- unsigned write_dreg = prev.add &&
- bi_class_props[prev.add->type] & BI_DATA_REG_DEST;
+ unsigned write_dreg = prev->add &&
+ bi_class_props[prev->add->type] & BI_DATA_REG_DEST;
/* First, assign reads */
- if (now.fma)
- bi_foreach_src(now.fma, src)
- bi_assign_port_read(®s, now.fma->src[src]);
+ if (now->fma)
+ bi_foreach_src(now->fma, src)
+ bi_assign_port_read(&now->regs, now->fma->src[src]);
- if (now.add) {
- bi_foreach_src(now.add, src) {
+ if (now->add) {
+ bi_foreach_src(now->add, src) {
if (!(src == 0 && read_dreg))
- bi_assign_port_read(®s, now.add->src[src]);
+ bi_assign_port_read(&now->regs, now->add->src[src]);
}
}
/* Next, assign writes */
- if (prev.add && prev.add->dest & BIR_INDEX_REGISTER && !write_dreg) {
- regs.port[2] = prev.add->dest & ~BIR_INDEX_REGISTER;
- regs.write_add = true;
+ if (prev->add && prev->add->dest & BIR_INDEX_REGISTER && !write_dreg) {
+ now->regs.port[2] = prev->add->dest & ~BIR_INDEX_REGISTER;
+ now->regs.write_add = true;
}
- if (prev.fma && prev.fma->dest & BIR_INDEX_REGISTER) {
- unsigned r = prev.fma->dest & ~BIR_INDEX_REGISTER;
+ if (prev->fma && prev->fma->dest & BIR_INDEX_REGISTER) {
+ unsigned r = prev->fma->dest & ~BIR_INDEX_REGISTER;
- if (regs.write_add) {
+ if (now->regs.write_add) {
/* Scheduler constraint: cannot read 3 and write 2 */
- assert(!regs.read_port3);
- regs.port[3] = r;
+ assert(!now->regs.read_port3);
+ now->regs.port[3] = r;
} else {
- regs.port[2] = r;
+ now->regs.port[2] = r;
}
- regs.write_fma = true;
+ now->regs.write_fma = true;
}
/* Finally, ensure port 1 > port 0 for the 63-x trick to function */
- if (regs.enabled[0] && regs.enabled[1] && regs.port[1] < regs.port[0]) {
- unsigned temp = regs.port[0];
- regs.port[0] = regs.port[1];
- regs.port[1] = temp;
+ if (now->regs.enabled[0] && now->regs.enabled[1] && now->regs.port[1] < now->regs.port[0]) {
+ unsigned temp = now->regs.port[0];
+ now->regs.port[0] = now->regs.port[1];
+ now->regs.port[1] = temp;
}
- return regs;
+ return now->regs;
}
/* Determines the register control field, ignoring the first? flag */
static enum bifrost_reg_control
-bi_pack_register_ctrl_lo(struct bi_registers r)
+bi_pack_register_ctrl_lo(bi_registers r)
{
if (r.write_fma) {
if (r.write_add) {
/* Ditto but account for the first? flag this time */
static enum bifrost_reg_control
-bi_pack_register_ctrl(struct bi_registers r)
+bi_pack_register_ctrl(bi_registers r)
{
enum bifrost_reg_control ctrl = bi_pack_register_ctrl_lo(r);
}
static uint64_t
-bi_pack_registers(struct bi_registers regs)
+bi_pack_registers(bi_registers regs)
{
enum bifrost_reg_control ctrl = bi_pack_register_ctrl(regs);
struct bifrost_regs s = { 0 };
}
static enum bifrost_packed_src
-bi_get_src_reg_port(struct bi_registers *regs, unsigned src)
+bi_get_src_reg_port(bi_registers *regs, unsigned src)
{
unsigned reg = src & ~BIR_INDEX_REGISTER;
}
static enum bifrost_packed_src
-bi_get_src(bi_instruction *ins, struct bi_registers *regs, unsigned s, bool is_fma)
+bi_get_src(bi_instruction *ins, bi_registers *regs, unsigned s)
{
