* IN THE SOFTWARE.
*/
-#include <stdio.h>
#include <inttypes.h>
#include "pipe/p_state.h"
#include "util/u_format.h"
#include "util/u_hash_table.h"
#include "util/u_hash.h"
#include "util/u_memory.h"
-#include "tgsi/tgsi_parse.h"
+#include "util/u_pack_color.h"
+#include "util/format_srgb.h"
+#include "util/ralloc.h"
#include "tgsi/tgsi_dump.h"
+#include "tgsi/tgsi_info.h"
+#include "tgsi/tgsi_lowering.h"
#include "vc4_context.h"
#include "vc4_qpu.h"
#include "vc4_qir.h"
-
-struct tgsi_to_qir {
- struct tgsi_parse_context parser;
- struct qcompile *c;
- struct qreg *temps;
- struct qreg *inputs;
- struct qreg *outputs;
- struct qreg *uniforms;
- struct qreg *consts;
- uint32_t num_consts;
-
- struct vc4_shader_state *shader_state;
- struct vc4_fs_key *fs_key;
- struct vc4_vs_key *vs_key;
-
- uint32_t *uniform_data;
- enum quniform_contents *uniform_contents;
- uint32_t num_uniforms;
- uint32_t num_inputs;
- uint32_t num_outputs;
-};
+#ifdef USE_VC4_SIMULATOR
+#include "simpenrose/simpenrose.h"
+#endif
struct vc4_key {
- struct vc4_shader_state *shader_state;
+ struct vc4_uncompiled_shader *shader_state;
+ struct {
+ enum pipe_format format;
+ unsigned compare_mode:1;
+ unsigned compare_func:3;
+ unsigned wrap_s:3;
+ unsigned wrap_t:3;
+ uint8_t swizzle[4];
+ } tex[VC4_MAX_TEXTURE_SAMPLERS];
};
struct vc4_fs_key {
struct vc4_key base;
enum pipe_format color_format;
+ bool depth_enabled;
+ bool stencil_enabled;
+ bool stencil_twoside;
+ bool stencil_full_writemasks;
+ bool is_points;
+ bool is_lines;
+ bool alpha_test;
+ bool point_coord_upper_left;
+ bool light_twoside;
+ uint8_t alpha_test_func;
+ uint32_t point_sprite_mask;
+
+ struct pipe_rt_blend_state blend;
};
struct vc4_vs_key {
struct vc4_key base;
enum pipe_format attr_formats[8];
+ bool per_vertex_point_size;
};
-static struct qreg
-get_temp_for_uniform(struct tgsi_to_qir *trans, uint32_t uniform)
+static void
+resize_qreg_array(struct vc4_compile *c,
+ struct qreg **regs,
+ uint32_t *size,
+ uint32_t decl_size)
{
- struct qcompile *c = trans->c;
- struct qreg u = { QFILE_UNIF, uniform };
+ if (*size >= decl_size)
+ return;
- struct qreg t = qir_MOV(c, u);
- trans->uniforms[uniform] = t;
- return t;
+ uint32_t old_size = *size;
+ *size = MAX2(*size * 2, decl_size);
+ *regs = reralloc(c, *regs, struct qreg, *size);
+ if (!*regs) {
+ fprintf(stderr, "Malloc failure\n");
+ abort();
+ }
+
+ for (uint32_t i = old_size; i < *size; i++)
+ (*regs)[i] = c->undef;
}
static struct qreg
-qir_uniform_ui(struct tgsi_to_qir *trans, uint32_t ui)
+add_uniform(struct vc4_compile *c,
+ enum quniform_contents contents,
+ uint32_t data)
{
- for (int i = 0; i < trans->num_uniforms; i++) {
- if (trans->uniform_contents[i] == QUNIFORM_CONSTANT &&
- trans->uniform_data[i] == ui)
- return trans->uniforms[i];
+ uint32_t uniform = c->num_uniforms++;
+ struct qreg u = { QFILE_UNIF, uniform };
+
+ if (uniform >= c->uniform_array_size) {
+ c->uniform_array_size = MAX2(MAX2(16, uniform + 1),
+ c->uniform_array_size * 2);
+
+ c->uniform_data = reralloc(c, c->uniform_data,
+ uint32_t,
+ c->uniform_array_size);
+ c->uniform_contents = reralloc(c, c->uniform_contents,
+ enum quniform_contents,
+ c->uniform_array_size);
}
- trans->uniform_contents[trans->num_uniforms] = QUNIFORM_CONSTANT;
- trans->uniform_data[trans->num_uniforms] = ui;
- return get_temp_for_uniform(trans, trans->num_uniforms++);
+ c->uniform_contents[uniform] = contents;
+ c->uniform_data[uniform] = data;
+
+ return u;
}
static struct qreg
-qir_uniform_f(struct tgsi_to_qir *trans, float f)
+get_temp_for_uniform(struct vc4_compile *c, enum quniform_contents contents,
+ uint32_t data)
{
- return qir_uniform_ui(trans, fui(f));
+ struct qreg u = add_uniform(c, contents, data);
+ struct qreg t = qir_MOV(c, u);
+ return t;
}
static struct qreg
-qir_uniform(struct tgsi_to_qir *trans, uint32_t index)
+qir_uniform_ui(struct vc4_compile *c, uint32_t ui)
{
- for (int i = 0; i < trans->num_uniforms; i++) {
- if (trans->uniform_contents[i] == QUNIFORM_UNIFORM &&
- trans->uniform_data[i] == index)
- return trans->uniforms[i];
- }
+ return get_temp_for_uniform(c, QUNIFORM_CONSTANT, ui);
+}
- trans->uniform_contents[trans->num_uniforms] = QUNIFORM_UNIFORM;
- trans->uniform_data[trans->num_uniforms] = index;
- return get_temp_for_uniform(trans, trans->num_uniforms++);
+static struct qreg
+qir_uniform_f(struct vc4_compile *c, float f)
+{
+ return qir_uniform_ui(c, fui(f));
}
static struct qreg
-get_src(struct tgsi_to_qir *trans, struct tgsi_src_register *src, int i)
+get_src(struct vc4_compile *c, unsigned tgsi_op,
+ struct tgsi_src_register *src, int i)
{
- struct qcompile *c = trans->c;
struct qreg r = c->undef;
uint32_t s = i;
case TGSI_FILE_NULL:
return r;
case TGSI_FILE_TEMPORARY:
- r = trans->temps[src->Index * 4 + s];
+ r = c->temps[src->Index * 4 + s];
break;
case TGSI_FILE_IMMEDIATE:
- r = trans->consts[src->Index * 4 + s];
+ r = c->consts[src->Index * 4 + s];
break;
case TGSI_FILE_CONSTANT:
- r = qir_uniform(trans, src->Index * 4 + s);
+ r = get_temp_for_uniform(c, QUNIFORM_UNIFORM,
+ src->Index * 4 + s);
break;
case TGSI_FILE_INPUT:
- r = trans->inputs[src->Index * 4 + s];
+ r = c->inputs[src->Index * 4 + s];
+ break;
+ case TGSI_FILE_SAMPLER:
+ case TGSI_FILE_SAMPLER_VIEW:
+ r = c->undef;
break;
default:
fprintf(stderr, "unknown src file %d\n", src->File);
if (src->Absolute)
r = qir_FMAXABS(c, r, r);
- if (src->Negate)
- r = qir_FSUB(c, qir_uniform_f(trans, 0), r);
+ if (src->Negate) {
+ switch (tgsi_opcode_infer_src_type(tgsi_op)) {
+ case TGSI_TYPE_SIGNED:
+ case TGSI_TYPE_UNSIGNED:
+ r = qir_SUB(c, qir_uniform_ui(c, 0), r);
+ break;
+ default:
+ r = qir_FSUB(c, qir_uniform_f(c, 0.0), r);
+ break;
+ }
+ }
return r;
};
static void
-update_dst(struct tgsi_to_qir *trans, struct tgsi_full_instruction *tgsi_inst,
+update_dst(struct vc4_compile *c, struct tgsi_full_instruction *tgsi_inst,
int i, struct qreg val)
{
struct tgsi_dst_register *tgsi_dst = &tgsi_inst->Dst[0].Register;
switch (tgsi_dst->File) {
case TGSI_FILE_TEMPORARY:
- trans->temps[tgsi_dst->Index * 4 + i] = val;
+ c->temps[tgsi_dst->Index * 4 + i] = val;
break;
case TGSI_FILE_OUTPUT:
- trans->outputs[tgsi_dst->Index * 4 + i] = val;
- trans->num_outputs = MAX2(trans->num_outputs,
- tgsi_dst->Index * 4 + i + 1);
+ c->outputs[tgsi_dst->Index * 4 + i] = val;
+ c->num_outputs = MAX2(c->num_outputs,
+ tgsi_dst->Index * 4 + i + 1);
break;
default:
fprintf(stderr, "unknown dst file %d\n", tgsi_dst->File);
};
static struct qreg
-tgsi_to_qir_alu(struct tgsi_to_qir *trans,
+get_swizzled_channel(struct vc4_compile *c,
+ struct qreg *srcs, int swiz)
+{
+ switch (swiz) {
+ default:
+ case UTIL_FORMAT_SWIZZLE_NONE:
+ fprintf(stderr, "warning: unknown swizzle\n");
+ /* FALLTHROUGH */
+ case UTIL_FORMAT_SWIZZLE_0:
+ return qir_uniform_f(c, 0.0);
+ case UTIL_FORMAT_SWIZZLE_1:
+ return qir_uniform_f(c, 1.0);
+ case UTIL_FORMAT_SWIZZLE_X:
+ case UTIL_FORMAT_SWIZZLE_Y:
+ case UTIL_FORMAT_SWIZZLE_Z:
+ case UTIL_FORMAT_SWIZZLE_W:
+ return srcs[swiz];
+ }
+}
+
+static struct qreg
+tgsi_to_qir_alu(struct vc4_compile *c,
struct tgsi_full_instruction *tgsi_inst,
enum qop op, struct qreg *src, int i)
{
- struct qcompile *c = trans->c;
struct qreg dst = qir_get_temp(c);
- qir_emit(c, qir_inst(op, dst, src[0 * 4 + i], src[1 * 4 + i]));
+ qir_emit(c, qir_inst4(op, dst,
+ src[0 * 4 + i],
+ src[1 * 4 + i],
+ src[2 * 4 + i],
+ c->undef));
return dst;
}
static struct qreg
-tgsi_to_qir_mad(struct tgsi_to_qir *trans,
+tgsi_to_qir_scalar(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ struct qreg dst = qir_get_temp(c);
+ qir_emit(c, qir_inst(op, dst,
+ src[0 * 4 + 0],
+ c->undef));
+ return dst;
+}
+
+static struct qreg
+qir_srgb_decode(struct vc4_compile *c, struct qreg srgb)
+{
+ struct qreg low = qir_FMUL(c, srgb, qir_uniform_f(c, 1.0 / 12.92));
+ struct qreg high = qir_POW(c,
+ qir_FMUL(c,
+ qir_FADD(c,
+ srgb,
+ qir_uniform_f(c, 0.055)),
+ qir_uniform_f(c, 1.0 / 1.055)),
+ qir_uniform_f(c, 2.4));
+
+ qir_SF(c, qir_FSUB(c, srgb, qir_uniform_f(c, 0.04045)));
+ return qir_SEL_X_Y_NS(c, low, high);
+}
+
+static struct qreg
+qir_srgb_encode(struct vc4_compile *c, struct qreg linear)
+{
+ struct qreg low = qir_FMUL(c, linear, qir_uniform_f(c, 12.92));
+ struct qreg high = qir_FSUB(c,
+ qir_FMUL(c,
+ qir_uniform_f(c, 1.055),
+ qir_POW(c,
+ linear,
+ qir_uniform_f(c, 0.41666))),
+ qir_uniform_f(c, 0.055));
+
+ qir_SF(c, qir_FSUB(c, linear, qir_uniform_f(c, 0.0031308)));
+ return qir_SEL_X_Y_NS(c, low, high);
+}
+
+static struct qreg
+tgsi_to_qir_umul(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ struct qreg src0_hi = qir_SHR(c, src[0 * 4 + i],
+ qir_uniform_ui(c, 16));
+ struct qreg src0_lo = qir_AND(c, src[0 * 4 + i],
+ qir_uniform_ui(c, 0xffff));
+ struct qreg src1_hi = qir_SHR(c, src[1 * 4 + i],
+ qir_uniform_ui(c, 16));
+ struct qreg src1_lo = qir_AND(c, src[1 * 4 + i],
+ qir_uniform_ui(c, 0xffff));
+
+ struct qreg hilo = qir_MUL24(c, src0_hi, src1_lo);
+ struct qreg lohi = qir_MUL24(c, src0_lo, src1_hi);
+ struct qreg lolo = qir_MUL24(c, src0_lo, src1_lo);
+
+ return qir_ADD(c, lolo, qir_SHL(c,
+ qir_ADD(c, hilo, lohi),
+ qir_uniform_ui(c, 16)));
+}
+
+static struct qreg
