+ lp_build_tgsi_soa(gallivm, tokens, type, &mask,
+ consts_ptr, &system_values,
+ interp->inputs,
+ outputs, sampler, &shader->info.base, NULL);
+
+ /* Alpha test */
+ if (key->alpha.enabled) {
+ int color0 = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_COLOR,
+ 0);
+
+ if (color0 != -1 && outputs[color0][3]) {
+ const struct util_format_description *cbuf_format_desc;
+ LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
+ LLVMValueRef alpha_ref_value;
+
+ alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
+ alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
+
+ cbuf_format_desc = util_format_description(key->cbuf_format[0]);
+
+ lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
+ &mask, alpha, alpha_ref_value,
+ (depth_mode & LATE_DEPTH_TEST) != 0);
+ }
+ }
+
+ /* Late Z test */
+ if (depth_mode & LATE_DEPTH_TEST) {
+ int pos0 = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_POSITION,
+ 0);
+
+ if (pos0 != -1 && outputs[pos0][2]) {
+ z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
+ }
+
+ lp_build_depth_stencil_load_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ depth_ptr, depth_stride,
+ &z_fb, &s_fb, loop_state.counter);
+
+ lp_build_depth_stencil_test(gallivm,
+ &key->depth,
+ key->stencil,
+ type,
+ zs_format_desc,
+ &mask,
+ stencil_refs,
+ z, z_fb, s_fb,
+ facing,
+ &z_value, &s_value,
+ !simple_shader);
+ /* Late Z write */
+ if (depth_mode & LATE_DEPTH_WRITE) {
+ lp_build_depth_stencil_write_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ NULL, NULL, NULL, loop_state.counter,
+ depth_ptr, depth_stride,
+ z_value, s_value);
+ }
+ }
+ else if ((depth_mode & EARLY_DEPTH_TEST) &&
+ (depth_mode & LATE_DEPTH_WRITE))
+ {
+ /* Need to apply a reduced mask to the depth write. Reload the
+ * depth value, update from zs_value with the new mask value and
+ * write that out.
+ */
+ lp_build_depth_stencil_write_swizzled(gallivm, type,
+ zs_format_desc, key->resource_1d,
+ &mask, z_fb, s_fb, loop_state.counter,
+ depth_ptr, depth_stride,
+ z_value, s_value);
+ }
+
+
+ /* Color write */
+ for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
+ {
+ unsigned cbuf = shader->info.base.output_semantic_index[attrib];
+ if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
+ ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
+ {
+ for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
+ if(outputs[attrib][chan]) {
+ /* XXX: just initialize outputs to point at colors[] and
+ * skip this.
+ */
+ LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
+ LLVMValueRef color_ptr;
+ color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
+ &loop_state.counter, 1, "");
+ lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
+ LLVMBuildStore(builder, out, color_ptr);
+ }
+ }
+ }
+ }
+
+ if (key->occlusion_count) {
+ LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
+ lp_build_name(counter, "counter");
+ lp_build_occlusion_count(gallivm, type,
+ lp_build_mask_value(&mask), counter);
+ }
+
+ mask_val = lp_build_mask_end(&mask);
+ LLVMBuildStore(builder, mask_val, mask_ptr);
+ lp_build_for_loop_end(&loop_state);
+}
+
+
+/**
+ * This function will reorder pixels from the fragment shader SoA to memory layout AoS
+ *
+ * Fragment Shader outputs pixels in small 2x2 blocks
+ * e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
+ *
+ * However in memory pixels are stored in rows
+ * e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
+ *
+ * @param type fragment shader type (4x or 8x float)
+ * @param num_fs number of fs_src
+ * @param is_1d whether we're outputting to a 1d resource
+ * @param dst_channels number of output channels
+ * @param fs_src output from fragment shader
+ * @param dst pointer to store result
+ * @param pad_inline is channel padding inline or at end of row
+ * @return the number of dsts
+ */
+static int
+generate_fs_twiddle(struct gallivm_state *gallivm,
+ struct lp_type type,
+ unsigned num_fs,
+ unsigned dst_channels,
+ LLVMValueRef fs_src[][4],
+ LLVMValueRef* dst,
+ bool pad_inline)
+{
+ LLVMValueRef src[16];
+
+ bool swizzle_pad;
+ bool twiddle;
+ bool split;
+
+ unsigned pixels = type.length / 4;
+ unsigned reorder_group;
+ unsigned src_channels;
+ unsigned src_count;
+ unsigned i;
+
+ src_channels = dst_channels < 3 ? dst_channels : 4;
+ src_count = num_fs * src_channels;
+
+ assert(pixels == 2 || pixels == 1);
+ assert(num_fs * src_channels <= Elements(src));
+
+ /*
+ * Transpose from SoA -> AoS
+ */
+ for (i = 0; i < num_fs; ++i) {
+ lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
+ }
+
+ /*
+ * Pick transformation options
+ */
+ swizzle_pad = false;
+ twiddle = false;
+ split = false;
+ reorder_group = 0;
+
+ if (dst_channels == 1) {