unsigned src = ins->src[s];
if (src & BIR_INDEX_REGISTER)
return bi_get_src_reg_port(regs, src);
- else if (src & BIR_INDEX_ZERO && is_fma)
- return BIFROST_SRC_STAGE;
else if (src & BIR_INDEX_PASS)
return src & ~BIR_INDEX_PASS;
- else
- unreachable("Unknown src");
+ else {
+ bi_print_instruction(ins, stderr);
+ unreachable("Unknown src in above instruction");
+ }
}
/* Constructs a packed 2-bit swizzle for a 16-bit vec2 source. Source must be
}
static unsigned
-bi_pack_fma_fma(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_fma_fma(bi_instruction *ins, bi_registers *regs)
{
/* (-a)(-b) = ab, so we only need one negate bit */
bool negate_mul = ins->src_neg[0] ^ ins->src_neg[1];
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),
+ .src0 = bi_get_src(ins, regs, flip_ab ? 1 : 0),
+ .src1 = bi_get_src(ins, regs, flip_ab ? 0 : 1),
+ .src2 = bi_get_src(ins, regs, 2),
+ .src3 = bi_get_src(ins, regs, 3),
.mscale_mode = 0,
.mode = ins->outmod,
.src0_abs = ins->src_abs[0] || ins->src_abs[1],
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),
- .src2 = bi_get_src(ins, regs, 2, true),
+ .src0 = bi_get_src(ins, regs, 0),
+ .src1 = bi_get_src(ins, regs, 1),
+ .src2 = bi_get_src(ins, regs, 2),
.src0_abs = ins->src_abs[0],
.src1_abs = ins->src_abs[1],
.src2_abs = ins->src_abs[2],
RETURN_PACKED(pack);
} else if (ins->dest_type == nir_type_float16) {
struct bifrost_fma_fma16 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),
+ .src0 = bi_get_src(ins, regs, 0),
+ .src1 = bi_get_src(ins, regs, 1),
+ .src2 = bi_get_src(ins, regs, 2),
.swizzle_0 = bi_swiz16(ins, 0),
.swizzle_1 = bi_swiz16(ins, 1),
.swizzle_2 = bi_swiz16(ins, 2),
}
static unsigned
-bi_pack_fma_addmin_f32(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_fma_addmin_f32(bi_instruction *ins, bi_registers *regs)
{
unsigned op =
(ins->type == BI_ADD) ? BIFROST_FMA_OP_FADD32 :
BIFROST_FMA_OP_FMAX32;
struct bifrost_fma_add pack = {
- .src0 = bi_get_src(ins, regs, 0, true),
- .src1 = bi_get_src(ins, regs, 1, true),
+ .src0 = bi_get_src(ins, regs, 0),
+ .src1 = bi_get_src(ins, regs, 1),
.src0_abs = ins->src_abs[0],
.src1_abs = ins->src_abs[1],
.src0_neg = ins->src_neg[0],
RETURN_PACKED(pack);
}
-static unsigned
-bi_pack_fmadd_min_f16(bi_instruction *ins, struct bi_registers *regs, bool FMA)
+static bool
+bi_pack_fp16_abs(bi_instruction *ins, bi_registers *regs, bool *flip)
{
- 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;
-
/* 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;
+ unsigned src_0 = bi_get_src(ins, regs, 0);
+ unsigned src_1 = bi_get_src(ins, regs, 1);
- assert(!(abs_0 && abs_1));
+ 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_1 < src_0);
+ return true;
}
+}
+
+static unsigned
+bi_pack_fmadd_min_f16(bi_instruction *ins, 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;
+
+ bool flip = false;
+ bool l = bi_pack_fp16_abs(ins, regs, &flip);
+ unsigned src_0 = bi_get_src(ins, regs, 0);
+ unsigned src_1 = bi_get_src(ins, regs, 1);
if (FMA) {
struct bifrost_fma_add_minmax16 pack = {
.src1 = flip ? src_0 : src_1,