+tgsi_to_qir_idiv(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ return qir_FTOI(c, qir_FMUL(c,
+ qir_ITOF(c, src[0 * 4 + i]),
+ qir_RCP(c, qir_ITOF(c, src[1 * 4 + i]))));
+}
+
+static struct qreg
+tgsi_to_qir_ineg(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ return qir_SUB(c, qir_uniform_ui(c, 0), src[0 * 4 + i]);
+}
+
+static struct qreg
+tgsi_to_qir_seq(struct vc4_compile *c,
struct tgsi_full_instruction *tgsi_inst,
enum qop op, struct qreg *src, int i)
{
- struct qcompile *c = trans->c;
- return qir_FADD(c,
- qir_FMUL(c,
- src[0 * 4 + i],
- src[1 * 4 + i]),
- src[2 * 4 + i]);
+ qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_ZS(c, qir_uniform_f(c, 1.0));
}
static struct qreg
-tgsi_to_qir_dp(struct tgsi_to_qir *trans,
- struct tgsi_full_instruction *tgsi_inst,
- int num, struct qreg *src, int i)
+tgsi_to_qir_sne(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
{
- struct qcompile *c = trans->c;
+ qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_ZC(c, qir_uniform_f(c, 1.0));
+}
- struct qreg sum = qir_FMUL(c, src[0 * 4 + 0], src[1 * 4 + 0]);
- for (int j = 1; j < num; j++) {
- sum = qir_FADD(c, sum, qir_FMUL(c,
- src[0 * 4 + j],
- src[1 * 4 + j]));
- }
- return sum;
+static struct qreg
+tgsi_to_qir_slt(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_NS(c, qir_uniform_f(c, 1.0));
+}
+
+static struct qreg
+tgsi_to_qir_sge(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_NC(c, qir_uniform_f(c, 1.0));
+}
+
+static struct qreg
+tgsi_to_qir_fseq(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_ZS(c, qir_uniform_ui(c, ~0));
+}
+
+static struct qreg
+tgsi_to_qir_fsne(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_ZC(c, qir_uniform_ui(c, ~0));
+}
+
+static struct qreg
+tgsi_to_qir_fslt(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_NS(c, qir_uniform_ui(c, ~0));
+}
+
+static struct qreg
+tgsi_to_qir_fsge(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ qir_SF(c, qir_FSUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_NC(c, qir_uniform_ui(c, ~0));
+}
+
+static struct qreg
+tgsi_to_qir_useq(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ qir_SF(c, qir_SUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_ZS(c, qir_uniform_ui(c, ~0));
+}
+
+static struct qreg
+tgsi_to_qir_usne(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ qir_SF(c, qir_SUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_ZC(c, qir_uniform_ui(c, ~0));
}
static struct qreg
-tgsi_to_qir_dp2(struct tgsi_to_qir *trans,
+tgsi_to_qir_islt(struct vc4_compile *c,
struct tgsi_full_instruction *tgsi_inst,
enum qop op, struct qreg *src, int i)
{
- return tgsi_to_qir_dp(trans, tgsi_inst, 2, src, i);
+ qir_SF(c, qir_SUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_NS(c, qir_uniform_ui(c, ~0));
}
static struct qreg
-tgsi_to_qir_dp3(struct tgsi_to_qir *trans,
+tgsi_to_qir_isge(struct vc4_compile *c,
struct tgsi_full_instruction *tgsi_inst,
enum qop op, struct qreg *src, int i)
{
- return tgsi_to_qir_dp(trans, tgsi_inst, 3, src, i);
+ qir_SF(c, qir_SUB(c, src[0 * 4 + i], src[1 * 4 + i]));
+ return qir_SEL_X_0_NC(c, qir_uniform_ui(c, ~0));
+}
+
+static struct qreg
+tgsi_to_qir_cmp(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ qir_SF(c, src[0 * 4 + i]);
+ return qir_SEL_X_Y_NS(c,
+ src[1 * 4 + i],
+ src[2 * 4 + i]);
}
static struct qreg
-tgsi_to_qir_dp4(struct tgsi_to_qir *trans,
+tgsi_to_qir_mad(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ return qir_FADD(c,
+ qir_FMUL(c,
+ src[0 * 4 + i],
+ src[1 * 4 + i]),
+ src[2 * 4 + i]);
+}
+
+static struct qreg
+tgsi_to_qir_lrp(struct vc4_compile *c,
struct tgsi_full_instruction *tgsi_inst,
enum qop op, struct qreg *src, int i)
{
- return tgsi_to_qir_dp(trans, tgsi_inst, 4, src, i);
+ struct qreg src0 = src[0 * 4 + i];
+ struct qreg src1 = src[1 * 4 + i];
+ struct qreg src2 = src[2 * 4 + i];
+
+ /* LRP is:
+ * src0 * src1 + (1 - src0) * src2.
+ * -> src0 * src1 + src2 - src0 * src2
+ * -> src2 + src0 * (src1 - src2)
+ */
+ return qir_FADD(c, src2, qir_FMUL(c, src0, qir_FSUB(c, src1, src2)));
+
+}
+
+static void
+tgsi_to_qir_tex(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src)
+{
+ assert(!tgsi_inst->Instruction.Saturate);
+
+ struct qreg s = src[0 * 4 + 0];
+ struct qreg t = src[0 * 4 + 1];
+ struct qreg r = src[0 * 4 + 2];
+ uint32_t unit = tgsi_inst->Src[1].Register.Index;
+
+ struct qreg proj = c->undef;
+ if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_TXP) {
+ proj = qir_RCP(c, src[0 * 4 + 3]);
+ s = qir_FMUL(c, s, proj);
+ t = qir_FMUL(c, t, proj);
+ }
+
+ struct qreg texture_u[] = {
+ add_uniform(c, QUNIFORM_TEXTURE_CONFIG_P0, unit),
+ add_uniform(c, QUNIFORM_TEXTURE_CONFIG_P1, unit),
+ add_uniform(c, QUNIFORM_CONSTANT, 0),
+ add_uniform(c, QUNIFORM_CONSTANT, 0),
+ };
+ uint32_t next_texture_u = 0;
+
+ /* There is no native support for GL texture rectangle coordinates, so
+ * we have to rescale from ([0, width], [0, height]) to ([0, 1], [0,
+ * 1]).
+ */
+ if (tgsi_inst->Texture.Texture == TGSI_TEXTURE_RECT ||
+ tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOWRECT) {
+ s = qir_FMUL(c, s,
+ get_temp_for_uniform(c,
+ QUNIFORM_TEXRECT_SCALE_X,
+ unit));
+ t = qir_FMUL(c, t,
+ get_temp_for_uniform(c,
+ QUNIFORM_TEXRECT_SCALE_Y,
+ unit));
+ }
+
+ if (tgsi_inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
+ tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE) {
+ struct qreg ma = qir_FMAXABS(c, qir_FMAXABS(c, s, t), r);
+ struct qreg rcp_ma = qir_RCP(c, ma);
+ s = qir_FMUL(c, s, rcp_ma);
+ t = qir_FMUL(c, t, rcp_ma);
+ r = qir_FMUL(c, r, rcp_ma);
+
+ texture_u[2] = add_uniform(c, QUNIFORM_TEXTURE_CONFIG_P2, unit);
+
+ qir_TEX_R(c, r, texture_u[next_texture_u++]);
+ } else if (c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
+ c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP ||
+ c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
+ c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP) {
+ qir_TEX_R(c, get_temp_for_uniform(c, QUNIFORM_TEXTURE_BORDER_COLOR, unit),
+ texture_u[next_texture_u++]);
+ }
+
+ if (c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP) {
+ s = qir_FMIN(c, qir_FMAX(c, s, qir_uniform_f(c, 0.0)),
+ qir_uniform_f(c, 1.0));
+ }
+
+ if (c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP) {
+ t = qir_FMIN(c, qir_FMAX(c, t, qir_uniform_f(c, 0.0)),
+ qir_uniform_f(c, 1.0));
+ }
+
+ qir_TEX_T(c, t, texture_u[next_texture_u++]);
+
+ if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_TXB)
+ qir_TEX_B(c, src[0 * 4 + 3], texture_u[next_texture_u++]);
+
+ qir_TEX_S(c, s, texture_u[next_texture_u++]);
+
+ c->num_texture_samples++;
+ struct qreg r4 = qir_TEX_RESULT(c);
+
+ enum pipe_format format = c->key->tex[unit].format;
+
+ struct qreg unpacked[4];
+ if (util_format_is_depth_or_stencil(format)) {
+ struct qreg depthf = qir_ITOF(c, qir_SHR(c, r4,
+ qir_uniform_ui(c, 8)));
+ struct qreg normalized = qir_FMUL(c, depthf,
+ qir_uniform_f(c, 1.0f/0xffffff));
+
+ struct qreg depth_output;
+
+ struct qreg one = qir_uniform_f(c, 1.0f);
+ if (c->key->tex[unit].compare_mode) {
+ struct qreg compare = src[0 * 4 + 2];
+
+ if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_TXP)
+ compare = qir_FMUL(c, compare, proj);
+
+ switch (c->key->tex[unit].compare_func) {
+ case PIPE_FUNC_NEVER:
+ depth_output = qir_uniform_f(c, 0.0f);
+ break;
+ case PIPE_FUNC_ALWAYS:
+ depth_output = one;
+ break;
+ case PIPE_FUNC_EQUAL:
+ qir_SF(c, qir_FSUB(c, compare, normalized));
+ depth_output = qir_SEL_X_0_ZS(c, one);
+ break;
+ case PIPE_FUNC_NOTEQUAL:
+ qir_SF(c, qir_FSUB(c, compare, normalized));
+ depth_output = qir_SEL_X_0_ZC(c, one);
+ break;
+ case PIPE_FUNC_GREATER:
+ qir_SF(c, qir_FSUB(c, compare, normalized));
+ depth_output = qir_SEL_X_0_NC(c, one);
+ break;
+ case PIPE_FUNC_GEQUAL:
+ qir_SF(c, qir_FSUB(c, normalized, compare));
+ depth_output = qir_SEL_X_0_NS(c, one);
+ break;
+ case PIPE_FUNC_LESS:
+ qir_SF(c, qir_FSUB(c, compare, normalized));
+ depth_output = qir_SEL_X_0_NS(c, one);
+ break;
+ case PIPE_FUNC_LEQUAL:
+ qir_SF(c, qir_FSUB(c, normalized, compare));
+ depth_output = qir_SEL_X_0_NC(c, one);
+ break;
+ }
+ } else {
+ depth_output = normalized;
+ }
+
+ for (int i = 0; i < 4; i++)
+ unpacked[i] = depth_output;
+ } else {
+ for (int i = 0; i < 4; i++)
+ unpacked[i] = qir_R4_UNPACK(c, r4, i);
+ }
+
+ const uint8_t *format_swiz = vc4_get_format_swizzle(format);
+ struct qreg texture_output[4];
+ for (int i = 0; i < 4; i++) {
+ texture_output[i] = get_swizzled_channel(c, unpacked,
+ format_swiz[i]);
+ }
+
+ if (util_format_is_srgb(format)) {
+ for (int i = 0; i < 3; i++)
+ texture_output[i] = qir_srgb_decode(c,
+ texture_output[i]);
+ }
+
+ for (int i = 0; i < 4; i++) {
+ if (!(tgsi_inst->Dst[0].Register.WriteMask & (1 << i)))
+ continue;
+
+ update_dst(c, tgsi_inst, i,
+ get_swizzled_channel(c, texture_output,
+ c->key->tex[unit].swizzle[i]));
+ }
}
static struct qreg
-tgsi_to_qir_abs(struct tgsi_to_qir *trans,
+tgsi_to_qir_trunc(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ return qir_ITOF(c, qir_FTOI(c, src[0 * 4 + i]));
+}
+
+/**
+ * Computes x - floor(x), which is tricky because our FTOI truncates (rounds
+ * to zero).
