/* * Copyright © 2025 Red Hat * * SPDX-License-Identifier: MIT */ #include "lvp_nir.h" extern unsigned lp_native_vector_width; #define MAX_CMAT_LEN 16 #define CMAT_LEN (lp_native_vector_width / 32) /* This pass lowers cooperative matrix. * * for lavapipe we advertise 8x8 matrix. * This means we can store vec8[8] and get the backend to do the right thing. */ static unsigned get_cmat_size(struct glsl_cmat_description matrix_desc) { return matrix_desc.cols * matrix_desc.rows; } static unsigned get_cmat_length(struct glsl_cmat_description matrix_desc) { return get_cmat_size(matrix_desc) / CMAT_LEN; } static const struct glsl_type * remap_matrix_type(struct hash_table *mapping, const struct glsl_type *orig) { struct hash_entry *entry = _mesa_hash_table_search(mapping, orig); if (entry) return entry->data; const struct glsl_type *new_type = orig; if (glsl_type_is_cmat(orig)) { struct glsl_cmat_description matrix_desc = *glsl_get_cmat_description(orig); new_type = glsl_vector_type(matrix_desc.element_type, get_cmat_length(matrix_desc)); } else if (glsl_type_is_array(orig)) { const struct glsl_type *elem_type = glsl_get_array_element(orig); const struct glsl_type *new_elem_type = remap_matrix_type(mapping, elem_type); if (elem_type != new_elem_type) { new_type = glsl_array_type(new_elem_type, glsl_get_length(orig), glsl_get_explicit_stride(orig)); } } _mesa_hash_table_insert(mapping, orig, (void *)new_type); return new_type; } static nir_def * load_cmat_deref(nir_builder *b, nir_deref_instr *src) { struct glsl_cmat_description matrix_desc = *glsl_get_cmat_description(src->type); return nir_build_load_deref( b, get_cmat_length(matrix_desc), glsl_base_type_bit_size(matrix_desc.element_type), &src->def, 0); } static ALWAYS_INLINE nir_def * load_cmat_src(nir_builder *b, nir_src src) { return load_cmat_deref(b, nir_src_as_deref(src)); } static ALWAYS_INLINE struct glsl_cmat_description cmat_src_desc(nir_src src) { nir_deref_instr *deref = nir_src_as_deref(src); return *glsl_get_cmat_description(deref->type); } static void store_cmat_deref(nir_builder *b, nir_deref_instr *dst, nir_def *val) { ASSERTED struct glsl_cmat_description matrix_desc = *glsl_get_cmat_description(dst->type); assert(val->bit_size == glsl_base_type_bit_size(matrix_desc.element_type)); assert(val->num_components == get_cmat_length(matrix_desc)); nir_store_deref(b, dst, val, ~0); } static ALWAYS_INLINE void store_cmat_src(nir_builder *b, nir_src dst_src, nir_def *val) { store_cmat_deref(b, nir_src_as_deref(dst_src), val); } static bool lower_cmat_copy(nir_builder *b, nir_intrinsic_instr *intr) { nir_build_copy_deref(b, intr->src[0].ssa, intr->src[1].ssa); nir_instr_remove(&intr->instr); return true; } static nir_def * convert_base_type(nir_builder *b, nir_def *src, enum glsl_base_type src_type, enum glsl_base_type dst_type) { if (dst_type == src_type) return src; nir_op op = nir_type_conversion_op(nir_get_nir_type_for_glsl_base_type(src_type), nir_get_nir_type_for_glsl_base_type(dst_type), nir_rounding_mode_undef); return nir_build_alu1(b, op, src); } static bool lower_cmat_convert(nir_builder *b, nir_intrinsic_instr *intr) { struct glsl_cmat_description dst_desc = cmat_src_desc(intr->src[0]); struct glsl_cmat_description src_desc = cmat_src_desc(intr->src[1]); enum glsl_base_type dst_element_type = dst_desc.element_type; enum glsl_base_type src_element_type = src_desc.element_type; nir_def *cmat = load_cmat_src(b, intr->src[1]); nir_def *ret = convert_base_type(b, cmat, src_element_type, dst_element_type); store_cmat_src(b, intr->src[0], ret); nir_instr_remove(&intr->instr); return true; } static bool lower_cmat_load_store(nir_builder *b, struct hash_table *type_mapping, nir_intrinsic_instr *intr) { const bool is_load = intr->intrinsic == nir_intrinsic_cmat_load; const struct glsl_cmat_description desc = cmat_src_desc(intr->src[!is_load]); enum glsl_matrix_layout layout = nir_intrinsic_matrix_layout(intr); nir_deref_instr *cmat_deref = nir_src_as_deref(intr->src[!is_load]); nir_deref_instr *deref = nir_src_as_deref(intr->src[is_load]); nir_def *stride = intr->src[2].ssa; nir_def *lane_id = nir_load_subgroup_invocation(b); unsigned type_size_B = glsl_base_type_bit_size(desc.element_type) / 8; const uint32_t ptr_stride = glsl_get_bit_size(deref->type) / 8 * glsl_get_vector_elements(deref->type); deref = nir_build_deref_cast(b, &deref->def, deref->modes, deref->type, ptr_stride); const struct glsl_type *cmat_type = remap_matrix_type(type_mapping, cmat_deref->type); cmat_deref = nir_build_deref_cast(b, &cmat_deref->def, cmat_deref->modes, cmat_type, 0); /* store B matrix transposed */ if (desc.use == GLSL_CMAT_USE_B) layout = layout == GLSL_MATRIX_LAYOUT_COLUMN_MAJOR ? GLSL_MATRIX_LAYOUT_ROW_MAJOR : GLSL_MATRIX_LAYOUT_COLUMN_MAJOR; unsigned idx_bits = deref->def.bit_size; nir_def *vars[MAX_CMAT_LEN]; if (!is_load) { nir_def *src = load_cmat_src(b, intr->src[!is_load]); for (unsigned i = 0; i < CMAT_LEN; i++) { vars[i] = nir_channel(b, src, i); } } for (unsigned i = 0; i < CMAT_LEN; i++) { nir_def *col_offset = lane_id; nir_def *row_offset = nir_imm_int(b, i); if (layout == GLSL_MATRIX_LAYOUT_ROW_MAJOR) { SWAP(col_offset, row_offset); } col_offset = nir_imul(b, col_offset, stride); col_offset = nir_u2uN(b, col_offset, idx_bits); row_offset = nir_u2uN(b, row_offset, idx_bits); nir_deref_instr *iter_deref = nir_build_deref_ptr_as_array(b, deref, col_offset); iter_deref = nir_build_deref_cast(b, &iter_deref->def, deref->modes, glsl_scalar_type(desc.element_type), type_size_B); iter_deref = nir_build_deref_ptr_as_array(b, iter_deref, row_offset); if (is_load) { vars[i] = nir_load_deref(b, iter_deref); } else { nir_store_deref(b, iter_deref, vars[i], ~0); } } if (is_load) { nir_def *mat = nir_vec(b, vars, CMAT_LEN); nir_store_deref(b, cmat_deref, mat, nir_component_mask(mat->num_components)); } nir_instr_remove(&intr->instr); return true; } static bool lower_cmat_construct(nir_builder *b, nir_intrinsic_instr *intr) { nir_deref_instr *dst_deref = nir_src_as_deref(intr->src[0]); struct glsl_cmat_description desc = *glsl_get_cmat_description(dst_deref->type); nir_def *elem = intr->src[1].ssa; nir_def *r = nir_replicate(b, elem, get_cmat_length(desc)); nir_store_deref(b, dst_deref, r, nir_component_mask(r->num_components)); nir_instr_remove(&intr->instr); return true; } static bool lower_cmat_extract(nir_builder *b, nir_intrinsic_instr *intr) { nir_def *mat = load_cmat_src(b, intr->src[0]); nir_def *index = intr->src[1].ssa; nir_def *elem = nir_vector_extract(b, mat, index); nir_def_replace(&intr->def, elem); return true; } static bool lower_cmat_insert(nir_builder *b, nir_intrinsic_instr *intr) { nir_def *elem = intr->src[1].ssa; nir_def *mat = load_cmat_src(b, intr->src[2]); nir_def *index = intr->src[3].ssa; nir_def *r = nir_vector_insert(b, mat, elem, index); store_cmat_src(b, intr->src[0], r); nir_instr_remove(&intr->instr); return