.src0_neg = ins->src_neg[flip ? 1 : 0],
.src1_neg = ins->src_neg[flip ? 0 : 1],
+ .src0_swizzle = bi_swiz16(ins, flip ? 1 : 0),
+ .src1_swizzle = bi_swiz16(ins, flip ? 0 : 1),
.abs1 = l,
.outmod = ins->outmod,
.mode = (ins->type == BI_ADD) ? ins->roundmode : ins->minmax,
.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),
+ .src0_swizzle = bi_swiz16(ins, flip ? 1 : 0),
+ .src1_swizzle = bi_swiz16(ins, flip ? 0 : 1),
.mode = ins->minmax,
.op = op
};
}
static unsigned
-bi_pack_fma_addmin(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_fma_addmin(bi_instruction *ins, bi_registers *regs)
{
if (ins->dest_type == nir_type_float32)
return bi_pack_fma_addmin_f32(ins, regs);
}
static unsigned
-bi_pack_fma_1src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
+bi_pack_fma_1src(bi_instruction *ins, bi_registers *regs, unsigned op)
{
struct bifrost_fma_inst pack = {
- .src0 = bi_get_src(ins, regs, 0, true),
+ .src0 = bi_get_src(ins, regs, 0),
.op = op
};
}
static unsigned
-bi_pack_fma_2src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
+bi_pack_fma_2src(bi_instruction *ins, 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),
+ .src0 = bi_get_src(ins, regs, 0),
+ .src1 = bi_get_src(ins, regs, 1),
.op = op
};
}
static unsigned
-bi_pack_add_1src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
+bi_pack_add_1src(bi_instruction *ins, bi_registers *regs, unsigned op)
{
struct bifrost_add_inst pack = {
- .src0 = bi_get_src(ins, regs, 0, true),
+ .src0 = bi_get_src(ins, regs, 0),
.op = op
};
}
static unsigned
-bi_pack_fma_csel(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_fma_csel(bi_instruction *ins, bi_registers *regs)
{
/* TODO: Use csel3 as well */
bool flip = false, invert = false;
unsigned res_1 = (invert ? 2 : 3);
struct bifrost_csel4 pack = {
- .src0 = bi_get_src(ins, regs, cmp_0, true),
- .src1 = bi_get_src(ins, regs, cmp_1, true),
- .src2 = bi_get_src(ins, regs, res_0, true),
- .src3 = bi_get_src(ins, regs, res_1, true),
+ .src0 = bi_get_src(ins, regs, cmp_0),
+ .src1 = bi_get_src(ins, regs, cmp_1),
+ .src2 = bi_get_src(ins, regs, res_0),
+ .src3 = bi_get_src(ins, regs, res_1),
.cond = cond,
.op = (size == 16) ? BIFROST_FMA_OP_CSEL4_V16 :
BIFROST_FMA_OP_CSEL4
}
static unsigned
-bi_pack_fma_frexp(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_fma_frexp(bi_instruction *ins, 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)
+bi_pack_fma_reduce(bi_instruction *ins, bi_registers *regs)
{
if (ins->op.reduce == BI_REDUCE_ADD_FREXPM) {
return bi_pack_fma_2src(ins, regs, BIFROST_FMA_OP_ADD_FREXPM);
*/
static unsigned
-bi_pack_convert(bi_instruction *ins, struct bi_registers *regs, bool FMA)
+bi_pack_convert(bi_instruction *ins, 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]);
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),
+ .src0 = bi_get_src(ins, regs, 0),
.src1 = BIFROST_SRC_STAGE, /* 0 */
.op = BIFROST_FMA_FLOAT32_TO_16
};
struct bifrost_add_2src padd = {
- .src0 = bi_get_src(ins, regs, 0, true),
+ .src0 = bi_get_src(ins, regs, 0),