+ */
+static struct qreg
+tgsi_to_qir_frc(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src[0 * 4 + i]));
+ struct qreg diff = qir_FSUB(c, src[0 * 4 + i], trunc);
+ qir_SF(c, diff);
+ return qir_SEL_X_Y_NS(c,
+ qir_FADD(c, diff, qir_uniform_f(c, 1.0)),
+ diff);
+}
+
+/**
+ * Computes floor(x), which is tricky because our FTOI truncates (rounds to
+ * zero).
+ */
+static struct qreg
+tgsi_to_qir_flr(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src[0 * 4 + i]));
+
+ /* This will be < 0 if we truncated and the truncation was of a value
+ * that was < 0 in the first place.
+ */
+ qir_SF(c, qir_FSUB(c, src[0 * 4 + i], trunc));
+
+ return qir_SEL_X_Y_NS(c,
+ qir_FSUB(c, trunc, qir_uniform_f(c, 1.0)),
+ trunc);
+}
+
+static struct qreg
+tgsi_to_qir_abs(struct vc4_compile *c,
struct tgsi_full_instruction *tgsi_inst,
enum qop op, struct qreg *src, int i)
{
- struct qcompile *c = trans->c;
struct qreg arg = src[0 * 4 + i];
return qir_FMAXABS(c, arg, arg);
}
+/* Note that this instruction replicates its result from the x channel */
+static struct qreg
+tgsi_to_qir_sin(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ float coeff[] = {
+ 2.0 * M_PI,
+ -pow(2.0 * M_PI, 3) / (3 * 2 * 1),
+ pow(2.0 * M_PI, 5) / (5 * 4 * 3 * 2 * 1),
+ -pow(2.0 * M_PI, 7) / (7 * 6 * 5 * 4 * 3 * 2 * 1),
+ };
+
+ struct qreg scaled_x =
+ qir_FMUL(c,
+ src[0 * 4 + 0],
+ qir_uniform_f(c, 1.0f / (M_PI * 2.0f)));
+
+
+ struct qreg x = tgsi_to_qir_frc(c, NULL, 0, &scaled_x, 0);
+ struct qreg x2 = qir_FMUL(c, x, x);
+ struct qreg sum = qir_FMUL(c, x, qir_uniform_f(c, coeff[0]));
+ for (int i = 1; i < ARRAY_SIZE(coeff); i++) {
+ x = qir_FMUL(c, x, x2);
+ sum = qir_FADD(c,
+ sum,
+ qir_FMUL(c,
+ x,
+ qir_uniform_f(c, coeff[i])));
+ }
+ return sum;
+}
+
+/* Note that this instruction replicates its result from the x channel */
+static struct qreg
+tgsi_to_qir_cos(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ float coeff[] = {
+ 1.0f,
+ -pow(2.0 * M_PI, 2) / (2 * 1),
+ pow(2.0 * M_PI, 4) / (4 * 3 * 2 * 1),
+ -pow(2.0 * M_PI, 6) / (6 * 5 * 4 * 3 * 2 * 1),
+ };
+
+ struct qreg scaled_x =
+ qir_FMUL(c, src[0 * 4 + 0],
+ qir_uniform_f(c, 1.0f / (M_PI * 2.0f)));
+ struct qreg x_frac = tgsi_to_qir_frc(c, NULL, 0, &scaled_x, 0);
+
+ struct qreg sum = qir_uniform_f(c, coeff[0]);
+ struct qreg x2 = qir_FMUL(c, x_frac, x_frac);
+ struct qreg x = x2; /* Current x^2, x^4, or x^6 */
+ for (int i = 1; i < ARRAY_SIZE(coeff); i++) {
+ if (i != 1)
+ x = qir_FMUL(c, x, x2);
+
+ struct qreg mul = qir_FMUL(c,
+ x,
+ qir_uniform_f(c, coeff[i]));
+ if (i == 0)
+ sum = mul;
+ else
+ sum = qir_FADD(c, sum, mul);
+ }
+ return sum;
+}
+
+static struct qreg
+tgsi_to_qir_clamp(struct vc4_compile *c,
+ struct tgsi_full_instruction *tgsi_inst,
+ enum qop op, struct qreg *src, int i)
+{
+ return qir_FMAX(c, qir_FMIN(c,
+ src[0 * 4 + i],
+ src[2 * 4 + i]),
+ src[1 * 4 + i]);
+}
+
static void
-emit_tgsi_declaration(struct tgsi_to_qir *trans,
- struct tgsi_full_declaration *decl)
+emit_vertex_input(struct vc4_compile *c, int attr)
+{
+ enum pipe_format format = c->vs_key->attr_formats[attr];
+ struct qreg vpm_reads[4];
+
+ /* Right now, we're setting the VPM offsets to be 16 bytes wide every
+ * time, so we always read 4 32-bit VPM entries.
+ */
+ for (int i = 0; i < 4; i++) {
+ vpm_reads[i] = qir_get_temp(c);
+ qir_emit(c, qir_inst(QOP_VPM_READ,
+ vpm_reads[i],
+ c->undef,
+ c->undef));
+ c->num_inputs++;
+ }
+
+ bool format_warned = false;
+ const struct util_format_description *desc =
+ util_format_description(format);
+
+ for (int i = 0; i < 4; i++) {
+ uint8_t swiz = desc->swizzle[i];
+ struct qreg result;
+
+ if (swiz > UTIL_FORMAT_SWIZZLE_W)
+ result = get_swizzled_channel(c, vpm_reads, swiz);
+ else if (desc->channel[swiz].size == 32 &&
+ desc->channel[swiz].type == UTIL_FORMAT_TYPE_FLOAT) {
+ result = get_swizzled_channel(c, vpm_reads, swiz);
+ } else if (desc->channel[swiz].size == 8 &&
+ (desc->channel[swiz].type == UTIL_FORMAT_TYPE_UNSIGNED ||
+ desc->channel[swiz].type == UTIL_FORMAT_TYPE_SIGNED) &&
+ desc->channel[swiz].normalized) {
+ struct qreg vpm = vpm_reads[0];
+ if (desc->channel[swiz].type == UTIL_FORMAT_TYPE_SIGNED)
+ vpm = qir_XOR(c, vpm, qir_uniform_ui(c, 0x80808080));
+ result = qir_UNPACK_8(c, vpm, swiz);
+ } else {
+ if (!format_warned) {
+ fprintf(stderr,
+ "vtx element %d unsupported type: %s\n",
+ attr, util_format_name(format));
+ format_warned = true;
+ }
+ result = qir_uniform_f(c, 0.0);
+ }
+
+ if (desc->channel[swiz].normalized &&
+ desc->channel[swiz].type == UTIL_FORMAT_TYPE_SIGNED) {
+ result = qir_FSUB(c,
+ qir_FMUL(c,
+ result,
+ qir_uniform_f(c, 2.0)),
+ qir_uniform_f(c, 1.0));
+ }
+
+ c->inputs[attr * 4 + i] = result;
+ }
+}
+
+static void
+tgsi_to_qir_kill_if(struct vc4_compile *c, struct qreg *src, int i)
+{
+ if (c->discard.file == QFILE_NULL)
+ c->discard = qir_uniform_f(c, 0.0);
+ qir_SF(c, src[0 * 4 + i]);
+ c->discard = qir_SEL_X_Y_NS(c, qir_uniform_f(c, 1.0),
+ c->discard);
+}
+
+static void
+emit_fragcoord_input(struct vc4_compile *c, int attr)
+{
+ c->inputs[attr * 4 + 0] = qir_FRAG_X(c);
+ c->inputs[attr * 4 + 1] = qir_FRAG_Y(c);
+ c->inputs[attr * 4 + 2] =
+ qir_FMUL(c,
+ qir_ITOF(c, qir_FRAG_Z(c)),
+ qir_uniform_f(c, 1.0 / 0xffffff));
+ c->inputs[attr * 4 + 3] = qir_RCP(c, qir_FRAG_W(c));
+}
+
+static void
+emit_point_coord_input(struct vc4_compile *c, int attr)
+{
+ if (c->point_x.file == QFILE_NULL) {
+ c->point_x = qir_uniform_f(c, 0.0);
+ c->point_y = qir_uniform_f(c, 0.0);
+ }
+
+ c->inputs[attr * 4 + 0] = c->point_x;
+ if (c->fs_key->point_coord_upper_left) {
+ c->inputs[attr * 4 + 1] = qir_FSUB(c,
+ qir_uniform_f(c, 1.0),
+ c->point_y);
+ } else {
+ c->inputs[attr * 4 + 1] = c->point_y;
+ }
+ c->inputs[attr * 4 + 2] = qir_uniform_f(c, 0.0);
+ c->inputs[attr * 4 + 3] = qir_uniform_f(c, 1.0);
+}
+
+static struct qreg
+emit_fragment_varying(struct vc4_compile *c, int index)
{
- struct qcompile *c = trans->c;
+ struct qreg vary = {
+ QFILE_VARY,
+ index
+ };
+
+ return qir_VARY_ADD_C(c, qir_FMUL(c, vary, qir_FRAG_W(c)));
+}
+static void
+emit_fragment_input(struct vc4_compile *c, int attr,
+ struct tgsi_full_declaration *decl)
+{
+ for (int i = 0; i < 4; i++) {
+ c->inputs[attr * 4 + i] =
+ emit_fragment_varying(c, attr * 4 + i);
+ c->num_inputs++;
+
+ if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR ||
+ decl->Semantic.Name == TGSI_SEMANTIC_BCOLOR)
+ c->color_inputs |= 1 << i;
+ }
+}
+
+static void
+emit_face_input(struct vc4_compile *c, int attr)
+{
+ c->inputs[attr * 4 + 0] = qir_FSUB(c,
+ qir_uniform_f(c, 1.0),
+ qir_FMUL(c,
+ qir_ITOF(c, qir_FRAG_REV_FLAG(c)),
+ qir_uniform_f(c, 2.0)));
+ c->inputs[attr * 4 + 1] = qir_uniform_f(c, 0.0);
+ c->inputs[attr * 4 + 2] = qir_uniform_f(c, 0.0);
+ c->inputs[attr * 4 + 3] = qir_uniform_f(c, 1.0);
+}
+
+static void
+emit_tgsi_declaration(struct vc4_compile *c,
+ struct tgsi_full_declaration *decl)
+{
switch (decl->Declaration.File) {
+ case TGSI_FILE_TEMPORARY: {
+ uint32_t old_size = c->temps_array_size;
+ resize_qreg_array(c, &c->temps, &c->temps_array_size,
+ (decl->Range.Last + 1) * 4);
+
+ for (int i = old_size; i < c->temps_array_size; i++)
+ c->temps[i] = qir_uniform_ui(c, 0);
+ break;
+ }
+
case TGSI_FILE_INPUT:
- if (c->stage == QSTAGE_FRAG) {
- for (int index = decl->Range.First;
- index <= decl->Range.Last;
- index++) {
- for (int i = 0; i < 4; i++) {
