broadcom/vc5: Move V3D 3.3 texturing to a separate file.
authorEric Anholt <eric@anholt.net>
Wed, 10 Jan 2018 20:48:05 +0000 (12:48 -0800)
committerEric Anholt <eric@anholt.net>
Sat, 13 Jan 2018 05:56:37 +0000 (21:56 -0800)
V3D 4.x texturing changes enough that #ifdefs would just make a mess of
it.

src/broadcom/Makefile.sources
src/broadcom/compiler/meson.build
src/broadcom/compiler/nir_to_vir.c
src/broadcom/compiler/v3d33_tex.c [new file with mode: 0644]
src/broadcom/compiler/v3d_compiler.h

index 827c1944af6c59b803c59ce846704d07876ae6b4..e8f82fbeaac378cda49cbac0a02274866bf18cc5 100644 (file)
@@ -29,6 +29,7 @@ BROADCOM_FILES = \
        compiler/vir_to_qpu.c \
        compiler/qpu_schedule.c \
        compiler/qpu_validate.c \
+       compiler/v3d33_tex.c \
        compiler/v3d33_vpm_setup.c \
        compiler/v3d_compiler.h \
        compiler/v3d_nir_lower_io.c \
index 8a7211b649ced7b2d24ae4a1123a1632bb4621c8..2c1b8d40b74cc5b55eba593c9547dcc3a45a4851 100644 (file)
@@ -30,6 +30,7 @@ libbroadcom_compiler_files = files(
   'vir_to_qpu.c',
   'qpu_schedule.c',
   'qpu_validate.c',
+  'v3d33_tex.c',
   'v3d33_vpm_setup.c',
   'v3d_compiler.h',
   'v3d_nir_lower_io.c',
index e76a87ca4cb283a1b51a30e887055679f7c89f24..cec00cb55f756e15e81e4943d3968d483b79b2ac 100644 (file)
 #include "common/v3d_device_info.h"
 #include "v3d_compiler.h"
 
-/* We don't do any address packing. */
-#define __gen_user_data void
-#define __gen_address_type uint32_t
-#define __gen_address_offset(reloc) (*reloc)
-#define __gen_emit_reloc(cl, reloc)
-#include "cle/v3d_packet_v33_pack.h"
-
-static struct qreg
-ntq_get_src(struct v3d_compile *c, nir_src src, int i);
 static void
 ntq_emit_cf_list(struct v3d_compile *c, struct exec_list *list);
 
@@ -65,7 +56,7 @@ resize_qreg_array(struct v3d_compile *c,
                 (*regs)[i] = c->undef;
 }
 
-static void
+void
 vir_emit_thrsw(struct v3d_compile *c)
 {
         if (c->threads == 1)
@@ -89,13 +80,6 @@ vir_SFU(struct v3d_compile *c, int waddr, struct qreg src)
         return vir_FMOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4));
 }
 
-static struct qreg
-vir_LDTMU(struct v3d_compile *c)
-{
-        vir_NOP(c)->qpu.sig.ldtmu = true;
-        return vir_MOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4));
-}
-
 static struct qreg
 indirect_uniform_load(struct v3d_compile *c, nir_intrinsic_instr *intr)
 {
@@ -163,7 +147,7 @@ ntq_init_ssa_def(struct v3d_compile *c, nir_ssa_def *def)
  * (knowing that the previous instruction doesn't depend on flags) and rewrite
  * its destination to be the NIR reg's destination
  */
-static void
+void
 ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan,
                struct qreg result)
 {
@@ -229,7 +213,7 @@ ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan,
         }
 }
 
-static struct qreg
+struct qreg
 ntq_get_src(struct v3d_compile *c, nir_src src, int i)
 {
         struct hash_entry *entry;
@@ -350,216 +334,7 @@ ntq_emit_tex(struct v3d_compile *c, nir_tex_instr *instr)
                 break;
         }
 