.src1 = BIFROST_SRC_STAGE, /* 0 */
.op = BIFROST_ADD_FLOAT32_TO_16
};
}
static unsigned
-bi_pack_fma_select(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_fma_select(bi_instruction *ins, bi_registers *regs)
{
unsigned size = nir_alu_type_get_type_size(ins->src_types[0]);
}
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),
+ .src0 = bi_get_src(ins, regs, 0),
+ .src1 = bi_get_src(ins, regs, 1),
+ .src2 = bi_get_src(ins, regs, 2),
+ .src3 = bi_get_src(ins, regs, 3),
.swizzle = swiz,
.op = BIFROST_FMA_OP_SEL8
};
}
}
+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, 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),
+ .src1 = bi_get_src(ins, regs, 1),
+ .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),
+ .src1 = bi_get_src(ins, regs, flip ? 0 : 1),
+ .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,
+ };
+
+ 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, 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),
+ .src1 = bi_get_src(ins, regs, 1),
+ .src2 = bi_get_src(ins, regs, 2),
+ .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_round(bi_instruction *ins, bi_registers *regs)
+{
+ bool fp16 = ins->dest_type == nir_type_float16;
+ assert(fp16 || ins->dest_type == nir_type_float32);
+
+ unsigned op = fp16
+ ? BIFROST_FMA_ROUND_16(ins->roundmode, bi_swiz16(ins, 0))
+ : BIFROST_FMA_ROUND_32(ins->roundmode);
+
+ return bi_pack_fma_1src(ins, regs, op);
+}
+
+static unsigned
+bi_pack_fma_imath(bi_instruction *ins, bi_registers *regs)
+{
+ /* Scheduler: only ADD can have 8/16-bit imath */
+ assert(ins->dest_type == nir_type_int32 || ins->dest_type == nir_type_uint32);
+
+ unsigned op = ins->op.imath == BI_IMATH_ADD
+ ? BIFROST_FMA_IADD_32
+ : BIFROST_FMA_ISUB_32;
+
+ return bi_pack_fma_2src(ins, regs, op);
+}
+
static unsigned
-bi_pack_fma(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
+bi_pack_fma(bi_clause *clause, bi_bundle bundle, bi_registers *regs)
{
if (!bundle.fma)
return BIFROST_FMA_NOP;
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_convert(bundle.fma, regs, true);
case BI_CSEL:
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_IMATH:
+ return bi_pack_fma_imath(bundle.fma, regs);
case BI_MINMAX:
return bi_pack_fma_addmin(bundle.fma, regs);
case BI_MOV:
return bi_pack_fma_1src(bundle.fma, regs, BIFROST_FMA_OP_MOV);
case BI_SHIFT:
- return BIFROST_FMA_NOP;
+ unreachable("Packing todo");
case BI_SELECT:
return bi_pack_fma_select(bundle.fma, regs);
case BI_ROUND:
- return BIFROST_FMA_NOP;
+ return bi_pack_fma_round(bundle.fma, regs);
case BI_REDUCE_FMA:
return bi_pack_fma_reduce(bundle.fma, regs);
default:
}
static unsigned
-bi_pack_add_ld_vary(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
+bi_pack_add_ld_vary(bi_clause *clause, bi_instruction *ins, bi_registers *regs)
{
unsigned size = nir_alu_type_get_type_size(ins->dest_type);
assert(size == 32 || size == 16);
if (ins->src[0] & BIR_INDEX_CONSTANT) {
/* Direct uses address field directly */
packed_addr = bi_get_immediate(ins, 0);
- assert(packed_addr < 0b1000);
} else {