true; } static bool lower_cmat_binary_op(nir_builder *b, nir_intrinsic_instr *intr) { nir_def *src_a = load_cmat_src(b, intr->src[1]); nir_def *src_b = load_cmat_src(b, intr->src[2]); nir_op op = nir_intrinsic_alu_op(intr); nir_def *ret = nir_build_alu2(b, op, src_a, src_b); store_cmat_src(b, intr->src[0], ret); nir_instr_remove(&intr->instr); return true; } static bool lower_cmat_unary_op(nir_builder *b, nir_intrinsic_instr *intr) { nir_def *src = load_cmat_src(b, intr->src[1]); nir_op op = nir_intrinsic_alu_op(intr); nir_def *ret = nir_build_alu1(b, op, src); store_cmat_src(b, intr->src[0], ret); nir_instr_remove(&intr->instr); return true; } static bool lower_cmat_scalar_op(nir_builder *b, nir_intrinsic_instr *intr) { nir_def *src_a = load_cmat_src(b, intr->src[1]); nir_op op = nir_intrinsic_alu_op(intr); nir_def *ret = nir_build_alu2(b, op, src_a, intr->src[2].ssa); store_cmat_src(b, intr->src[0], ret); nir_instr_remove(&intr->instr); return true; } static bool lower_cmat_length(nir_builder *b, nir_intrinsic_instr *intr) { nir_def_replace(&intr->def, nir_imm_int(b, CMAT_LEN)); return true; } static bool lower_cmat_muladd(nir_builder *b, nir_intrinsic_instr *intr) { const struct glsl_cmat_description a_desc = cmat_src_desc(intr->src[1]); const struct glsl_cmat_description b_desc = cmat_src_desc(intr->src[2]); const struct glsl_cmat_description c_desc = cmat_src_desc(intr->src[3]); nir_def *cmat_a = load_cmat_src(b, intr->src[1]); nir_def *cmat_b = load_cmat_src(b, intr->src[2]); nir_def *cmat_c = load_cmat_src(b, intr->src[3]); unsigned a_length = get_cmat_length(a_desc); unsigned b_length = get_cmat_length(b_desc); unsigned c_length = get_cmat_length(c_desc); nir_def *a_comps[NIR_MAX_VEC_COMPONENTS]; nir_def *b_comps[NIR_MAX_VEC_COMPONENTS]; nir_def *c_comps[NIR_MAX_VEC_COMPONENTS]; nir_def *d_comps[NIR_MAX_VEC_COMPONENTS]; const nir_cmat_signed cmat_signed_mask = nir_intrinsic_cmat_signed_mask(intr); enum glsl_base_type c_element_type = glsl_apply_signedness_to_base_type(c_desc.element_type, cmat_signed_mask & NIR_CMAT_C_SIGNED); for (unsigned i = 0; i < a_length; i++) a_comps[i] = nir_channel(b, cmat_a, i); for (unsigned i = 0; i < b_length; i++) b_comps[i] = nir_channel(b, cmat_b, i); for (unsigned i = 0; i < c_length; i++) c_comps[i] = nir_channel(b, cmat_c, i); nir_def *lane_id = nir_load_subgroup_invocation(b); int accum_bit_size = glsl_base_type_bit_size(c_desc.element_type); for (unsigned i = 0; i < CMAT_LEN; i++) { nir_def *ref = nir_imm_zero(b, 1, glsl_base_type_bit_size(c_desc.element_type)); for (unsigned j = 0; j < CMAT_LEN; j++) { nir_def *outer_else_val = ref; ref = nir_imm_zero(b, 1, glsl_base_type_bit_size(c_desc.element_type)); nir_def *a_i = a_comps[i]; nir_def *b_j = b_comps[j]; /* B is stored transposed */ nir_def *val; if (glsl_base_type_is_integer(c_desc.element_type)) { if (c_element_type == GLSL_TYPE_INT) a_i = nir_i2iN(b, a_i, accum_bit_size); else a_i = nir_u2uN(b, a_i, accum_bit_size); if (c_element_type == GLSL_TYPE_INT) b_j = nir_i2iN(b, b_j, accum_bit_size); else b_j = nir_u2uN(b, b_j, accum_bit_size); val = nir_imul(b, a_i, b_j); ref = nir_iadd(b, ref, val); } else { a_i = nir_f2fN(b, a_i, accum_bit_size); b_j = nir_f2fN(b, b_j, accum_bit_size); val = nir_fmul(b, a_i, b_j); ref = nir_fadd(b, ref, val); } if (glsl_base_type_is_integer(c_desc.element_type)) { ref = nir_reduce(b, ref, .reduction_op = nir_op_iadd); } else { ref = nir_reduce(b, ref, .reduction_op = nir_op_fadd); } nir_def *lane = nir_ieq_imm(b, lane_id, j); ref = nir_bcsel(b, lane, ref, outer_else_val); } if (glsl_base_type_is_integer(c_desc.element_type)) { ref = nir_iadd(b, ref, c_comps[i]); } else { ref = nir_fadd(b, ref, c_comps[i]); } d_comps[i] = ref; } nir_def *ret = nir_vec(b, d_comps, CMAT_LEN); store_cmat_src(b, intr->src[0], ret); nir_instr_remove(&intr->instr); return true; } static bool lower_cmat_bitcast(nir_builder *b, nir_intrinsic_instr *intr) { nir_def *src1 = load_cmat_src(b, intr->src[1]); nir_store_deref(b, nir_src_as_deref(intr->src[0]), src1, nir_component_mask(src1->num_components)); nir_instr_remove(&intr->instr); return true; } static bool lower_impl(nir_function_impl *impl, struct hash_table *type_mapping) { bool progress = false; /* Remap all cmat temp var to array of scalars */ nir_foreach_function_temp_variable(var, impl) { const struct glsl_type *new_type = remap_matrix_type(type_mapping, var->type); if (new_type != var->type) { var->type = new_type; progress = true; } } /* Iterate in reverse order so that lowering can still use the matrix types from the derefs before we change it. */ nir_builder b = nir_builder_create(impl); nir_foreach_block_reverse_safe (block, impl) { nir_foreach_instr_reverse_safe (instr, block) { b.cursor = nir_before_instr(instr); switch (instr->type) { case nir_instr_type_intrinsic: { nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); switch (intr->intrinsic) { case nir_intrinsic_cmat_length: progress |= lower_cmat_length(&b, intr); break; case nir_intrinsic_cmat_construct: progress |= lower_cmat_construct(&b, intr); break; case nir_intrinsic_cmat_extract: progress |= lower_cmat_extract(&b, intr); break; case nir_intrinsic_cmat_insert: progress |= lower_cmat_insert(&b, intr); break; case nir_intrinsic_cmat_load: case nir_intrinsic_cmat_store: progress |= lower_cmat_load_store(&b, type_mapping, intr); break; case nir_intrinsic_cmat_binary_op: progress |= lower_cmat_binary_op(&b, intr); break; case nir_intrinsic_cmat_unary_op: progress |= lower_cmat_unary_op(&b, intr); break; case nir_intrinsic_cmat_scalar_op: progress |= lower_cmat_scalar_op(&b, intr); break; case nir_intrinsic_cmat_muladd: progress |= lower_cmat_muladd(&b, intr); break; case nir_intrinsic_cmat_copy: progress |= lower_cmat_copy(&b, intr); break; case nir_intrinsic_cmat_convert: progress |= lower_cmat_convert(&b, intr); break; case nir_intrinsic_cmat_bitcast: progress |= lower_cmat_bitcast(&b, intr); break; default: break; } break; } case nir_instr_type_deref: { nir_deref_instr *deref = nir_instr_as_deref(instr); const struct glsl_type *new_type = remap_matrix_type(type_mapping, deref->type); if (new_type != deref->type) { deref->type = new_type; progress = true; } break; } default: break; } } } return nir_progress(progress, impl, nir_metadata_none); } bool lvp_nir_lower_cooperative_matrix(nir_shader *shader) { bool progress = false; if (!shader->info.cs.has_cooperative_matrix) return false; struct hash_table *type_mapping = _mesa_pointer_hash_table_create(NULL); /* Remap all cmat shader temp var to array of vectors */ nir_foreach_variable_with_modes(var, shader, nir_var_shader_temp) { const struct glsl_type *new_type = remap_matrix_type(type_mapping, var->type); if (new_type != var->type) { var->type = new_type; progress = true; } } progress |= lower_impl(nir_shader_get_entrypoint(shader), type_mapping); _mesa_hash_table_destroy(type_mapping, NULL); return progress; }