- struct qreg vary = {
- QFILE_VARY,
- index * 4 + i
- };
-
- /* XXX: multiply by W */
- trans->inputs[index * 4 + i] =
- qir_VARY_ADD_C(c,
- qir_MOV(c,
- vary));
-
- trans->num_inputs++;
+ resize_qreg_array(c, &c->inputs, &c->inputs_array_size,
+ (decl->Range.Last + 1) * 4);
+
+ for (int i = decl->Range.First;
+ i <= decl->Range.Last;
+ i++) {
+ if (c->stage == QSTAGE_FRAG) {
+ if (decl->Semantic.Name ==
+ TGSI_SEMANTIC_POSITION) {
+ emit_fragcoord_input(c, i);
+ } else if (decl->Semantic.Name == TGSI_SEMANTIC_FACE) {
+ emit_face_input(c, i);
+ } else if (decl->Semantic.Name == TGSI_SEMANTIC_GENERIC &&
+ (c->fs_key->point_sprite_mask &
+ (1 << decl->Semantic.Index))) {
+ emit_point_coord_input(c, i);
+ } else {
+ emit_fragment_input(c, i, decl);
}
+ } else {
+ emit_vertex_input(c, i);
}
}
break;
+
+ case TGSI_FILE_OUTPUT:
+ resize_qreg_array(c, &c->outputs, &c->outputs_array_size,
+ (decl->Range.Last + 1) * 4);
+
+ switch (decl->Semantic.Name) {
+ case TGSI_SEMANTIC_POSITION:
+ c->output_position_index = decl->Range.First * 4;
+ break;
+ case TGSI_SEMANTIC_COLOR:
+ c->output_color_index = decl->Range.First * 4;
+ break;
+ case TGSI_SEMANTIC_PSIZE:
+ c->output_point_size_index = decl->Range.First * 4;
+ break;
+ }
+
+ break;
}
}
static void
-emit_tgsi_instruction(struct tgsi_to_qir *trans,
+emit_tgsi_instruction(struct vc4_compile *c,
struct tgsi_full_instruction *tgsi_inst)
{
- struct qcompile *c = trans->c;
struct {
enum qop op;
- struct qreg (*func)(struct tgsi_to_qir *trans,
+ struct qreg (*func)(struct vc4_compile *c,
struct tgsi_full_instruction *tgsi_inst,
enum qop op,
struct qreg *src, int i);
[TGSI_OPCODE_SUB] = { QOP_FSUB, tgsi_to_qir_alu },
[TGSI_OPCODE_MIN] = { QOP_FMIN, tgsi_to_qir_alu },
[TGSI_OPCODE_MAX] = { QOP_FMAX, tgsi_to_qir_alu },
+ [TGSI_OPCODE_F2I] = { QOP_FTOI, tgsi_to_qir_alu },
+ [TGSI_OPCODE_I2F] = { QOP_ITOF, tgsi_to_qir_alu },
+ [TGSI_OPCODE_UADD] = { QOP_ADD, tgsi_to_qir_alu },
+ [TGSI_OPCODE_USHR] = { QOP_SHR, tgsi_to_qir_alu },
+ [TGSI_OPCODE_ISHR] = { QOP_ASR, tgsi_to_qir_alu },
+ [TGSI_OPCODE_SHL] = { QOP_SHL, tgsi_to_qir_alu },
+ [TGSI_OPCODE_IMIN] = { QOP_MIN, tgsi_to_qir_alu },
+ [TGSI_OPCODE_IMAX] = { QOP_MAX, tgsi_to_qir_alu },
+ [TGSI_OPCODE_AND] = { QOP_AND, tgsi_to_qir_alu },
+ [TGSI_OPCODE_OR] = { QOP_OR, tgsi_to_qir_alu },
+ [TGSI_OPCODE_XOR] = { QOP_XOR, tgsi_to_qir_alu },
+ [TGSI_OPCODE_NOT] = { QOP_NOT, tgsi_to_qir_alu },
+
+ [TGSI_OPCODE_UMUL] = { 0, tgsi_to_qir_umul },
+ [TGSI_OPCODE_IDIV] = { 0, tgsi_to_qir_idiv },
+ [TGSI_OPCODE_INEG] = { 0, tgsi_to_qir_ineg },
+
[TGSI_OPCODE_RSQ] = { QOP_RSQ, tgsi_to_qir_alu },
- [TGSI_OPCODE_SEQ] = { QOP_SEQ, tgsi_to_qir_alu },
- [TGSI_OPCODE_SNE] = { QOP_SNE, tgsi_to_qir_alu },
- [TGSI_OPCODE_SGE] = { QOP_SGE, tgsi_to_qir_alu },
- [TGSI_OPCODE_SLT] = { QOP_SLT, tgsi_to_qir_alu },
+ [TGSI_OPCODE_SEQ] = { 0, tgsi_to_qir_seq },
+ [TGSI_OPCODE_SNE] = { 0, tgsi_to_qir_sne },
+ [TGSI_OPCODE_SGE] = { 0, tgsi_to_qir_sge },
+ [TGSI_OPCODE_SLT] = { 0, tgsi_to_qir_slt },
+ [TGSI_OPCODE_FSEQ] = { 0, tgsi_to_qir_fseq },
+ [TGSI_OPCODE_FSNE] = { 0, tgsi_to_qir_fsne },
+ [TGSI_OPCODE_FSGE] = { 0, tgsi_to_qir_fsge },
+ [TGSI_OPCODE_FSLT] = { 0, tgsi_to_qir_fslt },
+ [TGSI_OPCODE_USEQ] = { 0, tgsi_to_qir_useq },
+ [TGSI_OPCODE_USNE] = { 0, tgsi_to_qir_usne },
+ [TGSI_OPCODE_ISGE] = { 0, tgsi_to_qir_isge },
+ [TGSI_OPCODE_ISLT] = { 0, tgsi_to_qir_islt },
+
+ [TGSI_OPCODE_CMP] = { 0, tgsi_to_qir_cmp },
[TGSI_OPCODE_MAD] = { 0, tgsi_to_qir_mad },
- [TGSI_OPCODE_DP2] = { 0, tgsi_to_qir_dp2 },
- [TGSI_OPCODE_DP3] = { 0, tgsi_to_qir_dp3 },
- [TGSI_OPCODE_DP4] = { 0, tgsi_to_qir_dp4 },
- [TGSI_OPCODE_RCP] = { QOP_RCP, tgsi_to_qir_alu },
- [TGSI_OPCODE_RSQ] = { QOP_RSQ, tgsi_to_qir_alu },
- [TGSI_OPCODE_EX2] = { QOP_EXP2, tgsi_to_qir_alu },
- [TGSI_OPCODE_LG2] = { QOP_LOG2, tgsi_to_qir_alu },
- [TGSI_OPCODE_LIT] = { QOP_MOV, tgsi_to_qir_alu }, /* XXX */
+ [TGSI_OPCODE_RCP] = { QOP_RCP, tgsi_to_qir_scalar },
+ [TGSI_OPCODE_RSQ] = { QOP_RSQ, tgsi_to_qir_scalar },
+ [TGSI_OPCODE_EX2] = { QOP_EXP2, tgsi_to_qir_scalar },
+ [TGSI_OPCODE_LG2] = { QOP_LOG2, tgsi_to_qir_scalar },
+ [TGSI_OPCODE_LRP] = { 0, tgsi_to_qir_lrp },
+ [TGSI_OPCODE_TRUNC] = { 0, tgsi_to_qir_trunc },
+ [TGSI_OPCODE_FRC] = { 0, tgsi_to_qir_frc },
+ [TGSI_OPCODE_FLR] = { 0, tgsi_to_qir_flr },
+ [TGSI_OPCODE_SIN] = { 0, tgsi_to_qir_sin },
+ [TGSI_OPCODE_COS] = { 0, tgsi_to_qir_cos },
+ [TGSI_OPCODE_CLAMP] = { 0, tgsi_to_qir_clamp },
};
static int asdf = 0;
uint32_t tgsi_op = tgsi_inst->Instruction.Opcode;
if (tgsi_op == TGSI_OPCODE_END)
return;
- if (tgsi_op > ARRAY_SIZE(op_trans) || !op_trans[tgsi_op].func) {
- fprintf(stderr, "unknown tgsi inst: ");
- tgsi_dump_instruction(tgsi_inst, asdf++);
- fprintf(stderr, "\n");
- abort();
- }
-
struct qreg src_regs[12];
for (int s = 0; s < 3; s++) {
for (int i = 0; i < 4; i++) {
src_regs[4 * s + i] =
- get_src(trans, &tgsi_inst->Src[s].Register, i);
+ get_src(c, tgsi_inst->Instruction.Opcode,
+ &tgsi_inst->Src[s].Register, i);
}
}
+ switch (tgsi_op) {
+ case TGSI_OPCODE_TEX:
+ case TGSI_OPCODE_TXP:
+ case TGSI_OPCODE_TXB:
+ tgsi_to_qir_tex(c, tgsi_inst,
+ op_trans[tgsi_op].op, src_regs);
+ return;
+ case TGSI_OPCODE_KILL:
+ c->discard = qir_uniform_f(c, 1.0);
+ return;
+ case TGSI_OPCODE_KILL_IF:
+ for (int i = 0; i < 4; i++)
+ tgsi_to_qir_kill_if(c, src_regs, i);
+ return;
+ default:
+ break;
+ }
+
+ if (tgsi_op > ARRAY_SIZE(op_trans) || !(op_trans[tgsi_op].func)) {
+ fprintf(stderr, "unknown tgsi inst: ");
+ tgsi_dump_instruction(tgsi_inst, asdf++);
+ fprintf(stderr, "\n");
+ abort();
+ }
+
for (int i = 0; i < 4; i++) {
if (!(tgsi_inst->Dst[0].Register.WriteMask & (1 << i)))
continue;
struct qreg result;
- result = op_trans[tgsi_op].func(trans, tgsi_inst,
+ result = op_trans[tgsi_op].func(c, tgsi_inst,
op_trans[tgsi_op].op,
src_regs, i);
result = qir_FMAX(c,
qir_FMIN(c,
result,
- qir_uniform_f(trans, 1.0)),
- qir_uniform_f(trans, low));
+ qir_uniform_f(c, 1.0)),
+ qir_uniform_f(c, low));
}
- update_dst(trans, tgsi_inst, i, result);
+ update_dst(c, tgsi_inst, i, result);
}
}
static void
-parse_tgsi_immediate(struct tgsi_to_qir *trans, struct tgsi_full_immediate *imm)
+parse_tgsi_immediate(struct vc4_compile *c, struct tgsi_full_immediate *imm)
{
for (int i = 0; i < 4; i++) {
- unsigned n = trans->num_consts++;
- trans->consts[n] = qir_uniform_ui(trans, imm->u[i].Uint);
+ unsigned n = c->num_consts++;
+ resize_qreg_array(c, &c->consts, &c->consts_array_size, n + 1);
+ c->consts[n] = qir_uniform_ui(c, imm->u[i].Uint);
}
}
-static void
-emit_frag_init(struct tgsi_to_qir *trans)
+static struct qreg
+vc4_blend_channel(struct vc4_compile *c,
+ struct qreg *dst,
+ struct qreg *src,
+ struct qreg val,
+ unsigned factor,
+ int channel)
{
+ switch(factor) {
+ case PIPE_BLENDFACTOR_ONE:
+ return val;
+ case PIPE_BLENDFACTOR_SRC_COLOR:
+ return qir_FMUL(c, val, src[channel]);
+ case PIPE_BLENDFACTOR_SRC_ALPHA:
+ return qir_FMUL(c, val, src[3]);
+ case PIPE_BLENDFACTOR_DST_ALPHA:
+ return qir_FMUL(c, val, dst[3]);
+ case PIPE_BLENDFACTOR_DST_COLOR:
+ return qir_FMUL(c, val, dst[channel]);
+ case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
+ return qir_FMIN(c, src[3], qir_FSUB(c,
+ qir_uniform_f(c, 1.0),
+ dst[3]));
+ case PIPE_BLENDFACTOR_CONST_COLOR:
+ return qir_FMUL(c, val,
+ get_temp_for_uniform(c,
+ QUNIFORM_BLEND_CONST_COLOR,
+ channel));
+ case PIPE_BLENDFACTOR_CONST_ALPHA:
+ return qir_FMUL(c, val,
+ get_temp_for_uniform(c,
+ QUNIFORM_BLEND_CONST_COLOR,
+ 3));
+ case PIPE_BLENDFACTOR_ZERO:
+ return qir_uniform_f(c, 0.0);
+ case PIPE_BLENDFACTOR_INV_SRC_COLOR:
+ return qir_FMUL(c, val, qir_FSUB(c, qir_uniform_f(c, 1.0),
+ src[channel]));
+ case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
+ return qir_FMUL(c, val, qir_FSUB(c, qir_uniform_f(c, 1.0),
+ src[3]));
+ case PIPE_BLENDFACTOR_INV_DST_ALPHA:
+ return qir_FMUL(c, val, qir_FSUB(c, qir_uniform_f(c, 1.0),
+ dst[3]));
+ case PIPE_BLENDFACTOR_INV_DST_COLOR:
+ return qir_FMUL(c, val, qir_FSUB(c, qir_uniform_f(c, 1.0),
+ dst[channel]));
+ case PIPE_BLENDFACTOR_INV_CONST_COLOR:
+ return qir_FMUL(c, val,
+ qir_FSUB(c, qir_uniform_f(c, 1.0),
+ get_temp_for_uniform(c,
+ QUNIFORM_BLEND_CONST_COLOR,
+ channel)));
+ case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
+ return qir_FMUL(c, val,
+ qir_FSUB(c, qir_uniform_f(c, 1.0),
+ get_temp_for_uniform(c,
+ QUNIFORM_BLEND_CONST_COLOR,
+ 3)));
+
+ default:
+ case PIPE_BLENDFACTOR_SRC1_COLOR:
+ case PIPE_BLENDFACTOR_SRC1_ALPHA:
+ case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
+ case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
+ /* Unsupported. */
+ fprintf(stderr, "Unknown blend factor %d\n", factor);
+ return val;
+ }
}
-static void
-emit_vert_init(struct tgsi_to_qir *trans)
+static struct qreg
+vc4_blend_func(struct vc4_compile *c,
+ struct qreg src, struct qreg dst,
+ unsigned func)
{
- struct qcompile *c = trans->c;
+ switch (func) {
+ case PIPE_BLEND_ADD:
+ return qir_FADD(c, src, dst);
+ case PIPE_BLEND_SUBTRACT:
+ return qir_FSUB(c, src, dst);
+ case PIPE_BLEND_REVERSE_SUBTRACT:
+ return qir_FSUB(c, dst, src);
+ case PIPE_BLEND_MIN:
+ return qir_FMIN(c, src, dst);
+ case PIPE_BLEND_MAX:
+ return qir_FMAX(c, src, dst);
+
+ default:
+ /* Unsupported. */
+ fprintf(stderr, "Unknown blend func %d\n", func);
+ return src;
- /* XXX: attribute type/size/count */
- for (int i = 0; i < 4; i++) {
- trans->inputs[i] = qir_get_temp(c);
- qir_emit(c, qir_inst(QOP_VPM_READ, trans->inputs[i],
- c->undef, c->undef));
}
}
+/**
+ * Implements fixed function blending in shader code.
+ *
+ * VC4 doesn't have any hardware support for blending. Instead, you read the
+ * current contents of the destination from the tile buffer after having
+ * waited for the scoreboard (which is handled by vc4_qpu_emit.c), then do
+ * math using your output color and that destination value, and update the
+ * output color appropriately.
+ */
static void
-emit_coord_init(struct tgsi_to_qir *trans)
+vc4_blend(struct vc4_compile *c, struct qreg *result,
+ struct qreg *dst_color, struct qreg *src_color)
{
- struct qcompile *c = trans->c;
+ struct pipe_rt_blend_state *blend = &c->fs_key->blend;
- /* XXX: attribute type/size/count */
- for (int i = 0; i < 4; i++) {
- trans->inputs[i] = qir_get_temp(c);
- qir_emit(c, qir_inst(QOP_VPM_READ, trans->inputs[i],
- c->undef, c->undef));
+ if (!blend->blend_enable) {
+ for (int i = 0; i < 4; i++)
+ result[i] = src_color[i];
+ return;
+ }
+
+ struct qreg src_blend[4], dst_blend[4];
+ for (int i = 0; i < 3; i++) {
+ src_blend[i] = vc4_blend_channel(c,
+ dst_color, src_color,
+ src_color[i],
+ blend->rgb_src_factor, i);
+ dst_blend[i] = vc4_blend_channel(c,
+ dst_color, src_color,
+ dst_color[i],
+ blend->rgb_dst_factor, i);
}
+ src_blend[3] = vc4_blend_channel(c,
+ dst_color, src_color,
+ src_color[3],
+ blend->alpha_src_factor, 3);
+ dst_blend[3] = vc4_blend_channel(c,
+ dst_color, src_color,
+ dst_color[3],
+ blend->alpha_dst_factor, 3);
+
+ for (int i = 0; i < 3; i++) {
+ result[i] = vc4_blend_func(c,
+ src_blend[i], dst_blend[i],
+ blend->rgb_func);
+ }
+ result[3] = vc4_blend_func(c,
+ src_blend[3], dst_blend[3],
+ blend->alpha_func);
}
static void
-emit_frag_end(struct tgsi_to_qir *trans)
+alpha_test_discard(struct vc4_compile *c)
{
- struct qcompile *c = trans->c;
+ struct qreg src_alpha;
+ struct qreg alpha_ref = get_temp_for_uniform(c, QUNIFORM_ALPHA_REF, 0);
+
+ if (!c->fs_key->alpha_test)
+ return;
+
+ if (c->output_color_index != -1)
+ src_alpha = c->outputs[c->output_color_index + 3];
+ else
+ src_alpha = qir_uniform_f(c, 1.0);
- struct qreg t = qir_get_temp(c);
+ if (c->discard.file == QFILE_NULL)
+ c->discard = qir_uniform_f(c, 0.0);
+
+ switch (c->fs_key->alpha_test_func) {
+ case PIPE_FUNC_NEVER:
+ c->discard = qir_uniform_f(c, 1.0);
+ break;
+ case PIPE_FUNC_ALWAYS:
+ break;
+ case PIPE_FUNC_EQUAL:
+ qir_SF(c, qir_FSUB(c, src_alpha, alpha_ref));
+ c->discard = qir_SEL_X_Y_ZS(c, c->discard,
+ qir_uniform_f(c, 1.0));
+ break;
+ case PIPE_FUNC_NOTEQUAL:
+ qir_SF(c, qir_FSUB(c, src_alpha, alpha_ref));
+ c->discard = qir_SEL_X_Y_ZC(c, c->discard,
+ qir_uniform_f(c, 1.0));
+ break;
+ case PIPE_FUNC_GREATER:
+ qir_SF(c, qir_FSUB(c, src_alpha, alpha_ref));
+ c->discard = qir_SEL_X_Y_NC(c, c->discard,
+ qir_uniform_f(c, 1.0));
+ break;
+ case PIPE_FUNC_GEQUAL:
+ qir_SF(c, qir_FSUB(c, alpha_ref, src_alpha));
+ c->discard = qir_SEL_X_Y_NS(c, c->discard,
+ qir_uniform_f(c, 1.0));
+ break;
+ case PIPE_FUNC_LESS:
+ qir_SF(c, qir_FSUB(c, src_alpha, alpha_ref));
+ c->discard = qir_SEL_X_Y_NS(c, c->discard,
+ qir_uniform_f(c, 1.0));
+ break;
+ case PIPE_FUNC_LEQUAL:
+ qir_SF(c, qir_FSUB(c, alpha_ref, src_alpha));
+ c->discard = qir_SEL_X_Y_NC(c, c->discard,
+ qir_uniform_f(c, 1.0));
+ break;
+ }
+}
- const struct util_format_description *format_desc =
- util_format_description(trans->fs_key->color_format);
+static void
+emit_frag_end(struct vc4_compile *c)
+{
+ alpha_test_discard(c);
+
+ enum pipe_format color_format = c->fs_key->color_format;
+ const uint8_t *format_swiz = vc4_get_format_swizzle(color_format);
+ struct qreg tlb_read_color[4] = { c->undef, c->undef, c->undef, c->undef };
+ struct qreg dst_color[4] = { c->undef, c->undef, c->undef, c->undef };
+ struct qreg linear_dst_color[4] = { c->undef, c->undef, c->undef, c->undef };
+ if (c->fs_key->blend.blend_enable ||
+ c->fs_key->blend.colormask != 0xf) {
+ struct qreg r4 = qir_TLB_COLOR_READ(c);
+ for (int i = 0; i < 4; i++)
+ tlb_read_color[i] = qir_R4_UNPACK(c, r4, i);
+ for (int i = 0; i < 4; i++) {
+ dst_color[i] = get_swizzled_channel(c,
+ tlb_read_color,
+ format_swiz[i]);
+ if (util_format_is_srgb(color_format) && i != 3) {
+ linear_dst_color[i] =
+ qir_srgb_decode(c, dst_color[i]);
+ } else {
+ linear_dst_color[i] = dst_color[i];
+ }
+ }
+ }
- struct qreg swizzled_outputs[4] = {
- trans->outputs[format_desc->swizzle[0]],
- trans->outputs[format_desc->swizzle[1]],
- trans->outputs[format_desc->swizzle[2]],
- trans->outputs[format_desc->swizzle[3]],
+ struct qreg blend_color[4];
+ struct qreg undef_array[4] = {
+ c->undef, c->undef, c->undef, c->undef
};
+ vc4_blend(c, blend_color, linear_dst_color,
+ (c->output_color_index != -1 ?
+ c->outputs + c->output_color_index :
+ undef_array));
+
+ if (util_format_is_srgb(color_format)) {
+ for (int i = 0; i < 3; i++)
+ blend_color[i] = qir_srgb_encode(c, blend_color[i]);
+ }
+
+ /* If the bit isn't set in the color mask, then just return the
+ * original dst color, instead.
+ */
+ for (int i = 0; i < 4; i++) {
+ if (!(c->fs_key->blend.colormask & (1 << i))) {
+ blend_color[i] = dst_color[i];
+ }
+ }
+
+ /* Debug: Sometimes you're getting a black output and just want to see
+ * if the FS is getting executed at all. Spam magenta into the color
+ * output.