/* Indirect gets an extra source */
- packed_addr = bi_get_src(ins, regs, 0, false) | 0b11000;
+ packed_addr = bi_get_src(ins, regs, 0) | 0b11000;
}
/* The destination is thrown in the data register */
assert(channels >= 1 && channels <= 4);
struct bifrost_ld_var pack = {
- .src0 = bi_get_src(ins, regs, 1, false),
+ .src0 = bi_get_src(ins, regs, 1),
.addr = packed_addr,
.channels = MALI_POSITIVE(channels),
.interp_mode = ins->load_vary.interp_mode,
}
static unsigned
-bi_pack_add_2src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
+bi_pack_add_2src(bi_instruction *ins, bi_registers *regs, unsigned op)
{
struct bifrost_add_2src pack = {
- .src0 = bi_get_src(ins, regs, 0, true),
- .src1 = bi_get_src(ins, regs, 1, true),
+ .src0 = bi_get_src(ins, regs, 0),
+ .src1 = bi_get_src(ins, regs, 1),
.op = op
};
}
static unsigned
-bi_pack_add_addmin_f32(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_add_addmin_f32(bi_instruction *ins, bi_registers *regs)
{
unsigned op =
(ins->type == BI_ADD) ? BIFROST_ADD_OP_FADD32 :
BIFROST_ADD_OP_FMAX32;
struct bifrost_add_faddmin pack = {
- .src0 = bi_get_src(ins, regs, 0, true),
- .src1 = bi_get_src(ins, regs, 1, true),
+ .src0 = bi_get_src(ins, regs, 0),
+ .src1 = bi_get_src(ins, regs, 1),
.src0_abs = ins->src_abs[0],
.src1_abs = ins->src_abs[1],
.src0_neg = ins->src_neg[0],
}
static unsigned
-bi_pack_add_add_f16(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_add_add_f16(bi_instruction *ins, 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 = bi_get_src(ins, regs, 0),
+ .src1 = bi_get_src(ins, regs, 1),
.src0_abs = ins->src_abs[0],
.src1_abs = ins->src_abs[1],
.src0_neg = ins->src_neg[0],
}
static unsigned
-bi_pack_add_addmin(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_add_addmin(bi_instruction *ins, bi_registers *regs)
{
if (ins->dest_type == nir_type_float32)
return bi_pack_add_addmin_f32(ins, regs);
}
static unsigned
-bi_pack_add_ld_ubo(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
+bi_pack_add_ld_ubo(bi_clause *clause, bi_instruction *ins, bi_registers *regs)
{
assert(ins->vector_channels >= 1 && ins->vector_channels <= 4);
}
static unsigned
-bi_pack_add_ld_var_addr(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
+bi_pack_add_ld_var_addr(bi_clause *clause, bi_instruction *ins, bi_registers *regs)
{
struct bifrost_ld_var_addr pack = {
- .src0 = bi_get_src(ins, regs, 1, false),
- .src1 = bi_get_src(ins, regs, 2, false),
+ .src0 = bi_get_src(ins, regs, 1),
+ .src1 = bi_get_src(ins, regs, 2),
.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)
+bi_pack_add_ld_attr(bi_clause *clause, bi_instruction *ins, 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),
+ .src0 = bi_get_src(ins, regs, 1),
+ .src1 = bi_get_src(ins, regs, 2),
.location = bi_get_immediate(ins, 0),
.channels = MALI_POSITIVE(ins->vector_channels),
.type = bi_pack_ldst_type(ins->dest_type),
}
static unsigned
-bi_pack_add_st_vary(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
+bi_pack_add_st_vary(bi_clause *clause, bi_instruction *ins, bi_registers *regs)