+ twiddle = true;
+
+ if (pixels == 2) {
+ split = true;
+ }
+ } else if (dst_channels == 2) {
+ if (pixels == 1) {
+ reorder_group = 1;
+ }
+ } else if (dst_channels > 2) {
+ if (pixels == 1) {
+ reorder_group = 2;
+ } else {
+ twiddle = true;
+ }
+
+ if (!pad_inline && dst_channels == 3 && pixels > 1) {
+ swizzle_pad = true;
+ }
+ }
+
+ /*
+ * Split the src in half
+ */
+ if (split) {
+ for (i = num_fs; i > 0; --i) {
+ src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
+ src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
+ }
+
+ src_count *= 2;
+ type.length = 4;
+ }
+
+ /*
+ * Ensure pixels are in memory order
+ */
+ if (reorder_group) {
+ /* Twiddle pixels by reordering the array, e.g.:
+ *
+ * src_count = 8 -> 0 2 1 3 4 6 5 7
+ * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
+ */
+ const unsigned reorder_sw[] = { 0, 2, 1, 3 };
+
+ for (i = 0; i < src_count; ++i) {
+ unsigned group = i / reorder_group;
+ unsigned block = (group / 4) * 4 * reorder_group;
+ unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
+ dst[i] = src[j];
+ }
+ } else if (twiddle) {
+ /* Twiddle pixels across elements of array */
+ lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
+ } else {
+ /* Do nothing */
+ memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
+ }
+
+ /*
+ * Moves any padding between pixels to the end
+ * e.g. RGBXRGBX -> RGBRGBXX
+ */
+ if (swizzle_pad) {
+ unsigned char swizzles[16];
+ unsigned elems = pixels * dst_channels;
+
+ for (i = 0; i < type.length; ++i) {
+ if (i < elems)
+ swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
+ else
+ swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
+ }
+
+ for (i = 0; i < src_count; ++i) {
+ dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
+ }
+ }
+
+ return src_count;
+}
+
+
+/**
+ * Load an unswizzled block of pixels from memory
+ */
+static void
+load_unswizzled_block(struct gallivm_state *gallivm,
+ LLVMValueRef base_ptr,
+ LLVMValueRef stride,
+ unsigned block_width,
+ unsigned block_height,
+ LLVMValueRef* dst,
+ struct lp_type dst_type,
+ unsigned dst_count,
+ unsigned dst_alignment)
+{
+ LLVMBuilderRef builder = gallivm->builder;
+ unsigned row_size = dst_count / block_height;
+ unsigned i;
+
+ /* Ensure block exactly fits into dst */
+ assert((block_width * block_height) % dst_count == 0);
+
+ for (i = 0; i < dst_count; ++i) {
+ unsigned x = i % row_size;
+ unsigned y = i / row_size;
+
+ LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
+ LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
+
+ LLVMValueRef gep[2];
+ LLVMValueRef dst_ptr;
+
+ gep[0] = lp_build_const_int32(gallivm, 0);
+ gep[1] = LLVMBuildAdd(builder, bx, by, "");
+
+ dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
+ dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
+
+ dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
+
+ lp_set_load_alignment(dst[i], dst_alignment);
+ }
+}
+
+
+/**
+ * Store an unswizzled block of pixels to memory
+ */
+static void
+store_unswizzled_block(struct gallivm_state *gallivm,
+ LLVMValueRef base_ptr,
+ LLVMValueRef stride,
+ unsigned block_width,
+ unsigned block_height,
+ LLVMValueRef* src,
+ struct lp_type src_type,
+ unsigned src_count,
+ unsigned src_alignment)
+{
+ LLVMBuilderRef builder = gallivm->builder;
+ unsigned row_size = src_count / block_height;
+ unsigned i;
+
+ /* Ensure src exactly fits into block */
+ assert((block_width * block_height) % src_count == 0);
+
+ for (i = 0; i < src_count; ++i) {
+ unsigned x = i % row_size;
+ unsigned y = i / row_size;
+
+ LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
+ LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
+
+ LLVMValueRef gep[2];
+ LLVMValueRef src_ptr;
+
+ gep[0] = lp_build_const_int32(gallivm, 0);
+ gep[1] = LLVMBuildAdd(builder, bx, by, "");
+
+ src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
+ src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
+
+ src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
+
+ lp_set_store_alignment(src_ptr, src_alignment);
+ }
+}
+
+
+/**
+ * Checks if a format description is an arithmetic format
+ *
+ * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
+ */
+static INLINE boolean
+is_arithmetic_format(const struct util_format_description *format_desc)
+{
+ boolean arith = false;
+ unsigned i;
+
+ for (i = 0; i < format_desc->nr_channels; ++i) {
+ arith |= format_desc->channel[i].size != format_desc->channel[0].size;
+ arith |= (format_desc->channel[i].size % 8) != 0;
+ }
+
+ return arith;
+}
+
+
+/**
+ * Checks if this format requires special handling due to required expansion
+ * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
+ * SoA conversion.