-        struct V3D33_TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1 p0_unpacked = {
-                V3D33_TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1_header,
-
-                .fetch_sample_mode = instr->op == nir_texop_txf,
-        };
-
-        struct V3D33_TEXTURE_UNIFORM_PARAMETER_1_CFG_MODE1 p1_unpacked = {
-        };
-
-        switch (instr->sampler_dim) {
-        case GLSL_SAMPLER_DIM_1D:
-                if (instr->is_array)
-                        p0_unpacked.lookup_type = TEXTURE_1D_ARRAY;
-                else
-                        p0_unpacked.lookup_type = TEXTURE_1D;
-                break;
-        case GLSL_SAMPLER_DIM_2D:
-        case GLSL_SAMPLER_DIM_RECT:
-                if (instr->is_array)
-                        p0_unpacked.lookup_type = TEXTURE_2D_ARRAY;
-                else
-                        p0_unpacked.lookup_type = TEXTURE_2D;
-                break;
-        case GLSL_SAMPLER_DIM_3D:
-                p0_unpacked.lookup_type = TEXTURE_3D;
-                break;
-        case GLSL_SAMPLER_DIM_CUBE:
-                p0_unpacked.lookup_type = TEXTURE_CUBE_MAP;
-                break;
-        default:
-                unreachable("Bad sampler type");
-        }
-
-        struct qreg coords[5];
-        int next_coord = 0;
-        for (unsigned i = 0; i < instr->num_srcs; i++) {
-                switch (instr->src[i].src_type) {
-                case nir_tex_src_coord:
-                        for (int j = 0; j < instr->coord_components; j++) {
-                                coords[next_coord++] =
-                                        ntq_get_src(c, instr->src[i].src, j);
-                        }
-                        if (instr->coord_components < 2)
-                                coords[next_coord++] = vir_uniform_f(c, 0.5);
-                        break;
-                case nir_tex_src_bias:
-                        coords[next_coord++] =
-                                ntq_get_src(c, instr->src[i].src, 0);
-
-                        p0_unpacked.bias_supplied = true;
-                        break;
-                case nir_tex_src_lod:
-                        coords[next_coord++] =
-                                vir_FADD(c,
-                                         ntq_get_src(c, instr->src[i].src, 0),
-                                         vir_uniform(c, QUNIFORM_TEXTURE_FIRST_LEVEL,
-                                                     unit));
-
-                        if (instr->op != nir_texop_txf &&
-                            instr->op != nir_texop_tg4) {
-                                p0_unpacked.disable_autolod_use_bias_only = true;
-                        }
-                        break;
-                case nir_tex_src_comparator:
-                        coords[next_coord++] =
-                                ntq_get_src(c, instr->src[i].src, 0);
-
-                        p0_unpacked.shadow = true;
-                        break;
-
-                case nir_tex_src_offset: {
-                        nir_const_value *offset =
-                                nir_src_as_const_value(instr->src[i].src);
-                        p0_unpacked.texel_offset_for_s_coordinate =
-                                offset->i32[0];
-
-                        if (instr->coord_components >= 2)
-                                p0_unpacked.texel_offset_for_t_coordinate =
-                                        offset->i32[1];
-
-                        if (instr->coord_components >= 3)
-                                p0_unpacked.texel_offset_for_r_coordinate =
-                                        offset->i32[2];
-                        break;
-                }
-
-                default:
-                        unreachable("unknown texture source");
-                }
-        }
-
-        bool return_16 = (c->key->tex[unit].return_size == 16 ||
-                          p0_unpacked.shadow);
-
-        /* Limit the number of channels returned to both how many the NIR
-         * instruction writes and how many the instruction could produce.
-         */
-        uint32_t instr_return_channels = nir_tex_instr_dest_size(instr);
-        if (return_16)
-                instr_return_channels = (instr_return_channels + 1) / 2;
-
-        p1_unpacked.return_words_of_texture_data =
-                (1 << MIN2(instr_return_channels,
-                           c->key->tex[unit].return_channels)) - 1;
-
-        uint32_t p0_packed;
-        V3D33_TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1_pack(NULL,