{
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),
+ .src0 = bi_get_src(ins, regs, 1),
+ .src1 = bi_get_src(ins, regs, 2),
+ .src2 = bi_get_src(ins, regs, 3),
.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)
+bi_pack_add_atest(bi_clause *clause, bi_instruction *ins, bi_registers *regs)
{
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),
+ .src0 = bi_get_src(ins, regs, 0),
+ .src1 = bi_get_src(ins, regs, 1),
.half = fp16,
.component = fp16 ? ins->swizzle[1][0] : 1, /* Set for fp32 */
.op = BIFROST_ADD_OP_ATEST,
}
static unsigned
-bi_pack_add_blend(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
+bi_pack_add_blend(bi_clause *clause, bi_instruction *ins, bi_registers *regs)
{
struct bifrost_add_inst pack = {
- .src0 = bi_get_src(ins, regs, 1, false),
+ .src0 = bi_get_src(ins, regs, 1),
.op = BIFROST_ADD_OP_BLEND
};
}
static unsigned
-bi_pack_add_special(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_add_special(bi_instruction *ins, bi_registers *regs)
{
unsigned op = 0;
bool fp16 = ins->dest_type == nir_type_float16;
}
static unsigned
-bi_pack_add_table(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_add_table(bi_instruction *ins, bi_registers *regs)
{
unsigned op = 0;
assert(ins->dest_type == nir_type_float32);
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)
+bi_pack_add_tex_compact(bi_clause *clause, bi_instruction *ins, bi_registers *regs, gl_shader_stage stage)
{
bool f16 = ins->dest_type == nir_type_float16;
+ bool vtx = stage != MESA_SHADER_FRAGMENT;
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
+ .src0 = bi_get_src(ins, regs, 0),
+ .src1 = bi_get_src(ins, regs, 1),
+ .op = f16 ? BIFROST_ADD_OP_TEX_COMPACT_F16(vtx) :
+ BIFROST_ADD_OP_TEX_COMPACT_F32(vtx),
+ .compute_lod = !vtx,
+ .tex_index = ins->texture.texture_index,
+ .sampler_index = ins->texture.sampler_index
};
bi_write_data_register(clause, ins);
}
static unsigned
-bi_pack_add_select(bi_instruction *ins, struct bi_registers *regs)
+bi_pack_add_select(bi_instruction *ins, bi_registers *regs)
{
unsigned size = nir_alu_type_get_type_size(ins->src_types[0]);
assert(size == 16);
return bi_pack_add_2src(ins, regs, op);
}
+static enum bifrost_discard_cond
+bi_cond_to_discard(enum bi_cond cond, bool *flip)
+{
+ switch (cond){
+ case BI_COND_GT:
+ *flip = true;
+ /* fallthrough */
+ case BI_COND_LT:
+ return BIFROST_DISCARD_FLT;
+ case BI_COND_GE:
+ *flip = true;
+ /* fallthrough */
+ case BI_COND_LE:
+ return BIFROST_DISCARD_FLE;
+ case BI_COND_NE:
+ return BIFROST_DISCARD_FNE;
+ case BI_COND_EQ:
+ return BIFROST_DISCARD_FEQ;
+ default:
+ unreachable("Invalid op for discard");
+ }
+}
+
+static unsigned
+bi_pack_add_discard(bi_instruction *ins, bi_registers *regs)
+{
+ bool fp16 = ins->src_types[0] == nir_type_float16;
+ assert(fp16 || ins->src_types[0] == nir_type_float32);
+
+ bool flip = false;
+ enum bifrost_discard_cond cond = bi_cond_to_discard(ins->cond, &flip);
+
+ struct bifrost_add_discard pack = {
+ .src0 = bi_get_src(ins, regs, flip ? 1 : 0),