+ */
+static INLINE boolean
+format_expands_to_float_soa(const struct util_format_description *format_desc)
+{
+ if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
+ format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
+ return true;
+ }
+ return false;
+}
+
+
+/**
+ * Retrieves the type representing the memory layout for a format
+ *
+ * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
+ */
+static INLINE void
+lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
+ struct lp_type* type)
+{
+ unsigned i;
+ unsigned chan;
+
+ if (format_expands_to_float_soa(format_desc)) {
+ /* just make this a 32bit uint */
+ type->floating = false;
+ type->fixed = false;
+ type->sign = false;
+ type->norm = false;
+ type->width = 32;
+ type->length = 1;
+ return;
+ }
+
+ for (i = 0; i < 4; i++)
+ if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
+ break;
+ chan = i;
+
+ memset(type, 0, sizeof(struct lp_type));
+ type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
+ type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
+ type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
+ type->norm = format_desc->channel[chan].normalized;
+
+ if (is_arithmetic_format(format_desc)) {
+ type->width = 0;
+ type->length = 1;
+
+ for (i = 0; i < format_desc->nr_channels; ++i) {
+ type->width += format_desc->channel[i].size;
+ }
+ } else {
+ type->width = format_desc->channel[chan].size;
+ type->length = format_desc->nr_channels;
+ }
+}
+
+
+/**
+ * Retrieves the type for a format which is usable in the blending code.
+ *
+ * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
+ */
+static INLINE void
+lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
+ struct lp_type* type)
+{
+ unsigned i;
+ unsigned chan;
+
+ if (format_expands_to_float_soa(format_desc)) {
+ /* always use ordinary floats for blending */
+ type->floating = true;
+ type->fixed = false;
+ type->sign = true;
+ type->norm = false;
+ type->width = 32;
+ type->length = 4;
+ return;
+ }
+
+ for (i = 0; i < 4; i++)
+ if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
+ break;
+ chan = i;
+
+ memset(type, 0, sizeof(struct lp_type));
+ type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
+ type->fixed = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
+ type->sign = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
+ type->norm = format_desc->channel[chan].normalized;
+ type->width = format_desc->channel[chan].size;
+ type->length = format_desc->nr_channels;
+
+ for (i = 1; i < format_desc->nr_channels; ++i) {
+ if (format_desc->channel[i].size > type->width)
+ type->width = format_desc->channel[i].size;
+ }
+
+ if (type->floating) {
+ type->width = 32;
+ } else {
+ if (type->width <= 8) {
+ type->width = 8;
+ } else if (type->width <= 16) {
+ type->width = 16;
+ } else {
+ type->width = 32;
+ }
+ }
+
+ if (is_arithmetic_format(format_desc) && type->length == 3) {
+ type->length = 4;
+ }
+}
+
+
+/**
+ * Scale a normalized value from src_bits to dst_bits.
+ *
+ * The exact calculation is
+ *
+ * dst = iround(src * dst_mask / src_mask)
+ *
+ * or with integer rounding
+ *
+ * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
+ *
+ * where
+ *
+ * src_mask = (1 << src_bits) - 1
+ * dst_mask = (1 << dst_bits) - 1
+ *
+ * but we try to avoid division and multiplication through shifts.
+ */
+static INLINE LLVMValueRef
+scale_bits(struct gallivm_state *gallivm,
+ int src_bits,
+ int dst_bits,
+ LLVMValueRef src,
+ struct lp_type src_type)
+{
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef result = src;
+
+ if (dst_bits < src_bits) {
+ int delta_bits = src_bits - dst_bits;
+
+ if (delta_bits <= dst_bits) {
+ /*
+ * Approximate the rescaling with a single shift.
+ *
+ * This gives the wrong rounding.
+ */
+
+ result = LLVMBuildLShr(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type, delta_bits),
+ "");
+
+ } else {
+ /*
+ * Try more accurate rescaling.
+ */
+
+ /*
+ * Drop the least significant bits to make space for the multiplication.
+ *
+ * XXX: A better approach would be to use a wider integer type as intermediate. But
+ * this is enough to convert alpha from 16bits -> 2 when rendering to
+ * PIPE_FORMAT_R10G10B10A2_UNORM.
+ */
+ result = LLVMBuildLShr(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type, dst_bits),
+ "");
+
+
+ result = LLVMBuildMul(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
+ "");
+
+ /*
+ * Add a rounding term before the division.
+ *
+ * TODO: Handle signed integers too.
+ */
+ if (!src_type.sign) {
+ result = LLVMBuildAdd(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
+ "");
+ }
+
+ /*
+ * Approximate the division by src_mask with a src_bits shift.
+ *
+ * Given the src has already been shifted by dst_bits, all we need
+ * to do is to shift by the difference.