-                                                         (uint8_t *)&p0_packed,
-                                                         &p0_unpacked);
-
-        uint32_t p1_packed;
-        V3D33_TEXTURE_UNIFORM_PARAMETER_1_CFG_MODE1_pack(NULL,
-                                                         (uint8_t *)&p1_packed,
-                                                         &p1_unpacked);
-        /* Load unit number into the address field, which will be be used by
-         * the driver to decide which texture to put in the actual address
-         * field.
-         */
-        p1_packed |= unit << 5;
-
-        /* 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 (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
-                coords[0] = vir_FMUL(c, coords[0],
-                                     vir_uniform(c, QUNIFORM_TEXRECT_SCALE_X,
-                                                 unit));
-                coords[1] = vir_FMUL(c, coords[1],
-                                     vir_uniform(c, QUNIFORM_TEXRECT_SCALE_Y,
-                                                 unit));
-        }
-
-        struct qreg texture_u[] = {
-                vir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P0_0 + unit, p0_packed),
-                vir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P1, p1_packed),
-        };
-        uint32_t next_texture_u = 0;
-
-        for (int i = 0; i < next_coord; i++) {
-                struct qreg dst;
-
-                if (i == next_coord - 1)
-                        dst = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUL);
-                else
-                        dst = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMU);
-
-                struct qinst *tmu = vir_MOV_dest(c, dst, coords[i]);
-
-                if (i < 2) {
-                        tmu->has_implicit_uniform = true;
-                        tmu->src[vir_get_implicit_uniform_src(tmu)] =
-                                texture_u[next_texture_u++];
-                }
-        }
-
-        vir_emit_thrsw(c);
-
-        struct qreg return_values[4];
-        for (int i = 0; i < 4; i++) {
-                /* Swizzling .zw of an RG texture should give undefined
-                 * results, not crash the compiler.
-                 */
-                if (p1_unpacked.return_words_of_texture_data & (1 << i))
-                        return_values[i] = vir_LDTMU(c);
-                else
-                        return_values[i] = c->undef;
-        }
-
-        for (int i = 0; i < nir_tex_instr_dest_size(instr); i++) {
-                struct qreg chan;
-
-                if (return_16) {
-                        STATIC_ASSERT(PIPE_SWIZZLE_X == 0);
-                        chan = return_values[i / 2];
-
-                        if (nir_alu_type_get_base_type(instr->dest_type) ==
-                            nir_type_float) {
-                                enum v3d_qpu_input_unpack unpack;
-                                if (i & 1)
-                                        unpack = V3D_QPU_UNPACK_H;
-                                else
-                                        unpack = V3D_QPU_UNPACK_L;
-
-                                chan = vir_FMOV(c, chan);
-                                vir_set_unpack(c->defs[chan.index], 0, unpack);
-                        } else {
-                                /* If we're unpacking the low field, shift it
-                                 * up to the top first.
-                                 */
-                                if ((i & 1) == 0) {
-                                        chan = vir_SHL(c, chan,
-                                                       vir_uniform_ui(c, 16));
-                                }
-
-                                /* Do proper sign extension to a 32-bit int. */
-                                if (nir_alu_type_get_base_type(instr->dest_type) ==
-                                    nir_type_int) {
-                                        chan = vir_ASR(c, chan,
-                                                       vir_uniform_ui(c, 16));
-                                } else {
-                                        chan = vir_SHR(c, chan,
-                                                       vir_uniform_ui(c, 16));
-                                }
-                        }
-                } else {
-                        chan = vir_MOV(c, return_values[i]);
-                }
-                ntq_store_dest(c, &instr->dest, i, chan);
-        }
+        v3d33_vir_emit_tex(c, instr);
 }
 