+ .src1 = bi_get_src(ins, regs, flip ? 0 : 1),
+ .cond = cond,
+ .src0_select = fp16 ? ins->swizzle[0][0] : 0,
+ .src1_select = fp16 ? ins->swizzle[1][0] : 0,
+ .fp32 = fp16 ? 0 : 1,
+ .op = BIFROST_ADD_OP_DISCARD
+ };
+
+ RETURN_PACKED(pack);
+}
+
+static enum bifrost_icmp_cond
+bi_cond_to_icmp(enum bi_cond cond, bool *flip, bool is_unsigned, bool is_16)
+{
+ switch (cond){
+ case BI_COND_LT:
+ *flip = true;
+ /* fallthrough */
+ case BI_COND_GT:
+ return is_unsigned ? (is_16 ? BIFROST_ICMP_IGE : BIFROST_ICMP_UGT)
+ : BIFROST_ICMP_IGT;
+ case BI_COND_LE:
+ *flip = true;
+ /* fallthrough */
+ case BI_COND_GE:
+ return is_unsigned ? BIFROST_ICMP_UGE :
+ (is_16 ? BIFROST_ICMP_UGT : BIFROST_ICMP_IGE);
+ case BI_COND_NE:
+ return BIFROST_ICMP_NEQ;
+ case BI_COND_EQ:
+ return BIFROST_ICMP_EQ;
+ default:
+ unreachable("Invalid op for icmp");
+ }
+}
+
+static unsigned
+bi_pack_add_icmp32(bi_instruction *ins, bi_registers *regs, bool flip,
+ enum bifrost_icmp_cond cond)
+{
+ struct bifrost_add_icmp pack = {
+ .src0 = bi_get_src(ins, regs, flip ? 1 : 0),
+ .src1 = bi_get_src(ins, regs, flip ? 0 : 1),
+ .cond = cond,
+ .sz = 1,
+ .d3d = false,
+ .op = BIFROST_ADD_OP_ICMP_32
+ };
+
+ RETURN_PACKED(pack);
+}
+
+static unsigned
+bi_pack_add_icmp16(bi_instruction *ins, bi_registers *regs, bool flip,
+ enum bifrost_icmp_cond cond)
+{
+ struct bifrost_add_icmp16 pack = {
+ .src0 = bi_get_src(ins, regs, flip ? 1 : 0),
+ .src1 = bi_get_src(ins, regs, flip ? 0 : 1),
+ .src0_swizzle = bi_swiz16(ins, flip ? 1 : 0),
+ .src1_swizzle = bi_swiz16(ins, flip ? 0 : 1),
+ .cond = cond,
+ .d3d = false,
+ .op = BIFROST_ADD_OP_ICMP_16
+ };
+
+ RETURN_PACKED(pack);
+}
+
+static unsigned
+bi_pack_add_cmp(bi_instruction *ins, bi_registers *regs)
+{
+ nir_alu_type Tl = ins->src_types[0];
+ nir_alu_type Tr = ins->src_types[1];
+ nir_alu_type Bl = nir_alu_type_get_base_type(Tl);
+
+ if (Bl == nir_type_uint || Bl == nir_type_int) {
+ assert(Tl == Tr);
+ unsigned sz = nir_alu_type_get_type_size(Tl);
+
+ bool flip = false;
+
+ enum bifrost_icmp_cond cond = bi_cond_to_icmp(
+ sz == 16 ? /*bi_invert_cond*/(ins->cond) : ins->cond,
+ &flip, Bl == nir_type_uint, sz == 16);
+
+ if (sz == 32)
+ return bi_pack_add_icmp32(ins, regs, flip, cond);
+ else if (sz == 16)
+ return bi_pack_add_icmp16(ins, regs, flip, cond);
+ else
+ unreachable("TODO");
+ } else {
+ unreachable("TODO");
+ }
+}
+
+static unsigned
+bi_pack_add_imath(bi_instruction *ins, bi_registers *regs)
+{
+ /* TODO: 32+16 add */
+ assert(ins->src_types[0] == ins->src_types[1]);
+ unsigned sz = nir_alu_type_get_type_size(ins->src_types[0]);
+ enum bi_imath_op p = ins->op.imath;
+
+ unsigned op = 0;
+
+ if (sz == 8) {
+ op = (p == BI_IMATH_ADD) ? BIFROST_ADD_IADD_8 :
+ BIFROST_ADD_ISUB_8;
+ } else if (sz == 16) {
+ op = (p == BI_IMATH_ADD) ? BIFROST_ADD_IADD_16 :
+ BIFROST_ADD_ISUB_16;
+ } else if (sz == 32) {
+ op = (p == BI_IMATH_ADD) ? BIFROST_ADD_IADD_32 :
+ BIFROST_ADD_ISUB_32;
+ } else {
+ unreachable("64-bit todo");
+ }
+
+ return bi_pack_add_2src(ins, regs, op);
+}
+
static unsigned