+ */
+
+ result = LLVMBuildLShr(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, delta_bits),
+ "");
+ }
+
+ } else if (dst_bits > src_bits) {
+ /* Scale up bits */
+ int db = dst_bits - src_bits;
+
+ /* Shift left by difference in bits */
+ result = LLVMBuildShl(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type, db),
+ "");
+
+ if (db < src_bits) {
+ /* Enough bits in src to fill the remainder */
+ LLVMValueRef lower = LLVMBuildLShr(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type, src_bits - db),
+ "");
+
+ result = LLVMBuildOr(builder, result, lower, "");
+ } else if (db > src_bits) {
+ /* Need to repeatedly copy src bits to fill remainder in dst */
+ unsigned n;
+
+ for (n = src_bits; n < dst_bits; n *= 2) {
+ LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
+
+ result = LLVMBuildOr(builder,
+ result,
+ LLVMBuildLShr(builder, result, shuv, ""),
+ "");
+ }
+ }
+ }
+
+ return result;
+}
+
+
+/**
+ * Convert from memory format to blending format
+ *
+ * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
+ */
+static void
+convert_to_blend_type(struct gallivm_state *gallivm,
+ unsigned block_size,
+ const struct util_format_description *src_fmt,
+ struct lp_type src_type,
+ struct lp_type dst_type,
+ LLVMValueRef* src, // and dst
+ unsigned num_srcs)
+{
+ LLVMValueRef *dst = src;
+ LLVMBuilderRef builder = gallivm->builder;
+ struct lp_type blend_type;
+ struct lp_type mem_type;
+ unsigned i, j, k;
+ unsigned pixels = block_size / num_srcs;
+ bool is_arith;
+
+ /*
+ * full custom path for packed floats and srgb formats - none of the later
+ * functions would do anything useful, and given the lp_type representation they
+ * can't be fixed. Should really have some SoA blend path for these kind of
+ * formats rather than hacking them in here.
+ */
+ if (format_expands_to_float_soa(src_fmt)) {
+ LLVMValueRef tmpsrc[4];
+ /*
+ * This is pretty suboptimal for this case blending in SoA would be much
+ * better, since conversion gets us SoA values so need to convert back.
+ */
+ assert(src_type.width == 32);
+ assert(dst_type.floating);
+ assert(dst_type.width == 32);
+ assert(dst_type.length % 4 == 0);
+ assert(num_srcs % 4 == 0);
+
+ for (i = 0; i < 4; i++) {
+ tmpsrc[i] = src[i];
+ }
+ for (i = 0; i < num_srcs / 4; i++) {
+ LLVMValueRef tmpsoa[4];
+ LLVMValueRef tmps = tmpsrc[i];
+ if (dst_type.length == 8) {
+ LLVMValueRef shuffles[8];
+ unsigned j;
+ /* fetch was 4 values but need 8-wide output values */
+ tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
+ /*
+ * for 8-wide aos transpose would give us wrong order not matching
+ * incoming converted fs values and mask. ARGH.
+ */
+ for (j = 0; j < 4; j++) {
+ shuffles[j] = lp_build_const_int32(gallivm, j * 2);
+ shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
+ }
+ tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
+ LLVMConstVector(shuffles, 8), "");
+ }
+ if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
+ }
+ else {
+ lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
+ }
+ lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
+ }
+ return;
+ }
+
+ lp_mem_type_from_format_desc(src_fmt, &mem_type);
+ lp_blend_type_from_format_desc(src_fmt, &blend_type);
+
+ /* Is the format arithmetic */
+ is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
+ is_arith &= !(mem_type.width == 16 && mem_type.floating);
+
+ /* Pad if necessary */
+ if (!is_arith && src_type.length < dst_type.length) {
+ for (i = 0; i < num_srcs; ++i) {
+ dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
+ }
+
+ src_type.length = dst_type.length;
+ }
+
+ /* Special case for half-floats */
+ if (mem_type.width == 16 && mem_type.floating) {
+ assert(blend_type.width == 32 && blend_type.floating);
+ lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
+ is_arith = false;
+ }
+
+ if (!is_arith) {
+ return;
+ }
+
+ src_type.width = blend_type.width * blend_type.length;
+ blend_type.length *= pixels;
+ src_type.length *= pixels / (src_type.length / mem_type.length);
+
+ for (i = 0; i < num_srcs; ++i) {
+ LLVMValueRef chans[4];
+ LLVMValueRef res = NULL;
+
+ dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
+
+ for (j = 0; j < src_fmt->nr_channels; ++j) {
+ unsigned mask = 0;
+ unsigned sa = src_fmt->channel[j].shift;
+#ifdef PIPE_ARCH_LITTLE_ENDIAN
+ unsigned from_lsb = j;
+#else