 static struct qreg
diff --git a/src/broadcom/compiler/v3d33_tex.c b/src/broadcom/compiler/v3d33_tex.c
new file mode 100644 (file)
index 0000000..9af9285
--- /dev/null
@@ -0,0 +1,248 @@
+/*
+ * Copyright © 2016-2018 Broadcom
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "v3d_compiler.h"
+
+/* We don't do any address packing. */
+#define __gen_user_data void
+#define __gen_address_type uint32_t
+#define __gen_address_offset(reloc) (*reloc)
+#define __gen_emit_reloc(cl, reloc)
+#include "cle/v3d_packet_v33_pack.h"
+
+void
+v3d33_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr)
+{
+        unsigned unit = instr->texture_index;
+
+        struct V3D33_TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1 p0_unpacked = {
+                V3D33_TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1_header,
+
+                .fetch_sample_mode = instr->op == nir_texop_txf,
+        };
+
+        struct V3D33_TEXTURE_UNIFORM_PARAMETER_1_CFG_MODE1 p1_unpacked = {
+        };
+
+        switch (instr->sampler_dim) {
+        case GLSL_SAMPLER_DIM_1D:
+                if (instr->is_array)
+                        p0_unpacked.lookup_type = TEXTURE_1D_ARRAY;
+                else
+                        p0_unpacked.lookup_type = TEXTURE_1D;
+                break;
+        case GLSL_SAMPLER_DIM_2D:
+        case GLSL_SAMPLER_DIM_RECT:
+                if (instr->is_array)
+                        p0_unpacked.lookup_type = TEXTURE_2D_ARRAY;
+                else
+                        p0_unpacked.lookup_type = TEXTURE_2D;
+                break;
+        case GLSL_SAMPLER_DIM_3D:
+                p0_unpacked.lookup_type = TEXTURE_3D;
+                break;
+        case GLSL_SAMPLER_DIM_CUBE:
+                p0_unpacked.lookup_type = TEXTURE_CUBE_MAP;
+                break;
+        default:
+                unreachable("Bad sampler type");
+        }
+
+        struct qreg coords[5];
+        int next_coord = 0;
+        for (unsigned i = 0; i < instr->num_srcs; i++) {
+                switch (instr->src[i].src_type) {
+                case nir_tex_src_coord:
+                        for (int j = 0; j < instr->coord_components; j++) {
+                                coords[next_coord++] =
+                                        ntq_get_src(c, instr->src[i].src, j);
+                        }
+                        if (instr->coord_components < 2)
+                                coords[next_coord++] = vir_uniform_f(c, 0.5);
+                        break;
+                case nir_tex_src_bias:
+                        coords[next_coord++] =
+                                ntq_get_src(c, instr->src[i].src, 0);
+
+                        p0_unpacked.bias_supplied = true;
+                        break;
+                case nir_tex_src_lod:
+                        coords[next_coord++] =
+                                vir_FADD(c,
+                                         ntq_get_src(c, instr->src[i].src, 0),
+                                         vir_uniform(c, QUNIFORM_TEXTURE_FIRST_LEVEL,
+                                                     unit));
+
+                        if (instr->op != nir_texop_txf &&
+                            instr->op != nir_texop_tg4) {
+                                p0_unpacked.disable_autolod_use_bias_only = true;
+                        }
+                        break;
+                case nir_tex_src_comparator:
+                        coords[next_coord++] =
+                                ntq_get_src(c, instr->src[i].src, 0);
+
+                        p0_unpacked.shadow = true;
+                        break;
+
+                case nir_tex_src_offset: {
+                        nir_const_value *offset =
+                                nir_src_as_const_value(instr->src[i].src);
+                        p0_unpacked.texel_offset_for_s_coordinate =
+                                offset->i32[0];
+
+                        if (instr->coord_components >= 2)
+                                p0_unpacked.texel_offset_for_t_coordinate =
+                                        offset->i32[1];
+
+                        if (instr->coord_components >= 3)
+                                p0_unpacked.texel_offset_for_r_coordinate =
+                                        offset->i32[2];
+                        break;
+                }
+
+                default:
+                        unreachable("unknown texture source");
+                }
+        }
+
+        bool return_16 = (c->key->tex[unit].return_size == 16 ||
+                          p0_unpacked.shadow);
+
+        /* Limit the number of channels returned to both how many the NIR
+         * instruction writes and how many the instruction could produce.
+         */
+        uint32_t instr_return_channels = nir_tex_instr_dest_size(instr);
+        if (return_16)
+                instr_return_channels = (instr_return_channels + 1) / 2;
+
+        p1_unpacked.return_words_of_texture_data =
+                (1 << MIN2(instr_return_channels,
+                           c->key->tex[unit].return_channels)) - 1;
+
+        uint32_t p0_packed;
+        V3D33_TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1_pack(NULL,
+                                                         (uint8_t *)&p0_packed,
+                                                         &p0_unpacked);
+
+        uint32_t p1_packed;
+        V3D33_TEXTURE_UNIFORM_PARAMETER_1_CFG_MODE1_pack(NULL,
+                                                         (uint8_t *)&p1_packed,
+                                                         &p1_unpacked);
+        /* Load unit number into the address field, which will be be used by