+ */
+ if (0) {
+ blend_color[0] = qir_uniform_f(c, 1.0);
+ blend_color[1] = qir_uniform_f(c, 0.0);
+ blend_color[2] = qir_uniform_f(c, 1.0);
+ blend_color[3] = qir_uniform_f(c, 0.5);
+ }
+
+ struct qreg swizzled_outputs[4];
+ for (int i = 0; i < 4; i++) {
+ swizzled_outputs[i] = get_swizzled_channel(c, blend_color,
+ format_swiz[i]);
+ }
+
+ if (c->discard.file != QFILE_NULL)
+ qir_TLB_DISCARD_SETUP(c, c->discard);
+
+ if (c->fs_key->stencil_enabled) {
+ qir_TLB_STENCIL_SETUP(c, add_uniform(c, QUNIFORM_STENCIL, 0));
+ if (c->fs_key->stencil_twoside) {
+ qir_TLB_STENCIL_SETUP(c, add_uniform(c, QUNIFORM_STENCIL, 1));
+ }
+ if (c->fs_key->stencil_full_writemasks) {
+ qir_TLB_STENCIL_SETUP(c, add_uniform(c, QUNIFORM_STENCIL, 2));
+ }
+ }
+
+ if (c->fs_key->depth_enabled) {
+ struct qreg z;
+ if (c->output_position_index != -1) {
+ z = qir_FTOI(c, qir_FMUL(c, c->outputs[c->output_position_index + 2],
+ qir_uniform_f(c, 0xffffff)));
+ } else {
+ z = qir_FRAG_Z(c);
+ }
+ qir_TLB_Z_WRITE(c, z);
+ }
+
+ bool color_written = false;
+ for (int i = 0; i < 4; i++) {
+ if (swizzled_outputs[i].file != QFILE_NULL)
+ color_written = true;
+ }
+
+ struct qreg packed_color;
+ if (color_written) {
+ /* Fill in any undefined colors. The simulator will assertion
+ * fail if we read something that wasn't written, and I don't
+ * know what hardware does.
+ */
+ for (int i = 0; i < 4; i++) {
+ if (swizzled_outputs[i].file == QFILE_NULL)
+ swizzled_outputs[i] = qir_uniform_f(c, 0.0);
+ }
+ packed_color = qir_get_temp(c);
+ qir_emit(c, qir_inst4(QOP_PACK_COLORS, packed_color,
+ swizzled_outputs[0],
+ swizzled_outputs[1],
+ swizzled_outputs[2],
+ swizzled_outputs[3]));
+ } else {
+ packed_color = qir_uniform_ui(c, 0);
+ }
- qir_emit(c, qir_inst4(QOP_PACK_COLORS, t,
- swizzled_outputs[0],
- swizzled_outputs[1],
- swizzled_outputs[2],
- swizzled_outputs[3]));
qir_emit(c, qir_inst(QOP_TLB_COLOR_WRITE, c->undef,
- t, c->undef));
+ packed_color, c->undef));
}
static void
-emit_scaled_viewport_write(struct tgsi_to_qir *trans)
+emit_scaled_viewport_write(struct vc4_compile *c, struct qreg rcp_w)
{
- struct qcompile *c = trans->c;
struct qreg xyi[2];
for (int i = 0; i < 2; i++) {
- trans->uniform_contents[trans->num_uniforms] =
- QUNIFORM_VIEWPORT_X_SCALE + i;
- struct qreg scale = { QFILE_UNIF, trans->num_uniforms++ };
-
- xyi[i] = qir_FTOI(c, qir_FMUL(c, trans->outputs[i], scale));
+ struct qreg scale =
+ add_uniform(c, QUNIFORM_VIEWPORT_X_SCALE + i, 0);
+
+ xyi[i] = qir_FTOI(c, qir_FMUL(c,
+ qir_FMUL(c,
+ c->outputs[i],
+ scale),
+ rcp_w));
}
qir_VPM_WRITE(c, qir_PACK_SCALED(c, xyi[0], xyi[1]));
}
static void
-emit_zs_write(struct tgsi_to_qir *trans)
+emit_zs_write(struct vc4_compile *c, struct qreg rcp_w)
{
- struct qcompile *c = trans->c;
+ struct qreg zscale = add_uniform(c, QUNIFORM_VIEWPORT_Z_SCALE, 0);
+ struct qreg zoffset = add_uniform(c, QUNIFORM_VIEWPORT_Z_OFFSET, 0);
+
+ qir_VPM_WRITE(c, qir_FMUL(c, qir_FADD(c, qir_FMUL(c,
+ c->outputs[2],
+ zscale),
+ zoffset),
+ rcp_w));
+}
- /* XXX: rescale */
- qir_VPM_WRITE(c, trans->outputs[2]);
+static void
+emit_rcp_wc_write(struct vc4_compile *c, struct qreg rcp_w)
+{
+ qir_VPM_WRITE(c, rcp_w);
}
static void
-emit_1_wc_write(struct tgsi_to_qir *trans)
+emit_point_size_write(struct vc4_compile *c)
{
- struct qcompile *c = trans->c;
+ struct qreg point_size;
+
+ if (c->output_point_size_index)
+ point_size = c->outputs[c->output_point_size_index + 3];
+ else
+ point_size = qir_uniform_f(c, 1.0);
- /* XXX: RCP */
- qir_VPM_WRITE(c, trans->outputs[3]);
+ /* Workaround: HW-2726 PTB does not handle zero-size points (BCM2835,
+ * BCM21553).
+ */
+ point_size = qir_FMAX(c, point_size, qir_uniform_f(c, .125));
+
+ qir_VPM_WRITE(c, point_size);
}
static void
-emit_vert_end(struct tgsi_to_qir *trans)
+emit_vert_end(struct vc4_compile *c)
{
- struct qcompile *c = trans->c;
+ struct qreg rcp_w = qir_RCP(c, c->outputs[3]);
- emit_scaled_viewport_write(trans);
- emit_zs_write(trans);
- emit_1_wc_write(trans);
+ emit_scaled_viewport_write(c, rcp_w);
+ emit_zs_write(c, rcp_w);
+ emit_rcp_wc_write(c, rcp_w);
+ if (c->vs_key->per_vertex_point_size)
+ emit_point_size_write(c);
- for (int i = 4; i < trans->num_outputs; i++) {
- qir_VPM_WRITE(c, trans->outputs[i]);
+ for (int i = 4; i < c->num_outputs; i++) {
+ qir_VPM_WRITE(c, c->outputs[i]);
}
}
static void
-emit_coord_end(struct tgsi_to_qir *trans)
+emit_coord_end(struct vc4_compile *c)
{
- struct qcompile *c = trans->c;
+ struct qreg rcp_w = qir_RCP(c, c->outputs[3]);
for (int i = 0; i < 4; i++)
- qir_VPM_WRITE(c, trans->outputs[i]);
+ qir_VPM_WRITE(c, c->outputs[i]);
- emit_scaled_viewport_write(trans);
- emit_zs_write(trans);
- emit_1_wc_write(trans);
+ emit_scaled_viewport_write(c, rcp_w);
+ emit_zs_write(c, rcp_w);
+ emit_rcp_wc_write(c, rcp_w);
+ if (c->vs_key->per_vertex_point_size)
+ emit_point_size_write(c);
}
-static struct tgsi_to_qir *
-vc4_shader_tgsi_to_qir(struct vc4_compiled_shader *shader, enum qstage stage,
+static struct vc4_compile *
+vc4_shader_tgsi_to_qir(struct vc4_context *vc4,
+ struct vc4_compiled_shader *shader, enum qstage stage,
struct vc4_key *key)
{
- struct tgsi_to_qir *trans = CALLOC_STRUCT(tgsi_to_qir);
- struct qcompile *c;
+ struct vc4_compile *c = qir_compile_init();
int ret;
- c = qir_compile_init();
c->stage = stage;
+ c->shader_state = &key->shader_state->base;
- memset(trans, 0, sizeof(*trans));
- /* XXX sizing */
- trans->temps = calloc(sizeof(struct qreg), 1024);
- trans->inputs = calloc(sizeof(struct qreg), 8 * 4);
- trans->outputs = calloc(sizeof(struct qreg), 1024);
- trans->uniforms = calloc(sizeof(struct qreg), 1024);
- trans->consts = calloc(sizeof(struct qreg), 1024);
-
- trans->uniform_data = calloc(sizeof(uint32_t), 1024);
- trans->uniform_contents = calloc(sizeof(enum quniform_contents), 1024);
-
- trans->shader_state = key->shader_state;
- trans->c = c;
- ret = tgsi_parse_init(&trans->parser, trans->shader_state->base.tokens);
- assert(ret == TGSI_PARSE_OK);
-
- if (vc4_debug & VC4_DEBUG_TGSI) {
- fprintf(stderr, "TGSI:\n");
- tgsi_dump(trans->shader_state->base.tokens, 0);
- }
-
+ c->key = key;
switch (stage) {
case QSTAGE_FRAG:
- trans->fs_key = (struct vc4_fs_key *)key;
- emit_frag_init(trans);
+ c->fs_key = (struct vc4_fs_key *)key;
+ if (c->fs_key->is_points) {
+ c->point_x = emit_fragment_varying(c, 0);
+ c->point_y = emit_fragment_varying(c, 0);
+ } else if (c->fs_key->is_lines) {
+ c->line_x = emit_fragment_varying(c, 0);
+ }
break;
case QSTAGE_VERT:
- trans->vs_key = (struct vc4_vs_key *)key;
- emit_vert_init(trans);
+ c->vs_key = (struct vc4_vs_key *)key;
break;
case QSTAGE_COORD:
- trans->vs_key = (struct vc4_vs_key *)key;
- emit_coord_init(trans);
+ c->vs_key = (struct vc4_vs_key *)key;
break;
}
- while (!tgsi_parse_end_of_tokens(&trans->parser)) {
- tgsi_parse_token(&trans->parser);
+ const struct tgsi_token *tokens = key->shader_state->base.tokens;
+ if (c->fs_key && c->fs_key->light_twoside) {
+ if (!key->shader_state->twoside_tokens) {
+ const struct tgsi_lowering_config lowering_config = {
+ .color_two_side = true,
+ };
+ struct tgsi_shader_info info;
+ key->shader_state->twoside_tokens =
+ tgsi_transform_lowering(&lowering_config,
+ key->shader_state->base.tokens,
+ &info);
+
+ /* If no transformation occurred, then NULL is
+ * returned and we just use our original tokens.