+ unsigned from_lsb = src_fmt->nr_channels - j - 1;
+#endif
+
+ for (k = 0; k < src_fmt->channel[j].size; ++k) {
+ mask |= 1 << k;
+ }
+
+ /* Extract bits from source */
+ chans[j] = LLVMBuildLShr(builder,
+ dst[i],
+ lp_build_const_int_vec(gallivm, src_type, sa),
+ "");
+
+ chans[j] = LLVMBuildAnd(builder,
+ chans[j],
+ lp_build_const_int_vec(gallivm, src_type, mask),
+ "");
+
+ /* Scale bits */
+ if (src_type.norm) {
+ chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
+ blend_type.width, chans[j], src_type);
+ }
+
+ /* Insert bits into correct position */
+ chans[j] = LLVMBuildShl(builder,
+ chans[j],
+ lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
+ "");
+
+ if (j == 0) {
+ res = chans[j];
+ } else {
+ res = LLVMBuildOr(builder, res, chans[j], "");
+ }
+ }
+
+ dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
+ }
+}
+
+
+/**
+ * Convert from blending format to memory format
+ *
+ * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
+ */
+static void
+convert_from_blend_type(struct gallivm_state *gallivm,
+ unsigned block_size,
+ const struct util_format_description *src_fmt,
+ struct lp_type src_type,
+ struct lp_type dst_type,
+ LLVMValueRef* src, // and dst
+ unsigned num_srcs)
+{
+ LLVMValueRef* dst = src;
+ unsigned i, j, k;
+ struct lp_type mem_type;
+ struct lp_type blend_type;
+ LLVMBuilderRef builder = gallivm->builder;
+ unsigned pixels = block_size / num_srcs;
+ bool is_arith;
+
+ /*
+ * full custom path for packed floats and srgb formats - none of the later
+ * functions would do anything useful, and given the lp_type representation they
+ * can't be fixed. Should really have some SoA blend path for these kind of
+ * formats rather than hacking them in here.
+ */
+ if (format_expands_to_float_soa(src_fmt)) {
+ /*
+ * This is pretty suboptimal for this case blending in SoA would be much
+ * better - we need to transpose the AoS values back to SoA values for
+ * conversion/packing.
+ */
+ assert(src_type.floating);
+ assert(src_type.width == 32);
+ assert(src_type.length % 4 == 0);
+ assert(dst_type.width == 32);
+
+ for (i = 0; i < num_srcs / 4; i++) {
+ LLVMValueRef tmpsoa[4], tmpdst;
+ lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
+ /* really really need SoA here */
+
+ if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
+ tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
+ }
+ else {
+ tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
+ src_type, tmpsoa);
+ }
+
+ if (src_type.length == 8) {
+ LLVMValueRef tmpaos, shuffles[8];
+ unsigned j;
+ /*
+ * for 8-wide aos transpose has given us wrong order not matching
+ * output order. HMPF. Also need to split the output values manually.
+ */
+ for (j = 0; j < 4; j++) {
+ shuffles[j * 2] = lp_build_const_int32(gallivm, j);
+ shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
+ }
+ tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
+ LLVMConstVector(shuffles, 8), "");
+ src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
+ src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
+ }
+ else {
+ src[i] = tmpdst;
+ }
+ }
+ return;
+ }
+
+ lp_mem_type_from_format_desc(src_fmt, &mem_type);
+ lp_blend_type_from_format_desc(src_fmt, &blend_type);
+
+ is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
+
+ /* Special case for half-floats */
+ if (mem_type.width == 16 && mem_type.floating) {
+ int length = dst_type.length;
+ assert(blend_type.width == 32 && blend_type.floating);
+
+ dst_type.length = src_type.length;
+
+ lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
+
+ dst_type.length = length;
+ is_arith = false;
+ }
+
+ /* Remove any padding */
+ if (!is_arith && (src_type.length % mem_type.length)) {
+ src_type.length -= (src_type.length % mem_type.length);
+
+ for (i = 0; i < num_srcs; ++i) {
+ dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
+ }
+ }
+
+ /* No bit arithmetic to do */
+ if (!is_arith) {
+ return;
+ }
+
+ src_type.length = pixels;
+ src_type.width = blend_type.length * blend_type.width;
+ dst_type.length = pixels;
+
+ for (i = 0; i < num_srcs; ++i) {
+ LLVMValueRef chans[4];
+ LLVMValueRef res = NULL;
+
+ dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
+
+ for (j = 0; j < src_fmt->nr_channels; ++j) {
+ unsigned mask = 0;
+ unsigned sa = src_fmt->channel[j].shift;
+#ifdef PIPE_ARCH_LITTLE_ENDIAN
+ unsigned from_lsb = j;
+#else