-bi_pack_add(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
+bi_pack_add(bi_clause *clause, bi_bundle bundle, bi_registers *regs, gl_shader_stage stage)
{
if (!bundle.add)
return BIFROST_ADD_NOP;
case BI_ATEST:
return bi_pack_add_atest(clause, bundle.add, regs);
case BI_BRANCH:
+ unreachable("Packing todo");
case BI_CMP:
- return BIFROST_ADD_NOP;
+ return bi_pack_add_cmp(bundle.add, regs);
case BI_BLEND:
return bi_pack_add_blend(clause, bundle.add, regs);
case BI_BITWISE:
- return BIFROST_ADD_NOP;
+ unreachable("Packing todo");
case BI_CONVERT:
return bi_pack_convert(bundle.add, regs, false);
case BI_DISCARD:
+ return bi_pack_add_discard(bundle.add, regs);
case BI_FREXP:
- case BI_ISUB:
+ unreachable("Packing todo");
+ case BI_IMATH:
+ return bi_pack_add_imath(bundle.add, regs);
case BI_LOAD:
- return BIFROST_ADD_NOP;
+ unreachable("Packing todo");
case BI_LOAD_ATTR:
return bi_pack_add_ld_attr(clause, bundle.add, regs);
case BI_LOAD_UNIFORM:
case BI_MOV:
case BI_SHIFT:
case BI_STORE:
- return BIFROST_ADD_NOP;
+ unreachable("Packing todo");
case BI_STORE_VAR:
return bi_pack_add_st_vary(clause, bundle.add, regs);
case BI_SPECIAL:
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);
+ return bi_pack_add_tex_compact(clause, bundle.add, regs, stage);
else
unreachable("Unknown tex type");
case BI_ROUND:
- return BIFROST_ADD_NOP;
+ unreachable("Packing todo");
default:
unreachable("Cannot encode class as ADD");
}
};
static struct bi_packed_bundle
-bi_pack_bundle(bi_clause *clause, bi_bundle bundle, bi_bundle prev, bool first_bundle)
+bi_pack_bundle(bi_clause *clause, bi_bundle bundle, bi_bundle prev, bool first_bundle, gl_shader_stage stage)
{
- struct bi_registers regs = bi_assign_ports(bundle, prev);
- bi_assign_uniform_constant(clause, ®s, bundle);
- regs.first_instruction = first_bundle;
+ bi_assign_ports(&bundle, &prev);
+ bi_assign_uniform_constant(clause, &bundle.regs, bundle);
+ bundle.regs.first_instruction = first_bundle;
- uint64_t reg = bi_pack_registers(regs);
- uint64_t fma = bi_pack_fma(clause, bundle, ®s);
- uint64_t add = bi_pack_add(clause, bundle, ®s);
+ uint64_t reg = bi_pack_registers(bundle.regs);
+ uint64_t fma = bi_pack_fma(clause, bundle, &bundle.regs);
+ uint64_t add = bi_pack_add(clause, bundle, &bundle.regs, stage);
struct bi_packed_bundle packed = {
.lo = reg | (fma << 35) | ((add & 0b111111) << 58),
static void
bi_pack_clause(bi_context *ctx, bi_clause *clause, bi_clause *next,
- struct util_dynarray *emission)
+ struct util_dynarray *emission, gl_shader_stage stage)
{
- struct bi_packed_bundle ins_1 = bi_pack_bundle(clause, clause->bundles[0], clause->bundles[0], true);
+ struct bi_packed_bundle ins_1 = bi_pack_bundle(clause, clause->bundles[0], clause->bundles[0], true, stage);
assert(clause->bundle_count == 1);
/* Used to decide if we elide writes */
bi_foreach_clause_in_block(block, clause) {
bi_clause *next = bi_next_clause(ctx, _block, clause);
- bi_pack_clause(ctx, clause, next, emission);
+ bi_pack_clause(ctx, clause, next, emission, ctx->stage);
}
}
}