+         * the driver to decide which texture to put in the actual address
+         * field.
+         */
+        p1_packed |= unit << 5;
+
+        /* 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 (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
+                coords[0] = vir_FMUL(c, coords[0],
+                                     vir_uniform(c, QUNIFORM_TEXRECT_SCALE_X,
+                                                 unit));
+                coords[1] = vir_FMUL(c, coords[1],
+                                     vir_uniform(c, QUNIFORM_TEXRECT_SCALE_Y,
+                                                 unit));
+        }
+
+        struct qreg texture_u[] = {
+                vir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P0_0 + unit, p0_packed),
+                vir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P1, p1_packed),
+        };
+        uint32_t next_texture_u = 0;
+
+        for (int i = 0; i < next_coord; i++) {
+                struct qreg dst;
+
+                if (i == next_coord - 1)
+                        dst = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUL);
+                else
+                        dst = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMU);
+
+                struct qinst *tmu = vir_MOV_dest(c, dst, coords[i]);
+
+                if (i < 2) {
+                        tmu->has_implicit_uniform = true;
+                        tmu->src[vir_get_implicit_uniform_src(tmu)] =
+                                texture_u[next_texture_u++];
+                }
+        }
+
+        vir_emit_thrsw(c);
+
+        struct qreg return_values[4];
+        for (int i = 0; i < 4; i++) {
+                /* Swizzling .zw of an RG texture should give undefined
+                 * results, not crash the compiler.
+                 */
+                if (p1_unpacked.return_words_of_texture_data & (1 << i))
+                        return_values[i] = vir_LDTMU(c);
+                else
+                        return_values[i] = c->undef;
+        }
+
+        for (int i = 0; i < nir_tex_instr_dest_size(instr); i++) {
+                struct qreg chan;
+
+                if (return_16) {
+                        STATIC_ASSERT(PIPE_SWIZZLE_X == 0);
+                        chan = return_values[i / 2];
+
+                        if (nir_alu_type_get_base_type(instr->dest_type) ==
+                            nir_type_float) {
+                                enum v3d_qpu_input_unpack unpack;
+                                if (i & 1)
+                                        unpack = V3D_QPU_UNPACK_H;
+                                else
+                                        unpack = V3D_QPU_UNPACK_L;
+
+                                chan = vir_FMOV(c, chan);
+                                vir_set_unpack(c->defs[chan.index], 0, unpack);
+                        } else {
+                                /* If we're unpacking the low field, shift it
+                                 * up to the top first.
+                                 */
+                                if ((i & 1) == 0) {
+                                        chan = vir_SHL(c, chan,
+                                                       vir_uniform_ui(c, 16));
+                                }
+
+                                /* Do proper sign extension to a 32-bit int. */
+                                if (nir_alu_type_get_base_type(instr->dest_type) ==
+                                    nir_type_int) {
+                                        chan = vir_ASR(c, chan,
+                                                       vir_uniform_ui(c, 16));
+                                } else {
+                                        chan = vir_SHR(c, chan,
+                                                       vir_uniform_ui(c, 16));
+                                }
+                        }
+                } else {
+                        chan = vir_MOV(c, return_values[i]);
+                }
+                ntq_store_dest(c, &instr->dest, i, chan);
+        }
+}
index 15cc16ceb6067f723719c31b4009fbfcdd85673c..cb501695aed77446c340fc6d1c497c60f23e567e 100644 (file)
@@ -665,6 +665,10 @@ bool vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst);
 bool vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst);
 struct qreg vir_follow_movs(struct v3d_compile *c, struct qreg reg);
 uint8_t vir_channels_written(struct qinst *inst);
+struct qreg ntq_get_src(struct v3d_compile *c, nir_src src, int i);
+void ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan,
+                    struct qreg result);
+void vir_emit_thrsw(struct v3d_compile *c);
 
 void vir_dump(struct v3d_compile *c);
 void vir_dump_inst(struct v3d_compile *c, struct qinst *inst);
@@ -686,6 +690,7 @@ void vir_lower_uniforms(struct v3d_compile *c);
 
 void v3d33_vir_vpm_read_setup(struct v3d_compile *c, int num_components);
 void v3d33_vir_vpm_write_setup(struct v3d_compile *c);
+void v3d33_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
 
 void v3d_vir_to_qpu(struct v3d_compile *c, struct qpu_reg *temp_registers);
 uint32_t v3d_qpu_schedule_instructions(struct v3d_compile *c);
@@ -887,6 +892,14 @@ vir_NOP(struct v3d_compile *c)
         return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_NOP,
                                                c->undef, c->undef, c->undef));
 }
+
+static inline struct qreg
+vir_LDTMU(struct v3d_compile *c)
+{
+        vir_NOP(c)->qpu.sig.ldtmu = true;
+        return vir_MOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4));
+}
+
 /*
 static inline struct qreg
 vir_LOAD_IMM(struct v3d_compile *c, uint32_t val)