+ unsigned from_lsb = src_fmt->nr_channels - j - 1;
+#endif
+
+ assert(blend_type.width > src_fmt->channel[j].size);
+
+ for (k = 0; k < blend_type.width; ++k) {
+ mask |= 1 << k;
+ }
+
+ /* Extract bits */
+ chans[j] = LLVMBuildLShr(builder,
+ dst[i],
+ lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
+ "");
+
+ chans[j] = LLVMBuildAnd(builder,
+ chans[j],
+ lp_build_const_int_vec(gallivm, src_type, mask),
+ "");
+
+ /* Scale down bits */
+ if (src_type.norm) {
+ chans[j] = scale_bits(gallivm, blend_type.width,
+ src_fmt->channel[j].size, chans[j], src_type);
+ }
+
+ /* Insert bits */
+ chans[j] = LLVMBuildShl(builder,
+ chans[j],
+ lp_build_const_int_vec(gallivm, src_type, sa),
+ "");
+
+ sa += src_fmt->channel[j].size;
+
+ if (j == 0) {
+ res = chans[j];
+ } else {
+ res = LLVMBuildOr(builder, res, chans[j], "");
+ }
+ }
+
+ assert (dst_type.width != 24);
+
+ dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
+ }
+}
+
+
+/**
+ * Convert alpha to same blend type as src
+ */
+static void
+convert_alpha(struct gallivm_state *gallivm,
+ struct lp_type row_type,
+ struct lp_type alpha_type,
+ const unsigned block_size,
+ const unsigned block_height,
+ const unsigned src_count,
+ const unsigned dst_channels,
+ const bool pad_inline,
+ LLVMValueRef* src_alpha)
+{
+ LLVMBuilderRef builder = gallivm->builder;
+ unsigned i, j;
+ unsigned length = row_type.length;
+ row_type.length = alpha_type.length;
+
+ /* Twiddle the alpha to match pixels */
+ lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
+
+ /*
+ * TODO this should use single lp_build_conv call for
+ * src_count == 1 && dst_channels == 1 case (dropping the concat below)
+ */
+ for (i = 0; i < block_height; ++i) {
+ lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
+ }
+
+ alpha_type = row_type;
+ row_type.length = length;
+
+ /* If only one channel we can only need the single alpha value per pixel */
+ if (src_count == 1 && dst_channels == 1) {
+
+ lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
+ } else {
+ /* If there are more srcs than rows then we need to split alpha up */
+ if (src_count > block_height) {
+ for (i = src_count; i > 0; --i) {
+ unsigned pixels = block_size / src_count;
+ unsigned idx = i - 1;
+
+ src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
+ (idx * pixels) % 4, pixels);
+ }
+ }
+
+ /* If there is a src for each pixel broadcast the alpha across whole row */
+ if (src_count == block_size) {
+ for (i = 0; i < src_count; ++i) {
+ src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
+ }
+ } else {
+ unsigned pixels = block_size / src_count;
+ unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
+ unsigned alpha_span = 1;
+ LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
+
+ /* Check if we need 2 src_alphas for our shuffles */
+ if (pixels > alpha_type.length) {
+ alpha_span = 2;
+ }
+
+ /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
+ for (j = 0; j < row_type.length; ++j) {
+ if (j < pixels * channels) {
+ shuffles[j] = lp_build_const_int32(gallivm, j / channels);
+ } else {
+ shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
+ }
+ }
+
+ for (i = 0; i < src_count; ++i) {
+ unsigned idx1 = i, idx2 = i;
+
+ if (alpha_span > 1){
+ idx1 *= alpha_span;
+ idx2 = idx1 + 1;
+ }
+
+ src_alpha[i] = LLVMBuildShuffleVector(builder,
+ src_alpha[idx1],
+ src_alpha[idx2],
+ LLVMConstVector(shuffles, row_type.length),
+ "");
+ }
+ }
+ }
+}
+
+
+/**
+ * Generates the blend function for unswizzled colour buffers
+ * Also generates the read & write from colour buffer
+ */
+static void
+generate_unswizzled_blend(struct gallivm_state *gallivm,
+ unsigned rt,
+ struct lp_fragment_shader_variant *variant,
+ enum pipe_format out_format,
+ unsigned int num_fs,
+ struct lp_type fs_type,
+ LLVMValueRef* fs_mask,
+ LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
+ LLVMValueRef context_ptr,
+ LLVMValueRef color_ptr,
+ LLVMValueRef stride,
+ unsigned partial_mask,
+ boolean do_branch)
+{
+ const unsigned alpha_channel = 3;
+ const unsigned block_width = LP_RASTER_BLOCK_SIZE;
+ const unsigned block_height = LP_RASTER_BLOCK_SIZE;
+ const unsigned block_size = block_width * block_height;
+ const unsigned lp_integer_vector_width = 128;
+
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
+ LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
+ LLVMValueRef src_alpha[4 * 4];