+ */
+ if (!key->shader_state->twoside_tokens) {
+ key->shader_state->twoside_tokens =
+ key->shader_state->base.tokens;
+ }
+ }
+ tokens = key->shader_state->twoside_tokens;
+ }
- switch (trans->parser.FullToken.Token.Type) {
+ ret = tgsi_parse_init(&c->parser, tokens);
+ assert(ret == TGSI_PARSE_OK);
+
+ if (vc4_debug & VC4_DEBUG_TGSI) {
+ fprintf(stderr, "TGSI:\n");
+ tgsi_dump(tokens, 0);
+ }
+
+ while (!tgsi_parse_end_of_tokens(&c->parser)) {
+ tgsi_parse_token(&c->parser);
+
+ switch (c->parser.FullToken.Token.Type) {
case TGSI_TOKEN_TYPE_DECLARATION:
- emit_tgsi_declaration(trans,
- &trans->parser.FullToken.FullDeclaration);
+ emit_tgsi_declaration(c,
+ &c->parser.FullToken.FullDeclaration);
break;
case TGSI_TOKEN_TYPE_INSTRUCTION:
- emit_tgsi_instruction(trans,
- &trans->parser.FullToken.FullInstruction);
+ emit_tgsi_instruction(c,
+ &c->parser.FullToken.FullInstruction);
break;
case TGSI_TOKEN_TYPE_IMMEDIATE:
- parse_tgsi_immediate(trans,
- &trans->parser.FullToken.FullImmediate);
+ parse_tgsi_immediate(c,
+ &c->parser.FullToken.FullImmediate);
break;
}
}
switch (stage) {
case QSTAGE_FRAG:
- emit_frag_end(trans);
+ emit_frag_end(c);
break;
case QSTAGE_VERT:
- emit_vert_end(trans);
+ emit_vert_end(c);
break;
case QSTAGE_COORD:
- emit_coord_end(trans);
+ emit_coord_end(c);
break;
}
+ tgsi_parse_free(&c->parser);
+
+ qir_optimize(c);
+
if (vc4_debug & VC4_DEBUG_QIR) {
fprintf(stderr, "QIR:\n");
qir_dump(c);
}
-
- tgsi_parse_free(&trans->parser);
- free(trans->temps);
-
- vc4_generate_code(c);
+ qir_reorder_uniforms(c);
+ vc4_generate_code(vc4, c);
if (vc4_debug & VC4_DEBUG_SHADERDB) {
fprintf(stderr, "SHADER-DB: %s: %d instructions\n",
qir_get_stage_name(c->stage), c->qpu_inst_count);
fprintf(stderr, "SHADER-DB: %s: %d uniforms\n",
- qir_get_stage_name(c->stage), trans->num_uniforms);
+ qir_get_stage_name(c->stage), c->num_uniforms);
}
- return trans;
+ return c;
}
static void *
vc4_shader_state_create(struct pipe_context *pctx,
const struct pipe_shader_state *cso)
{
- struct vc4_shader_state *so = CALLOC_STRUCT(vc4_shader_state);
+ struct vc4_uncompiled_shader *so = CALLOC_STRUCT(vc4_uncompiled_shader);
if (!so)
return NULL;
- so->base.tokens = tgsi_dup_tokens(cso->tokens);
+ const struct tgsi_lowering_config lowering_config = {
+ .lower_DST = true,
+ .lower_XPD = true,
+ .lower_SCS = true,
+ .lower_POW = true,
+ .lower_LIT = true,
+ .lower_EXP = true,
+ .lower_LOG = true,
+ .lower_DP4 = true,
+ .lower_DP3 = true,
+ .lower_DPH = true,
+ .lower_DP2 = true,
+ .lower_DP2A = true,
+ };
+
+ struct tgsi_shader_info info;
+ so->base.tokens = tgsi_transform_lowering(&lowering_config, cso->tokens, &info);
+ if (!so->base.tokens)
+ so->base.tokens = tgsi_dup_tokens(cso->tokens);
return so;
}
static void
copy_uniform_state_to_shader(struct vc4_compiled_shader *shader,
int shader_index,
- struct tgsi_to_qir *trans)
+ struct vc4_compile *c)
{
- int count = trans->num_uniforms;
+ int count = c->num_uniforms;
struct vc4_shader_uniform_info *uinfo = &shader->uniforms[shader_index];
uinfo->count = count;
uinfo->data = malloc(count * sizeof(*uinfo->data));
- memcpy(uinfo->data, trans->uniform_data,
+ memcpy(uinfo->data, c->uniform_data,
count * sizeof(*uinfo->data));
uinfo->contents = malloc(count * sizeof(*uinfo->contents));
- memcpy(uinfo->contents, trans->uniform_contents,
+ memcpy(uinfo->contents, c->uniform_contents,
count * sizeof(*uinfo->contents));
+ uinfo->num_texture_samples = c->num_texture_samples;
}
static void
vc4_fs_compile(struct vc4_context *vc4, struct vc4_compiled_shader *shader,
struct vc4_fs_key *key)
{
- struct tgsi_to_qir *trans = vc4_shader_tgsi_to_qir(shader, QSTAGE_FRAG,
- &key->base);
- shader->num_inputs = trans->num_inputs;
- copy_uniform_state_to_shader(shader, 0, trans);
- shader->bo = vc4_bo_alloc_mem(vc4->screen, trans->c->qpu_insts,
- trans->c->qpu_inst_count * sizeof(uint64_t),
+ struct vc4_compile *c = vc4_shader_tgsi_to_qir(vc4, shader,
+ QSTAGE_FRAG,
+ &key->base);
+ shader->num_inputs = c->num_inputs;
+ shader->color_inputs = c->color_inputs;
+ copy_uniform_state_to_shader(shader, 0, c);
+ shader->bo = vc4_bo_alloc_mem(vc4->screen, c->qpu_insts,
+ c->qpu_inst_count * sizeof(uint64_t),
"fs_code");
- qir_compile_destroy(trans->c);
- free(trans);
+ qir_compile_destroy(c);
}
static void
vc4_vs_compile(struct vc4_context *vc4, struct vc4_compiled_shader *shader,
struct vc4_vs_key *key)
{
- struct tgsi_to_qir *vs_trans = vc4_shader_tgsi_to_qir(shader,
- QSTAGE_VERT,
- &key->base);
- copy_uniform_state_to_shader(shader, 0, vs_trans);
-
- struct tgsi_to_qir *cs_trans = vc4_shader_tgsi_to_qir(shader,
- QSTAGE_COORD,
- &key->base);
- copy_uniform_state_to_shader(shader, 1, cs_trans);
-
- uint32_t vs_size = vs_trans->c->qpu_inst_count * sizeof(uint64_t);
- uint32_t cs_size = cs_trans->c->qpu_inst_count * sizeof(uint64_t);
+ struct vc4_compile *vs_c = vc4_shader_tgsi_to_qir(vc4, shader,
+ QSTAGE_VERT,
+ &key->base);
+ copy_uniform_state_to_shader(shader, 0, vs_c);
+
+ struct vc4_compile *cs_c = vc4_shader_tgsi_to_qir(vc4, shader,
+ QSTAGE_COORD,
+ &key->base);
+ copy_uniform_state_to_shader(shader, 1, cs_c);
+
+ uint32_t vs_size = vs_c->qpu_inst_count * sizeof(uint64_t);
+ uint32_t cs_size = cs_c->qpu_inst_count * sizeof(uint64_t);
shader->coord_shader_offset = vs_size; /* XXX: alignment? */
shader->bo = vc4_bo_alloc(vc4->screen,
shader->coord_shader_offset + cs_size,
"vs_code");
void *map = vc4_bo_map(shader->bo);
- memcpy(map, vs_trans->c->qpu_insts, vs_size);
+ memcpy(map, vs_c->qpu_insts, vs_size);
memcpy(map + shader->coord_shader_offset,
- cs_trans->c->qpu_insts, cs_size);
+ cs_c->qpu_insts, cs_size);
- qir_compile_destroy(vs_trans->c);
- qir_compile_destroy(cs_trans->c);
+ qir_compile_destroy(vs_c);
+ qir_compile_destroy(cs_c);
}
static void
-vc4_update_compiled_fs(struct vc4_context *vc4)
+vc4_setup_shared_key(struct vc4_key *key, struct vc4_texture_stateobj *texstate)
+{
+ for (int i = 0; i < texstate->num_textures; i++) {
+ struct pipe_sampler_view *sampler = texstate->textures[i];
+ struct pipe_sampler_state *sampler_state =
+ texstate->samplers[i];
+
+ if (sampler) {
+ key->tex[i].format = sampler->format;
+ key->tex[i].swizzle[0] = sampler->swizzle_r;
+ key->tex[i].swizzle[1] = sampler->swizzle_g;
+ key->tex[i].swizzle[2] = sampler->swizzle_b;
+ key->tex[i].swizzle[3] = sampler->swizzle_a;
+ key->tex[i].compare_mode = sampler_state->compare_mode;
+ key->tex[i].compare_func = sampler_state->compare_func;
+ key->tex[i].wrap_s = sampler_state->wrap_s;
+ key->tex[i].wrap_t = sampler_state->wrap_t;
+ }
+ }
+}
+
+static void
+vc4_update_compiled_fs(struct vc4_context *vc4, uint8_t prim_mode)
{
struct vc4_fs_key local_key;
struct vc4_fs_key *key = &local_key;
memset(key, 0, sizeof(*key));
+ vc4_setup_shared_key(&key->base, &vc4->fragtex);
key->base.shader_state = vc4->prog.bind_fs;
+ key->is_points = (prim_mode == PIPE_PRIM_POINTS);
+ key->is_lines = (prim_mode >= PIPE_PRIM_LINES &&
+ prim_mode <= PIPE_PRIM_LINE_STRIP);
+ key->blend = vc4->blend->rt[0];
if (vc4->framebuffer.cbufs[0])
key->color_format = vc4->framebuffer.cbufs[0]->format;
+ key->stencil_enabled = vc4->zsa->stencil_uniforms[0] != 0;
+ key->stencil_twoside = vc4->zsa->stencil_uniforms[1] != 0;
+ key->stencil_full_writemasks = vc4->zsa->stencil_uniforms[2] != 0;
+ key->depth_enabled = (vc4->zsa->base.depth.enabled ||
+ key->stencil_enabled);
+ if (vc4->zsa->base.alpha.enabled) {
+ key->alpha_test = true;
+ key->alpha_test_func = vc4->zsa->base.alpha.func;
+ }
+
+ if (key->is_points) {
+ key->point_sprite_mask =
+ vc4->rasterizer->base.sprite_coord_enable;
+ key->point_coord_upper_left =
+ (vc4->rasterizer->base.sprite_coord_mode ==
+ PIPE_SPRITE_COORD_UPPER_LEFT);
+ }
+
+ key->light_twoside = vc4->rasterizer->base.light_twoside;
+
vc4->prog.fs = util_hash_table_get(vc4->fs_cache, key);
if (vc4->prog.fs)
return;
vc4_fs_compile(vc4, shader, key);
util_hash_table_set(vc4->fs_cache, key, shader);
+ if (vc4->rasterizer->base.flatshade &&
+ vc4->prog.fs &&
+ vc4->prog.fs->color_inputs != shader->color_inputs) {
+ vc4->dirty |= VC4_DIRTY_FLAT_SHADE_FLAGS;
+ }
+
vc4->prog.fs = shader;
}
static void
-vc4_update_compiled_vs(struct vc4_context *vc4)
+vc4_update_compiled_vs(struct vc4_context *vc4, uint8_t prim_mode)
{
struct vc4_vs_key local_key;
struct vc4_vs_key *key = &local_key;
memset(key, 0, sizeof(*key));
+ vc4_setup_shared_key(&key->base, &vc4->verttex);
key->base.shader_state = vc4->prog.bind_vs;
+ for (int i = 0; i < ARRAY_SIZE(key->attr_formats); i++)
+ key->attr_formats[i] = vc4->vtx->pipe[i].src_format;
+
+ key->per_vertex_point_size =
+ (prim_mode == PIPE_PRIM_POINTS &&
+ vc4->rasterizer->base.point_size_per_vertex);
+
vc4->prog.vs = util_hash_table_get(vc4->vs_cache, key);
if (vc4->prog.vs)
return;
}
void
-vc4_update_compiled_shaders(struct vc4_context *vc4)
+vc4_update_compiled_shaders(struct vc4_context *vc4, uint8_t prim_mode)
{
- vc4_update_compiled_fs(vc4);
- vc4_update_compiled_vs(vc4);