+ LLVMValueRef src1_alpha[4 * 4];
+ LLVMValueRef src_mask[4 * 4];
+ LLVMValueRef src[4 * 4];
+ LLVMValueRef src1[4 * 4];
+ LLVMValueRef dst[4 * 4];
+ LLVMValueRef blend_color;
+ LLVMValueRef blend_alpha;
+ LLVMValueRef i32_zero;
+ LLVMValueRef check_mask;
+ LLVMValueRef undef_src_val;
+
+ struct lp_build_mask_context mask_ctx;
+ struct lp_type mask_type;
+ struct lp_type blend_type;
+ struct lp_type row_type;
+ struct lp_type dst_type;
+
+ unsigned char swizzle[TGSI_NUM_CHANNELS];
+ unsigned vector_width;
+ unsigned src_channels = TGSI_NUM_CHANNELS;
+ unsigned dst_channels;
+ unsigned dst_count;
+ unsigned src_count;
+ unsigned i, j;
+
+ const struct util_format_description* out_format_desc = util_format_description(out_format);
+
+ unsigned dst_alignment;
+
+ bool pad_inline = is_arithmetic_format(out_format_desc);
+ bool has_alpha = false;
+ const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
+ util_blend_state_is_dual(&variant->key.blend, 0);
+
+ const boolean is_1d = variant->key.resource_1d;
+ unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
+
+ mask_type = lp_int32_vec4_type();
+ mask_type.length = fs_type.length;
+
+ for (i = num_fs; i < num_fullblock_fs; i++) {
+ fs_mask[i] = lp_build_zero(gallivm, mask_type);
+ }
+
+ /* Do not bother executing code when mask is empty.. */
+ if (do_branch) {
+ check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
+
+ for (i = 0; i < num_fullblock_fs; ++i) {
+ check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
+ }
+
+ lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
+ lp_build_mask_check(&mask_ctx);
+ }
+
+ partial_mask |= !variant->opaque;
+ i32_zero = lp_build_const_int32(gallivm, 0);
+
+#if HAVE_LLVM < 0x0302
+ /*
+ * undef triggers a crash in LLVMBuildTrunc in convert_from_blend_type in some
+ * cases (seen with r10g10b10a2, 128bit wide vectors) (only used for 1d case).
+ */
+ undef_src_val = lp_build_zero(gallivm, fs_type);
+#else
+ undef_src_val = lp_build_undef(gallivm, fs_type);
+#endif
+
+
+ /* Get type from output format */
+ lp_blend_type_from_format_desc(out_format_desc, &row_type);
+ lp_mem_type_from_format_desc(out_format_desc, &dst_type);
+
+ row_type.length = fs_type.length;
+ vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
+
+ /* Compute correct swizzle and count channels */
+ memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
+ dst_channels = 0;
+
+ for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
+ /* Ensure channel is used */
+ if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
+ continue;
+ }
+
+ /* Ensure not already written to (happens in case with GL_ALPHA) */
+ if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
+ continue;
+ }
+
+ /* Ensure we havn't already found all channels */
+ if (dst_channels >= out_format_desc->nr_channels) {
+ continue;
+ }
+
+ swizzle[out_format_desc->swizzle[i]] = i;
+ ++dst_channels;
+
+ if (i == alpha_channel) {
+ has_alpha = true;
+ }
+ }
+
+ if (format_expands_to_float_soa(out_format_desc)) {
+ /*
+ * the code above can't work for layout_other
+ * for srgb it would sort of work but we short-circuit swizzles, etc.
+ * as that is done as part of unpack / pack.
+ */
+ dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
+ has_alpha = true;
+ swizzle[0] = 0;
+ swizzle[1] = 1;
+ swizzle[2] = 2;
+ swizzle[3] = 3;
+ pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
+ }
+
+ /* If 3 channels then pad to include alpha for 4 element transpose */
+ if (dst_channels == 3 && !has_alpha) {
+ for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
+ if (swizzle[i] > TGSI_NUM_CHANNELS)
+ swizzle[i] = 3;
+ }
+ if (out_format_desc->nr_channels == 4) {
+ dst_channels = 4;
+ }
+ }
+
+ /*
+ * Load shader output
+ */
+ for (i = 0; i < num_fullblock_fs; ++i) {
+ /* Always load alpha for use in blending */
+ LLVMValueRef alpha;
+ if (i < num_fs) {
+ alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
+ }
+ else {
+ alpha = undef_src_val;
+ }
+
+ /* Load each channel */
+ for (j = 0; j < dst_channels; ++j) {
+ assert(swizzle[j] < 4);
+ if (i < num_fs) {
+ fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
+ }
+ else {
+ fs_src[i][j] = undef_src_val;
+ }
+ }
+
+ /* If 3 channels then pad to include alpha for 4 element transpose */
+ /*
+ * XXX If we include that here maybe could actually use it instead of
+ * separate alpha for blending?
+ */
+ if (dst_channels == 3 && !has_alpha) {
+ fs_src[i][3] = alpha;
+ }
+
+ /* We split the row_mask and row_alpha as we want 128bit interleave */
+ if (fs_type.length == 8) {
+ src_mask[i*2 + 0] = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
+ src_mask[i*2 + 1] = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
+
+ src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
+ src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
+ } else {
+ src_mask[i] = fs_mask[i];
+ src_alpha[i] = alpha;
+ }
+ }
+ if (dual_source_blend) {
+ /* same as above except different src/dst, skip masks and comments... */
+ for (i = 0; i < num_fullblock_fs; ++i) {
+ LLVMValueRef alpha;
+ if (i < num_fs) {
+ alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
+ }
+ else {
+ alpha = undef_src_val;
+ }
+
+ for (j = 0; j < dst_channels; ++j) {
+ assert(swizzle[j] < 4);
+ if (i < num_fs) {
+ fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
+ }
+ else {
+ fs_src1[i][j] = undef_src_val;
+ }
+ }
+ if (dst_channels == 3 && !has_alpha) {
+ fs_src1[i][3] = alpha;
+ }
+ if (fs_type.length == 8) {
+ src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
+ src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
+ } else {
+ src1_alpha[i] = alpha;
+ }
+ }
+ }
+
+ if (util_format_is_pure_integer(out_format)) {
+ /*
+ * In this case fs_type was really ints or uints disguised as floats,
+ * fix that up now.
+ */
+ fs_type.floating = 0;
+ fs_type.sign = dst_type.sign;
+ for (i = 0; i < num_fullblock_fs; ++i) {
+ for (j = 0; j < dst_channels; ++j) {
+ fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
+ lp_build_vec_type(gallivm, fs_type), "");
+ }
+ if (dst_channels == 3 && !has_alpha) {
+ fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
+ lp_build_vec_type(gallivm, fs_type), "");
+ }
+ }
+ }
+
+ /*
+ * Pixel twiddle from fragment shader order to memory order
+ */
+ src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
+ dst_channels, fs_src, src, pad_inline);
+ if (dual_source_blend) {
+ generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
+ fs_src1, src1, pad_inline);
+ }
+
+ src_channels = dst_channels < 3 ? dst_channels : 4;
+ if (src_count != num_fullblock_fs * src_channels) {
+ unsigned ds = src_count / (num_fullblock_fs * src_channels);
+ row_type.length /= ds;
+ fs_type.length = row_type.length;
+ }
+
+ blend_type = row_type;
+ mask_type.length = 4;
+
+ /* Convert src to row_type */
+ if (dual_source_blend) {
+ struct lp_type old_row_type = row_type;
+ lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
+ src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
+ }
+ else {
+ src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
+ }
+
+ /* If the rows are not an SSE vector, combine them to become SSE size! */
+ if ((row_type.width * row_type.length) % 128) {
+ unsigned bits = row_type.width * row_type.length;
+ unsigned combined;
+
+ assert(src_count >= (vector_width / bits));
+
+ dst_count = src_count / (vector_width / bits);
+
+ combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
+ if (dual_source_blend) {
+ lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
+ }
+
+ row_type.length *= combined;
+ src_count /= combined;
+
+ bits = row_type.width * row_type.length;
+ assert(bits == 128 || bits == 256);
+ }
+
+
+ /*
+ * Blend Colour conversion
+ */
+ blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
+ blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
+ blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
+
+ /* Convert */
+ lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
+
+ if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
+ /*
+ * since blending is done with floats, there was no conversion.
+ * However, the rules according to fixed point renderbuffers still
+ * apply, that is we must clamp inputs to 0.0/1.0.
+ * (This would apply to separate alpha conversion too but we currently
+ * force has_alpha to be true.)
+ * TODO: should skip this with "fake" blend, since post-blend conversion
+ * will clamp anyway.
+ * TODO: could also skip this if fragment color clamping is enabled. We
+ * don't support it natively so it gets baked into the shader however, so
+ * can't really tell here.
+ */
+ struct lp_build_context f32_bld;
+ assert(row_type.floating);
+ lp_build_context_init(&f32_bld, gallivm, row_type);
+ for (i = 0; i < src_count; i++) {
+ src[i] = lp_build_clamp(&f32_bld, src[i], f32_bld.zero, f32_bld.one);
+ }
+ if (dual_source_blend) {
+ for (i = 0; i < src_count; i++) {
+ src1[i] = lp_build_clamp(&f32_bld, src1[i], f32_bld.zero, f32_bld.one);
+ }
+ }
+ /* probably can't be different than row_type but better safe than sorry... */
+ lp_build_context_init(&f32_bld, gallivm, blend_type);
+ blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
+ }
+
+ /* Extract alpha */
+ blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
+
+ /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
+ pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
+ if (pad_inline) {
+ /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
+ blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
+ } else {
+ /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
+ blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
+ }
+
+ /*
+ * Mask conversion
+ */
+ lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
+
+ if (src_count < block_height) {
+ lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
+ } else if (src_count > block_height) {
+ for (i = src_count; i > 0; --i) {
+ unsigned pixels = block_size / src_count;
+ unsigned idx = i - 1;
+
+ src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
+ (idx * pixels) % 4, pixels);
+ }
+ }