+ vc4_update_compiled_fs(vc4, prim_mode);
+ vc4_update_compiled_vs(vc4, prim_mode);
}
static unsigned
struct delete_state {
struct vc4_context *vc4;
- struct vc4_shader_state *shader_state;
+ struct vc4_uncompiled_shader *shader_state;
};
static enum pipe_error
vc4_shader_state_delete(struct pipe_context *pctx, void *hwcso)
{
struct vc4_context *vc4 = vc4_context(pctx);
- struct vc4_shader_state *so = hwcso;
+ struct vc4_uncompiled_shader *so = hwcso;
struct delete_state del;
del.vc4 = vc4;
util_hash_table_foreach(vc4->fs_cache, fs_delete_from_cache, &del);
util_hash_table_foreach(vc4->vs_cache, vs_delete_from_cache, &del);
+ if (so->twoside_tokens != so->base.tokens)
+ free((void *)so->twoside_tokens);
free((void *)so->base.tokens);
free(so);
}
+static uint32_t translate_wrap(uint32_t p_wrap, bool using_nearest)
+{
+ switch (p_wrap) {
+ case PIPE_TEX_WRAP_REPEAT:
+ return 0;
+ case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
+ return 1;
+ case PIPE_TEX_WRAP_MIRROR_REPEAT:
+ return 2;
+ case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
+ return 3;
+ case PIPE_TEX_WRAP_CLAMP:
+ return (using_nearest ? 1 : 3);
+ default:
+ fprintf(stderr, "Unknown wrap mode %d\n", p_wrap);
+ assert(!"not reached");
+ return 0;
+ }
+}
+
+static void
+write_texture_p0(struct vc4_context *vc4,
+ struct vc4_texture_stateobj *texstate,
+ uint32_t unit)
+{
+ struct pipe_sampler_view *texture = texstate->textures[unit];
+ struct vc4_resource *rsc = vc4_resource(texture->texture);
+
+ cl_reloc(vc4, &vc4->uniforms, rsc->bo,
+ VC4_SET_FIELD(rsc->slices[0].offset >> 12, VC4_TEX_P0_OFFSET) |
+ VC4_SET_FIELD(texture->u.tex.last_level, VC4_TEX_P0_MIPLVLS) |
+ VC4_SET_FIELD(texture->target == PIPE_TEXTURE_CUBE,
+ VC4_TEX_P0_CMMODE) |
+ VC4_SET_FIELD(rsc->vc4_format & 7, VC4_TEX_P0_TYPE));
+}
+
+static void
+write_texture_p1(struct vc4_context *vc4,
+ struct vc4_texture_stateobj *texstate,
+ uint32_t unit)
+{
+ struct pipe_sampler_view *texture = texstate->textures[unit];
+ struct vc4_resource *rsc = vc4_resource(texture->texture);
+ struct pipe_sampler_state *sampler = texstate->samplers[unit];
+ static const uint8_t minfilter_map[6] = {
+ VC4_TEX_P1_MINFILT_NEAR_MIP_NEAR,
+ VC4_TEX_P1_MINFILT_LIN_MIP_NEAR,
+ VC4_TEX_P1_MINFILT_NEAR_MIP_LIN,
+ VC4_TEX_P1_MINFILT_LIN_MIP_LIN,
+ VC4_TEX_P1_MINFILT_NEAREST,
+ VC4_TEX_P1_MINFILT_LINEAR,
+ };
+ static const uint32_t magfilter_map[] = {
+ [PIPE_TEX_FILTER_NEAREST] = VC4_TEX_P1_MAGFILT_NEAREST,
+ [PIPE_TEX_FILTER_LINEAR] = VC4_TEX_P1_MAGFILT_LINEAR,
+ };
+
+ bool either_nearest =
+ (sampler->mag_img_filter == PIPE_TEX_MIPFILTER_NEAREST ||
+ sampler->min_img_filter == PIPE_TEX_MIPFILTER_NEAREST);
+
+ cl_u32(&vc4->uniforms,
+ VC4_SET_FIELD(rsc->vc4_format >> 4, VC4_TEX_P1_TYPE4) |
+ VC4_SET_FIELD(texture->texture->height0 & 2047,
+ VC4_TEX_P1_HEIGHT) |
+ VC4_SET_FIELD(texture->texture->width0 & 2047,
+ VC4_TEX_P1_WIDTH) |
+ VC4_SET_FIELD(magfilter_map[sampler->mag_img_filter],
+ VC4_TEX_P1_MAGFILT) |
+ VC4_SET_FIELD(minfilter_map[sampler->min_mip_filter * 2 +
+ sampler->min_img_filter],
+ VC4_TEX_P1_MINFILT) |
+ VC4_SET_FIELD(translate_wrap(sampler->wrap_s, either_nearest),
+ VC4_TEX_P1_WRAP_S) |
+ VC4_SET_FIELD(translate_wrap(sampler->wrap_t, either_nearest),
+ VC4_TEX_P1_WRAP_T));
+}
+
+static void
+write_texture_p2(struct vc4_context *vc4,
+ struct vc4_texture_stateobj *texstate,
+ uint32_t unit)
+{
+ struct pipe_sampler_view *texture = texstate->textures[unit];
+ struct vc4_resource *rsc = vc4_resource(texture->texture);
+
+ cl_u32(&vc4->uniforms,
+ VC4_SET_FIELD(VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE,
+ VC4_TEX_P2_PTYPE) |
+ VC4_SET_FIELD(rsc->cube_map_stride >> 12, VC4_TEX_P2_CMST));
+}
+
+
+#define SWIZ(x,y,z,w) { \
+ UTIL_FORMAT_SWIZZLE_##x, \
+ UTIL_FORMAT_SWIZZLE_##y, \
+ UTIL_FORMAT_SWIZZLE_##z, \
+ UTIL_FORMAT_SWIZZLE_##w \
+}
+
+static void
+write_texture_border_color(struct vc4_context *vc4,
+ struct vc4_texture_stateobj *texstate,
+ uint32_t unit)
+{
+ struct pipe_sampler_state *sampler = texstate->samplers[unit];
+ struct pipe_sampler_view *texture = texstate->textures[unit];
+ struct vc4_resource *rsc = vc4_resource(texture->texture);
+ union util_color uc;
+
+ const struct util_format_description *tex_format_desc =
+ util_format_description(texture->format);
+
+ float border_color[4];
+ for (int i = 0; i < 4; i++)
+ border_color[i] = sampler->border_color.f[i];
+ if (util_format_is_srgb(texture->format)) {
+ for (int i = 0; i < 3; i++)
+ border_color[i] =
+ util_format_linear_to_srgb_float(border_color[i]);
+ }
+
+ /* Turn the border color into the layout of channels that it would
+ * have when stored as texture contents.
+ */
+ float storage_color[4];
+ util_format_unswizzle_4f(storage_color,
+ border_color,
+ tex_format_desc->swizzle);
+
+ /* Now, pack so that when the vc4_format-sampled texture contents are
+ * replaced with our border color, the vc4_get_format_swizzle()
+ * swizzling will get the right channels.
+ */
+ if (util_format_is_depth_or_stencil(texture->format)) {
+ uc.ui[0] = util_pack_z(PIPE_FORMAT_Z24X8_UNORM,
+ sampler->border_color.f[0]) << 8;
+ } else {
+ switch (rsc->vc4_format) {
+ default:
+ case VC4_TEXTURE_TYPE_RGBA8888:
+ util_pack_color(storage_color,
+ PIPE_FORMAT_R8G8B8A8_UNORM, &uc);
+ break;
+ case VC4_TEXTURE_TYPE_RGBA4444:
+ util_pack_color(storage_color,
+ PIPE_FORMAT_A8B8G8R8_UNORM, &uc);
+ break;
+ case VC4_TEXTURE_TYPE_RGB565:
+ util_pack_color(storage_color,
+ PIPE_FORMAT_B8G8R8A8_UNORM, &uc);
+ break;
+ case VC4_TEXTURE_TYPE_ALPHA:
+ uc.ui[0] = float_to_ubyte(storage_color[0]) << 24;
+ break;
+ case VC4_TEXTURE_TYPE_LUMALPHA:
+ uc.ui[0] = ((float_to_ubyte(storage_color[1]) << 24) |
+ (float_to_ubyte(storage_color[0]) << 0));
+ break;
+ }
+ }
+
+ cl_u32(&vc4->uniforms, uc.ui[0]);
+}
+
+static uint32_t
+get_texrect_scale(struct vc4_texture_stateobj *texstate,
+ enum quniform_contents contents,
+ uint32_t data)
+{
+ struct pipe_sampler_view *texture = texstate->textures[data];
+ uint32_t dim;
+
+ if (contents == QUNIFORM_TEXRECT_SCALE_X)
+ dim = texture->texture->width0;
+ else
+ dim = texture->texture->height0;
+
+ return fui(1.0f / dim);
+}
+
void
-vc4_get_uniform_bo(struct vc4_context *vc4, struct vc4_compiled_shader *shader,
+vc4_write_uniforms(struct vc4_context *vc4, struct vc4_compiled_shader *shader,
struct vc4_constbuf_stateobj *cb,
- int shader_index, struct vc4_bo **out_bo,
- uint32_t *out_offset)
+ struct vc4_texture_stateobj *texstate,
+ int shader_index)
{
struct vc4_shader_uniform_info *uinfo = &shader->uniforms[shader_index];
- struct vc4_bo *ubo = vc4_bo_alloc(vc4->screen,
- MAX2(1, uinfo->count * 4), "ubo");
- uint32_t *map = vc4_bo_map(ubo);
+ const uint32_t *gallium_uniforms = cb->cb[0].user_buffer;
+
+ cl_start_shader_reloc(&vc4->uniforms, uinfo->num_texture_samples);
for (int i = 0; i < uinfo->count; i++) {
+
switch (uinfo->contents[i]) {
case QUNIFORM_CONSTANT:
- map[i] = uinfo->data[i];
+ cl_u32(&vc4->uniforms, uinfo->data[i]);
break;
case QUNIFORM_UNIFORM:
- map[i] = ((uint32_t *)cb->cb[0].user_buffer)[uinfo->data[i]];
+ cl_u32(&vc4->uniforms,
+ gallium_uniforms[uinfo->data[i]]);
break;
case QUNIFORM_VIEWPORT_X_SCALE:
- map[i] = fui(vc4->framebuffer.width * 16.0f / 2.0f);
+ cl_f(&vc4->uniforms, vc4->viewport.scale[0] * 16.0f);
break;
case QUNIFORM_VIEWPORT_Y_SCALE:
- map[i] = fui(vc4->framebuffer.height * -16.0f / 2.0f);
+ cl_f(&vc4->uniforms, vc4->viewport.scale[1] * 16.0f);
+ break;
+
+ case QUNIFORM_VIEWPORT_Z_OFFSET:
+ cl_f(&vc4->uniforms, vc4->viewport.translate[2]);
+ break;
+ case QUNIFORM_VIEWPORT_Z_SCALE:
+ cl_f(&vc4->uniforms, vc4->viewport.scale[2]);
+ break;
+
+ case QUNIFORM_TEXTURE_CONFIG_P0:
+ write_texture_p0(vc4, texstate, uinfo->data[i]);
+ break;
+
+ case QUNIFORM_TEXTURE_CONFIG_P1:
+ write_texture_p1(vc4, texstate, uinfo->data[i]);
+ break;
+
+ case QUNIFORM_TEXTURE_CONFIG_P2:
+ write_texture_p2(vc4, texstate, uinfo->data[i]);
+ break;
+
+ case QUNIFORM_TEXTURE_BORDER_COLOR:
+ write_texture_border_color(vc4, texstate, uinfo->data[i]);
+ break;
+
+ case QUNIFORM_TEXRECT_SCALE_X:
+ case QUNIFORM_TEXRECT_SCALE_Y:
+ cl_u32(&vc4->uniforms,
+ get_texrect_scale(texstate,
+ uinfo->contents[i],
+ uinfo->data[i]));
+ break;
+
+ case QUNIFORM_BLEND_CONST_COLOR:
+ cl_f(&vc4->uniforms,
+ vc4->blend_color.color[uinfo->data[i]]);
+ break;
+
+ case QUNIFORM_STENCIL:
+ cl_u32(&vc4->uniforms,
+ vc4->zsa->stencil_uniforms[uinfo->data[i]] |
+ (uinfo->data[i] <= 1 ?
+ (vc4->stencil_ref.ref_value[uinfo->data[i]] << 8) :
+ 0));
+ break;
+
+ case QUNIFORM_ALPHA_REF:
+ cl_f(&vc4->uniforms, vc4->zsa->base.alpha.ref_value);
break;
}
#if 0
+ uint32_t written_val = *(uint32_t *)(vc4->uniforms.next - 4);
fprintf(stderr, "%p/%d: %d: 0x%08x (%f)\n",
- shader, shader_index, i, map[i], uif(map[i]));
+ shader, shader_index, i, written_val, uif(written_val));
#endif
}
-
- *out_bo = ubo;
- *out_offset = 0;
}
static void
projects / mesa.git / blobdiff